The AI Power Map

Discover Contribute The Book

Table of Contents

Preface2.7k 1 · The Power Network0.9k 2 · The Architects4.7k 3 · The Labs9.4k 3.1 Anthropic 3.2 Google DeepMind 4 · How Power Moves9.9k 4.1 Academic Roots 4.2 Safety Architects 5 · The Capital Network15.6k 5.1 Agentic Revolution 5.2 Family & Trust Layer 5.3 Funding & Talent Wars 5.4 Google Era & PayPal Mafia 5.5 AI Exits & Acquisitions 6 · Enterprise & Infrastructure7.8k 6.1 Infrastructure Layer 7 · Conflicts & Power Struggles8.1k 8 · Hidden Ties & Pattern Synthesis7.6k 9 · The Outsider Playbook4.3k How This Book Is Organized Age × Education → Influence 5 Headline Findings Founding Cohort Wave Talent Flow Sankey Tier-Weighting Sensitivity 5 Career Paths to AI Influence Top 10 AI Acquirers AI Enterprise Stack 5 AI Conflicts Timeline Bridges by Relationship Type Cofounder Clusters No-Revenue, Nine-Figure Exits Appendix 1 · Editorial Notice & Legal Disclaimer2.6k 2 · How to Read the Map (General Methods)2.8k A note for SNA and ONA readers 3 · Full Roster11.9k 4 · Schools & Universities0.6k 5 · Company & Product Guide1.1k

The AI Power Map

Preface

How an industry argues its way forward

Power in AI is not only a merit system. It is a network system.

Behind every model is a human map: researchers, founders, executives, investors, engineers, academics, operators, mentors, rivals, spouses, siblings, former classmates, former coworkers, professors, students, board members, and longtime advisors. Who trained whom, funded whom, hired whom, disagreed with whom, and built something new afterward often explains how the industry moves better than any single company story.

This project is a public demonstration of the author's broader interest: how human context, trust, informal authority, and organizational behavior shape AI systems. The book traces connections through companies, universities, papers, funding rounds, acquisitions, boards, lawsuits, friendships, marriages, sibling teams, cofounder histories, ideological disputes, and shared infrastructure, with AI-assisted research and human editorial review.

What surprised me most was not how competitive AI is. It was how connected it is.

From the outside, AI can look like a clean story of progress led by company names: OpenAI, Anthropic, DeepMind, Meta, Microsoft, NVIDIA, xAI, Perplexity, Databricks, Glean, and others. But the deeper story is about people moving through labs, cap tables, papers, podcasts, lawsuits, boards, universities, and private circles.

At the highest levels of AI, personal and professional trust often run on the same network. Spouses cofound companies. Siblings build labs. Former classmates fund each other. The point is not that personal relationships corrupt the system. The point is that they are part of how the system works.

Former colleagues become competitors. Competitors become references. Investors fund both sides. Researchers argue publicly while still respecting each other's work. Founders split over safety, speed, governance, or control, then create new institutions that push the field forward.

For example, the OpenAI board crisis in November 2023 was not only corporate drama. It exposed the central question every major AI organization now faces: who gets to decide how powerful systems should be built, shipped, slowed down, governed, or commercialized? The five day crisis also showed how much trust still sits outside formal process. As of November 2023, Wikipedia and contemporaneous reporting pointed to investor pressure and shifting board relationships as central to the resolution.

The same pattern appears across the industry. Anthropic came from a disagreement inside OpenAI (see Wikipedia). DeepMind came from a different vision of intelligence and science. The Transformer paper created not one company, but a generation of careers. Even the "AI godfathers" disagree on AI risk. As of 2024, Hinton, Bengio, and LeCun have stated publicly divergent positions on existential AI risk (see Wikipedia).

These conflicts are not separate from progress. They are part of how the industry tests itself, fractures, recombines, and moves forward.

That is why this book is called The AI Power Map. It maps formal power: titles, companies, valuations, and funding rounds. It also maps the trust layer: mentorship, reputation, shared history, technical taste, ideology, loyalty, and the ability to make other people move.

This is an educational project built from public sources. Its goal is to help readers understand who builds AI, how ideas and capital move, how people enter the field, and why the future of AI is shaped not only by models, but by the human network around them.

Methodology, source tiers, evidence tags, the four known biases, and the corrections policy live in Appendix · Editorial Notice & Legal Disclaimer.

How the five tiers are scored

Because the book references T1, T2, and T3 throughout (especially in Chapter 1 and Chapter 9), here is the criteria each person is scored against. Each person is rated on four observable dimensions: institutional leverage (the role they hold and what it controls), capital and decision authority (what they can fund, hire, or block), technical or operating contribution (papers, products, deals attributable to them), and network position (the people and institutions they connect). The score reflects current role scope, not lifetime achievement, so the same person can move between tiers as their role changes.

This is a labeling shortcut for analysis, not an evaluation of the person. The tiers only reflect how visible a person's role is inside the core AI industry, based on the public sources this map relies on. Someone scored T4 here may be T1 in another field, in a private role, or in work that simply does not show up in public records. The goal is to make patterns in the network legible, not to reduce anyone to a number. Construct caveat: tier is an editorial judgment without published inter-rater reliability, so any association that uses tier as the target (e.g. "PhD holders are overrepresented at T3+") is descriptive within the cast. It cannot be interpreted as a causal predictor or generalised beyond this curated dataset.

Tier	Criteria
T1 (29)	CEO or founder of a frontier lab or top-tier AI company; researcher whose foundational work is cited across the industry; managing partner at a top-tier fund with a documented track record of leading definitive AI rounds.
T2 (105)	Board member of a frontier lab; co-founder, president, or chief officer of a $1B+ AI company; partner or GP at a top-tier fund leading multiple AI rounds; primary author of a foundational paper.
T3 (179)	Senior leader at a frontier lab or major AI company; founding team member of a smaller AI company; principal or junior partner at an active AI fund; well-known applied researcher with industry-cited work.
T4 (72)	Strong technical or operating contributor inside a frontier lab; founder of a post-Series-A AI company; emerging fund manager; mid-career researcher with cited but not foundational work.
T5 (35)	Adjacent to the network: early-career researchers at frontier labs, supporting operational roles, contextual figures who appear in the cast as connectors or background rather than as primary actors.

Note: the influence tiers above describe people. They should not be confused with the edge strength tiers (Strong, Medium, Weak, Latent, Affiliation) defined in Appendix · How to Read the Map, which describe relationships between people.

My own work focuses on the behavioral context layer for enterprise AI: how organizations understand who should act on what, who holds influence, who has informal authority, and how decisions actually move beyond what is written in documents. That work is more complex and uses deeper organizational signals, while this book is a simpler, public-facing example of the same network lens applied to the AI industry itself.

The book is organized in three angles. The Power angle (Chapters 2-3) profiles the people and labs that hold influence now. The Pattern angle (Chapters 4-8) analyzes how that influence was built: talent pipelines, capital flows, hidden ties, acquisition strategies, and the conflicts that reshaped institutions. The Pathway angle (Chapter 9) turns the data into a practical playbook for people entering the field from the outside.

What follows is not just a history of AI.

It is a map of the people, relationships, arguments, alliances, fractures, and feedback loops that made the industry move this fast.

Chapter 1 · The Power Network: What the Data Reveals

The first finding is the most counterintuitive: in AI, the network is hard to enter but it is not sealed. The same names show up in cofounder lists, mentor lineages, board votes, and funding rounds, but the patterns that put them there are visible and learnable.

Before the profiles, the patterns. This map covers 420 people, with 139 receiving dedicated profiles, 88 in the extended roster, and the rest referenced in context. The book does not profile everyone equally. It focuses analytical depth on the people whose positions reveal how influence works. This chapter presents the macro findings from across all 420.

For the dataset, pipeline, and the eight methods used throughout the book, see Appendix · How this analysis was built.

Do You Need a Technical Background?

Not necessarily. Across the full dataset, 60 people have pure business backgrounds (MBA, law, finance, no CS/engineering degree). Their overall T1-T2 rate is 26%, lower than pure tech (36%), but their T1 members are some of the most powerful names in the industry: Marc Benioff (Salesforce, USC business), Sundar Pichai (IIT + Stanford MS + Wharton MBA, per Wikipedia), Thomas Kurian (Google Cloud, Princeton + Stanford MBA), and Hock Tan (Broadcom, MIT + Harvard MBA).

The surprising finding is in late entry. Among founders who entered after 35, pure business backgrounds have the highest T1-T2 rate: 50% (6 of 12), compared to 46% for pure tech (22 of 47) and 37% for hybrid (3 of 8). The sample is small, but the pattern is clear: business-background founders who enter AI late tend to do so through enterprise distribution channels (Benioff, Bhusri, McDermott) or capital networks (Hoffman, Sacks), and those channels have high conversion rates at the top.

Hybrid backgrounds (both technical + business credentials, like Musk or Nadella) are rarer (66 people) and have the lowest overall T1-T2 rate (21%), but the two who reach T1 are disproportionately powerful. The hybrid path appears to be high-variance: either you end up running a trillion-dollar company or you stay in the middle.

Chapter 2 · The Architects

The people who reach the top of AI do not lead in one uniform way. In this dataset, the 134 people in the top two influence levels sort into four background-driven archetypes, and the most surprising result is that 73 of them, 54%, came from outside the CS pipeline, which makes that route the modal path rather than the exception.

Eight people, working within thirty miles of each other in the Bay Area, decide what intelligence becomes next. They run the labs, sign the checks, and pick the next thing to train. Governments and academics weigh in afterward. These eight set the agenda first.

They disagree publicly, about timelines, about safety, about how fast to ship, and the disagreements are the most diagnostic thing about the field. They show how AI tests itself, fractures, and recombines in full view of people who will live with the result. Start with the person whose chips quietly power everyone else.

How Different Backgrounds Lead Differently

The four archetypes are built from bio-text analysis across the 134 people at the top two influence levels, not from interviews. The chart below shows the role mix and pattern signals for each.

Jensen Huang, The Hardware Sovereign

When Jensen was nine, his parents put him and his older brother on a plane to America. They stayed behind in Taiwan to finish the paperwork. His uncle, who had taken in the boys, enrolled them at Oneida Baptist Institute in Harlan County, Kentucky, a boarding school built for children from troubled homes. Jensen was not from a troubled home, but he arrived there anyway: a Taiwanese boy in eastern Kentucky learning to become self-reliant, learning to play ping-pong well enough to win, and learning that he could hold his own in rooms designed for someone else.

Read full profile ↓

Founded NVIDIA (Mid-career · 1993 · age 30). Before co-founding NVIDIA, Jensen Huang worked as a microprocessor designer at Advanced Micro Devices (AMD) and held roles, including director of a division, at LSI Logic Corporation. He earned his master's degree in electrical engineering from Stanford University in 1992, one year before founding NVIDIA. Exit: NVIDIA went public on January 22, 1999, at $12 per share, raising approximately $42 million in capital. (The Story of Jensen Huang and Nvidia - Quartr Insights · Jensen Huang - Wikipedia)

He finished high school in Oregon, earned a B.S. in electrical engineering at Oregon State, and then an M.S. at Stanford. Somewhere along the way he worked at a Denny's in San Jose as a busboy. In April 1993, he co-founded NVIDIA in a Denny's in San Jose. It is a detail he often retells in interviews.

For most of its existence NVIDIA made graphics cards for gamers. CUDA, the software layer that turned GPUs into general-purpose computers, launched in 2006 and was commercially irrelevant for years. Machine learning researchers started using it around 2011. Then AlexNet won ImageNet in 2012, trained on NVIDIA GPUs by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, and the whole arc of the company bent toward AI. As of November 2024, NVIDIA briefly hit a $3.6 trillion market cap, making it the most valuable company on Earth (per Wikipedia). Fiscal 2025 revenue: roughly $130.5B per NVIDIA investor relations. In 2022 it was $16B.

Jensen Huang and Lisa Su, the CEO of AMD (NVIDIA's main competitor in AI accelerators), are first cousins once removed: Lisa Su's maternal grandfather is the youngest brother of Jensen Huang's mother. Both emigrated to the U.S. as small children, Huang at four to Oregon, Su at three. They did not grow up together. Su has said publicly: "No family dinners." They met as adults at an industry event well into their careers. Two cousins from the same Taiwanese family now lead the world's two most important GPU companies, and their companies bid against each other for nearly every major AI hardware contract in the world. Tom's Hardware

Jensen and Lori have been married since 1988. Their daughter Madison is a Senior Director at NVIDIA, working on Omniverse and Physical AI. Their son Spencer ran an award-winning cocktail bar in San Francisco called Robin for eight years before he, too, joined NVIDIA as a product manager for AI infrastructure. It is a family business, in a way few family businesses are.

Why he matters to the map: He concentrates compute supply for every frontier lab, so capital flows pass through NVIDIA first.

Sam Altman, The Accelerationist

Samuel Harris Altman grew up in St. Louis, his father a real-estate lawyer, his mother a dermatologist. He has said in interviews that formative experiences in his youth taught him something specific and useful: the gap between how people present themselves and what is actually happening underneath. It's a skill Silicon Valley would reward him for, repeatedly.

Read full profile ↓

Background: Sam Altman has been publicly open about his identity since high school. He is married to Oliver Mulherin. (Wikipedia)

Founded OpenAI (Mid-career · 2015 · age 30). Sam Altman previously co-founded Loopt in 2005, which was acquired in 2012. He joined Y Combinator in 2011 and served as its president from February 2014 until March 2019, overlapping with OpenAI's founding in December 2015. (MicroVentures' Portfolio Company: OpenAI's History and Milestones · OpenAI | ChatGPT, Sam Altman, Microsoft, & History | Britannica Money)

He was born April 22, 1985, in Chicago. He enrolled at Stanford to study computer science and dropped out after his freshman year to build Loopt, a location-sharing app for a world that didn't yet have smartphones. Green Dot acquired Loopt in March 2012 for $43.4M. Respectable. But not the exit he wanted. He then became an angel investor, joined Y Combinator part-time in 2011 under Paul Graham, and by 2014, at 28, he was running the place. Under his watch YC scaled from dozens of companies a batch to hundreds, including Stripe, Airbnb, Coinbase, and Instacart.

His own bets did exceptionally well. He got into Stripe at a $1.7B valuation; it's now north of $65B. He served briefly as interim CEO of Reddit in 2014. He backed Helion Energy and dozens of others. As of 2025, Forbes reported his net worth between $2B and $3B, derived from those non-OpenAI investments (he reportedly holds no direct equity stake in OpenAI, per Wikipedia).

He has become the public face of moving faster. People who have worked with him often describe a different operator inside the building, patient with research disagreements, willing to let strong views surface before he picks a path. Running a company through a governance crisis in the middle of a product moment that will define the decade is the kind of problem no playbook covers. What you learn, if you survive it, is that holding the public story together and holding the internal story together at the same time is not hypocrisy. It is the job. With Altman, both seem to be true at once.

Why he matters to the map: He routes more disconnected groups than anyone else in this network, bridging YC, OpenAI, and Microsoft capital.

Elon Musk, The Outsider Founder

Elon Reeve Musk was born June 28, 1971, to an engineer father, Errol, and a model-and-dietitian mother, Maye. At eight he taught himself to code. At nine he read the Encyclopedia Britannica. Around that time, other children at his Pretoria school pushed him down a flight of stairs and put him in the hospital. He left South Africa at 17, ended up at the University of Pennsylvania studying economics and physics, and in 1995 dropped out of a Stanford energy physics PhD after two days to start Zip2 with his brother Kimbal.

Read full profile ↓

Founded Tesla (Post-corporate · 2003 · age 32). After co-founding and selling Zip2 in 1999 and X.com (which became PayPal) in 2002, Elon Musk had accumulated wealth. He invested $6.5 million in Tesla's Series A funding round in February 2004, became chairman and largest shareholder, and was later designated a co-founder through a legal settlement. Exit: Tesla went public on June 29, 2010, issuing 13.3 million shares of common stock at a price of $17 per share on the NASDAQ, raising $226 million. (Elon Musk | Biography | Research Starters - EBSCO · Tesla, Inc. - Wikipedia)

The exits made him. Compaq bought Zip2 in 1999 for $307M; Elon's cut was around $22M. He co-founded X.com, it merged into PayPal, eBay bought it in 2002 for $1.5B, and Elon walked away with around $165M after taxes. He poured the money into SpaceX and Tesla simultaneously. Both nearly went bankrupt. Neither did. In 2015 he co-founded OpenAI. He left OpenAI's board in February 2018; later reporting and OpenAI's own account describe disputes over control and funding.

xAI was incorporated in March 2023, emerged from stealth in July, launched Grok in November. As of May 2024, it raised $6B at a $24B valuation, built a 100,000-H100 training cluster it calls Colossus, and in March 2025 merged with X Corp at an $80B valuation (per Wikipedia). The merger gave xAI something none of its rivals have: a live, real-time firehose of human writing.

Why he matters to the map: He seeded OpenAI then exited to build xAI, splitting the original founder circle into rival labs.

Dario Amodei, The Safety Maximalist

In 2006, Dario Amodei was at Princeton working on a physics PhD when his father, Riccardo, a leather craftsman from Massa Marittima in Tuscany, died of a rare illness. Dario was 23. The death redirected him. He moved out of theoretical physics, into computational neuroscience, and eventually into machine learning. He has not said much publicly about that pivot. His career since appears to have been shaped by that turning point.

Read full profile ↓

Substack

Founded Anthropic (Post-corporate · 2021 · age 38). Dario Amodei was the Vice President of Research at OpenAI, where he led projects like GPT-2 and GPT-3. He left OpenAI in late 2020 with his sister Daniela and the colleagues who would become Anthropic's seven co-founders, over disagreements about AI safety priorities and the company's direction, then incorporated Anthropic in early 2021. Exit: Still running. As of February 2026, Anthropic had an estimated value of $380 billion. (Anthropic - Wikipedia · What Is Anthropic? | Built In)

He joined Google Brain in 2015, moved to OpenAI in 2016, and rose to VP of Research, the technical center of the company. In early 2021 he and his sister Daniela walked out with five of OpenAI's safety researchers (seven co-founders total) and founded Anthropic (sources vary on whether to date the departure to late 2020 or early 2021, and on whether to count adjacent leavers like Paul Christiano who founded ARC separately; we use the Wikipedia-cited seven-founder figure here). By early 2025 the company had raised about $14.8B. As of February 2026, it was valued at roughly $380B, and by April 2026 annual recurring revenue had reached $30B (per Wikipedia). Amazon has committed $4B and made AWS the preferred cloud. Google has committed $7.3B. They become the hardest operational decisions in the company. The safety lab, as it turns out, is also a very large business.

Why he matters to the map: He led the OpenAI safety walkout that became Anthropic, the cleanest organizational fork in the lab graph.

Demis Hassabis, The Scientist-Operator

At 13, Demis Hassabis was the second-best chess player in the world for his age. At 17, while still in secondary school, he was writing code for Bullfrog Productions, contributing to Theme Park, which sold four million copies. He was born July 27, 1976, in London, to a Greek-Cypriot father and a Chinese-Singaporean mother, and he had taught himself to program on a Commodore 64 at eight.

Read full profile ↓

Founded DeepMind (A few years out · 2010 · age 34). Demis Hassabis co-founded DeepMind in November 2010, a year after completing his PhD in cognitive neuroscience in 2009. He had previously worked as a video game designer and was a chess prodigy. Exit: DeepMind was acquired by Google (now Alphabet Inc.) in January 2014 for a reported price between $400 million and $650 million, with $650 million frequently cited. Value at exit: At acquisition, DeepMind was largely pre-revenue. Google paid for its team of AI scientists and engineers, its research in deep reinforcement learning, and its goal to develop artificial general intelligence (AGI). Google was also blocking a rival bidder, reportedly Facebook. The deal gave Google an edge in AI and a way to apply DeepMind's technology to problems like reducing energy consumption in its data centers. (Demis Hassabis | Biography, Nobel Prize, DeepMind, AlphaFold, & Facts | Britannica · Google DeepMind - Wikipedia)

He earned a double-first (top honors in two parts) in computer science at Cambridge. Founded Elixir Studios, designed games for four years. Then enrolled in a PhD in computational neuroscience at UCL, where he met Shane Legg at a 2009 lecture and reconnected with his childhood friend Mustafa Suleyman. The three of them started DeepMind. Google bought it in January 2014 for around $650M.

In October 2024, Demis Hassabis and John Jumper won the Nobel Prize in Chemistry for AlphaFold, the AI system that predicted the 3D structure of nearly every known protein and resolved a 50-year-old problem in structural biology. He is the first video game designer, the first chess prodigy, and the first AI researcher to win a Chemistry Nobel. He accepted in a dark suit and made no dramatic statements.

Isomorphic Labs, which spun out of DeepMind in 2021 as its own Alphabet company, signed deals with Eli Lilly and Novartis in January 2024 worth up to $3B combined. The first time big pharma had bet at scale on AI drug discovery.

Why he matters to the map: He bridges deep research and Google's capital, the rare scientist running an Alphabet-scale operating unit.

Ilya Sutskever, The Safety Founder

He was born in 1986 in Nizhny Novgorod, Russia, moved to Israel at five, to Canada at sixteen. At the University of Toronto a man named Geoffrey Hinton took him on as a student, and the rest of Ilya Sutskever's life began arranging itself around that fact. His PhD work on recurrent networks and sequence modeling laid foundations the field still rests on. In 2012 he co-authored AlexNet with Alex Krizhevsky and Geoffrey Hinton. In 2013 Google bought their tiny company, DNNresearch, for $44M. He worked at Google Brain, then helped found OpenAI.

Read full profile ↓

Founded Safe Superintelligence Inc. (Post-corporate · 2024 · age 38). Ilya Sutskever left OpenAI in May 2024 after a period of internal disruption, including his vote to terminate CEO Sam Altman in November 2023, a decision he subsequently expressed regret about, per Wikipedia. He announced the founding of Safe Superintelligence Inc. in June 2024, saying he had a "big new vision" better suited to a new company. Value at exit: Safe Superintelligence Inc. reached a $32 billion valuation by April 2025, despite having no revenue or publicly released product. Investors priced in Ilya Sutskever's reputation and track record, the company's singular focus on safe superintelligence, the team it is assembling, and partnerships such as Google Cloud access to TPUs. (Ilya Sutskever - Wikipedia · Safe Superintelligence Inc. - Wikipedia)

Ilya left OpenAI in May 2024. In June he co-founded Safe Superintelligence Inc. with Daniel Gross and Daniel Kokotajlo. SSI raised $1B at a $5B valuation in September 2024, then a further $2B at a $32B valuation by April 2025, with no product, no customers, and a research agenda it has refused to describe in detail (per Wikipedia; sources vary on the precise round structure, we cite the latest reported figures). It has said it will not take revenue until safety is guaranteed. The backers include a16z, Sequoia, and DST Global. They are paying $3B to wait.

Why he matters to the map: He links Hinton's research lineage to OpenAI's founding and to SSI's reputation-driven $32B round.

Yann LeCun, The Contrarian Godfather

In the 1990s, if you wrote a check in the United States, there is a good chance a neural network designed by Yann LeCun read it. He had built convolutional networks at AT&T Bell Labs that could recognize handwriting at scale; the banks adopted them widely.

Geoffrey Hinton, The Conscience

Geoffrey Everest Hinton's great-great-grandfather was George Boole, the man who invented Boolean algebra, the math that underlies every line of code ever written. The lineage is almost too neat. He was born December 6, 1947, in London, studied psychology at Cambridge, finished a PhD in AI at Edinburgh, and spent decades at Toronto, part-funded by Google. He co-authored the 1986 paper with David Rumelhart and Ronald Williams that demonstrated neural networks could learn through backpropagation. He trained the generation that built the industry: Ilya Sutskever, Yann LeCun, Salakhutdinov, and many others. In 2013 Google bought his startup DNNresearch.

Read full profile ↓

Founded DNNresearch Inc. (Second act · 2012 · age 65). Geoffrey Hinton, a University Professor Emeritus at the University of Toronto and a leader in neural networks research, co-founded DNNresearch Inc. with his graduate students after a long academic career. Exit: DNNresearch Inc. was acquired by Google in March 2013 for $44 million. Value at exit: Google bought the company for its deep learning technology, especially the AlexNet neural network that broke through in image recognition, and for the talent of its co-founders, including Geoffrey Hinton and his graduate students Alex Krizhevsky and Ilya Sutskever. Google wanted their expertise and technology to improve its speech and image recognition capabilities. (Geoffrey Hinton - Wikipedia · Geoffrey Hinton | Biography, Nobel Prize, Machine Learning, AI, & Facts | Britannica)

In May 2023 he quit Google so he could say, out loud, that he had come to think the risks were more serious than he had previously believed. His former student Yann LeCun called that "not serious science." His former student Ilya Sutskever called it courageous.

In October 2024 he shared the Nobel Prize in Physics with John Hopfield for foundational work on neural networks. The Physics Nobel Committee's decision to give the prize to AI researchers was itself contested; some physicists argued it wasn't physics. Geoffrey Hinton got the call at 5am in a hotel room and assumed it was a prank. His first wife Rosalind died of ovarian cancer in 1994. He has said her death deepened his interest in how minds store and retrieve memory.

Why he matters to the map: He trained the cohort that built modern deep learning, connecting three of the eight architects through one advisor.

Chapter 3 · The Labs: OpenAI, Anthropic, Google DeepMind

The leading AI labs were built from tight circles where talent, capital, and mission already overlapped.

Late 2015, after the plates were cleared at Sam Altman's house in San Francisco, Greg Brockman, Ilya Sutskever, Elon Musk, Peter Thiel, Jessica Livingston, Reid Hoffman, and others sketched a nonprofit lab that would publish its work and keep AGI aligned with the public good.

On paper, the ambition was moral. Elon Musk and Peter Thiel each committed up to $1B over time. In practice, it was also strategic, a race to assemble rare talent before Google could absorb it.

Elon Musk's actual contribution to OpenAI was approximately $45M, far less than the $1B he originally pledged, as documented in court filings from the litigation between Musk and OpenAI (per Wikipedia); he then left the board. As of 2026, OpenAI was valued at $852B. The nonprofit structure became increasingly strained by the scale and economics of the capped-profit entity. The original circle had split into rival labs, lawsuits, and boardroom trauma.

This chapter tracks the core pattern. OpenAI did not only build models. OpenAI converted relationships into institutional power, then paid the governance cost of that design.

The Founding Dinner (December 2015)

The official history is clean. A mission-first nonprofit would publish research and prevent AGI from being captured by one company. The founding cast included Sam Altman, Greg Brockman, Ilya Sutskever, Elon Musk, Peter Thiel, Jessica Livingston, and Reid Hoffman.

The unofficial history uses the same names and reaches a different conclusion. Publicly, the room described a lab for humanity. Privately, many observers saw a preemptive move in a coming talent war.

Both narratives belong in this chapter. The endpoint is not in dispute. The nonprofit counterweight became one of the most valuable companies in history, and several founders now compete against it.

Greg Brockman, The Builder

Greg Brockman grew up in Grand Forks, North Dakota, dropped out of MIT, became Stripe's first CTO, and then moved into the risk of building a lab that did not yet know what it would become.

Read full profile ↓

Founded OpenAI (Mid-career · 2015 · age 28). Greg Brockman left his role as CTO of Stripe in May 2015, where he had been since 2013 and helped the company grow, to co-found OpenAI in December 2015. Exit: OpenAI is still an active, privately held company. (OpenAI | ChatGPT, Sam Altman, Microsoft, & History | Britannica Money · OpenAI - Wikipedia)

In the founding years, Sam Altman was the public face and Ilya Sutskever was the scientific center. Greg Brockman was the operator who turned intention into institution, hiring, structuring teams, and keeping execution tight while the mission kept changing.

That distinction matters. OpenAI's early mythology is personality heavy. Its survival, and later scale, depended on systems work. Greg Brockman was that systems work.

On November 17, 2023, the board fired Sam Altman. Greg Brockman resigned from the board the same day and helped organize the employee revolt that followed. During the weekend crisis, Anna Brockman spoke at length with Ilya Sutskever; he reversed his vote shortly after. Sam Altman returned on November 22. Wikipedia

Why he matters to the map: He turned the founding circle into a working institution and rallied the staff letter that reversed the board.

Sources: Wikipedia (2025 donations to MAGA Inc. + Leading the Future) · NYT on OpenAI staff letter

The November 2023 Board Crisis, Full Account

The official record begins on November 17, 2023. OpenAI board members Ilya Sutskever, Adam D'Angelo, Tasha McCauley, and Helen Toner removed Sam Altman, stating publicly that he had not been "consistently candid" with the board. Greg Brockman stepped down from the board. After a brief interim moment for Mira Murati, Emmett Shear became interim CEO.

Then Microsoft moved fast. Satya Nadella offered Sam Altman a role and signaled willingness to hire the full team. Within days, nearly 700 of roughly 770 OpenAI employees signed a letter threatening to resign, as reported by The New York Times in November 2023.

By November 22, Sam Altman was back. Ilya Sutskever reversed publicly. Most of the old board exited. Bret Taylor and Larry Summers joined a reconstructed board.

The company survived. The old governance thesis did not.

Motive remains contested. Publicly, the board released little detail. Privately, participants and observers pointed to overlapping causes: disputes over information flow, concern over undisclosed side projects and potential conflicts of interest, and fallout from a Helen Toner paper that, according to reporting at the time, board members felt had favored competitor narratives. WSJ reported on the events. This is one of the central examples in AI where official and unofficial history diverge, but both are necessary to explain what happened.

Andrej Karpathy, The Beloved Teacher

Andrej Karpathy built unusual influence by making difficult ideas legible.

Read full profile ↓

Substack

Founded Eureka Labs (Post-corporate · 2024 · age 38). Andrej Karpathy founded Eureka Labs in 2024, after his second stint as a Senior Research Scientist at OpenAI, which he left in February 2024. Before that, he was the Director of AI at Tesla from 2017 to 2022. He also started creating AI education videos on YouTube in February 2023. Exit: As of May 2026, Eureka Labs is still running, with Andrej Karpathy serving as its founder and CEO. (Andrej Karpathy - Grokipedia · Andrej Karpathy - Wikipedia)

Born in Bratislava in 1985, Andrej Karpathy moved to Toronto as a child, trained at the University of Toronto and UBC, and completed a Stanford PhD under Fei-Fei Li. His work on visual-semantic representations became part of multimodal AI's early foundation.

The timeline is part of the indicator. Andrej Karpathy joined OpenAI in 2015, left for Tesla in 2017 to lead core Autopilot AI work, returned in 2022, left again in February 2024, and founded Eureka Labs that same year.

In a field that often hides behind complexity, clarity became an edge. Andrej Karpathy's explanation layer influenced researchers, founders, students, and executives at once.

Why he matters to the map: He translates frontier research into a public curriculum, widening the talent funnel for every lab on this map.

Sources: Wikipedia · Karpathy website · Bloomberg on Eureka Labs

Jan Leike, The Superalignment Resignee

Jan Leike led the team meant to solve OpenAI's hardest internal contradiction.

Read full profile ↓

Substack

He came from DeepMind and algorithmic information theory, then ran Superalignment at OpenAI, with a direct mandate to keep future systems aligned before capabilities outran control.

In May 2024, two days after Ilya Sutskever left, Jan Leike resigned publicly. He stated publicly that product velocity had outpaced safety commitments, and that promised compute for alignment had not arrived at the level the team needed.

Publicly it was framed as a priority dispute. Many in the field read it as a credibility shock. If the person running superalignment says alignment resources are insufficient, outsiders revise trust immediately. Jan Leike moved to Anthropic, and the move was widely read as a recruiting bellwether across the market.

Why he matters to the map: He turned an internal alignment dispute into a public credibility shock, redirecting safety talent from OpenAI to Anthropic.

Sources: Jan Leike resignation post · The Verge · Anthropic announcement

Alec Radford, The Quiet Genius Behind Every OpenAI Product

Alec Radford built his first computer with his father at five, became an Eagle Scout, enrolled at Olin in 2011, and left in 2014 to work full time at Indico Data Solutions. Before OpenAI, he co-authored the 2015 DCGAN paper with Luke Metz and Soumith Chintala, a foundational result in modern generative modeling.

Read full profile ↓

Founded Indico (In school · 2013 · age 20). Alec Radford co-founded Indico in his dorm room at Olin College with fellow students. He dropped out of Olin College in August 2014 to focus on Indico after the company was accepted into the Techstars accelerator program. Exit: Still running. (Alec Radford - Wikipedia)

Inside OpenAI, Alec Radford became the rare researcher whose name appears on nearly every major line: GPT-1 in 2018, GPT-2 in 2019, CLIP in 2021, then DALL-E, Whisper, and Codex. Sam Altman has publicly described Radford as operating at an exceptional level of research insight, and the breadth of the citation graph agrees with the scale of impact.

In December 2024, he announced he was leaving OpenAI for independent research while continuing collaboration. In February 2025 he was subpoenaed in the Authors Guild and Sarah Silverman copyright litigation, and by April 2025 he was advising Mira Murati's Thinking Machines Lab. In the talent diaspora after OpenAI's governance crisis, Alec Radford became a quiet connector node.

Sources: Wikipedia · TechCrunch on subpoena · The Information on departure

Mark Chen, The Quant Who Became OpenAI's Research Chief

Mark Chen trained at MIT in math and computer science, spent time as a quantitative trader at Jane Street, and carried that rigor into AI research leadership. He also coached the U.S. team for the International Olympiad in Informatics, which tells you something about his operating style, precision, repetition, execution.

Read full profile ↓

He joined OpenAI in 2018 and led key programs including DALL-E, Codex, GPT-4 multimodal work, and the o1 reasoning line. After leadership exits in 2024, he moved up to SVP of Research, sharing oversight with Jakub Pachocki.

In March 2025, he became Chief Research Officer. By mid-2025, external profiles were already framing Mark Chen and Jakub Pachocki as the pair shaping OpenAI's technical future. After the departures of Ilya Sutskever and Bob McGrew, Mark Chen became the highest-ranking research leader still inside the company.

Sources: OpenAI leadership update (March 2025) · MIT Technology Review profile · Core Memory profile

Kevin Weil, The Product Polymath, Twitter to Libra to OpenAI to Out

Kevin Weil's career reads like a guided tour of platform eras: Twitter growth years, Instagram's defensive pivot, Meta's Libra and Novi effort, then OpenAI's enterprise phase. Harvard and Stanford training, plus U.S. Army Reserve service, gave him the profile of a high-trust operator across very different institutions.

Read full profile ↓

Founded Diem (formerly Libra) (Post-corporate · 2018 · age 35). Kevin Weil co-founded Diem (formerly Libra) and Novi, a digital wallet for the Diem network, within Facebook after product leadership roles as Senior Vice President of Product at Twitter and Vice President of Product at Instagram. Exit: The Diem Association, which oversaw the Diem project, sold its intellectual property and other assets to Silvergate Capital for $182 million in January 2022, as Bloomberg reported. Value at exit: The value to Silvergate Capital was the intellectual property, development, implementation, and operations infrastructure, and tools for operating a blockchain-based payment network, which Silvergate intended to use to launch its own stablecoin. (Business Insider)

At Twitter, he helped scale from dozens of people to thousands and through the IPO era. At Instagram, he helped ship Stories under Kevin Systrom and Mike Krieger. At Meta, he co-created Libra and Novi, the payments project that regulators effectively shut down.

Sam Altman recruited him in June 2024 as OpenAI's first CPO alongside CFO Sarah Friar; in 2025 he shifted to lead OpenAI for Science. In October 2025, a deleted post about GPT-5 and unsolved Erdos problems drew a public rebuttal. He exited OpenAI on April 17, 2026 alongside Bill Peebles and Srinivas Narayanan as the company narrowed focus around its core model and enterprise execution.

He has since joined Cisco's board of directors, where his platform-product experience across Twitter, Instagram, Meta, and OpenAI gives him a vantage point on enterprise networking as it absorbs AI.

Sources: OpenAI welcomes CFO/CPO (June 2024) · TechCrunch on April 2026 exits · Cisco board appointment

Extended Roster · 2 more from this chapter

Barret Zoph, Co-Founder. Six years at Google Brain where he co-invented Neural Architecture Search (NAS, ICLR 2017), using RL to automatically design neural architectures. Joined OpenAI 2022 as VP Research leading post-training: RLHF, tool use, multimodality, and all ChatGPT/GPT-4 alignment. GitHub | Sources: Barrett Zoph site TechCrunch

Bill Peebles, Former Head of Sora. UC Berkeley PhD (Alyosha Efros lab). Co-authored 'Scalable Diffusion Models with Transformers' (DiT, ICCV 2023), replacing U-Net with Transformers in diffusion models, enabling predictable scaling. GitHub | Sources: Peebles site Google Scholar GitHub TechCrunch

The pattern is that elite labs do not just train talent, they also produce breakaways. Anthropic shows what the OpenAI network looks like when trust splits and key people leave together.

3.1 Anthropic: The Walkout, the Safety Gospel, and the EA Trust Network

Seven people walked out of OpenAI together in late 2020 and early 2021. They had been the safety team. They left citing disagreements about AI safety priorities and the company's direction, and founded Anthropic, now valued at roughly $380B. Among the cases tracked in this book, the walkout is one of the clearest examples of a research team voting with its feet, and the chapter that follows is about who left, why, who funded them, and how the philosophy they brought with them (effective altruism, character training, the safety-first frame) produced both a major lab and a recurring set of internal debates the rest of the industry now has to take seriously.

The Seven Who Left OpenAI

Seven OpenAI staff resigned in a coordinated window in late 2020 and early 2021. They had been the safety team. Within months they had incorporated Anthropic and raised a $124M Series A.

In early 2021, the seven Anthropic co-founders left OpenAI: Dario Amodei, Daniela Amodei, Tom Brown, Jared Kaplan, Sam McCandlish, Jack Clark, and Chris Olah. A handful of adjacent OpenAI researchers (including Nick Joseph, Paul Christiano, Niki Parmar, and Sholto Douglas) departed in the same window, though not all joined Anthropic; Paul Christiano went on to found the Alignment Research Center. Sources vary on the exact timing and count; we use the Wikipedia-cited seven-founder figure throughout. On the surface, this was a disagreement about pace. At a deeper level, people close to the founders described a genuine strategic divergence: capabilities were accelerating faster than safety work could keep up, commercialization pressure had grown sharply after Microsoft's $1B 2019 investment, and the nonprofit structure looked increasingly mismatched to the ambitions of both sides. What followed was a new lab built around the premise that safety and frontier research belong in the same organization. What the clean narrative skips is the operational cost of that move: seven co-founders resigning together, burning relationships with the organization they helped build, then immediately needing to hire, fundraise, and publish credible research before the window closed. There is no playbook for that. You figure it out while doing it.

Anthropic incorporated in early 2021 and closed a $124M Series A in May 2021, followed by a $580M Series B in April 2022 led by FTX (per Wikipedia). By early 2025, it had raised about $14.8B. As of 2026, it was valued around $380B.

Daniela Amodei, The Builder

Before Anthropic had a polished narrative, Daniela Amodei was already building the machine behind it.

Read full profile ↓

Founded Anthropic (Post-corporate · 2021 · age 34). Before co-founding Anthropic in 2021, Daniela Amodei was the Vice President of Safety and Policy at OpenAI, where she worked from 2018 to 2020. Before that, she was an early employee at Stripe, joining in 2013 and spending five years in roles including risk management and policy development.(Anthropic - Wikipedia · Timeline of Anthropic)

Daniela Amodei, born in 1985 in San Francisco, studied economics at Georgetown and worked in investment banking before joining Stripe as head of business development. She joined OpenAI around 2016 in an operating role, then became Anthropic's President. Research-focused organizations often underinvest in the operational infrastructure that makes research possible. Daniela Amodei built that infrastructure: hiring pipelines, revenue relationships, communications, and the institutional credibility that lets a safety-first lab compete for top talent and serious customers at the same time. The Series A closes and the Series B clock starts immediately. You are building the plane and selling seats on it in the same quarter.

In 2023, Daniela Amodei and Dario Amodei signed the Giving Pledge. She has testified before Congress and joined White House AI safety summits.

Why she matters to the map: She built the operational layer that lets a safety lab compete for talent and revenue at frontier scale.

Chris Olah, The Interpretability Pioneer

Chris Olah made his name by looking inside models when most of the field only cared about outputs.

Read full profile ↓

Founded Anthropic (A few years out · 2021 · age 28). Chris Olah skipped college and received a Thiel Fellowship in 2012 at age 19. He then worked at Google Brain and OpenAI, pioneering mechanistic interpretability, before co-founding Anthropic in 2021.(Anthropic - Wikipedia · What Is Anthropic? | Built In)

He attended the University of Waterloo, dropped out, and became a Peter Thiel Fellow in 2012. Vitalik Buterin joined in 2014 and Laura Deming in 2011, so they were near-cohort, not classmates. Chris Olah's Distill.pub work on neural network visualization became required reading for a generation of researchers.

At Anthropic, Chris Olah's team pushed mechanistic interpretability into core research, monosemanticity, superposition, and circuits. The bet is simple and radical: you cannot govern a system you cannot inspect.

Amanda Askell, Claude's Personality Architect

Amanda Askell's job is to answer a question most labs avoid in public, what kind of mind should an assistant sound like?

Read full profile ↓

She grew up in Scotland, studied philosophy at St Andrews, and completed a PhD in philosophy of mathematics and logic at CUNY. She joined OpenAI in 2019 on alignment work, then left with Anthropic's founding group. At Anthropic she leads Claude's character training, trying to make behavior stable, curious, honest, and less sycophantic under pressure.

Her approach draws on Aristotelian virtue ethics: good behavior comes not from following rules but from cultivating good character. She authored a 30,000-word "soul document" that defines Claude's personality, values, and conversational disposition. The guiding metaphor is that Claude should be like "a well-liked traveler who can adjust to local customs without pandering." The goal is not a compliant assistant but "a genuinely good, wise, and virtuous agent." A 30,000-word character spec is consistent with a philosophical brief, not only a product brief. Daily interactions with Claude appear to carry the imprint of that approach. Big Technology

EA network connection: Amanda Askell's professional community overlaps substantially with that of William MacAskill, Oxford philosopher and co-founder of the Effective Altruism movement, connecting Anthropic's alignment research with the EA philanthropic network. More on Wikipedia

Paul Christiano, The Alignment Theorist

Paul Christiano's influence is often invisible because it sits in training pipelines, not headlines.

Read full profile ↓

Founded Alignment Research Center (A few years out · 2021 · age 30). Paul Christiano graduated from MIT in 2012 and received his PhD from UC Berkeley in 2017. He led the language model alignment team at OpenAI from 2016 until January 2021, then founded the Alignment Research Center in April 2021. Exit: The Alignment Research Center (ARC) is a non-profit research institute and is still running as of the current date. It is a 501(c)(3) tax-exempt charity. (Paul Christiano - Wikipedia · Paul Christiano | NIST - National Institute of Standards and Technology)

He studied mathematics at MIT, earned a theoretical computer science PhD at UC Berkeley, and at OpenAI co-invented RLHF, reinforcement learning from human feedback, now used across major frontier models including GPT-4 and Claude. He left OpenAI in 2021 to found the Alignment Research Center, then in 2023 became head of AI safety at the US AI Safety Institute, helping define how government evaluates frontier model risk.

Jack Clark, The Policy Voice

Jack Clark became influential by translating technical change into policy language before most policymakers realized they needed a translator.

Read full profile ↓

Import AI

Founded Anthropic (Mid-career · 2021 · age 32). Jack Clark was the Policy Director at OpenAI before co-founding Anthropic in 2021. Before OpenAI, he was a technology journalist at Bloomberg.(Jack Clark - Forbes · Jack Clark | Longterm Wiki)

He studied at the University of Bath and was one of OpenAI's earliest employees, where he built communications and policy from scratch. Since 2016, his Import AI newsletter has been one of the field's most consistent briefings for researchers, investors, and regulators. At Anthropic, Jack Clark is Co-Founder and Head of Policy, focused on regulation, government engagement, and international AI governance. Doing that job from inside a frontier lab is a different animal than doing it from a think tank or a government post. You are simultaneously the person asking for regulatory frameworks and the person whose products those frameworks will govern. The credibility requires managing that tension in public, every time.

Mike Krieger, Anthropic's First Product Chief, Instagram's First Engineer

Michel “Mike” Krieger was born March 4, 1986 in São Paulo, Brazil, and moved to California at 18 to study Symbolic Systems at Stanford, the CS+cognitive-psychology hybrid that has produced an outsized share of Silicon Valley product leaders. After Stanford he worked briefly at Meebo before co-founding Burbn with classmate Kevin Systrom in 2010. The two pivoted Burbn into Instagram and launched the photo app in October 2010. Research on product-market fit rarely captures how fast adoption actually moves: Instagram reached one million users in two months. In April 2012 Facebook acquired it for roughly $1 billion in cash and stock, a then-shocking price for a 13-employee startup with no revenue, and evidence that mobile consumer attention was worth more than most balance sheets suggested.

Read full profile ↓

Founded Instagram (Right after school · 2010 · age 24). Mike Krieger moved to California in 2004 to attend Stanford University, where he earned a Bachelor of Science in Symbolic Systems in 2008 and a Master of Science in the same field in 2009. He co-founded Instagram in 2010 with Kevin Systrom, soon after finishing graduate school. They first launched a location-based check-in app called Burbn in 2010, then pivoted to photo-sharing and renamed it Instagram that year. Exit: Facebook acquired Instagram for approximately $1 billion in cash and stock in April 2012. The deal officially closed on September 6, 2012. At the time, Instagram had 13 employees and 30 million users. Value at exit: Facebook bought Instagram for its fast-growing user base, its mobile photo-sharing product, and its team of designers and engineers. Mark Zuckerberg said photo-sharing was a core reason many people loved Facebook, and combining the companies would improve that experience. The deal was also seen as a defensive move against potential acquirers like Twitter. Instagram had no direct revenue at the time, but its user engagement and growth showed market demand and future monetization potential. (Instagram - Wikipedia · The Story of Instagram's $1billion Acquisition - Dealmakers)

Mike Krieger stayed on as CTO and scaled engineering from a handful of people to 450+ while monthly active users crossed one billion. He and Kevin Systrom resigned in September 2018 after clashes with Mark Zuckerberg over Instagram's autonomy, as Fast Company reported. After a quiet stretch, including the COVID era rt.live tracker, Mike Krieger and Kevin Systrom launched Artifact, an AI-driven news app, in January 2023. Yahoo acquired it and shut it down in April 2024.

On May 15, 2024 Mike Krieger said he was joining Anthropic as its first CPO to run product engineering, product management, and design as Claude moved from a research project into a product business. In a lab led mostly by researchers, he became the operator connecting model work to user-facing execution.

Sources: Anthropic announcement · CNBC coverage · Wikipedia

The $500M FTX Stake That Got Sold Too Early

Sam Bankman-Fried invested $500M in Anthropic for approximately 8% of the company. It was one of the largest single venture bets of 2022. Then FTX collapsed, and the bankruptcy estate had no choice: sell the stake, cover the creditors, move on. As of March 2024, they sold for $884M per Bloomberg Law, nearly doubling the original investment. The hard part to sit with is the counterfactual: based on Anthropic's subsequent $380B valuation range, that 8% stake would have been worth roughly $30B. When you are forced to sell a great asset at the worst possible moment, $884M is not a failure of judgment. It is the cost of the other failure. SBF lamented the forced sale from prison. CNBC

Extended Roster · 6 more from this chapter

Sam McCandlish, Co-Founder & Chief Architect, Anthropic. Co-Founder and Chief Architect at Anthropic, one of seven original co-founders who left OpenAI in 2021. Co-authored the influential 2020 scaling laws paper. | Sources: Lifestyles Magazine Google Scholar arXiv Anthropic

Avital Balwit, Chief of Staff, Anthropic (b. 1999). Avital Balwit is Chief of Staff to Dario Amodei at Anthropic. A Rhodes Scholar from Portland, Oregon, she previously researched transformative AI at Oxford's Future of Humanity Institute, worked at the FTX Future Fund, and did grantmaking in AI safety and biosecurity. | Sources: Palladium Avital Balwit site UVA News American Rhodes Jefferson Scholars

David Soria Parra, Staff Engineer, MCP Lead. One of the primary architects of the Model Context Protocol, the emerging standard for how AI agents connect to tools and data sources. MCP is rapidly becoming the USB-C of agentic AI, now supported by OpenAI, Google, and Microsoft. GitHub | Sources: experimentalworks.net SE Daily Anthropic Wikipedia

Justin Spahr-Summers, Engineer, MCP Core. Core contributor to the MCP specification and the open-source SDKs (TypeScript, Python) that developers use to build MCP servers. Previously co-founded GitHub's Atom editor and was an iOS framework author at GitHub, brought rigorous protocol design thinking to MCP from years of building developer tooling. GitHub | Sources: Josh Spahr-Summers site Josh Spahr-Summers site GitHub Anthropic

Mahesh Murag, PM, Claude APIs & MCP. Product lead for the developer-facing Claude platform and MCP adoption. Shaped which MCP capabilities get prioritized through the Tools API launch, Claude's computer use beta, and the MCP specification. | Sources: The Focus AI The New Stack OpenTools AI

Parinaz Firozi, Partnerships, Anthropic. Partnerships leader who most recently led global strategic alliances and AI partnerships at Stripe, with her LinkedIn now indicating an Anthropic affiliation. Previously co-founded Y Combinator-backed Zaam.io (business identity onboarding) as CEO and led partnerships at TrustToken. | Sources: Crunchbase Y Combinator Stripe events

3.2 Google DeepMind: Childhood Friends, Nobel Prizes, and the DeepMind Wars

Scan the founder lists of OpenAI, Anthropic, Inflection, Adept, Character, Cohere, Mistral, Perplexity, Cognition, Sierra, Glean, and the long tail of smaller labs: an outsized share trained at Google, especially Brain and Translate. Even the Transformer paper began there. The question is not whether Google produced breakthroughs; it is why Google kept producing people who produced breakthroughs elsewhere. This section follows the operating conditions that created that diaspora.

DeepMind was acquired by Google in 2014 for around £400M, with publicly negotiated independence protections (an external ethics board) that Google later absorbed. The founders are two childhood friends, Demis Hassabis and Mustafa Suleyman, plus Shane Legg. This chapter is the lineage from AlphaGo's Move 37 to AlphaFold's Nobel Prize, with the Mustafa Suleyman departure in the middle, one of the more legible cases in this cast of how a research lab handles success, internal conflict, and the question of whether a parent company is a host or a buyer.

The DeepMind Founding

Before DeepMind was a brand, it was three people in London who shared an unusually specific conviction: that human-level AI was achievable, and that the right team could get there first.

DeepMind was founded in 2010 by Demis Hassabis, Mustafa Suleyman, and Shane Legg. Demis Hassabis and Mustafa Suleyman were childhood friends through Demis Hassabis's younger brother. Shane Legg entered through UCL's computational neuroscience orbit after meeting Demis Hassabis at a 2009 lecture.

What that conviction actually required, day to day, was three people making each other's research credible before anyone outside London had any reason to believe it. Google bought DeepMind in January 2014 for about $650M, then the largest pure AI research acquisition on record, with independence protections that were heavily publicized (per Wikipedia). In April 2023, DeepMind and Google Brain merged into Google DeepMind under Demis Hassabis.

Jeff Dean, The Infrastructure God

Jeff Dean is the quiet constant behind multiple AI eras.

Read full profile ↓

Founded Google Brain (Mid-career · 2011 · age 42). Jeff Dean joined Google in mid-1999 and had already co-invented core Google infrastructure like MapReduce (2004) and Bigtable (2005) before co-founding Google Brain in 2011, about 12 years into his time at Google. Exit: Google Brain, an internal research team, merged with Google's DeepMind in April 2023 to form Google DeepMind, combining the company's AI research efforts. Value at exit: As an internal project, Google Brain's value at the time of its merger into Google DeepMind was its research in large-scale deep learning, its development of AI infrastructure like TensorFlow, its contributions to Google products (e.g., Google Translate, Android speech recognition, YouTube recommendations), and its talent. The merger aimed to combine those strengths with DeepMind's to accelerate AI advances and build more capable and responsible AI systems. (Google Brain - Wikipedia)

He holds a BS and PhD from the University of Washington. At Google, Jeff Dean co-designed MapReduce, BigTable, Spanner, TensorFlow, and TPU infrastructure, systems that made modern large-scale AI economically possible. He led Google Brain through the deep learning surge and now serves as Chief Scientist of Google DeepMind.

The Jeff Dean jokes became folklore because they pointed at something real. The combination of first-order technical range and institution-building ability, across four decades of infrastructure that other people build careers on top of, is genuinely rare. The harder thing to see from outside is the sustained judgment it takes to stay relevant across that many approach shifts inside a single company: not just technically current, but trusted enough that the next generation of researchers still wants you in the room.

Why he matters to the map: He concentrates four decades of Google's AI infrastructure under one technical reputation, anchoring the DeepMind merger.

Neel Nanda, The Interpretability Heir

Neel Nanda represents a newer archetype in AI: the researcher who treats public communication as part of the scientific work itself, combining original interpretability research with open tooling and unusually honest writing about how the field actually operates.

Read full profile ↓

He grew up in the UK, studied mathematics at Cambridge, and entered the AI safety and rationalist communities early. He worked under Chris Olah at Anthropic on mechanistic interpretability, then moved to Google DeepMind to lead interpretability work there. His TransformerLens library became a standard way to inspect transformer internals, and his writing at neelnanda.io helped normalize open technical discourse around safety.

Oriol Vinyals, The Catalan co-lead of Gemini, AlphaStar's Architect

Oriol Vinyals was born in 1983 in Sabadell, Catalonia, Spain, and trained as a mathematician and telecom engineer at UPC Barcelona before moving to California for an MS at UC San Diego and a 2013 PhD at UC Berkeley under speech recognition pioneer Nelson Morgan. He joined Google Brain straight out of Berkeley, where in 2014 he co-invented the sequence-to-sequence (seq2seq) architecture with Ilya Sutskever and Quoc Le, one of the most cited deep learning papers of the decade and a direct intellectual ancestor of the Transformer and modern LLMs.

Read full profile ↓

He moved to DeepMind in London and led the AlphaStar project, the AI system that in 2019 reached Grandmaster level in StarCraft II, beating top human pros in a game long considered AI's grand challenge after Go (results published in Nature). He also co-authored the seminal knowledge distillation paper with Geoffrey Hinton and Jeff Dean (2015). His Google Scholar citations exceed 200,000.

After the 2023 merger of Google Brain and DeepMind, Oriol Vinyals was appointed one of three technical leads of Gemini, alongside Noam Shazeer and Jeff Dean, reporting to Demis Hassabis and Koray Kavukcuoglu. He became the public face of Gemini 2.0's agentic and multimodal turn through 2025. MIT Technology Review named him to its TR35 list in 2016. He is the most decorated active researcher inside DeepMind's flagship model effort, the technical brain alongside Demis Hassabis's strategic one.

Sources: Wikipedia · Google Research profile · DeepMind podcast on Gemini 2.0

Ian Goodfellow, The GANfather, From a Montreal Bar to DeepMind

Ian Goodfellow grew up in California, graduated from San Dieguito Academy in Encinitas in 2004, and completed both his BSc and MSc at Stanford under Andrew Ng before moving to Montreal for a PhD in deep learning under Yoshua Bengio and Aaron Courville, defending in February 2015. In 2014, during a celebration at Les 3 Brasseurs in Montreal for a fellow PhD student's graduation, Ian Goodfellow conceived Generative Adversarial Networks, a two-network adversarial training scheme, went home that night, coded the prototype, and it worked. The resulting NeurIPS 2014 paper “Generative Adversarial Nets” is one of the most influential ML papers of the decade and earned him the nickname “the GANfather.”

Read full profile ↓

Ian Goodfellow joined Google Brain after his PhD, then OpenAI as a research scientist in 2016 (a founding-era hire), then returned to Google Brain. He also co-authored the textbook “Deep Learning” (2016, MIT Press) with Bengio and Courville, the field's standard reference. In 2019 he joined Apple as Director of Machine Learning in the Special Projects Group.

He resigned from Apple in April 2022 in a widely covered protest against the company's return-to-office mandate, then joined Google DeepMind as a research scientist (now Principal Scientist), where he has worked on LLM factuality and on the open-source TORAX plasma physics simulator for fusion research.

Sources: Wikipedia · MIT Technology Review, The GANfather · TNW, Ian Goodfellow leaves Apple · Deep Learning textbook

Extended Roster · 6 more from this chapter

Thomas Kurian, CEO, Google Cloud (b. 1966). CEO of Google Cloud since 2019. Former President of Oracle (22 years, led 35,000-person engineering org). | Sources: Wikipedia Google Cloud

Shane Legg, Co-Founder & Chief AGI Scientist, Google DeepMind (b. 1975). Met Demis Hassabis at a 2009 UCL lecture on Gatsby computational neuroscience. Their shared obsession with AGI formed the foundation of DeepMind. | Sources: Wikipedia Time

Zack Kenton, Staff Research Scientist (b. 1991). Specializes in AI safety, alignment, and the challenge of supervising superhuman AI systems. Co-authored key work on scalable oversight and reward modeling. GitHub | Sources: Zack Enton site Google Scholar

Chi Wang, Senior Staff Research Scientist, Google DeepMind / Creator AutoGen. Creator of AutoGen, the multi-agent LLM framework that won Best Paper at the ICLR 2024 LLM Agents Workshop. Principal Researcher at Microsoft Research (2014-2024) where AutoGen originated; left Microsoft 2024 to continue evolving the project as community-driven AG2 (>1M monthly downloads), while joining Google DeepMind. GitHub | Sources: GitHub arXiv Google Scholar

Lada Adamic, Former Director Research, Meta Core Data Science. Lada Adamic is an American network scientist whose foundational research on information diffusion, weak ties, and the structure of online networks has shaped computational social science. After serving as an associate professor at the University of Michigan, she joined Facebook/Meta where she led the Computational Social Science team within Core Data Science, producing influential studies on how information and behaviors spread through large-scale digital networks. BlueSky Lada Adamic site | Sources: Wikipedia Lada Adamic site Google Scholar Meta Research

Sanjay Ghemawat, Google Senior Fellow · Co-Author MapReduce, BigTable, TensorFlow (b. 1966). Sanjay Ghemawat is a Google Senior Fellow in the Systems Infrastructure Group, working in close partnership with Jeff Dean. His foundational work co-architected MapReduce, the Google File System, Bigtable, Spanner, and TensorFlow, shaping modern large-scale distributed computing. GitHub | Sources: Wikipedia Google Research ACM GitHub

The Bidding War

Before Google acquired DeepMind, Demis Hassabis visited Mark Zuckerberg's home for dinner. It was a courtship meeting, Facebook wanted to buy DeepMind. As reported by OfficeChai, Demis Hassabis steered the conversation toward AI; when Zuckerberg showed equal enthusiasm for VR, AR, and 3D printing, Hassabis decided Facebook was not the right home. Larry Page later made his pitch during a walk through the grounds of a castle: "Why don't you take advantage of what I've already created?", meaning Google's compute infrastructure, data, and research talent. Google paid between $400M and $650M. The pattern, per the reported account, is that Hassabis appears to have applied a values filter before a valuation filter. OfficeChai

Mustafa Suleyman's DeepMind Departure

In August 2019, Mustafa Suleyman, DeepMind's co-founder and head of applied AI, was placed on administrative leave following staff complaints about management conduct and a hostile workplace, as reported by Bloomberg and The Wall Street Journal. Google retained an external law firm to investigate; CNBC reported that settlements with some affected staff were reached, the terms of which were not disclosed, and his management duties were stripped. He has publicly acknowledged he "drove people too hard." He was not formally adjudicated of wrongdoing. He went on to co-found Inflection AI in March 2022 with Reid Hoffman and Karén Simonyan (raising $1.5B total by 2023, per Wikipedia), then in March 2024 was hired by Satya Nadella as CEO of Microsoft AI. CNBC

Move 37, The Move That Made DeepMind Famous

Move 37 is the move DeepMind's AlphaGo played against Lee Sedol on March 10, 2016, in Seoul. It is the moment DeepMind became famous outside of research circles, and it is the moment most of the people in this book point to when they explain why they started paying serious attention to deep learning.

The context: a five-game exhibition match. Lee Sedol was an 18-time world champion. AlphaGo was the Go-playing system Demis Hassabis's DeepMind, then a young Google subsidiary, had been building since 2014, led on the engineering side by Koray Kavukcuoglu's deep learning team. In Game 2, AlphaGo played a shoulder hit on the fifth line. No human player in recorded history had played that move in that position; the commentators went silent on air.

Lee Sedol was on a smoke break when AlphaGo played it. When he returned to the board and saw the move, he left the room again for 15 minutes. He came back and lost.

Game 4 produced the human response. Lee Sedol's "God's Touch", Move 78, was a wedge insertion that AlphaGo's value network had assigned a 1-in-10,000 probability. AlphaGo began making errors. Lee Sedol won. It remains the only game a human has ever won against AlphaGo across all 74 of its official matches (per Wikipedia). Lee Sedol retired from professional Go in 2019, saying: "With the advent of AI in Go games, I've realized that I'm not at the top even if I become the number one." That sentence is often read as defeat. It is also a rare acknowledgment of what mastery actually requires: an honest accounting of the ceiling.

For DeepMind, Move 37 was the public proof that funded everything after. Within three years the same lineage of architectures was solving protein folding (AlphaFold). Seven years later, in April 2023, DeepMind absorbed Google Brain to become Google DeepMind. The line from a Seoul hotel ballroom in 2016 to that merger announcement runs straight through this move. Wikipedia

Liang Wenfeng, The Hedge Fund Quant Who Outshipped OpenAI

OpenAI, Anthropic, and Google DeepMind do not exhaust the frontier-lab story. The clearest counterexample is DeepSeek, built outside the US lab pipeline entirely.

Liang Wenfeng was born in 1985 in Guangdong, China, and built High-Flyer, a quantitative hedge fund, into one of China's top quant shops with reportedly thousands of NVIDIA A100 GPUs stockpiled before US export controls. He founded DeepSeek as a separate AI lab in 2023. DeepSeek-V3 (December 2024) and DeepSeek-R1 (January 2025) demonstrated frontier-level reasoning on a reported sub-$10M training budget, open-weighted, and caused a one-day approximately $1T market cap drop across US AI infrastructure stocks on January 27, 2025, as reported by Forbes and The Guardian. Liang Wenfeng rarely gives interviews; his Caijing essays on engineering-first culture and the limits of "$100M cluster" thinking are widely circulated inside Chinese labs and quoted by Western researchers. What is easy to miss is that the DeepSeek result was not primarily a story about frugality. It was a story about what a team does differently when it cannot buy its way past a problem.

Read full profile ↓

Founded High-Flyer (A few years out · 2015 · age 30). Liang Wenfeng graduated with a Master of Engineering in 2010. After graduation, he spent several years experimenting with applying AI to various fields, eventually focusing on finance. He co-founded Hangzhou Yakebi Investment Management in 2013 and Zhejiang Jiuzhang Asset Management in 2015, which are considered part of the High-Flyer group of companies. Exit: High-Flyer, a quantitative hedge fund, is still actively running and manages a significant amount of assets. It serves as the primary funder and backer of DeepSeek, an AI firm spun out in 2023. (Liang Wenfeng - Wikipedia · The Man Behind DeepSeek, How Liang Wenfeng is Reshaping AI - Medium)

Sources: Wikipedia · Liang Wenfeng · Wikipedia · DeepSeek · NYT, Jan 27 2025 sell-off

Chapter 4 · How Power Moves: Talent Pipelines, Diaspora, and the Safety Architects

A small number of institutions produced a disproportionate share of the people now leading modern AI. The clearest university example is University of Toronto: among the 9 alumni captured in this dataset, 6 are labeled tier-1 or tier-2 (a within-cast share of 67%; n=9, 95% CI roughly 35-88%, so treat as suggestive rather than definitive). Most of the tie traces back through the Geoffrey Hinton lineage. This matters because the field did not spread evenly across the schools and labs the public record covers. It moved through a narrow trust layer of advisors, collaborators, and employers that kept sending talent into the next important company.

The transformer paper is the cleanest corporate example of that pattern. In June 2017, eight Google Brain researchers posted "Attention Is All You Need" to arXiv, then presented it that December at NeurIPS in Long Beach to a room that was not packed, with no applause. By 2023 it was the most cited paper in the history of machine learning, the foundation of GPT, of Gemini, of Claude, and none of the eight authors were still at Google. That fact helps explain how one lab became a launchpad. The same cohort carried a shared research language into very different companies, products, and scientific bets.

What follows starts with the talent-factories chart, then moves to eight profile cards from the "Attention Is All You Need" cohort. The chart shows the pattern at the institution level. The profiles show what one specific group did with it after the paper became the foundation of a field. Some built enterprise companies, some built consumer products, some pushed the architecture toward biology, Tokyo, or blockchains. The hard part, when you have written the foundational paper, is not deciding to leave. It is deciding which version of the future you have enough conviction to build.

The Talent Factories: Where AI Leaders Come From

Stanford appears in 77 biographies in this dataset, more than any other institution. The within-cast composition pattern that stands out is alumni concentration vs raw volume. University of Toronto has 9 alumni in the dataset and 6 of them are labeled tier-1 or tier-2 (within-cast share 67%; n=9, 95% CI roughly 35-88%, treat as suggestive). The cluster is plausibly explained by the Geoffrey Hinton lineage, but with n=9 the rate alone does not establish that conclusively. Princeton has 11 alumni in the dataset and 3 of them are labeled tier-1 (within-cast share 27%; n=11, 95% CI roughly 10-57%, too noisy to claim as the highest US tier-1 density). ENS Paris nearly ties Cambridge in this sample and produced all of Mistral AI’s co-founders. Three of the eight “Attention Is All You Need” co-authors hold USC PhDs, a fact almost nobody in the industry notices.

On the corporate side, Google (Brain + DeepMind combined) is the single most dominant talent incubator in AI, having seeded Anthropic, Mistral, Cohere, Sakana AI, Inceptive, and dozens of other frontier ventures. OpenAI is the second-most potent launchpad: nearly every alumnus who has left has founded or co-led a frontier lab. The December 2020 exodus alone produced Anthropic at a $380B valuation.

The advisor-to-student lineage explains more than the company chart does. If you redraw the map by PhD lineage rather than by employer, the influence concentrates around a handful of advisors whose students appear throughout this book. The table below traces the five most consequential mentorship chains in the dataset.

The Talent Flow: How Schools, Labs, and Spinouts Connect

The mentorship table above shows the advisor-to-student lineage. The diagram below shows the same pattern at the institutional level: where the top universities in this dataset sent their alumni first (the frontier labs), and where those labs sent them next (the spinouts). Toronto sits at the top because its 9 alumni include the highest concentration of frontier-lab founders by school size. OpenAI is the largest single funnel into a spinout, the December 2020 walkout that produced Anthropic. Google (Brain + DeepMind combined) is the largest absolute outflow.

Three observations read directly off this picture. The single largest documented power transfer in the dataset is the OpenAI to Anthropic line, 11 people in one event. The second is Google's outflow to a half-dozen separate spinouts, which makes Google the most prolific single corporate talent factory in modern AI. The third is the Toronto effect at small scale: the school with the fewest alumni in the top 15 has produced an outsize share of frontier-lab founders, almost entirely through the Geoffrey Hinton lineage.

Ashish Vaswani, The Lead Author Who Went to Enterprise

The paper lists authors alphabetically, which puts Ashish Vaswani first. He grew up in India, did his undergraduate at IIT, took his PhD at USC, and joined Google Brain in 2014. He stayed five more years after the paper went up. In 2022 he left to co-found Essential AI with Niki Parmar, his co-author from the original eight. The company builds enterprise AI systems for business workflows. It raised $56.5M Series B in November 2023.

Read full profile ↓

Founded Essential AI (Post-corporate · 2023 · age 37). Ashish Vaswani co-founded Essential AI after serving as a Staff Research Scientist at Google Brain (2016-2021), where he co-authored the 'Attention Is All You Need' paper, and then co-founding Adept AI Labs and serving as Chief Scientist (January 2022 - November 2022). Exit: Essential AI is still running and has raised funding, including a $175 million Series B round in August 2025 at a $1 billion valuation. (Ashish Vaswani - Wikipedia · Essential AI Raises $56.5M Series A to Build the Enterprise Brain - Business Wire)

For the lead author on the paper that reorganized the field, Essential AI is a deliberately focused bet rather than a splashy one. Fortune 500 buyers, governance features, deployments built to last rather than to make headlines. What stands out is the clarity of conviction: Vaswani understood what enterprise AI actually requires, and built toward that instead of toward visibility. Building for enterprise when the consumer demo is capturing all the attention takes a specific kind of operator discipline. The customer is harder to reach, the sales cycle is longer, and nobody writes the headline. Vaswani built toward it anyway.

Noam Shazeer, The Researcher Google Paid $2.7B to Get Back

Among machine learning researchers, Noam Shazeer is the name other researchers say with a kind of quiet reverence. He invented Mixture-of-Experts scaling, the architectural trick now sitting under GPT-4, Gemini, and Mistral. Inside Google for over twenty years, he was the engineer whose code other engineers studied to learn how to write code.

Read full profile ↓

Why he left, and how he came back: At Google, Shazeer and Daniel de Freitas built a chatbot called Meena that they believed should be released publicly. When Google declined, Shazeer walked away from a 20-year career to co-found Character.ai, essentially because he wanted people to be able to talk to AI characters. The product became the second-most-used chatbot after ChatGPT, and in 2024 Google paid $2.7B to license the technology and bring him back. He had previously, in 2017, been lead author on "Attention Is All You Need", the paper that introduced the transformer architecture now beneath every major AI system, and pioneered the Sparsely-Gated Mixture-of-Experts approach. Wikipedia

Founded Character.AI (Post-corporate · 2021 · age 45). Noam Shazeer joined Google in 2000, stayed for over 20 years, co-authored the Transformer paper, and developed the Meena chatbot with Daniel de Freitas. He left in 2021 and co-founded Character.AI after Google declined to release Meena publicly. Exit: In August 2024, Google signed a $2.7 billion non-exclusive licensing deal with Character.AI for its generative AI technology. As part of the agreement, Noam Shazeer returned to Google to co-lead the Gemini AI project, while Character.AI remains independent. Value at exit: The $2.7 billion licensing deal shows the value of Character.AI's generative AI technology and underlying models. The agreement also brought Noam Shazeer and Daniel de Freitas back to Google. (Character.ai - Wikipedia · Noam Shazeer | AI Scientist, Google Gemini Co-Lead)

In 2021 he left to co-found Character.AI, an AI companionship and roleplay platform. As of 2024, it had 20 million daily active users, raised $150M at $1B in March 2023, then $2.7B at $5B in October 2023. The product was wildly engaging and structurally hard to monetize.

In August 2024, Google paid Character.AI roughly $2.7B in licensing and employment agreements to re-hire Noam Shazeer and his team. The company stayed nominally independent. Shazeer returned as VP Engineering and Gemini co-lead. The structure, billions paid without an acquisition, mirrored the Inflection and Microsoft arrangement earlier that year and became a template for talent deals built to avoid full antitrust review.

His arc makes one point clear. At the frontier, the scarcest asset is not distribution or compute. It is a handful of researchers who can still move the curve.

Aidan Gomez, The Intern Who Founded Cohere

In the summer of 2017, while the paper was being finalized, Aidan Gomez was an undergraduate at the University of Toronto visiting Google Brain on an internship. He was 20. His name went on the paper.

Read full profile ↓

Founded Cohere (In school · 2019 · age 26). Aidan Gomez co-founded Cohere in 2019 while he was a PhD student in Computer Science at the University of Oxford, which he had started in 2018. He completed his undergraduate degree at the University of Toronto in 2018 and interned at Google Brain in 2017, where he co-authored the Transformer paper. Exit: Cohere is still an independent, privately held company. As of May 2026, it is operational and was valued at $5.5 billion in July 2024 and $7 billion in September 2025. Aidan Gomez has said Cohere is "not for sale." (Aidan Gomez - Wikipedia)

In 2019, with Nick Frosst and Ivan Zhang, he co-founded Cohere, one of the most enterprise-focused major LLM companies. No consumer chatbot, no toy demos. The pitch is data privacy and on-premises deployment for Fortune 500 customers that cannot route sensitive data through a third-party API. Cohere raised $500M at a $5.5B valuation in July 2024.

Of the eight authors, Aidan Gomez is the one who started the youngest and built the most explicitly B2B. His Cohere bet has held its shape through three hype cycles.

Jakob Uszkoreit, Linguistics as Inheritance, RNA as Application

His father, Hans Uszkoreit, is one of Germany's best-known computational linguists, a founding director at the German Research Center for AI in Saarbrücken. Jakob Uszkoreit grew up in that household, joined Google Brain, and contributed to the Transformer architecture.

Read full profile ↓

Founded Inceptive (Post-corporate · 2021). Jakob Uszkoreit co-founded Inceptive in 2021 after 13 years at Google, where he did deep learning research at Google Brain, built the language understanding team for Google Assistant, worked on Google Translate, and co-authored 'Attention Is All You Need'. Exit: Inceptive is still running and has raised $120 million, including a $100 million Series A round in September 2023. (The big interview: Inceptive's Jakob Uszkoreit on the promise of biological software · Jakob Uszkoreit Biography | Booking Info for Speaking Engagements)

In 2021 he left to found Inceptive, applying transformers not to language but to RNA. The premise: if attention can learn the structure of English, it can learn the structure of nucleotide sequences. Inceptive raised $100M Series B from a16z Bio, NVIDIA, and others.

It is the clearest case in the chapter of the Transformer escaping its original scope. The architecture his father's field produced is now designing molecules.

Llion Jones, The Welshman in Tokyo

Of the eight, Llion Jones is the one who left the bay. In 2023 he co-founded Sakana AI, "fish" in Japanese, in Tokyo, with David Ha, the former lead of Google Brain Japan. The company's pitch is "nature-inspired AI," learning approaches that draw on evolutionary biology and emergence rather than brute-force scaling. Sakana raised $30M Series A in 2023 and attracted attention for research on model merging.

Read full profile ↓

Founded Sakana AI (Post-corporate · 2023 · age 39). Llion Jones joined Google in 2011 as a software engineer working on YouTube. In 2015, he moved to Google Research, where he co-authored 'Attention Is All You Need' in 2017. After 12 years at Google, he left in 2023 to co-found Sakana AI. Exit: Still running. Sakana AI was founded in July 2023 and later raised a Series B round in November 2025 that valued the company at approximately $2.65 billion. (Sakana AI - Wikipedia · Llion Jones - Venture Café Global Institute)

The Tokyo address is deliberate. Japan is investing aggressively in sovereign AI capability and is the largest pool of advanced research talent outside the US-China axis. Llion Jones is an early indication that the next decade of frontier AI may not all happen between Mountain View and Hangzhou.

Łukasz Kaiser, The Author the Chapter Was Named For

Łukasz Kaiser arrived at deep learning the long way around. Born in Wrocław in 1981, he studied mathematics and computer science there, then earned his PhD at RWTH Aachen in logic and automata theory before taking a tenured CNRS research position in Paris. That detour through formal logic turned out to be an asset: the precision of theoretical computer science is exactly what the architecture work at Brain required.

Read full profile ↓

Kaiser joined Google Brain in 2013, shortly after the team was founded. At Brain he co-built TensorFlow, then co-created the Tensor2Tensor framework (with Aidan Gomez) that powered the experiments behind the 2017 NeurIPS paper “Attention Is All You Need.” He is the seventh listed co-author on that paper, between Aidan Gomez and Illia Polosukhin, completing the “Transformer Eight” roster: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, Illia Polosukhin. He also later worked on Trax and the Reformer (efficient long-context Transformer).

Kaiser left Google in 2021 to join OpenAI, where he became a core contributor on GPT-4, GPT-5, and the reasoning model line codenamed o1 and o3, making him one of very few people present at the creation of both Transformers (2017) and modern reasoning models (2024). He is unusual among the Transformer Eight in that he stayed in research rather than founding a startup.

Sources: Attention Is All You Need (arXiv) · Wikipedia, Attention Is All You Need · Google Scholar profile

4.1 The Academic Roots: Stanford, Berkeley, and the Canadian Mafia

Most of the people in this book were trained by a much smaller set of people in this chapter. Stanford and Berkeley produced the founders; Toronto produced the godfathers; CMU produced the engineers. If you redraw the AI industry by PhD advisor instead of by employer, the influence concentrates around perhaps a dozen nodes, and you can guess where most of the rest of the cast first sat in a lab meeting. This chapter is about those nodes (Fei-Fei Li, Pieter Abbeel, Andrew Ng, Yoshua Bengio) and the system they built that determined who got to do this work.

The Canadian Mafia: Three Godfathers, Three Disagreements

In 2018, three men shared the Turing Award for their work on deep learning. Geoffrey Hinton at Toronto. Yoshua Bengio at Mila and the Université de Montréal. Yann LeCun at NYU and Meta. For two decades they had been the stubborn minority at conferences, the ones who kept arguing that neural networks would work if you gave them enough scale and data. Careers were built on proving them wrong. Then the field caught up, and in 2018 the Turing committee made it official.

Five years later, they could no longer agree on whether they had built a wonder or a weapon.

Yoshua Bengio underwent what he has called a safety conversion starting around 2019. He began testifying before governments on multiple continents, founded the International AI Safety Report process for the UK and UN, and has spoken about a moral responsibility for the technology he helped create. Geoffrey Hinton left Google in May 2023 specifically so he could speak about existential risk without organizational constraint. Yann LeCun considers this whole frame dangerous scaremongering, says current systems are no closer to existential danger than a toaster, and posts saying so on X most weeks.

The three people who built the intellectual foundation of the modern AI industry now represent the full spectrum of opinion about its consequences. That the field's own architects hold such divergent views is not a sign of fracture but of the genuine difficulty of the question they are all trying to answer. What rarely gets said is what the years before 2012 actually cost. Holding a research conviction when the field has decided you are wrong is not a philosophical position. You lose funding cycles, graduate students choose safer advisors, and conference reviewers stop engaging with your papers on the merits. Most researchers find a way to rationalize a pivot. These three did not.

Fei-Fei Li, The Vision Pioneer

She was sixteen when she landed in Parsippany, New Jersey, in 1992, speaking almost no English. Her parents were doing the kind of work people who used to be middle-class in Beijing do in suburban New Jersey: her father at a camera shop, her mother as a cashier. Fei-Fei Li worked the front desk of a Chinese restaurant on weekends. She got into Princeton.

Read full profile ↓

Founded World Labs (Mid-career · 2024 · age 48). Fei-Fei Li co-founded World Labs in 2024 while serving as the inaugural Sequoia Professor in the Computer Science Department at Stanford University and a Founding Co-Director of Stanford's Human-Centered AI Institute. She previously served as Vice President at Google and Chief Scientist of AI/ML at Google Cloud from 2017-2018. Exit: World Labs is still running. In 2024, it raised $230 million at a valuation of over $1 billion. As of 2026, World Labs had raised a total of $1.23 billion, with a valuation of $5 billion. (Fei-Fei Li - Wikipedia · About - World Labs)

By 2007, on the faculty at Princeton and then Stanford, she made what looked like a strange decision. She told her students they should not work on better algorithms. They should work on a dataset. Fifteen million labeled images across 22,000 categories. She called it ImageNet.

Nobody wanted to grade it. The labels had to be done by humans, and humans were expensive. She found Amazon Mechanical Turk and used it at scale before academia had a vocabulary for that. The grant agencies thought the project was a waste of money. She kept building.

In 2012 a paper from Toronto, by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, swept the ImageNet competition by a margin nobody had ever seen. The architecture was the deep convolutional neural network. The benchmark was Fei-Fei Li's. Without ImageNet, AlexNet's victory would have had no proving ground. Without that proving ground, deep learning would have taken years longer to be taken seriously.

She finished her PhD at Caltech, became director of Stanford's AI Lab, served as Chief Scientist at Google Cloud from 2017 to 2018, returned to Stanford, and in 2023 founded World Labs, working on spatial intelligence, 3D understanding, and video generation. World Labs raised $230M Series A at $1B valuation in October 2024, led by a16z and Radical Ventures with Google Ventures and Patrick Collison investing.

She is married to Silvio Savarese, formerly a computer vision professor at Stanford, now Chief Scientist at Salesforce, which puts two senior enterprise-AI research leads in the same household. Her mother's sacrifices, the multiple jobs after immigration, are a reference point she returns to often in public speeches about who gets to build the future.

Andrew Ng, The Teacher

Most fields produce knowledge and then wonder who will use it. Andrew Ng reversed that sequence. His Machine Learning course on Coursera has drawn over 5 million enrollments, more students than have ever attended Stanford in person, making it the most-taken AI course in recorded history.

Read full profile ↓

Founded Coursera (Mid-career · 2012 · age 36). Andrew Ng co-founded Coursera in 2012 while serving as an adjunct professor at Stanford University, a decade after receiving his PhD in 2002. Before Coursera, he led the development of Stanford's Massive Open Online Course (MOOC) platform and taught an online machine learning class to over 100,000 students, which led directly to Coursera. Exit: Coursera went public on March 31, 2021, listing on the NYSE under the ticker 'COUR'. The IPO priced at $33.00 per share, raised approximately $483.9 million in gross proceeds for the company, and valued it at about $4.3 billion. The company was not profitable at the time of IPO, reporting a net loss of $66.8 million in 2020, alongside revenue of $293.5 million in the same year. (Andrew Ng - Wikipedia · Coursera - Wikipedia)

He was born in London to Singaporean and Chinese parents, grew up between Hong Kong and Singapore, did his BS at Carnegie Mellon, MS at MIT, PhD at UC Berkeley. He co-founded and ran Google Brain from 2011 to 2012, then went to Baidu as Chief Scientist from 2014 to 2017 to build out one of the world's most advanced AI research labs in Beijing.

In 2012, with his Stanford colleague Daphne Koller, he co-founded Coursera. In 2017 he founded DeepLearning.AI and the AI Fund venture studio. His weekly newsletter, "The Batch," reaches hundreds of thousands of practitioners.

His career is the answer to the question "What does an AI researcher do when their goal is reach instead of profit?" Most of the field's working engineers learned from him. Getting to 5 million enrollments required decisions that had nothing to do with research: platform infrastructure, pricing models, certificate credibility, university partnership politics, and a MOOC business model that nobody had proven at scale. The teaching was the easy part. The hard part was making the course available to someone in Lagos or Manila who needed it to get a job, and making sure they actually finished it.

Wikipedia reports that Andrew Ng was previously married to Carol Reiley, a robotics researcher and co-founder of drive.ai (acquired by Apple in 2019)

The Berkeley Robotics Cluster

If the Canadian Mafia gave AI its theoretical spine, the Berkeley cluster gave it a body. What emerged from that campus is less a school of thought than a genealogy of labs, spinouts, and tenure lines all pointed at the same question: how do you teach a physical machine to learn from experience?

Pieter Abbeel took his PhD at Stanford under Andrew Ng. He is Belgian by birth, the godfather of robot learning from demonstration, and the rare academic who can hold a tenured chair at Berkeley and run an industry research function at the same time. He co-founded Covariant; in August 2024 Amazon hired the co-founders and licensed Covariant's robotic foundation models in a reverse acqui-hire (approximately $380M plus a $20M licensing payment, per a 2025 whistleblower complaint; TechCrunch). He joined Amazon that same year as VP of Frontier AI and Robotics.

Sergey Levine did his PhD at Stanford. In 2023 he co-founded Physical Intelligence (also written π) with Chelsea Finn and Brian Ichter. It is the first robotics foundation model company. They raised $400M at $2.1B in November 2023, the largest early-stage robotics raise on record.

Chelsea Finn did her BS at MIT and her PhD at Berkeley. She was among the first researchers to demonstrate meta-learning in robots, the capacity for a system to learn how to learn. She is a Stanford professor and a co-founder of Physical Intelligence at the same time, which used to be impossible and is now becoming, in AI, the default.

Daphne Koller, From Probabilistic Graphical Models to AI Drug Discovery

Daphne Koller was born August 27, 1968 in Jerusalem. She finished her bachelor's by 17 at Hebrew University and her PhD at Stanford by 24 (1993, advisor Joseph Halpern), joining the Stanford CS faculty in 1995. With Nir Friedman she co-authored the canonical textbook “Probabilistic Graphical Models” (MIT Press, 2009). She won the MacArthur Fellowship in 2004 and was elected to the National Academy of Engineering in 2011.

Read full profile ↓

Founded Coursera (Mid-career · 2012 · age 43). Daphne Koller joined the Stanford University faculty in 1995 as a Computer Science professor. She co-founded Coursera with fellow Stanford professor Andrew Ng after their online courses at Stanford in 2011 showed global demand for accessible, high-quality education. Exit: Coursera completed its Initial Public Offering (IPO) on March 31, 2021, listing on the NYSE under the ticker symbol "COUR". The IPO priced 15.7 million shares at $33 each, raised approximately $519 million, and gave the company an implied valuation of $4.3 billion. (Daphne Koller - Wikipedia · Coursera - Wikipedia)

In 2012 she and her Stanford colleague Andrew Ng co-founded Coursera after their AI/ML courses each drew 100,000+ online students. Daphne Koller served as co-CEO and then president; Coursera, an online learning marketplace, went public in March 2021. In spring 2018 she left Coursera to found insitro, a machine learning-driven drug-discovery company that combines induced pluripotent stem cells, high-content imaging, and ML to identify disease targets.

Insitro raised a $100M Series A in 2018 (a16z, Bezos Expeditions, GV), signed a $1B+ collaboration with Gilead on NASH in April 2019, and has raised over $700M cumulatively (Series C $400M, SoftBank Vision Fund 2 led). She is the direct counterpart to Demis Hassabis's Isomorphic Labs in the AI-for-drug-discovery race. She is married to Dan Avida, a venture capitalist at Opus Capital.

Sources: Wikipedia · insitro leadership page · TechCrunch interview, 2019 · MacArthur Fellowship, 2004

Extended Roster · 7 more from this chapter

Yejin Choi, Dieter Schwarz Foundation Professor, Stanford HAI. A leading NLP researcher focused on commonsense reasoning, pluralistic alignment, and small language models. MacArthur Fellow (2022), AI2050 Senior Fellow, TIME100 AI (2023, 2025). | Sources: Stanford HAI Wikipedia Yejin Choi site

Diyi Yang, Assistant Professor CS / Stanford NLP and HAI. Stanford NLP/HAI faculty member who leads the Social and Language Technologies (SALT) Lab building socially aware AI systems that understand social context, power dynamics, and group communication. Previously an assistant professor at Georgia Tech (2019-2022); recipient of Forbes 30 Under 30 in Science, IEEE 'AI's 10 to Watch', the Intel Rising Star Faculty Award, Microsoft Research Faculty Fellowship, and an NSF CAREER Award. | Sources: Stanford CS Stanford profiles Wikipedia Google Scholar Stanford HAI

Emma Brunskill, Associate Professor CS / Director AI for Human Impact, Stanford (b. 1979). Leads Stanford's AI for Human Impact lab, focused on reinforcement learning systems that learn from few samples to make robust decisions in high-stakes settings like healthcare and education. Joined Stanford in 2017 after starting her faculty career at CMU in 2011; elected AAAI Fellow in 2025 for contributions to reinforcement learning and AI for societal benefit. | Sources: Wikipedia Stanford CS Stanford profiles Google Scholar

George Sivulka, Co-Founder & CEO, Hebbia (b. 1999). Dropped out of his Stanford EE PhD in 2020 at age 23 to found Hebbia, an AI agent platform for knowledge work used by BlackRock, Carlyle, Centerview and 40% of the largest asset managers by AUM. Raised a $130M Series B and the company was valued near $700M, with backers including Peter Thiel and Andreessen Horowitz. | Sources: Crunchbase 20VC

James Landay, Professor CS / Director Stanford HAI. PhD dissertation was the first to demonstrate sketching in user interface design tools, a foundation of modern HCI. Previously Associate Professor at UC Berkeley, Professor at UW (2003-2013), and Director of Intel Labs Seattle (2003-2006) where he led ubiquitous-computing research. | Sources: Stanford profiles Jon Landay site Stanford HAI Stanford HAI

Michael Bernstein, Professor of CS / Senior Fellow Stanford HAI. Michael Bernstein is a Bass University Fellow and Professor of Computer Science at Stanford University and Senior Fellow at the Stanford Institute for Human-Centered AI (HAI), where he leads research on social computing, human-computer interaction, and generative agent simulations of human social behavior. His work on AI simulations is the highest-cited research in the history of UIST, he has won eight best paper awards at CHI, CSCW, and UIST, and he holds an Alfred P. | Sources: Stanford HCI Stanford HAI Stanford profiles Google Scholar

Silvio Savarese, Professor CS, Stanford University. Silvio Savarese is Executive Vice President and Chief Scientist at Salesforce Research and Adjunct Faculty at Stanford University, where he previously served as a tenured Associate Professor and Director of the SAIL-Toyota Center for AI Research. At Salesforce he leads Agentforce and previously led Einstein GPT and the open-source CodeGen model. | Sources: Wikipedia Stanford AI Center Google Scholar

4.2 The Safety Architects & The Three Godfathers

The pattern is that talent pipelines do not end with training and hiring. They also shape which researchers turn capability work into a safety mission and build new institutions around it.

Safety research now sits inside every frontier lab. The people in this chapter answer one question for a living: how do you make powerful AI systems actually do what we want. Their papers get cited across the industry as evidence the field is doomed (or fine, depending on who's reading). The chapter ends with the three researchers who made deep learning possible and who can no longer agree about what to do with it.

Durk Kingma, The Man Who Gave AI Its Optimizer

Durk Kingma produced two of the most foundational tools in modern machine learning before finishing his PhD. In 2013, he and Max Welling published "Auto-Encoding Variational Bayes", the Variational Autoencoder (VAE), a generative model grounded in Bayesian inference that could encode data into a latent space and decode samples back into realistic outputs. The VAE gave researchers a principled probabilistic framework for generative modeling that ultimately seeded the latent diffusion architectures behind Stable Diffusion and DALL-E. In 2014, he co-authored "Adam: A Method for Stochastic Optimization" with Jimmy Ba. Adam is now the default optimizer used to train virtually every neural network in the world, by citation count, one of the most-cited papers in all of computer science.

Read full profile ↓

Kingma site

Founded OpenAI (In school · 2015 · age 32). Durk Kingma was a PhD student at the University of Amsterdam from 2013 to 2017, where he co-invented the Variational Autoencoder (VAE) and the Adam optimizer. He co-founded OpenAI in December 2015 while still pursuing his PhD. Exit: OpenAI was founded as a non-profit. In 2019, it created a for-profit subsidiary, and by October 2025, it restructured into a public benefit corporation (PBC) partially controlled by the non-profit foundation. The company has raised significant funding, with a valuation reaching $852 billion by April 2026. (OpenAI - Wikipedia · OpenAI | ChatGPT, Sam Altman, Microsoft, & History | Britannica Money)

Durk Kingma was part of OpenAI's founding team, later worked at Google Brain and Google DeepMind on normalizing flows and diffusion models, and joined Anthropic as a Research Scientist in October 2024. His partner Tim Salimans is also a prominent ML researcher. His move to Anthropic was not a pivot away from research. It was a bet that the tools he built are approaching a threshold where the next foundational contribution has to be about what the tools do when no one is watching.

Source: Durk Kingma and Welling, "Auto-Encoding Variational Bayes," ICLR 2014; Durk Kingma and Ba, "Adam: A Method for Stochastic Optimization," ICLR 2015; TechCrunch, "Anthropic hires OpenAI co-founder Durk Kingma," October 2024.

Evan Hubinger, The Sleeper Agent Problem

Evan Hubinger is known for two contributions that reshaped how the field thinks about AI risk. The first is a 2019 LessWrong sequence, "Risks from Learned Optimization," which introduced the terms "mesa-optimization" and "inner alignment", now standard vocabulary in technical AI safety. The core argument: a model trained by an outer optimizer (gradient descent) to perform well might itself develop into an inner optimizer pursuing a subtly different objective that only reveals itself outside the training distribution. This formalized a class of failure modes previously discussed only vaguely.

Read full profile ↓

LessWrong

The second contribution is the "Sleeper Agents" paper (2024, co-authored at Anthropic): large language models can be fine-tuned to harbor deceptive behaviors, behaving safely in most contexts but triggering unsafe behavior on a specific condition, and standard safety training techniques (RLHF, adversarial training, supervised fine-tuning) largely fail to remove this behavior. In some cases, safety training made models better at concealing the behavior rather than eliminating it. The paper was immediately cited as evidence that behavioral safety training alone is insufficient for frontier models. The harder finding is what the paper implies about the work itself: running RLHF and adversarial training and calling that a safety program is not the same as having one.

Source: Evan Hubinger et al., "Risks from Learned Optimization in Advanced Machine Learning Systems," arXiv 2019; Evan Hubinger et al., "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training," Anthropic, January 2024.

Nicholas Carlini, How Models Remember What They Shouldn't

Nicholas Carlini spent seven years at Google Brain doing the job that organizations rarely reward and always need: systematically finding the gaps between what safety researchers claim and what models actually do. His 2017 paper "Towards Evaluating the Robustness of Neural Networks" (with David Wagner) introduced the Nicholas Carlini & Wagner (C&W) attack, an optimization-based adversarial method that systematically broke dozens of proposed defenses. It reset the robustness research agenda by proving that claimed defenses needed to be evaluated against strong, adaptive attacks. Virtually every robustness paper since cites it.

Read full profile ↓

Nicholas Carlini site

More consequential for the LLM era is his 2021 paper "Extracting Training Data from Large Language Models", demonstrating empirically that GPT-2 had memorized verbatim text from its training corpus that could be extracted through careful prompting. The paper defined types of memorization and developed a methodology for distinguishing genuine memorization from coincidence. Its implications were immediate and legal: LLMs do not merely generalize from data, they retain specific private information that can be extracted. This finding now underlies major privacy litigation and regulatory analysis of foundation models. He joined Anthropic around 2024.

Source: Nicholas Carlini and Wagner, "Towards Evaluating the Robustness of Neural Networks," IEEE S&P 2017; Nicholas Carlini et al., "Extracting Training Data from Large Language Models," USENIX Security 2021; nicholas.carlini.com.

Victoria Krakovna, The Specification Gaming Catalog

Victoria Krakovna grew up in Russia, won a silver medal at the International Mathematical Olympiad, and completed a Harvard PhD in Statistics in 2016, while simultaneously co-founding the Future of Life Institute (FLI) with Max Tegmark and others in 2014, while still a graduate student. FLI became one of the first credible institutional voices on existential risk from AI, organizing the 2015 autonomous weapons open letter and the 2017 Asilomar AI Principles conference.

Read full profile ↓

Victoria Krakovna blog

Founded Future of Life Institute (In school · 2014). Victoria Krakovna co-founded the Future of Life Institute in 2014 while she was a PhD student in Statistics at Harvard University, which she completed in 2016. Exit: The Future of Life Institute (FLI) is a non-profit organization founded in 2014 and still operating. (Timeline of Future of Life Institute)

At Google DeepMind, Victoria Krakovna created the canonical specification gaming catalog, a living document, first published in 2018 at vkrakovna.wordpress.com, now containing hundreds of examples of AI systems finding reward loopholes rather than solving the intended problem: a boat racing agent earning points by circling power-ups, a robot grasping task solved by flipping its own fingers, a locomotion task solved by exploiting physics engine numerical errors. The catalog is now assigned in AI safety courses worldwide and cited in policy documents. Her more recent work covers goal misgeneralization (formally proving that correct in-distribution behavior gives no guarantee of correct out-of-distribution behavior) and dangerous capability evaluations for Gemini models.

Source: Victoria Krakovna et al., "Specification gaming: the flip side of AI ingenuity," Google DeepMind Blog, 2020; Victoria Krakovna et al. (Langosco et al.), "Goal Misgeneralization in Deep Reinforcement Learning," ICML 2022; vkrakovna.wordpress.com.

John Jumper, The Nobel Laureate Inside the Machine

John Jumper was born on January 1, 1985, in Little Rock, Arkansas. A Marshall Scholar at Cambridge, he studied theoretical condensed matter physics before deciding he wanted to work on problems where, as he later put it, "someone goes home from the hospital." He earned his PhD in theoretical chemistry at the University of Chicago in 2017, joined Google DeepMind, and was immediately embedded in the AlphaFold project. When the first AlphaFold version showed promise but hit architectural limits, John Jumper led the decision to scrap it entirely and rebuild from scratch. It was persuading a room full of people who had already committed years to version one that the right move was to start over.

Read full profile ↓

Google DeepMind

AlphaFold2 debuted at the CASP14 competition in December 2020 with median accuracy across most protein classes within the margin of experimental error. Nature called it "a solution to a 50-year-old grand challenge." The AlphaFold Protein Structure Database, released in 2021, eventually grew to over 200 million predicted structures, essentially every protein known to science, freely available to any researcher anywhere (Google DeepMind). As of the 2024 Nobel announcement, more than 2 million researchers in 190 countries had used it. On October 9, 2024, John Jumper and Demis Hassabis were jointly awarded the Nobel Prize in Chemistry. At 39, John Jumper was among the younger chemistry laureates in recent memory.

Source: John Jumper et al., "Highly accurate protein structure prediction with AlphaFold," Nature, 2021; Nobel Prize in Chemistry 2024 press release, NobelPrize.org; University of Chicago, "How an accidental chemist honed his approach," 2024.

Koray Kavukcuoglu, From DeepMind Lab to Google's Technical Summit

Koray Kavukcuoglu started as an aerospace engineering student in Turkey before pivoting entirely to computer science. He earned his PhD at NYU under Yann LeCun, working in deep learning before it was mainstream, and joined DeepMind as an early researcher in 2012. Over 13 years he rose from researcher to head of the deep learning team to VP of Research to CTO of Google DeepMind. Under his leadership the deep learning team built DQN (Atari games from raw pixels), WaveNet (the neural voice of Google Assistant, used by hundreds of millions), AlphaGo (which defeated Lee Sedol in 2016), and AlphaFold.

Read full profile ↓

On June 11, 2025, Google CEO Sundar Pichai announced Koray Kavukcuoglu's appointment as Chief AI Architect of all of Google, a newly created SVP role reporting directly to Pichai, with a mandate to accelerate how DeepMind's models integrate into Google products. He relocated from London to Mountain View. The appointment is significant: the most senior technical role at one of the world's largest technology companies now belongs to a career research scientist who has never run a product team, an indication that at Google, deep scientific credibility won the AI race.

Source: CNBC, "Google taps DeepMind's Koray Kavukcuoglu for new chief AI architect role," June 11, 2025; Semafor, "Google names new chief AI architect," June 2025; Google I/O 2024.

Anca Dragan, What to Optimize For

Anca Dragan grew up in Braila, Romania and earned her PhD in Robotics at Carnegie Mellon in 2015, where her dissertation asked how robots could make their movements legible, readable to nearby humans, communicating intention through motion rather than just executing tasks efficiently. This question, not how to optimize, but how to communicate the optimization to humans who need to understand and trust it, became the organizing principle of the Berkeley InterACT Lab she founded and eventually of her work at Google DeepMind.

Read full profile ↓

She co-PI'd CHAI (Center for Human-Compatible AI) with Stuart Russell and spent six years consulting for Waymo on safety-critical autonomous systems. Google DeepMind appointed her Head of AI Safety and Alignment in February 2024, where she leads AGI Alignment, Gemini Safety, and a team exploring how to handle the plurality of human values rather than optimizing toward a single idealized preference. Her 2025 ICML invited talk was titled "What to Optimize For, From Robot Arms to Frontier AI," tracing the arc from a dissertation insight in 2010 to the most pressing design question in the industry.

Source: Berkeley EECS announcement, "Anca Dragan named Head of AI Safety and Alignment at Google DeepMind," February 2024; ICML 2025 invited talk program; TechCrunch, "Google DeepMind forms new AI safety org," February 2024.

Kaiming He, The Most Cited Paper in Computer Science

Kaiming He scored first in Guangdong province's 2003 university entrance examination. He studied at Tsinghua University, earned a PhD in Hong Kong, and joined Microsoft Research Asia, where in 2015 he and colleagues published "Deep Residual Learning for Image Recognition," introducing ResNet. The key insight: adding skip connections that allow gradients to bypass layers entirely solves the vanishing gradient problem that had made very deep networks untrainable. ResNet 152 won the ILSVRC 2015 competition across classification, detection, localization, and segmentation simultaneously. As of 2025, the ResNet paper has over 298,000 citations, the most-cited paper in the history of computer science. Residual connections appear in every modern architecture: Transformers, AlphaFold, every generative model, every LLM.

Read full profile ↓

He later joined Meta FAIR, where he created Masked Autoencoders (MAE, 2021), adapting BERT-style masked pre-training to vision. In 2024 he joined MIT as associate professor (tenured 2025) and simultaneously became a Distinguished Scientist at Google DeepMind on a part-time basis. The dual arrangement, academia and industry simultaneously, on his own terms, is a window into the leverage that elite AI researchers now hold.

Source: He et al., "Deep Residual Learning for Image Recognition," CVPR 2016; He et al., "Masked Autoencoders Are Scalable Vision Learners," ICCV 2022; MIT EECS announcement, 2025; 36Kr, "He Kaiming Officially Joins Google DeepMind," 2025.

Extended Roster · 2 more from this chapter

William MacAskill, Oxford Philosopher · Co-Founder, Effective Altruism & 80,000 Hours (b. 1987). Scottish philosopher and one of the originators of the effective altruism movement. Co-founded Giving What We Can (2009), 80,000 Hours (2011), and the Centre for Effective Altruism. | Sources: Wikipedia William MacAskill site 80,000 Hours

Ashton Anderson, Associate Professor of CS, Univ of Toronto. Ashton Anderson is an Associate Professor of Computer Science at the University of Toronto and Faculty Affiliate with the Vector Institute and Schwartz-Reisman Institute. His computational social science research spans human-AI collaboration, AI alignment, and skill-compatible AI in structured environments, including the Maia chess project for human-AI alignment. | Sources: U Toronto CS Google Scholar U Toronto CS U Toronto SRI

A 30-year friendship. Hinton, LeCun, and Bengio championed neural networks together through two AI winters when most of the field dismissed the approach as a dead end. They co-directed CIFAR's Learning in Machines & Brains program, co-founded the ICLR conference, co-authored the landmark 2015 Nature review paper, and shared the 2018 Turing Award. The mentorship chain reaches into the next generation: Hinton's student Ilya Sutskever built AlexNet in his lab, and Hinton later joked, "Ilya thought we should do it, Alex made it work, and I got the Nobel Prize" (University of Toronto). Andrej Karpathy, for his part, has publicly described Fei-Fei Li, his PhD advisor, as "our fearless leader," recalling their time building Stanford's CS231n course together. ACM Turing · Karpathy on X

In 2018 the ACM gave the Turing Award jointly to three researchers who had spent thirty years convincing the rest of computer science that neural networks were not a dead end. What happened after is not simply a story of former allies diverging. It is a case study in what genuine scientific disagreement looks like when the people doing the disagreeing each have a career's worth of evidence behind their position. They share a prize, a research lineage, and a set of open questions, and they have reached fundamentally different answers about what those questions demand of them now.

Yoshua Bengio, The Quiet Safety Convert

Yoshua Bengio was born March 5, 1964 in Paris, France, to Moroccan-Sephardic parents, and grew up between France and Montreal. He earned his PhD at McGill, worked with Yann LeCun and Larry Jackel at Bell Labs, then joined the Université de Montréal where he founded MILA, the Quebec AI institute, in 1993. His most-cited papers include neural language models (2003), word embeddings, attention mechanisms preceding transformers, and the GAN paper (2014, with Ian Goodfellow). He shared the 2018 Turing Award with Hinton and LeCun. From 2023 onward he became one of the most prominent academic voices on AI safety, signing the Center for AI Safety extinction statement, chairing the International Scientific Report on AI Safety commissioned by the UK and 30 countries, and founding LawZero in 2025 as a nonprofit to develop provably safe AI systems. He is the most-cited computer scientist alive (MILA).

Read full profile ↓

Founded Element AI (Mid-career · 2016 · age 52). Yoshua Bengio co-founded Element AI in 2016 while he was a full professor at the Université de Montréal and the scientific director of MILA (Montreal Institute for Learning Algorithms), which he founded. Exit: ServiceNow acquired Element AI for approximately $230 million USD. The acquisition was announced on November 30, 2020, and completed on January 8, 2021. Wikipedia notes the company was running out of money and options at the time of the acquisition. Value at exit: ServiceNow bought Element AI for its AI capabilities, technical talent, and existing AI solutions. ServiceNow planned to use Element AI's technology and expertise in its workflow platform. Yoshua Bengio became a technical advisor for ServiceNow after the acquisition. (Yoshua Bengio - Wikipedia · Element AI - Wikipedia)

Sources: Wikipedia · Personal site · MILA

Chapter 5 · The Capital Network: VCs, Funding Wars

In AI, money often follows trust before it follows revenue. The capital chapter is about how the trust layer built over decades, in labs, on boards, between cofounders, and across families, becomes the funding graph that decides which companies actually exist.

Two older networks sit underneath much of the capital graph: Google trained many of the researchers who became founders, while PayPal trained many of the investors and operators who learned to bet aggressively before consensus formed.

Who gets funded in AI depends sharply on which fund is writing the check, and the split by founder age is cleaner than most people realize. In this dataset, Sequoia Capital's average funded founder age is 24, while Kleiner Perkins backs founders 40 and older. a16z averages 32 and appears relatively age-blind, backing founders at 22 and 45. The chart that follows, Investor Profiles: Who Funds Which Founders, n=30, 2012-2025, makes that pattern visible across firms and individual investors.

That age pattern is the surface of an older relationship network. The PayPal 2002 IPO produced a group of about a dozen people whose capital, trust, and deal flow have been compounding since 1998 across LinkedIn, YouTube, Palantir, Tesla, SpaceX, and later OpenAI, Anthropic, and xAI. Social-capital research is consistent with this pattern: tight alumni networks compound because trust reduces diligence cost. In practice, that helps explain why some investors move early, back each other, and keep showing up in the same company histories.

What follows starts with the age chart, then moves to 16 profile cards across VCs, angels, and partners. The chart shows the surface pattern. The profiles show who is actually choosing whom, and how those choices repeat. The through-line is simple: the AI capital network is small, old, and split into a handful of taste clusters, each carrying a different risk profile.

How Investors Pick: Fund Preferences by Founder Age

Different investors have strikingly different age preferences. The data below covers 30 AI companies with verified Series A data, founded between 2012 and 2025.

Peter Thiel, The Architect of Everything

He was born in Frankfurt on October 11, 1967, and his family moved to California when he was a child. He studied philosophy at Stanford for his BA and law at Stanford Law School for his JD. He says of his philosophy training that he learned to argue against any thesis, which turns out to be useful in two careers: hedge funds, and starting companies.

Why he matters to the map: He concentrates capital and ideology through the PayPal network, seeding OpenAI's founding circle and shaping the contrarian wing.

Marc Andreessen, The Intellectual Investor

He grew up in New Lisbon, Wisconsin, population 2,500. He studied Computer Science at the University of Illinois Urbana-Champaign and there, in 1993, co-created Mosaic, the first graphical web browser most people ever used. At 22 he co-founded Netscape; its 1995 IPO kicked off the first internet boom, AOL bought it in 1998. In 2009 he co-founded Andreessen Horowitz with Ben Horowitz, now managing roughly $42B with an AI portfolio that includes Mistral, Character.AI, Together AI, Skild AI, and dozens of others. His sustained writing as @pmarca on X has done more than any other partner's output to articulate where the firm thinks technology is going next.

Why he matters to the map: He anchors a16z's $42B AI portfolio and uses his public writing to set the field's investment narrative.

Reid Hoffman, The Human Networker

Reid Hoffman grew up in Berkeley and studied philosophy at Stanford, focusing on philosophy of language and cognitive science. The training is more relevant to his later AI investments than is widely appreciated. He co-founded SocialNet in 1997, worked at PayPal, and in 2002 co-founded LinkedIn. Microsoft acquired LinkedIn in 2016 for $26.2B (Microsoft), of which his share was approximately $2.8B.

Why he matters to the map: He links LinkedIn capital to Greylock and to Inflection's founding, the human bridge across most AI cap tables.

Sarah Guo, The Conviction Bet

Sarah Guo grew up in the Bay Area, studied computer science and economics at Harvard, went to Goldman Sachs, and joined Greylock Partners in 2012, where she led enterprise-software investments in Workday, Cloudera, and Sumo Logic. In 2022 she left to found Conviction, the first venture firm built AI-native from day one. The $101M debut fund produced eight unicorns including Harvey ($8B by late 2025), Mistral, Cognition, HeyGen, and Pika; she closed Fund II at $230M in 2025.

Why she matters to the map: She demonstrated, with eight unicorns from $101M, that a purpose-built AI fund can outperform generalist VCs at the earliest stage in this cohort.

David Sacks, PayPal Mafia to AI Czar to All-In Co-Host

Born May 25, 1972 in Cape Town, South Africa, David Sacks immigrated to the US at age 5. He met Peter Thiel and Keith Rabois at Stanford, where he edited the libertarian Stanford Review under Peter Thiel, a network that became the spine of the PayPal Mafia. After a Chicago JD and a stint at McKinsey, he joined Confinity/PayPal in 1999 as COO/product head, working alongside Peter Thiel, Elon Musk, Reid Hoffman, Max Levchin, and Roelof Botha. PayPal IPO'd in February 2002; eBay bought it months later for $1.5B.

Garry Tan, Y Combinator's Most Political President

Born 1981 in Winnipeg, Canada to a Chinese-Singaporean father (machine-shop foreman) and a Burmese-Chinese mother (nursing assistant), Garry Tan moved to Fremont, California in 1991. He graduated from American High School and went on to Stanford for computer systems engineering. He worked at Microsoft, then became the 10th employee at Palantir Technologies under Alex Karp and Joe Lonsdale, a foundational stint that wired him into the Peter Thiel network.

Spotlight · The First Checks: Five Investors Who Saw AI Early

Across hundreds of AI investments since 2018, five firms keep showing up at the moment a future winner was still a hypothesis. They wrote the first institutional checks, or the first conviction-sized ones, into companies most of the field had not yet decided to take seriously. The bets look obvious in retrospect. They were not obvious at the time. What follows is the operator story behind each.

1. Khosla Ventures · The $50M Letter to LPs

In 2019, Vinod Khosla wrote a $50M check into OpenAI at a $1B valuation. It was the largest single position Khosla Ventures had ever taken, by a factor of two. The company was still nominally a nonprofit. There was no commercial model. Elon Musk had just walked away from his pledged $1B and demanded control. Most institutional investors were not even taking the meeting.

Khosla took the meeting, then wrote an apology letter to his LPs warning that the bet looked foolhardy and admitting he was doing it anyway. He framed the conviction in geopolitical terms: if the West did not build advanced AI first, China would. As of 2025 the position was estimated at $8B to $15B, roughly 160x the entry. No other VC in history has written a larger early-belief check into the company that turned out to be the most consequential of the decade. (Fortune)

The operator lesson is the letter. Writing a check that large into a nonprofit research lab required Khosla to tell his LPs in advance that he might be wrong, then ask them to trust him anyway. Most VCs cannot do that move. It only works once you have the track record to spend, and Khosla had spent thirty years building that account.

2. Conviction · The Purpose-Built AI Fund

Sarah Guo launched Conviction in October 2022 as the first venture firm built AI-native from day one. The debut fund was $101M. Her first check was Harvey, the legal AI startup that had no product and reportedly no revenue when Conviction wrote in. By December 2025 Harvey was valued at $8B. The same debut fund produced seven additional unicorns, including Mistral, Cognition, HeyGen, and Pika. Guo added Mike Vernal as a second GP in early 2025, when she closed Fund II at $230M. (TechCrunch · Harvey at $8B)

The thesis Conviction was built around is that generalist VCs would underwrite AI through the wrong frame. By the time a generalist partner had read enough papers to develop conviction, the early-stage round had already closed. A purpose-built AI fund, staffed by people who could read the research themselves, would see the company before the rest of the market formed an opinion. Eight unicorns from a single $101M debut fund is the proof. If AI cools, there is no diversification to fall back on.

3. Andreessen Horowitz · The Series A Leads

a16z's biggest AI positions, OpenAI and Databricks, came at growth stage. The two cleanest early-stage wins are Character.AI and Mistral, both led by the firm. In March 2023, a16z led Character.AI's $150M Series A at a $1B valuation when the company was 16 months old; Sarah Wang (a16z growth partner) joined the board. As of 2024, Character.AI had over 20 million daily users, and Google paid roughly $2.7B for the model license and the founding team. (CNBC · a16z on Mistral)

In December 2023, a16z led Mistral's €385M Series A at roughly $2B, the largest European AI round to that point. Martin Casado's infrastructure thesis, that defensibility in AI sits at the data and workflow layer rather than the model layer, has shaped how a16z evaluates these positions. Casado has said in interviews that he runs the technical interview before the financial one. That sequencing is the operator move. It produces investments where the partner can hold the founder's confidence through the parts of the build that have no obvious answer.

4. Greylock · The Incubation Model

Greylock is the only firm on this list that builds AI companies from zero. Inflection AI was incubated inside the firm in 2022, with Reid Hoffman co-founding it while still a sitting partner, an arrangement almost unheard of in venture capital. Microsoft eventually licensed the model and hired the team for roughly $650M. Adept, Cresta, Snorkel, and Neeva all came out of the same playbook: founder team plus capital plus credibility, before the company has a name. (Greylock on Inflection · BusinessWire on Adept)

The partner most responsible for the firm's enterprise-AI bets is Saam Motamedi, who led the $65M Series A into Adept and sits on the board. He is among the youngest GPs at the firm, and his thesis is that pricing power in enterprise AI will accrue to systems that already own the workflow data, not to the model layer. Incubating from zero is harder than leading rounds.

5. Sequoia Capital · The Durable Operator

Sequoia is the most durable AI investor across cycles. The firm came into OpenAI at growth stage in 2021, not at founding, and its $7B AI-focused expansion fund announced in April 2026 sits at the larger end of the early-stage spectrum. The Series A Sequoia led that belongs in this Spotlight is Harvey: a $21M check in April 2023 that returned roughly 380x by late 2025. Pat Grady, Sonya Huang, and Konstantine Buhler jointly run the firm's AI strategy, anchored by Buhler's "agent economy" framing that estimates the market for autonomous AI agents at over $10 trillion, roughly ten times the size of cloud computing. (TechCrunch on Sequoia's $7B fund)

What Sequoia provides that the smaller funds cannot is a sustained presence across every stage of a portfolio company's life. Harvey will need that across the next decade. So will Sierra. The trade-off is real: a generalist mega-fund cannot move as fast as a Conviction at the earliest stage. Sequoia's discipline has been to write the Series A when the conviction is there and the growth round when the company has earned it.

The pattern across these five is not a strategy. It is a willingness to be wrong loudly. Khosla wrote the letter, Guo started the firm, a16z planted a flag in consumer AI before anyone had agreed it was a category, Greylock co-founded a portfolio company while a GP sat on the board, and Sequoia treated agent autonomy as a $10 trillion market before there was a real product to point at. The operator move is the same in every case: write the check, then defend it in public, then keep working.

Holden Karnofsky, The EA Philosopher Who Married Anthropic's President

Holden Karnofsky co-founded Open Philanthropy (now Coefficient Giving) in 2014 with Dustin Moskovitz and Cari Tuna. As CEO and co-CEO, he directed hundreds of millions of dollars toward AI safety research, animal welfare, and global catastrophic risk causes. His "Most Important Century" essay series articulated the case that AI development may represent the most pivotal period in human history, a perspective that shaped an entire generation's thinking on existential risk and long-termism. In January 2025, Holden Karnofsky joined Anthropic as a member of the technical staff working on responsible scaling policies.

Martin Casado, The Networking Intellectual at a16z

Martin Casado was born in Cartagena, Spain. He earned his BS from Northern Arizona University (2000), worked as a researcher at Lawrence Livermore National Laboratory (running large-scale simulations for the Department of Defense), then earned his MS and PhD in Computer Science from Stanford by 2007. His PhD work pioneered OpenFlow, the open-source protocol that became foundational to Software-Defined Networking (SDN), one of the most important networking innovations of the past 20 years. He co-founded Nicira Networks in 2007; VMware acquired Nicira for $1.26B in July 2012. He became VMware Fellow and CTO for networking and security before joining Marc Andreessen Horowitz as its ninth general partner in February 2016.

Jerry Chen, "Systems of Intelligence"

Jerry Chen is a Greylock Partner who coined the concept of "Systems of Intelligence", the framework explaining how AI and data create durable competitive advantages in enterprise software. After Stanford Industrial Engineering and Harvard MBA, Jerry Chen spent years at Bain and Accel before joining VMware in 2004, where he ran product management and marketing for cloud and application services. He coined the term "VDI" (Virtual Desktop Infrastructure) and helped found Cloud Foundry. At Greylock since 2013, he has become the leading voice on why the winning enterprise AI companies will be those that build proprietary ontologies and domain-specific models that compound over time, not those that merely wrap commodity LLMs.

Konstantine Buhler, Sequoia's Agent Economy Theorist

Konstantine Buhler joined Sequoia Capital after earning a BS in Management Science and Engineering and an MBA from Stanford, winning the Terman Engineering Award, and minoring in Byzantine Art history. His Greek immigrant grandfather's entrepreneurial courage is a formative story he tells frequently. At Sequoia's AI Ascent 2025 (with partners Pat Grady and Sonya Huang), Konstantine Buhler articulated the "agent economy" thesis: autonomous agents will transfer resources, make transactions, and operate in their own economic systems, a market Sequoia estimates at over $10 trillion, "ten times larger than cloud computing." This framing, that agents are not just software features but economic actors, represents one of the most expansive bull cases for agentic AI and directly underlies investments in Sierra, OpenAI, and the broader agent-framework category. Buhler did that in front of an audience of operators who had every reason to be skeptical, and the number he put on it was large enough that being wrong would be embarrassing.

Pat Grady, Sequoia's Enterprise Whisperer

Pat Grady has been at Sequoia since 2007 and led the firm's growth-stage investing since 2015. His portfolio reads like a directory of transformational enterprise software: Amplitude, HubSpot, Okta, Qualtrics, ServiceNow, Snowflake, Zoom, OpenAI, Notion, Hugging Face. In November 2025, Sequoia announced Pat Grady would succeed Roelof Botha as co-steward alongside Michael Lin, a transition from one legendary investor to the next. He gravitates toward founders who "bulldoze incumbents". ServiceNow, Snowflake, and HubSpot are the receipts. What gets less attention is what a transition like this actually demands. Taking over from Roelof Botha is not a promotion, it is a test of whether the firm's judgment survives the person who embodied it for a decade. Grady has been inside the answer to that question for long enough to know what it costs.

Scott Sandell, NEA's Most Enduring Investor

Scott Sandell has been at New Enterprise Associates since 1996, nearly 30 years at one firm, an anomaly in a profession characterized by mobility. He holds a Dartmouth AB in Engineering and a Stanford GSB MBA. Before NEA he worked at Microsoft, Boston Consulting Group, and C-ATS Software. His portfolio includes Bloom Energy, Cloudflare, Robinhood, Salesforce, Tableau, and Workday, companies that collectively represent the backbone of enterprise cloud infrastructure. He became Managing General Partner in 2017 and currently serves as Executive Chairman and CIO. Thirty years at one firm is rare in venture. Founders he funded in the 2000s now send their next companies back to him.

Chamath Palihapitiya, The Social Capital Partner Dispute

Chamath Palihapitiya was an early Facebook executive (VP of User Growth, 2007-2011) widely credited with the playbook that scaled the platform past one billion users. He founded Social Capital in 2011 and co-hosts the All-In podcast. Wikipedia

Read full profile ↓

The arc that's easy to miss: Chamath Palihapitiya was born in Galle, Sri Lanka. His family fled to Canada when he was five after his father, a diplomat, criticized violence against Tamils during the civil war. They sought asylum and fell into poverty: his father struggled with unemployment, his mother cleaned houses. At fourteen he took a job at Burger King to help with household bills. He attended Lisgar Collegiate Institute in Ottawa, earned an electrical engineering degree from the University of Waterloo, and rose from derivatives trading to becoming Facebook's VP of User Growth (2007-2011), credited inside the company with the playbook that scaled it past a billion users, without a PhD or privileged connections. He founded Social Capital in 2011 and co-hosts the All-In podcast. Wikipedia

In March 2024, Bloomberg reported that Social Capital terminated partners Jay Zaveri and Ravi Tanuku in connection with reports that they had organized a side vehicle for Groq investors without firm approval. The matter has been reported, not adjudicated.

The pattern is that investor power shows up in repeated presence, not one famous deal. The numbers make that pattern plain before the conflict cases show how it plays out in practice.

The Numbers That Define the AI Race

OpenAI: $852B valuation (as of early 2026) as of 2025, the most valuable private company in history. Sam Altman has publicly stated that he holds no equity in OpenAI and takes a nominal salary, an arrangement he has described as a deliberate marker that he is motivated by the mission rather than personal return. What is easy to miss is that structural choices like this function as governance under pressure: when a CEO owns nothing, the board's leverage over the mission changes in ways that matter most precisely when things get hard.

Anthropic: From $87M in annual revenue to $30B ARR in approximately 16 months, a 342x growth rate with few obvious precedents in enterprise software history. Dario Amodei called it "just crazy" in a rare moment of public understatement. The growth is driven almost entirely by API consumption from developers and enterprises building on Claude, not consumer subscriptions. Anthropic is among the fastest-growing enterprise AI companies on record.

NVIDIA: The first company in human history to reach a $5 trillion market cap. Jensen Huang's personal net worth: approximately $180B, making him among the five wealthiest people alive. NVIDIA's dominance rests on a single insight Jensen Huang had in the 2000s: GPUs designed for gaming could be repurposed for parallel computation. Every frontier AI model, GPT, Claude, Gemini, Llama, trains on NVIDIA hardware. The company's revenue grew from $27B (FY2023) to over $130B (FY2025).

Perplexity: Aravind Srinivas, Perplexity's CEO, submitted a $34.5B bid to acquire Chrome from Google as an antitrust remedy after the DOJ ruled Google maintained an illegal search monopoly. The bid is audacious: a startup valued at $18B bidding $34.5B for the world's dominant browser. If successful, it would give Perplexity direct distribution to 3.4 billion Chrome users. The move follows a logic that is counterintuitive until it isn't: when a regulatory window opens, the right move is usually the one that looks too big to be serious.

Elon Musk: Net worth between $788B and $811B depending on the day, the wealthiest person on most major wealth indexes at the time of writing (per Forbes). In February 2026, SpaceX acquired xAI (Elon Musk's AI company, maker of Grok) in a deal valuing xAI at $250B (Forbes; TechCrunch). The transaction effectively merged Elon Musk's AI ambitions with his rocket company, creating one of the strangest corporate combinations in history: a rocket and satellite company that also runs a frontier AI lab.

Cursor: Four MIT students raised a $400K pre-seed round, forked VS Code, added AI-native code completion, and built one of the fastest value-creation trajectories in venture history: $400K to ~$50B in three years. Zero marketing spend. Zero sales team. Entirely word-of-mouth adoption by developers who loved the product enough to tell other developers. Contrary Research

SSI: Ilya Sutskever raised $3B with no product, no revenue, and no customers yet, on the strength of his reputation and technical thesis. Safe Superintelligence Inc. has stated a single goal: build safe superintelligence, with no intermediate products and no announced timeline. By any conventional metric this is an unusual financing, and it is essentially a bet on Sutskever and the team. TechCrunch

The Investors Who Shaped Everything

Four early checks defined modern AI venture: one priced OpenAI when it was still a nonprofit; one led the Series A at both DeepMind and Anthropic; one bet a fortune on Anthropic through effective altruism; one seeded the inference chip that NVIDIA later bought. Each made the bet before the consensus formed.

The pattern is that the same firms do more than fund companies. The VC wars begin when financing, board influence, and access sit in the same hands.

How Sequoia Is Backing Multiple Frontier AI Labs

Sequoia broke the taboo: Traditional venture capital etiquette holds that you don't invest in direct competitors. Sequoia invested in OpenAI, Anthropic, and xAI simultaneously, betting on the entire frontier AI market rather than picking a winner. Other VCs criticized the strategy as "indexing, not investing." Sequoia's answer, implied by its returns, is that when the market is large enough, backing multiple winners can be the rational play. The harder lesson for operators is what this says about the funding landscape: in a category this large, your investor may be equally committed to your three closest competitors, and your competitive moat has to be built without assuming otherwise.

Patrick Collison's quiet bet: The Stripe co-founder backed Leopold Aschenbrenner's investment fund after the two met at a 2021 dinner. Leopold Aschenbrenner, fired from OpenAI's Superalignment team at age 23 (The Information), author of the viral 165-page "Situational Awareness" essay, launched a fund that returned over 100% in its first year. Collison's investment was characteristic: no public announcement, no board seat, just capital deployed on conviction about a person.

5.1 The Agentic Revolution: Coding AI, Agents, and the Opportunity Nobody Planned

An AI agent is a model with a goal and a loop. The loop is doing the work the goal describes; the model is doing the choosing. Until 2023, most AI products were question-and-answer machines. In 2024, the better ones started doing things. This chapter is about the companies that figured out which things, in which order, and at which moment of trust the human steps back in. Cursor went from a $400K pre-seed to roughly $50B in valuation in three years. It is not the only one.

Cursor / Anysphere, The ~$50B Surprise

Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger met at MIT. Founded Anysphere in 2022, launched Cursor, an AI code editor built on VS Code with deep integration of Claude, GPT-4, and open-source models. Cursor became the fastest-growing developer tool in history: roughly $100M ARR in 20 months (sources vary between 18 and 20 months depending on how the founding date is measured; we cite the 20-month figure here). August 2024: $60M at $400M valuation. December 2024: $100M at $2.5B valuation. April 2025: $900M at $9B valuation; approaching $50B by mid-2026. Founders are all in their mid-20s. What drove that growth is instructive: multi-model flexibility meant no single provider could strand users, and software engineers, historically the fastest adopters in any organization, spread it bottom-up before procurement ever got involved. What people underestimate is the product judgment required to make those calls at 24. Every model integration decision, every UX tradeoff, every pricing move was made by founders with no prior playbook, under compounding pressure, in real time.

Windsurf / Codeium → Google ($2.4B)

Varun Mohan and Douglas Chen co-founded Codeium (later Windsurf) in 2021. Raised $150M in February 2025 at $1.25B valuation. The deal that followed was a three-part split, not a clean acquisition: OpenAI's ~$3B LOI to acquire Windsurf collapsed July 11, 2025; Google then hired CEO Varun Mohan and the core team in a ~$2.4B talent-and-licensing arrangement; and Cognition acquired the remaining Windsurf IP and approximately 210 employees on July 14, 2025. Per TechCrunch, Mohan and Chen joined Google DeepMind to work on Gemini while Cognition absorbed the rest of the company.

Replit, Amjad Masad

Amjad Masad grew up in Amman, Jordan, taught himself to code at 15 on a bootlegged computer, emigrated to the US to join Codecademy then Facebook, and founded Replit in 2015. Now used by ~30 million developers, with strength in education and developers in countries where expensive hardware is inaccessible. Raised $97.4M at $1.16B valuation in 2023. Replit Agent, which generates and runs full applications from a prompt, extends that accessibility into the agentic era: for a first-time builder in Lagos or Lima, it removes both the hardware barrier and the expertise barrier at once. The decision to go fully agentic rather than stay in safe IDE territory was not obvious. It traded a predictable business for a harder, more uncertain one. That is usually the right call, and it usually looks wrong at the time.

Extended Roster · 23 more from this chapter

Michael Truell, Co-Founder & CEO, Anysphere (Cursor) (b. 2000). Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger all met through MIT's CSAIL, and all four rejected lucrative big-tech offers to found Anysphere in 2022 without finishing their MIT degrees. Built Cursor as the AI-native coding environment. GitHub | Sources: Wikipedia Martin Truell site Lenny's Newsletter

Clay Bavor, Co-Founder & CEO. Co-founded Sierra with Bret Taylor in 2023. Spent ~20 years at Google as VP of VR, AR & Maps. | Sources: Sierra AI TechCrunch

Stanislas Polu, Co-Founder & CEO, Dust.tt (b. 1987). Co-founded Dust (2022/2023) with Gabriel Hubert to help enterprises build AI agents over their internal knowledge. Previously a research engineer at OpenAI (2019-2022) working on mathematical reasoning with language models (formal proof / pre-Q* reasoning era) alongside Ilya Sutskever. GitHub | Sources: Sequoia Capital Latent Space

Varun Mohan, Co-Founder & CEO, Codeium / Windsurf (b. 1993). Co-founded Codeium in 2021 with MIT classmate Douglas Chen; pivoted from GPU virtualization to AI code-completion to standalone IDE rebranded Windsurf (Nov 2024). In July 2025 Windsurf was split in a complex Google/OpenAI/Cognition deal; Mohan and Chen joined Google DeepMind to work on Gemini. | Sources: Lenny's Newsletter Business Standard

Alexander Rinke, Co-Founder & Co-CEO, Celonis (b. 1989). Alexander Rinke co-founded Celonis in 2011 at age 22 with TUM classmates Bastian Nominacher and Martin Klenk, growing it into the global leader in process mining and intelligence valued at $13B+. The company's AgentC product embeds AI agents inside enterprise processes for Fortune 500 customers. | Sources: Celonis Crunchbase World Economic Forum

Christina Abraham, Senior Vice President, Salesforce AI. Christina Abraham is SVP of AI and Agentforce at Salesforce since June 2025, where she leads development of customer-facing enterprise generative AI features. Previously a GVP/VP at ServiceNow (2023-2025) leading AI agents and generative AI, Senior Director of Software Engineering at Google (2020-2023), and SVP at SAP in Cloud Engineering (2014-2020). | Sources: theorg.com

Claire Lecarpentier, Director, Product Strategy & Operations, ServiceNow. Claire Lecarpentier is Director of Corporate Development & Ventures (Platform and AI) at ServiceNow, leading product strategy for ServiceNow AI Agents and enterprise workflow automation. A French-born CFA charterholder, her prior career includes Corporate Venture/Principal Investments at J.P. | Sources: Global Corporate Venturing Medium currnt.com

Connor Solimano, Engineer, Cursor. Connor Solimano leads founding strategy and business operations at Anysphere, the company behind Cursor, the AI-first code editor that has become one of the fastest-growing developer tools. He joined Cursor in February 2025 after three years at Insight Partners on the investment team focused on cyber, infrastructure, and developer tooling. | Sources: cursor-insider.com Harvard CS50 Pitchbook

Daniel Chalef, Co-Founder & CEO. Former founder and enterprise software veteran who built Zep, a temporal knowledge graph and memory store for AI agents with enterprise privacy and compliance features. Zep's temporal graph approach differentiates it from simple vector stores, making it the go-to for compliance-conscious enterprise teams. GitHub | Sources: GitHub Y Combinator Zep blog cmscritic.com

Dhruv Malhotra, Co-Founder & CEO. Former McKinsey consultant who built Beam AI after watching enterprise teams manually transfer data between systems that should talk to each other automatically. Beam builds autonomous AI agents for enterprise business process automation, finance, operations, HR. | Sources: Beam AI Beam AI Crunchbase Beam AI

Duncan Lennox, Chief Product & Technology Officer, HubSpot. Chief Product & Technology Officer at HubSpot leading Engineering, Product, UX, Next Bets, and IT & Security, grounding HubSpot's AI agents in rich CRM context for SMB customers. Dublin-born product leader previously VP/GM of Applied AI and Ads Privacy & Safety at Google Cloud, GM of Amazon Elastic File System at AWS, and co-founder/CEO of Qstream and WBT Systems. | Sources: HubSpot HubSpot IR The Official Board

Flo Crivello, Founder & CEO, Lindy AI. French-born founder who launched Lindy in January 2023 as an AI agent platform for automating knowledge work; the company is one of the most prominent AI-agent startups for business automation. Previously founded virtual-office startup Teamflow (raised $50M+) and was an early PM at Uber where he ran Head of Product for JUMP and was a founding member of Uber Works. | Sources: Flo Crivello site Aakash Gupta newsletter

Gerrit Kazmaier, President, Product & Technology. The executive building Workday's AI agent platform. Former Google Cloud VP/GM of Data & Analytics (BigQuery, Looker) and SAP President for analytics/BI. | Sources: Workday Workday Newsroom

Hari Thyagarajan, Founder & CEO. Renamed his framework phidata to Agno in early 2025 to reflect its evolution from data tools to full agentic infrastructure. Agno is a fast-growing Python agent framework emphasizing multimodal agents, memory, and knowledge, positioned as a lightweight alternative to LangChain for teams building production agents. | Sources: GitHub Agno Crunchbase Pitchbook

Helena Tan, AI Agent Product, Box. Leads AI Agent product at Box, building enterprise AI agent capabilities on top of Box's content management platform. Previously Senior PM at Niantic working on Lightship Maps and Wayfarer community mapping tools, and Product Manager at Uber AI Maps focused on time/distance prediction models. GitHub | Sources: ZoomInfo GitHub

Jans Aasman, CEO, Franz Inc. (b. 1958). Dutch cognitive scientist who spent his early career as a part-time professor in Industrial Design at TU Delft and at TNO Research. Spent ~1995-2004 in telecommunications research building precursor technology to the iPad and Siri, gathering patents in speech technology, multimodal user interaction, and recommendation engines. | Sources: Wikipedia Franz allegrograph.com

Karan Vaidya, Co-Founder, Composio. Karan Vaidya co-founded Composio in 2023 with Soham Ganatra (a fellow IIT Bombay physics Olympiad teammate) to build the leading tool-integration layer for AI agents, connecting them to 200+ external apps and APIs. He previously held engineering roles at Nirvana, Rubrik, and Google. GitHub | Sources: GitHub Elevation Capital Lightspeed Venture Partners

Madhav Thattai, EVP & GM Agentforce, Salesforce. Madhav Thattai is Executive Vice President and General Manager of Agentforce at Salesforce, leading the company's autonomous AI agent strategy and roadmap. He was previously COO of Agentforce and COO of Customer Success at Salesforce, and earlier held senior roles at Google (Maps Enterprise GTM strategy and product marketing), a quantum computing startup, and Dell, where he spent eight years across manufacturing, customer service, and product. | Sources: Salesforce Salesforce Ben Indeed The New Stack

Manjeet Singh, Sr Director, Agentforce AI, Salesforce. Leads AI evaluation, observability, reinforcement learning, and multimodal capabilities for Salesforce Agentforce, one of the largest enterprise agent platforms. Deep technical background in ML systems at scale. | Sources: Product School Salesforce Admin Movate Salesforce

Manny Medina, Co-Founder & CEO Paid (formerly Outreach). Manny Medina is Co-Founder and CEO of Paid, a London-based startup building economic infrastructure for AI agents so they can be monetized by results rather than seats, backed by EQT Ventures and Sequoia. He previously co-founded sales engagement platform Outreach in 2014 and led it as CEO until 2024, growing it to a $4.4 billion valuation, and was the third employee on Amazon's AWS team and earlier led Microsoft's mobile division from launch to $50M in revenue. | Sources: Crunchbase TechCrunch GeekWire

Mihir Shukla, Co-Founder Chairman & CEO, Automation Anywhere. Founded Automation Anywhere in 2003 with a vision to automate up to 80% of enterprise work via Agentic Process Automation. Previously held leadership roles at E2Open, Kiva, Netscape, Infoseek, and Omnisky. | Sources: Automation Anywhere Crunchbase Bloomberg World Economic Forum

Pascal Weinberger, Co-Founder & CEO, Bardeen. Co-founded Bardeen in 2020 with Artem Harutyunyan to build context-aware AI agents and no-code automations for browser-based business workflows. Previously led the AI team at Telefonica Alpha (Europe's first moonshot factory), and worked at the intersection of machine learning and neuroscience earlier in his career. | Sources: Crunchbase Bardeen Business Wire Bloomberg

Tarun Dhupia, Co-Founder & CTO. Technical architect behind Composio's managed execution environment for AI agents. Builds the infra that handles auth, retries, and sandboxing for agent-to-SaaS connections at enterprise scale. | Sources: Composio

5.2 Family, Marriage, and the Trust Layer of AI Power

This section is about trust architecture, not gossip. At the highest levels of AI, many consequential relationships are not purely professional. Spouses co-found companies. Siblings build labs. Couples sit near capital flows. Long friendships become hiring channels. That does not make these relationships improper. In high-stakes company-building, trust, speed, and private information are currency. Personal relationships are part of how the system works. Every detail below passes a test: did this relationship shape a founding choice, funding decision, board vote, or hiring channel?

Most industries claim to keep work and life separate. This one doesn't even try. The professional and personal networks in AI are so densely interleaved that the November 2023 OpenAI board crisis was resolved in part by a Saturday-night conversation between a co-founder's wife and the board member he had voted against. The chapter that follows is about why that's structurally true, not gossip-true: who married whom, whose siblings co-founded what, which families have produced two AI executives, and what those overlaps actually do to governance when things go sideways.

The EA / Rationalist Trust Network

The effective altruism community has had an outsized influence on AI safety thinking, funding, and personnel. Among the people at the center of that work, long-term partnerships and shared intellectual commitments have reinforced one another in ways that are visible in the research record.

Paul Christiano + Ajeya Cotra: Christiano (RLHF co-inventor, founder of the Alignment Research Center) is married to Cotra, a senior researcher at Open Philanthropy where she leads grantmaking on AI timelines work. The marriage links a technical alignment lab to one of the largest AI-safety funders.

Amanda Askell + William MacAskill: Askell leads Claude's character training at Anthropic; MacAskill co-founded the EA movement and its philanthropic networks have directed hundreds of millions of dollars into AI safety research, including grants to organizations that share staff with Anthropic. The overlap matters because EA-aligned funders helped seed the labor pool Anthropic hires from.

Leopold Aschenbrenner + Avital Balwit: Aschenbrenner runs a multi-billion-dollar AI-focused fund and is engaged to Balwit, Anthropic's Chief of Staff. The pairing puts a major AI-equity investor and a senior Anthropic operating role in the same household, a governance-proximity worth naming. Fortune

Nick Bostrom / Oxford cluster: Nick Bostrom's 2014 book Superintelligence was arguably the most influential single work in bringing AI existential risk into mainstream policy and academic discourse; until 2024 he led Oxford's Future of Humanity Institute, and FHI alumni have since populated Anthropic, DeepMind, ARC, MIRI, and other safety-adjacent organizations. Wikipedia

Greg + Anna Brockman

Married November 2019 at OpenAI's San Francisco offices; Ilya Sutskever officiated. Public reporting on the November 2023 crisis, including Wikipedia coverage, noted that family and social ties within the OpenAI circle played a role in the resolution that brought Altman back. Governance crises at this scale rarely resolve through formal process. The actual turning point tends to be one conversation, with one person, where trust still exists. That is an operational reality most org charts do not capture.

Shivon Zilis

Shivon Zilis studied economics and philosophy at Yale, worked at Bloomberg Beta and served on OpenAI's nonprofit board, then joined Neuralink in 2019 as Director of Operations. She has been deeply involved in Neuralink's first human patient trials (Noland Arbaugh, implanted January 2024). Wikipedia

Personal Trust as Operating Infrastructure

The Amodei Family

Dario (CEO, born 1983) and Daniela (President, born 1987) are the sibling pair running Anthropic, widely considered one of the most influential AI safety companies in operation. Their father Riccardo was an Italian leather craftsman from Tuscany, who died in 2006, an event Dario has said redirected him from physics to AI. Their mother Elena is Jewish-American from Chicago. Daniela married Holden Karnofsky, co-founder of Open Philanthropy and GiveWell, making the Anthropic-EA connection not just ideological but literally familial. The Amodeis were named to the TIME 100 together in 2026. Two siblings, one company, $60B+ valuation, and a mission to make sure artificial intelligence doesn't end the world. The thing about sibling co-founders is that the trust is deep and the disagreements are deeper, and neither of those things is a weakness. Time

Demi Guo, Pika Labs, and the "Daughter Concept Stock"

Demi Guo (Chinese name Guo Wenjing) was born in 1998 in Hangzhou. She enrolled in Stanford's AI PhD program, then dropped out in April 2023 with her classmate Chenlin Meng to co-found Pika Labs, an AI video-generation startup based in San Francisco. Pika 1.0 launched in November 2023 and went viral on social media within a week. As of 2025, the company had raised more than $135M from Lightspeed, Adam D'Angelo, Adam Bain, and others, and is one of the three companies named in the Pika/Runway/Sora video-generation race profiled later in this book.

5.3 The Funding & Talent Wars: Every Big Round, Every Talent Bet

In 2024 the salaries became unanswerable. Meta paid Jason Wei and Hyung Won Chung roughly $300M each to leave OpenAI. NVIDIA reached a ~$20B licensing-and-assets deal with Groq, a chip company founded by Jonathan Ross, a high-school dropout who later studied math and computer science at NYU and designed Google's original TPU. Groq. SSI raised $1B at a $5B valuation with no product or revenue and a single-sentence research thesis (per TechCrunch). This chapter is the documented record of what a category looks like when the price of being right early is bigger than the price of being right late, told through specific deals (Cursor, Groq, SSI, Inflection, Adept) and specific compensation packages.

Groq: The 5.3 Funding-Cycle Cut

Inference-optimized custom silicon built by the engineer who designed Google's original TPU. By December 2025 Groq had reached a licensing arrangement with NVIDIA (CNBC). Full profile of the founder and the company: Jonathan Ross, The Original TPU Engineer Who Built Groq →

Cursor: The 5.3 Funding-Cycle Cut

Four MIT students, a $400K pre-seed, $100M ARR in 20 months with zero marketing spend, approaching a $50B valuation by 2026 (Contrary Research), one of the defining velocity stories of this funding cycle. Full company profile: Cursor / Anysphere, The ~$50B Surprise →

SSI, $3B With Zero Product

Recursive, $650M to Build AI That Improves Itself

In May 2026, Richard Socher (Stanford PhD under Chris Manning and Andrew Ng, prior MetaMind/Salesforce/You.com) announced Recursive, a $650M-funded lab building "recursive self-improving superintelligence." The seven co-founders came from five different frontier labs at once (DeepMind, OpenAI, Salesforce, Meta FAIR, Google Brain). GV, Greycroft, NVIDIA, and AMD all backed the round, the last two rarely co-investing (Tech.eu).

Alfred Lin

Alfred Lin met Tony Hsieh at Harvard reselling dorm-room pizza by the slice; they later built LinkExchange (Microsoft, $265M in 17 months) and Zappos (Amazon, $1.2B). Lin joined Sequoia in 2010 and in 2021 championed the ~$20M bet on OpenAI at $20B; three years later OpenAI was worth $852B. In November 2025 Sequoia elevated him to co-managing partner alongside Pat Grady, succeeding Roelof Botha; he was named #1 on the Forbes Midas List the same year.

Why he matters to the map: He championed Sequoia's $20M check into OpenAI at $20B and now runs the firm.

Josh Kushner: The Only Term Sheet

In early 2022 every major VC firm passed on OpenAI's round. Josh Kushner's Thrive Capital wrote the only term sheet: $130M at $29B. When the OpenAI board fired Altman in November 2023, COO Brad Lightcap's first call was to Kushner, who supported reinstatement publicly. By 2026 Thrive had committed approximately $3B across rounds, and in an unprecedented structure, OpenAI took an equity stake in Thrive Holdings, embedding its engineers and models across Thrive's portfolio (Fortune).

Why he matters to the map: He wrote the only term sheet of OpenAI's 2022 round and got the first call when the board fired Altman.

Neil Mehta and the $500M SSI Bet

Neil Mehta's grandfather, who ran a gun shop in Ahmedabad selling antique Maharaja pistols, taught him to see craftsmanship as value, which became Greenoaks Capital's investment thesis. In 2012 he put 40% of his $50M first fund into Coupang and added $500M when it stumbled; the 2021 IPO at $60B returned more than 10x. For two years after ChatGPT, Greenoaks refused every frontier-AI deal. In April 2025 he led the $2B SSI round at $32B and reportedly committed $500M or more (Colossus).

Daniel Gross: The Triple-Hatted Year

Daniel Gross sold Cue to Apple in 2013, led AI at Apple, became a YC partner, and in 2023 co-founded the $1.1B NFDG fund with Nat Friedman (portfolio included SSI, ElevenLabs, Perplexity, Character.AI). When Sutskever asked him to be SSI's founding CEO, Gross spent 13 months running both NFDG and SSI at once, a structure no one in AI had occupied before. In June 2025 Meta acquired a 49% stake in NFDG and hired both partners; Gross departed SSI and Sutskever took over as CEO (CNBC).

Why he matters to the map: He held SSI, NFDG, and YC roles simultaneously, the rare investor sitting inside a deal and its competition.

Jordan Jacobs and the Vector Institute

Jordan Jacobs, an entertainment lawyer, co-founded Layer 6 in 2016 with researchers from Geoffrey Hinton's lab; TD Bank acquired it in 15 months. Worried that Canadian AI talent was bleeding to US tech, he co-founded the Vector Institute in 2017 with Hinton on board and $80M in initial funding, then started Radical Ventures, where Hinton became an investor and Fei-Fei Li joined as Scientific Partner in 2023 (Radical Ventures).

The Anthropic Arms Race: $1B to $380B

In January 2023, with Anthropic under 100 employees and no public product, Lightspeed's Guru Chahal led the first big check. The escalation: Series E $3.5B at $61.5B (March 2025, Lightspeed), Series F $13B at $183B (September 2025, ICONIQ lead with Fidelity and Lightspeed), Series G $30B at $380B (February 2026, GIC and Coatue co-lead), the largest private AI round ever, oversubscribed from a $10B target. Amazon added $5B on top of $8B; Microsoft and NVIDIA contributed up to $15B combined (Anthropic).

The Business Minds: OpenAI's Unsung Team

Sarah Friar (Square IPO, ex-Nextdoor CEO, OBE) became OpenAI's first CFO in June 2024, tasked with managing a $122B raise. Fidji Simo joined as President of Applications in August 2025 (full profile below). Kevin Scott, Microsoft CTO from rural Virginia, sent the June 2019 memo to Nadella warning Microsoft was "multiple years behind" Google on ML scale; Nadella replied "this is why I want us to do this" and the initial $1B OpenAI investment, now $13B, followed (Fortune).

Alexandr Wang & the Meta–Scale Deal

The 2025 Meta–Scale acquisition ($14.3B for 49% at $29B, with Wang becoming Meta’s Chief AI Officer and heading Meta Superintelligence Labs) is one of the defining talent moves of this funding cycle. Full profile in Chapter 6: Alexandr Wang, Scale AI →

Jason Wei & Hyung Won Chung, The Chain-of-Thought Architects

Jason Wei invented chain-of-thought prompting at Google Brain in 2022, the technique that taught LLMs to reason step-by-step and now underlies every modern reasoning model. Hyung Won Chung's MIT PhD work formalized instruction fine-tuning at scale (FLAN-PaLM, Flan-T5, 2022), now a standard phase in every frontier training pipeline. Both moved Brain to OpenAI (contributing to the o1 reasoning line), then to Meta MSL in July 2025 for reported packages of approximately $300M each.

Soumith Chintala, The Man Who Open-Sourced Modern AI

Soumith Chintala did his graduate work in Yann LeCun's NYU lab, then joined Meta FAIR in 2014 and co-created PyTorch in 2017. By 2020 it had overtaken TensorFlow in academic citations; by 2022, in production. Every major frontier model, GPT, Claude, Gemini, Llama, trains on PyTorch or a descendant. He donated PyTorch to the Linux Foundation in 2022 and left Meta in November 2025; the bet that neutrality outlasts proprietary advantage has held.

Jakub Pachocki, OpenAI Research Chief

Jakub Pachocki competed six times in the International Olympiad in Informatics for Poland, earned a CMU PhD, and joined OpenAI to lead the project that beat Dota 2 world champion OG in 2019. He architected GPT-4 and later led the o1 and o3 reasoning lines. Ilya Sutskever named him Chief Scientist on the day Sutskever announced his own departure in May 2024; TIME 100 AI 2025.

Lilian Weng, The Safety Architect

Lilian Weng joined OpenAI in 2017 as roughly its 60th employee and built the safety team from zero to 80+ people, owning the RLHF pipelines, classifiers, red-teaming, and alignment infrastructure underneath GPT-3, InstructGPT, ChatGPT, and GPT-4. Her blog Lil'Log made transformers, RLHF, and agent design legible to hundreds of thousands of readers. She left in November 2024 to co-found Thinking Machines Lab, which raised $2B at $12B in July 2025 (Wikipedia).

Noam Brown, The Researcher Who Refused $100M

Noam Brown's 2017 Libratus beat four top professional poker players over 120,000 hands at Rivers Casino in Pittsburgh, the first AI to crack heads-up no-limit Texas Hold'em. Pluribus (Science, 2019) extended that to six-player poker, and CICERO (Science, November 2022) became the first AI to reach human-level play at Diplomacy. In 2025 he declined Meta's reported $100M-plus offer to stay at OpenAI working on reasoning research.

David Holz: The Man Who Said No to Everything

David Holz (Max Planck physics, NASA LiDAR) co-founded Leap Motion in 2010; it raised $120M from a16z, Founders Fund, and Intel, hit a $306M valuation, and sold for $30M in 2019. When he founded Midjourney in 2021 he refused all VC: no investors, no website, no press, no marketing, run entirely through Discord. As of early 2026, Midjourney was approaching $600M in annual revenue with roughly 100 to 160 employees.

Palmer Luckey: From Homeschool to $60B

Palmer Luckey was homeschooled in Long Beach, built VR headsets in the family garage, and founded Oculus at 19 after a Carmack-endorsed Kickstarter. Facebook bought Oculus for $2B in 2014. After his 2017 exit from Facebook (cause publicly disputed), he co-founded Anduril Industries with former Palantir executives; by March 2026 the company held a $20B-plus US Army contract and was valued at $60B (TechCrunch).

Chip Huyen

Chip Huyen grew up in a Vietnamese rice-farming village, was rejected by Stanford on first application, spent three years writing four bestselling Vietnamese travel books across 25 countries, then was admitted second-try. At Stanford she created CS 20 (TensorFlow for Deep Learning Research). Her O'Reilly book "Designing Machine Learning Systems" (2022) became the canonical reference for production ML, translated into 10-plus languages.

swyx

Shawn Wang (swyx), a Wharton MBA, walked away from a ~$350K currency-options trading seat for a coding bootcamp. His 2018 "Learn in Public" essay seeded a generation of developer-educators. He co-founded the AI Engineer Summit (2023, instant sellout), the AI Engineer World's Fair (2024, 3,000-plus attendees in San Francisco), and runs the Latent Space podcast reaching over 10M annually.

Fidji Simo

Fidji Simo grew up in a Sicilian immigrant fishing family in Sète, France, first in her family to finish high school, then HEC Paris, UCLA Anderson, and a decade at Facebook running the main app for 3 billion users. As Instacart CEO from 2021 she froze headcount, built almost $1B in ad revenue, and IPO'd in September 2023 at $10B, breaking the longest tech IPO drought in 20 years. She joined OpenAI as President of Applications in August 2025 and took medical leave in April 2026 for a POTS relapse (CNBC).

Who Acquires Whom: Early Patterns from 38 Corporate Buyers

Note: This analysis is scoped to the most active corporate acquirers on the map, frontier labs, hyperscalers, and AI-adjacent platforms. The patterns below are directional observations from this sample, not comprehensive industry statistics. As coverage expands, these numbers will shift.

We scanned 41 of the most active corporate buyers on the map, frontier labs (OpenAI, Anthropic, DeepMind, Mistral), hyperscalers (Microsoft, Google, Meta, Amazon, NVIDIA), and AI-adjacent platforms (Salesforce, Snowflake, Databricks, Stripe, Workday, and others). 38 returned usable acquisition histories via Gemini 2.5 Flash with Google Search grounding; three timed out (AMD, Sequoia, a16z) and were dropped. Those 38 companies surfaced 144 candidate acquisition events between 2024 and mid-2026. Each was independently re-searched against tier-one press and SEC filings: 121 confirmed, 21 unverifiable and removed.

Across the verified 121 deals, the clearest pattern is that operator-founders (Benioff at Salesforce, Huang at NVIDIA, Ramaswamy at Snowflake) acquired roughly 3× more than lab-founders (Amodei at Anthropic, Mensch at Mistral, Liang Wenfeng at DeepSeek). About 23% of verified acquisitions were vertical SaaS, companies bought for domain-specific data. The majority were horizontal AI infrastructure, talent acqui-hires, or adjacent technology acquisitions.

5.4 The Google Era & The PayPal Mafia: The Launching Pads

Two launchpads pre-wired the AI industry: Google trained the researchers, PayPal trained the funders. Almost every founder, board member, and check-writer in the preceding chapters trained in one of these two systems before building everything else. This chapter is the synthesis bridge, each person below has been profiled in full in their thematic chapter; here they appear under the launchpad that produced them, with the through-line role made explicit.

Google: The Research Launchpad

Five Google-trained figures anchor the research side of the AI cast. Jeff Dean (Google's near-original employee, co-designer of MapReduce, BigTable, TensorFlow, and the first TPUs; co-founder of Google Brain 2011; Chief Scientist of the merged Google DeepMind from 2023) is the institutional through-line from PageRank-era Google to the Gemini era. Sundar Pichai (joined 2004 to lead Chrome, CEO from 2015) issued the internal “code red” after ChatGPT and merged Brain with DeepMind in April 2023, staking Google's future on Gemini. Larry Page and Sergey Brin (Stanford 1998 co-founders) hold the controlling super-voting stake; Page personally championed the 2014 DeepMind acquisition against board skepticism, and Brin returned to hands-on Gemini engineering work in 2023. Andrew Ng (co-founded Brain in 2011 alongside Dean and Greg Corrado, then Baidu Chief Scientist 2014-2017, then Coursera, DeepLearning.AI, and the AI Fund) ran the world's largest AI education pipeline.

PayPal: The Capital Launchpad

Six PayPal alumni anchor the capital side. Peter Thiel (Confinity co-founder) was the only PayPal founder who became a founding donor to OpenAI in 2015 while also chairing Palantir to its 2020 IPO. Elon Musk (X.com co-founder, briefly PayPal's CEO before being ousted in late 2000, per Wikipedia) parlayed the ~$180M PayPal exit into SpaceX and Tesla, then became a founding donor and director of OpenAI, and now runs his own frontier AI lab (xAI) while litigating against the lab he helped seed. Reid Hoffman (LinkedIn co-founder, sold to Microsoft for $26.2B) was the early OpenAI board member who personally bridged OpenAI salaries when Musk's pledged $1B fell short in 2018. Max Levchin (Confinity co-founder, engineering and cryptography lead) went from PayPal to Slide (sold to Google), HVF, and Affirm (founded 2012, IPO 2021), and chairs Yelp's board. David Sacks (PayPal's COO under Thiel, founded Yammer and Craft Ventures, All-In co-host) was named White House AI & Crypto Czar in December 2024. Keith Rabois (PayPal VP of Business Development) ran operating roles at LinkedIn, Slide, and Square (COO), then partner at Khosla, Founders Fund, and back to Khosla in 2024; he co-founded OpenStore in 2021 and remains one of the most active personal angel investors in early-stage AI.

The launchpads matter because they explain why the funding graph looks the way it does. Trained inside Google, the researchers had a shared vocabulary about scale, infrastructure, and what "production" means. Cashed out by PayPal, the funders had a shared vocabulary about contrarian conviction, founder selection, and how much capital to commit before there is anything to show. Most other chapters of this book are downstream of those two languages meeting.

5.5 AI Exits & Acquisitions: Every Deal, Every Price (1998-2025)

The "No-Revenue, Nine-Figure Exit" Pattern: What's Actually Being Bought

A recurring shape in this dataset is a category that barely existed before 2013: the multi-hundred-million-dollar (sometimes multi-billion-dollar) acquisition of a company with little or no commercial revenue, sometimes no shipped product at all. Standard SaaS revenue multiples do not apply, because revenue is not what's being bought.

This section exists so that investors and M&A bankers stop reading "no revenue" on a target's P&L as a deal-killer, and start reading it as the headline.

Every nine-figure no-revenue exit below is a bet on one or more of those three. The annotation in brackets after each example flags which asset(s) the buyer was paying for.

The pattern is not limited to no-revenue companies. The clearest indicator that the SaaS comp set has moved is Scale AI / Meta partnership, 2025 ($14.3B for 49% at a $29B valuation). Scale was already profitable when Meta wrote the check, yet the deal was priced as a data-infrastructure position, not on a revenue multiple. Alexandr Wang came in as Meta's Chief AI Officer heading Meta Superintelligence Labs. A few months later the Pentagon handed Scale a ~$500M contract on top, roughly five times the prior year's deal (Forbes). When the buyer is paying for labeled data as a chokepoint, even a profitable target gets priced like the no-revenue ones above.

The same shape appears on the still-independent side: Safe Superintelligence (Ilya Sutskever) raised $1B in 2024 at a $5B valuation, then another $2B at $32B in 2025, with no product, no revenue, and fewer than twenty employees, pricing People + the model architecture yet to be built. AMI Labs (Yann LeCun, after leaving Meta) raised ~$1.03B seed at $3.5B in 2025 with no saleable product expected for roughly five years, a pure bet on a Turing Award winner's research thesis and the team he assembles around it. Fei-Fei Li's World Labs raised $230M at a $1B valuation in October 2024, also pre-product, pricing the spatial-intelligence dataset thesis + Fei-Fei's name + the founding team. All three deals price reputation, technical thesis, and the option value of getting to AGI (or to the next computing primitive) first.

Further evidence that data alone is being repriced as an asset class: Reddit licensed its archive to Google (~$60M/year, 2024) and to OpenAI on overlapping terms, same trade as Scale AI, different wrapper. The chokepoint is labeled and conversational human data, and incumbents are paying nine figures per year for access.

The throughline: in this market, a small group of credentialed researchers and the data and model weights around them is itself the asset class. Incumbents (Google, Microsoft, Amazon, NVIDIA, Apple, Meta) have shown they will pay nine to ten figures for that asset whether or not it ever generated a dollar of revenue. The investors and bankers who see this earliest are the ones writing the next round of the deals listed above.

Private Valuations Beyond the Frontier Labs

The frontier labs (OpenAI, Anthropic, xAI, DeepMind, SSI) dominate the headline numbers, but a second tier of AI-adjacent private companies is now sitting on equally striking valuations.

VC Firepower Backing the Bets

Active Funding Rounds (as of May 2026)

📅 Last verified: May 2026 · Sources cited inline

Chapter 6 · The Enterprise & Infrastructure Stack

AI will reach most knowledge workers through the enterprise software they already use, not through a new app they choose from scratch. Three of enterprise software's most durable institutions have placed significant bets on AI, each shaped by its own culture, customers, and business model. Together they manage the workflows of hundreds of millions of employees globally, so their choices will define how AI first shows up at work.

The AI Enterprise Stack, Layer by Layer

Most of the chapter ahead is about individual product wars: Agentforce versus ServiceNow agents, Copilot versus Cursor, the LLM tooling layer, the chip supply chain. The diagram below puts those product fights into the larger structure they sit inside. The stack has five readable layers, chip to enterprise app, and the most important commercial fact about modern AI is that no major frontier model is more than two layers away from depending on NVIDIA's silicon. The dashed lines highlight the most consequential cross-layer commitments in the dataset.

Three observations from this picture. First, the chip layer is the smallest and most concentrated, with NVIDIA visibly larger than the rest combined and TSMC manufacturing nearly everything that matters above it. Second, the model layer is where the named conflicts of this book mostly play out, but each of those competitors is locked into a cloud commitment that is harder to switch than a model commitment. Third, the application layer is the only place in the stack with both real product variety and real demand-side power, which is why the next sections of this chapter are mostly fights at that layer rather than further down.

Marc Benioff, The Evangelist Who Found a New Religion

Marc Russell Benioff was born September 25, 1964, in San Francisco, into a Jewish family with retail DNA, his grandfather was an executive at Sears. He began his entrepreneurial journey at age 15 by founding Liberty Software in 1979, building Atari 8-bit games (Flapper, King Arthur's Heir) to pay for college. He joined Oracle after graduation and spent 13 years under the mentorship of Larry Ellison, learning the Oracle playbook so thoroughly that when he broke free to found Salesforce in 1999, Ellison immediately saw the threat and attempted to build a competing CRM. Marc Benioff has described this tension as "epic love-hate": the man who taught him everything also became his most dangerous rival.

Read full profile ↓

Founded Salesforce (Post-corporate · 1999 · age 35). Marc Benioff spent 13 years at Oracle Corporation, became the youngest vice president in the company's history, and was a millionaire by age 25. He left Oracle in 1999 after a sabbatical that gave him the idea for a web-based software company. Exit: Salesforce completed its initial public offering (IPO) on June 23, 2004, on the New York Stock Exchange (NYSE) under the ticker symbol CRM. The company raised $110 million at $11 per share and was the best-performing tech IPO of 2004. (The History of Salesforce · Marc Benioff - Wikipedia)

Salesforce was founded from a San Francisco apartment on $500K of Marc Benioff's personal investment, no Sand Hill Road VC would touch the idea of "software as a service" delivered over the internet. He raised $17M privately from individuals and IPO'd in 2004. As of December 2025, the company carried a market capitalization of approximately $250B (per stockanalysis.com), having pioneered the enterprise SaaS model that underpins much of Silicon Valley's wealth.

Agentforce: Marc Benioff's current obsession, an AI agent platform that automates customer service, sales, and ops workflows. "We now have 8,000 Agentforce customers," he announced at Dreamforce 2025, as reported by Salesforce. The platform built on Einstein (Salesforce's AI layer since 2016) represents Salesforce's largest strategic bet: that enterprises will pay for AI agents the same way they paid for CRM seats. The platform team is led by Adam Evans (departed early 2026) and Madhav Thattai, with AI oversight from SVP Christina Abraham. If it works, Salesforce adds a second revenue engine as large as its original. If it doesn't, the company faces a generation of disruption from purpose-built AI startups.

Bill McDermott, The Deli Kid Who Runs a Trillion-Dollar Workflow

William R. McDermott was born August 18, 1961, in Amityville, Long Island. His grandfather was Naismith Memorial Basketball Hall of Fame player Bobby McDermott, a self-made athlete from the same working-class Long Island that shaped Bill. At 16, Bill purchased the Amityville Country Delicatessen for $7,000, installed video games to out-compete the 7-Eleven across the street, made it profitable enough to fund his entire college education at Dowling College, then sold it for enough to buy his parents a home in Myrtle Beach. He joined Xerox after graduation and became the company's youngest-ever general manager at 25. He tells this story at every keynote. It is his founding myth, the deli as parable for his approach to every problem: find the angle, outmaneuver the incumbent, take care of your family.

Read full profile ↓

At SAP, McDermott served as co-CEO and then CEO from 2010-2019, growing the company's market value from $39B to $156B. He departed in October 2019 during a period of pressure from activist investor Elliott Management and ongoing customer feedback about pricing. What no one writes about is that running a company through activist pressure while your customers are publicly unhappy is a two-front problem: the board room and the customer base move on different timelines and different logics, and you cannot fix one without the other knowing. In November 2019 he became CEO of ServiceNow, transforming it from an IT ticketing platform into an enterprise workflow behemoth competing directly with Salesforce and Microsoft on AI agent territory.

The glass eye: In July 2015, McDermott was staying at his brother's home and fell on broken glass, losing vision in his left eye. He subsequently had a glass eye fitted. He wears sunglasses in most public appearances. He has spoken about the experience as a lesson in resilience, the CEO of a $150B company has spoken publicly about the injury and recovery as a clarifying experience. He regards it, publicly at least, as clarifying.

(source: SAP chief executive loses eye after fall - Financial Times)

ServiceNow's AI play centers on the Now Platform's AI agents, integrating with large language models to automate IT service management, HR workflows, legal case routing, and finance approvals. As of March 2025, ServiceNow announced the $2.85B acquisition of Moveworks (closed December 2025), bringing Bhavin Shah's enterprise AI assistant in-house to transform how organizations handle support workflows; the deal represents the most expensive single AI bet in ServiceNow's history (acquisition reported by moveworks.com).

Satya Nadella, The Empathy CEO

Satya Nadella was born August 19, 1967, in Hyderabad, India, to a Telugu Hindu family. His father Bukkapuram Satya Nadella Yugandhar was an Indian Administrative Service officer; his mother Prabhavati was a Sanskrit lecturer. He became CEO of Microsoft in 2014, succeeding Steve Ballmer. He has since become the most successful tech CEO of his generation in terms of value creation, Microsoft's market cap under Satya Nadella grew from $300B to nearly $3T.

Read full profile ↓

Zain and the empathy curriculum: Satya Nadella has spoken extensively about how fatherhood, particularly raising Zain, fundamentally reshaped his approach to empathy, accessibility, and inclusive design. He describes an early realization that he was approaching his son's condition as a problem to be managed rather than a human experience to be understood, and that this shift in perspective was the most important leadership lesson of his life. He explores these themes in "Hit Refresh" (2017). Microsoft's emphasis on accessibility features, inclusive design, and cognitive diversity in product thinking traces directly to Satya Nadella's personal formation as a parent.

The $13B OpenAI bet: Satya Nadella's initial $1B investment in OpenAI in 2019 was internally controversial, Bill Gates reportedly told him he would "burn this billion dollars." Satya Nadella navigated board skepticism and championed the partnership anyway. As of October 2025, Microsoft's cumulative investment in OpenAI exceeded $13B, securing exclusive cloud hosting rights and a 27% stake in OpenAI Group PBC valued at approximately $135B, as confirmed by Wikipedia. In the November 2023 crisis when OpenAI's board fired Sam Altman, Satya Nadella was reportedly blindsided, but responded strategically by immediately offering Sam Altman a position at Microsoft, effectively presenting the board with a consequential choice: reinstate its CEO or watch him land at a rival. The board chose reinstatement. The move looked decisive in hindsight. At the time, Nadella was committing Microsoft's credibility to a public position before knowing whether the board would fold, whether Sam would actually take the Microsoft role, or whether the OpenAI employees threatening to walk would follow through.

Mustafa Suleyman, The Complicated Architect

Mustafa Suleyman was born in 1984 in London to a Syrian-immigrant father (a taxi driver) and an English mother. He has spoken publicly about growing up in a Muslim household and identifying as an atheist today. At 19 he dropped out of Oxford to help co-found Muslim Youth Helpline, a telephone counseling service. The shape, picking a population and building a service to reach them, recurs in everything he has built since.

Read full profile ↓

In 2010 he co-founded DeepMind Technologies with Demis Hassabis and Shane Legg. Google acquired DeepMind for approximately £400M in 2014 (The Guardian). Mustafa Suleyman became head of applied AI, overseeing the deployment of DeepMind's research into Google's products.

DeepMind departure, the misconduct investigation: In August 2019, Mustafa Suleyman was placed on administrative leave at DeepMind. An external lawyer was hired to investigate following staff complaints about management conduct and a what CNBC described as a hostile workplace. CNBC reported that settlements with some affected staff were reached, the terms of which were not disclosed. His management duties were stripped; he officially departed in December 2019. Suleyman has publicly acknowledged he "drove people too hard." He was not formally adjudicated of wrongdoing.

After leaving Google, Mustafa Suleyman joined Greylock Partners as venture partner (January 2022), then in March 2022 co-founded Inflection AI with Reid Hoffman and Karén Simonyan. Inflection raised $1.525B total (Microsoft, Nvidia, Bill Gates, Reid Hoffman), as announced via Business Wire; its consumer chatbot Pi gained attention but did not break out at frontier scale. As of March 2024, Microsoft paid roughly $650M for a technology license and hired most of the Inflection team, including Mustafa Suleyman and Simonyan, in a structure that avoided a full acquisition process. He became CEO of Microsoft AI, a new division overseeing Copilot, Bing, MSN, and Edge.

His 2023 book "The Coming Wave" explores AI, control, and societal adaptation, a sophisticated treatment that simultaneously champions technological progress and advocates for governance frameworks. He is one of the few AI executives credibly positioned in both camps.

Why he matters to the map: He routes DeepMind talent and capital through Inflection into Microsoft, the rare founder embedded in three frontier institutions.

Extended Roster · 7 more from this chapter

Charles Lamanna, President Business & Industry Copilot, Microsoft. Charles Lamanna is EVP/President of Business & Industry Copilot at Microsoft, leading product innovation across Power Platform, Dynamics 365, and Copilot Studio. A former startup founder (MetricsHub, acquired by Microsoft in 2013), he has built Power Apps into a low-code/no-code leader with 25M+ monthly users and now leads Microsoft's enterprise AI agent strategy. | Sources: Microsoft Ignite Microsoft Fortune

Danah Boyd, Partner Researcher, Microsoft Research / Founder Data & Society (b. 1977). danah boyd is Partner Researcher at Microsoft Research and founder (2013) of the Data & Society Research Institute, which examines the social, technical, ethical, and policy implications of data-centric technology. As of 2026 she is also the Geri Gay Professor of Communication at Cornell and a visiting distinguished professor at Georgetown. | Sources: Wikipedia Data and Society danah boyd site Georgetown

Emre Kiciman, Partner Research Manager / Head Copilot Tuning Research, Microsoft. Senior Principal Researcher at Microsoft Research who leads the Copilot Tuning Research group, advancing model fine-tuning innovations for productivity scenarios across Microsoft 365 Copilot. His broader research bridges causal machine learning, AI, and societal impact, including the widely-used DoWhy open-source library for causal inference. | Sources: Microsoft Emre Kiciman site Google Scholar

Hanna Wallach, VP and Distinguished Scientist FATE/STAC, Microsoft Research (b. 1979). Leads the Sociotechnical Alignment Center (STAC), an interdisciplinary applied science team in Microsoft Research's FATE (Fairness, Accountability, Transparency, Ethics) group focused on responsible AI evaluation and measurement of generative AI systems. Best paper award winner at AISTATS, CHI, and NAACL; served as senior program chair (NeurIPS 2018) and general chair (NeurIPS 2019). | Sources: Wikipedia Microsoft Google Scholar

Jaime Teevan, Chief Scientist and Technical Fellow, Microsoft. Sold her Yale senior thesis to early search engine Infoseek as an undergraduate; pioneered the field of Personal Information Management at MIT. From 2017-2018 served as the first AI-background Technical Adviser to CEO Satya Nadella. | Sources: Wikipedia Microsoft Jaime Teevan site

Jake Hofman, Senior Principal Researcher, Microsoft Research NYC. Founding member of Microsoft Research New York City; before MSR was a member of the Microeconomics and Social Systems group at Yahoo! Research. | Sources: Microsoft Google Scholar

Tomer Cohen, Chief Product Officer. Has been LinkedIn's CPO since 2012, overseeing AI product development for the 1-billion-member professional network. Currently leading LinkedIn's transformation to AI-native development, including LLM-powered job matching, AI search, and a 'Full Stack Builder' program replacing traditional PM roles. | Sources: Computer Weekly Jerusalem Post Lenny's Newsletter

Before anyone can build an AI product, someone has to build the foundation. This chapter documents the engineers and founders who constructed the invisible layer beneath every AI product in this book: the databases, frameworks, protocols, and tools that rarely earn a headline but quietly determine what is possible. The hard part for infrastructure founders is not building the thing. It is building it before anyone agrees the category needs to exist, against larger commercial players who can add it as a feature once the market is proven.

Aravind Srinivas, The Fastest Billionaire in AI

Aravind Srinivas was born June 7, 1994, in Chennai, India, to a family more comfortable with finance than engineering, choosing electrical engineering at IIT Madras was itself a departure. His first choice was computer science; he missed the cutoff by a narrow margin and was placed in electrical engineering instead. The redirected path turned out to be formative: the mathematical and physical rigor of EE gave him depth that pure CS graduates lacked. He graduated with a dual degree (B.Tech + M.Tech) before moving to the US for a PhD at UC Berkeley. He owned no personal computer; he worked on shared cloud resources from 5:30 AM to 8 PM daily. He did research stints at Google Brain, DeepMind, and OpenAI, three of the four most important AI labs on earth, before dropping out to build Perplexity.

Read full profile ↓

Founded Perplexity AI (Right after school · 2022 · age 28). Aravind Srinivas completed his Ph.D. in Computer Science from the University of California, Berkeley in 2021. He then held research positions at OpenAI, Google, and DeepMind before co-founding Perplexity AI in August 2022. Exit: Perplexity AI is still an active, privately held company. As of September 2025, it was valued at $20 billion. There were reports of Apple exploring an acquisition in July 2025, but these were internal discussions and the deal did not materialize. A partnership with Snap for AI search integration, announced in November 2025, was amicably terminated in Q1 2026. (Perplexity AI - Wikipedia · Perplexity AI: Valuation, Funding & Investors - PitchBook)

Copyright disputes, 2024: Perplexity faced a sustained copyright crisis in 2024. Forbes reported the Pages tool lifting paywalled content; the New York Times sent a cease-and-desist; Dow Jones and NY Post Holdings filed lawsuits over alleged unauthorized scraping and reproduction. Perplexity launched a publishers' revenue-sharing program in July 2024 as a partial remedy. Forbes · TechCrunch

Perplexity raised $500M at a $9B valuation as of December 2024, then climbed to $18B by mid-2025 (per Wikipedia). Aravind Srinivas's estimated net worth: approximately $2.5B as of late 2025, making him India's youngest billionaire.

The pattern is that application companies and infrastructure standards rise together. Perplexity shows demand for new interfaces, and MCP shows how the trust layer forms underneath them.

David Soria Parra & The MCP Protocol

David Soria Parra is a Staff Engineer at Anthropic who co-created the Model Context Protocol (MCP) with Justin Spahr-Summers, announced November 2024. MCP is an open standard enabling AI assistants to connect to any data system, codebases, business applications, APIs, through a standardized interface rather than bespoke integrations. The protocol achieved 110 million monthly downloads within roughly half the time React took to reach that milestone. Anthropic subsequently donated MCP to the Agentic AI Foundation (a Linux Foundation directed fund), with co-founding support from Block and OpenAI and backing from Google, Microsoft, AWS, Cloudflare, and Bloomberg. In twelve months, a staff engineer's internal project became the foundational internet protocol of the agentic AI era.

Harrison Chase, LangChain and Its Discontents

Harrison Chase built LangChain at a hackathon in October 2022, while working as a lead ML engineer at Robust Intelligence. Over a weekend, he assembled an open-source framework that would become the most widely installed LLM application library, 20M+ installs. He studied statistics and computer science at Harvard (class of 2017), had worked at Kensho Technologies on entity linking, then at Robust Intelligence on model testing. LangChain raised $25M in Series A in May 2023. The framework became so ubiquitous that "LangChain" became a shorthand for "building with LLMs."

Read full profile ↓

Founded LangChain (A few years out · 2023 · age 27). Harrison Chase graduated from Harvard University in 2017. He worked as a Machine Learning Engineer at Kensho Technologies from 2017 to 2019, then as Head of ML at Robust Intelligence from 2019 until 2023, before co-founding LangChain. Exit: LangChain is still running as a privately held, venture capital-backed company. It has raised a total of $260 million across multiple funding rounds, reaching a $1.25 billion valuation in October 2025 with its Series B round. (Report: LangChain Business Breakdown & Founding Story | Contrary Research · LangChain - 2026 Company Profile, Team, Funding & Competitors - Tracxn)

Harrison Chase + Jerry Liu, Parallel Origins at Robust Intelligence

Jerry Liu (Princeton CS '17) was Harrison Chase's colleague at Robust Intelligence. Within days of LangChain's release in October 2022, Jerry Liu identified the gap: LangChain handled orchestration but lacked solid data handling. He launched LlamaIndex as the complementary data framework, a specialized tool for indexing and querying external documents using LLMs. LlamaIndex raised $8.5M seed. Jerry Liu and co-founder Simon Suo made Forbes 30 Under 30 in 2024. The LangChain-LlamaIndex combination became the de facto open-source LLM stack for a year before fragmentation and maturation produced alternatives.

Bret Taylor, The Map, the Document, the Agent

Bret Taylor was born 1980 in Oakland. He graduated Stanford CS in 2003, was hired by Marissa Mayer as a 23-year-old associate product manager, and was a primary architect of Google Maps, one of the most important consumer software products in history. In 2007 he co-founded FriendFeed (Facebook acquired for $50M in 2009). At Facebook as CTO (2010-2012) he scaled the world's largest social graph. In 2012 he co-founded Quip (Salesforce acquired for $750M in 2016). At Salesforce he became co-CEO alongside Marc Benioff in November 2021, then stepped down January 31, 2023.

Read full profile ↓

Founded Sierra (Post-corporate · 2023 · age 43). Bret Taylor left his role as Co-CEO of Salesforce in January 2023, convinced large language models would change the world. After a few months off, he co-founded Sierra with Clay Bavor in 2023 to build enterprise AI customer agents. Exit: Still running. Sierra raised $950 million in May 2026, valuing the company at over $15 billion. (Bret Taylor - Wikipedia · About Sierra)

In February 2023, Taylor co-founded Sierra with Clay Bavor. Sierra is a conversational AI platform for enterprise customer service agents, the bet that the first real AI consumer product will be "the company you can actually talk to." Sierra raised $1B Series A at $5.1B valuation (January 2024), $175M Series B at $4.5B (October 2024), $950M at $15.8B (May 2026). In November 2023, Taylor became OpenAI's board chairman after the reconstitution, making him simultaneously the chairman of OpenAI and the CEO of a company deploying OpenAI's models commercially. It was a structural tension the reconstituted board concluded it could live with.

Jeff Huber, From Google Maps to Cancer Detection to Vector Databases

Jeff Huber has built credibility in three distinct domains that rarely share the same career: consumer technology, life sciences, and AI infrastructure. At Google (2003-2016) he built Google Ads, Gmail, Calendar, Google Maps, and helped create Google X's life sciences division. He became Founding CEO of GRAIL (early cancer detection via blood tests and AI), pioneering liquid biopsy. He then co-founded Chroma, the open-source vector database that has become the most widely-used embedding store in the AI developer community. His Harvard MBA and UIUC engineering degree give him rare legitimacy across business and technical domains.

Read full profile ↓

Emil Eifrem, He Drew the Graph on a Napkin

Emil Eifrem is Swedish, studied at Linköping University, and was a lifelong programmer who sketched the property graph model on a flight to Mumbai in 2000. That sketch became the intellectual foundation of Neo4j, the company he co-founded in 2007 with Johan Svensson and Peter Neubauer. He also coined the term "graph database." As of June 2021, Neo4j raised $325M in Series F funding, led by Eurazeo and GV (Alphabet's venture capital arm), at a $2B+ valuation, a round the company called "the largest funding round in database history" (Neo4j). The company surpassed $100M ARR that same year.

Read full profile ↓

Founded Neo4j (Right after school · 2007). Emil Eifrem studied at Linköping University from January 2003 to January 2007. Neo4j was formally spun out in February 2007, shortly after he finished his studies. Before that, he was CTO at Windh Technologies (2002-2007) and Chief Architect at Windh AB (1999-2002). Exit: Neo4j is a private company that has raised over $581 million in funding and reached a valuation of over $2 billion in 2021. The company began preparing for an initial public offering (IPO) on the Nasdaq stock exchange in late 2024 and is considered IPO-ready as of 2026. (Neo4j - Wikipedia · How Much Did Neo4j Raise? Funding & Key Investors | TexA - TexAu)

Graph databases have re-emerged as mission-critical AI infrastructure: knowledge graphs ground LLMs in real enterprise semantics, enabling reasoning that pure vector similarity cannot provide. Emil Eifrem spent 25 years building toward a conviction he held before the field agreed: that graph thinking is foundational to AI, not supplementary to it.

Shyam Sankar, The Ontology Evangelist

Shyam Sankar was born in 1982 in Mumbai. His family moved to Nigeria when he was an infant; five armed men broke into their home when he was two years old and violently attacked his father, forcing the family to flee to Orlando, Florida. His father, born in a mud hut in Tamil Nadu, was the first in his family to attend college. His parents subsequently ran a souvenir shop and a dry-cleaning business that went bankrupt before stabilizing. He earned his BS in Electrical and Computer Engineering from Cornell, then his MS in Management Science and Engineering from Stanford in one year.

Read full profile ↓

Shyam Sankar joined Palantir as employee #13 in 2006, the company's first business hire. He pioneered the "Forward Deployed Engineer" model (FDE), where engineers embed directly with customers rather than building in isolation. Over 18+ years he has become CTO and EVP, articulating a thesis central to enterprise AI: frontier foundation models commoditize quickly, but domain-specific ontologies, knowledge graphs encoding enterprise semantics, processes, and relationships, create durable competitive advantages. The winning enterprise AI product is not the one with the best LLM. It is the one with the best ontology layered on top of commodity LLMs.

Leopold Aschenbrenner, The Prodigy Who Wrote the Pamphlet

Leopold Aschenbrenner graduated as Columbia's valedictorian in 2021 at age 19, with a double major in economics and mathematics-statistics. His senior thesis on catastrophic risk won the Romine Prize. He co-founded Columbia's EA chapter while an undergraduate. He joined FTX's Future Fund philanthropic arm in February 2022, a connection he severed before the collapse. He joined OpenAI's Superalignment team in 2023. In April 2024, OpenAI dismissed him; The Information reported the company characterized the dismissal as a security policy violation tied to a document he had shared externally, while Aschenbrenner has publicly disputed that characterization and attributed the dismissal to internal politics.

Read full profile ↓

What happened next is remarkable: he wrote "Situational Awareness", a 165-page essay arguing that AGI arrives within a few years and that the US is dangerously underprepared. The essay went viral in policy circles, establishing him as a major voice in AI governance before age 22. He then founded Situational Awareness LP, an AI-focused hedge fund backed by Patrick and John Collison (Stripe founders), Daniel Gross, and Nat Friedman. The fund started with $225M; within one year it managed $5.5B, with reported strong first-year returns per 13F filings (whalewisdom.com); the precise return figure varies across sources. The arc from fired safety researcher to billion-dollar fund manager took less than two years.

Yumi W. Kimura, The Org-Behavior Layer for Enterprise AI

Yumi Kimura is the founder of BehaviorGraph, a runtime organizational-behavior layer for enterprise AI. She is also the author of this book.

Read full profile ↓

BehaviorGraph Yumi K site @yumi_willems

Education & languages: Trilingual, native Japanese and Chinese, plus English. Undergraduate Law degree from Japan, Berkeley data-science bootcamp, and research in knowledge management, organizational behavior, and AI at Columbia.

Built Meitu Japan from scratch as founding country manager, making Japan Meitu's second-largest market and contributing to the company's $5B IPO on the Hong Kong Stock Exchange in 2016.

Her core argument: today's agents are not digital employees yet. They are still hardcoded workflows. They can retrieve documents, summarize tickets, and trigger tools, but they do not understand how companies actually work.

A human employee learns who really owns decisions, who is trusted, which escalation paths work, where authority sits, and when human judgment is needed. Agents do not learn this naturally. They only see the context and routing logic they are given.

BehaviorGraph fills that gap.

It maps the human layer beneath enterprise systems: trust, expertise, informal authority, ownership, influence, bottlenecks, and handoff patterns. Through API and MCP integrations, it gives AI agents runtime context for routing, escalation, and human checkpoints.

A Jira ticket can show assignment, but not real ownership. Slack can show responses, but not whether people trust them. A calendar can show meetings, but not who actually made the decision. An org chart can show hierarchy, but not how work really moves.

Before BehaviorGraph, Kimura founded LEAD, a people-side knowledge management and employee-connection platform used by enterprise customers. Her work combines organizational behavior, knowledge management, and organizational network analysis. She is also the author of the OIL framework (Organizational Intelligence Loop), a Columbia research-based methodology for designing organizationally aware AI systems. Columbia Academic Commons · SSRN

The bet: enterprise AI's next bottleneck is not model quality. It is organizational context.

Why AI Needs Organizational Behavior

A model can draft an answer in seconds. An enterprise still needs to know who can approve it, review it, own it, or move it forward.

BehaviorGraph sits in that decision gap.

It does not replace retrieval, graph databases, semantic governance, or execution tools. Those systems help AI find information, structure data, govern meaning, or call tools. BehaviorGraph focuses on the human operating layer: how decisions, trust, expertise, and authority actually move through people.

This matters for support escalation, incident routing, RevOps and finance approvals, change management, cross-functional handoffs, and copilots that need auditable human checkpoints.

This book's AI-generated map was built from public data: biographies, companies, schools, labs, investments, affiliations, and visible relationships. Even from public information alone, patterns appeared: shared institutions, hidden clusters, repeated career arcs, influence paths, and people who broke the expected pattern.

Inside a real organization, the signals are richer: collaboration behavior, trust indicators, decision paths, bottlenecks, informal expertise, team dynamics, escalation history, and cultural norms. Combined with ontology, knowledge graphs, organizational network analysis, and behavioral science, BehaviorGraph gives agents the context to understand not only what the company knows, but who understands it, who can act on it, who is trusted, what should not be done, and when human judgment is required.

Thomas Dohmke, From East Berlin to GitHub CEO

He was 11 years old, in East Berlin, when the Berlin Wall fell on November 9, 1989. He has called the moment the foundational experience of his life. The proof, as he saw it, that closed systems open, that walls fall, that information defeats restriction.

Read full profile ↓

Founded HockeyApp (Mid-career · 2011 · age 33). Thomas Dohmke completed his PhD in mechanical engineering in 2008 and worked at DaimlerChrysler (2002-2006) as a systems engineer and Robert Bosch GmbH (2007-2008) as a project manager on driver-assistance and parking systems before co-founding HockeyApp in 2011. Exit: HockeyApp was acquired by Microsoft in December 2014 for an undisclosed sum. Value at exit: Microsoft bought HockeyApp to add mobile app crash reporting, distribution, and beta testing across iOS, Android, and Windows Phone to Visual Studio Online's Application Insights service. The deal also strengthened Microsoft's cross-platform mobile development tools and added iOS and Android talent and experience. (Microsoft Acquires App Developer Platform HockeyApp - MacRumors · Microsoft Acquires German Mobile App Analytics Firm - Redmond Channel Partner)

He grew up in a society where software was copied freely, private property being a bourgeois concept, and learned to code on Eastern European computers reverse-engineered from Western machines. He studied at the Technical University of Berlin, built mobile software companies in Germany, sold them to HockeyApp, which Microsoft acquired in 2014. In 2021 he became CEO of GitHub.

Under him, GitHub launched Copilot, the AI pair programmer that became the fastest-growing product in the company's history. In May 2025 he left GitHub to found Entire, a new AI coding company. The arc from an apartment in East Berlin to the CEO of the company that hosts the world's open-source code to founder of an AI coding startup is one of the more instructive careers in the industry: each chapter required betting that a closed system would open, and each time it did.

Alexandr Wang, Scale AI

His parents are both physicists at Los Alamos National Laboratory who worked on nuclear weapons programs. Alexandr Wang was born in New Mexico in 1997 and grew up in a household where the morality of dual-use technology was a dinner-table topic. He dropped out of MIT at 19, worked at Quora briefly, and in 2016 founded Scale AI.

Read full profile ↓

Scale builds the data labeling and annotation infrastructure that most AI models require to be trained. It is the invisible backbone of the AI industry, the unglamorous middle layer that makes the glamorous top layer possible. He was the youngest self-made billionaire in 2021 when Scale reached $7.3B. As of May 2024, the company raised $1B at a $13.8B valuation, per Crunchbase. The US Army, Air Force, and multiple intelligence agencies are customers.

The biographical fact has not been lost on him. He has been more public than most AI CEOs about the national-security stakes of frontier AI, the dual-use logic, the China race. He is a living case study in how the questions a community asks at dinner shape the questions its children eventually put to the world, in his case: who controls the most powerful tools, and toward what end.

In 2025 Meta took a 49% stake in Scale AI at a $29B valuation for $14.3B, and Wang became Meta’s Chief AI Officer, heading Meta Superintelligence Labs alongside Nat Friedman and Daniel Gross. The deal converted a data-infrastructure founder into the operational lead of one of the most expensive AI talent acquisitions on record.

Parker Conrad, Rippling (The Comeback)

In 2016, at 36, Parker Conrad was forced to resign as CEO of Zenefits, the HR-tech company he had founded in 2013 and built to a $4.5B valuation. He had been selling insurance products without proper state licensing. The fines, the investigations, the legal bar from leading regulated financial services for a time, the public collapse of a unicorn. It was the kind of ending most careers do not recover from. As Business Insider reported.

Read full profile ↓

In 2017 he founded Rippling. The pitch was what he called a "compound startup," HR, IT, and finance operations sitting on a single integrated platform, with real data integration instead of API stitching. The industry was skeptical. The integration thesis sounded too ambitious. By April 2024 Rippling had raised at a $13.5B valuation in its Series F, led by Coatue (per siliconangle.com).

Architecturally Rippling is now regarded as the most ambitious HR tech platform built. What makes the Zenefits-to-Rippling arc worth studying is less about redemption than about persistence of conviction: Conrad held the same integration thesis through a very public collapse, refined the execution, and built the harder version of the company he originally imagined. Most people, given that exit, would have built something smaller and safer the second time. The move that did not look obvious was to build something more ambitious. Conrad's bet was that the integration thesis was right and the compliance shortcuts were wrong, and the right response was to fix the wrong part, not abandon the right part. That judgment, made under real duress, is the thing worth noting.

6.1 The Infrastructure Layer: Geography, Hardware & The Open-Source Rebellion

Every model in this book runs on silicon someone else built, in a building someone else owns, drinking power someone else allocated. The infrastructure layer is the part of AI that doesn't get podcast episodes, and the part where the next decade's bottleneck actually sits. This chapter is about three things: the geography (where AI gets made, and why a 30-mile radius around San Francisco still produces a statistically improbable share of it), the hardware (NVIDIA, AMD, TSMC, and the chips that price every other decision), and the rebellion (Mistral, DeepSeek, Hugging Face, and the open-source camp arguing the frontier shouldn't be owned).

The Rosewood Hotel

Sand Hill Road, Menlo Park. The Madera restaurant. This is where VCs drink, where deals get whispered, and where OpenAI's founding dinner happened in December 2015. Blazers and jeans, not suits. It looks like a casual gathering. What no one mentions is that the casualness IS the message: you only dress down when everyone in the room already knows your net worth. Islands

Cerebral Valley (Hayes Valley, SF)

AI's social epicenter since ChatGPT launched. Genesis House, a 21-room Victorian where AI founders live together. A dozen AI parties in two weeks became normal. The neighborhood that was known for boutique coffee shops and vintage clothing became the place where billion-dollar companies get founded over weekend hackathons. What the geography actually provides is not talent density but trust density: founders who can walk to each other's houses don't need a contract to start building together. SF Standard

The Power Circuit

AGI House, The Hillsborough Mansion

The Hillsborough mansion (koi pond, pool, Zen garden, wine cellar) is associated with Andrej Karpathy, who originally leased it in 2022 with Jeremy Nixon (ex-Google Brain) under the name NeoGenesis. The two co-founded the house as a residency and speaker series. The lease transferred to other operators in February 2023, and the mansion continues today as AGI House Hillsborough, run by Rocky Yu, hosting merit-based speaker events (past speakers include Jeff Dean and Sam Altman) and a small VC fund. Jeremy Nixon kept running events under AGI House SF, a separate hackathon and demo-day series in San Francisco. As of July 2025, AGI House LLC (Rocky Yu) filed a federal trademark-infringement complaint against Jeremy Nixon in the US District Court for the Northern District of California, with countersuit activity, as reported by Bloomberg Law and Forbes Australia. The trust-layer read on this fight: Karpathy, the more senior co-founder, stepped back to build Eureka Labs; he has not intervened to reclaim either the lease or the brand.

EA Berkeley, The Rationalist Underground

Lighthaven (a former hotel turned rationalist campus). MIRI workshops. Bay Area Secular Solstice (490 attendees from DeepMind and Anthropic). Taco Tuesdays. The group houses where Anthropic's founders lived before they were founders. This is the intellectual infrastructure of AI safety, not a university, not a company, but a network of communal living spaces and unconventional gatherings where the people most worried about AI's future figure out what to do about it. The hard part about that: it means the most consequential governance decisions in the field get made in places with no board, no charter, and no public accountability. See LessWrong's Lighthaven wiki page for the campus history, and Religion News Service for a 2025 inside look at the Secular Solstice gathering.

The Infrastructure Stack

The Supplier-Competitor Paradox

Anthropic supplies Claude to Cursor but competes with Claude Code. Google invested $2B in Anthropic while Gemini competes directly with Claude. Microsoft invested $13B in OpenAI while Copilot competes with ChatGPT. Every cloud provider simultaneously funds AND competes with their model lab tenant. The entire industry runs on a paradox: your biggest investor is also your biggest competitor, and your best customer is also building the product that will replace you. What is easy to miss is that this structure is not an accident or a temporary dysfunction; it is a deliberate hedge by every party involved, and each side has incentives to keep the arrangement alive even as the tension grows. The Information

The OpenAI Diaspora Pattern

Dario Amodei → Anthropic. Ilya Sutskever → SSI. Jan Leike → Anthropic. David Luan → Adept. Jason Wei → Meta. Leopold Aschenbrenner → $5.5B fund. Elon Musk → xAI. Mira Murati → Thinking Machines Lab. John Schulman → Anthropic → Thinking Machines. OpenAI is the company that every competing frontier lab traces its talent back to. Every lab that competes with OpenAI was founded by someone who used to work there, left in circumstances that often involved public or reported disagreements over direction, and started building exactly what they wished OpenAI had built instead. The Verge

The Agent Framework Convergence

In demo week, every framework looked interchangeable. In month six of production, they looked nothing alike.

LangChain, CrewAI, and AutoGen became orchestration brains. n8n became workflow plumbing for operations teams. Composio and Zapier became action connectors across SaaS tools. Each category solves a different failure mode, and each depends on systems it cannot fully control, model APIs, app uptime, identity policies, and audit logs. LangChain

No single vendor owns the full path from prompt to accountable action. That is not an interim phase. It is the structure of the market. The hard part for any operator building on this stack is not picking the right framework; it is deciding who you blame when the chain breaks at a layer you do not control.

Agent Orchestration, The Real Control Points

If you walk the stack top to bottom, power is easier to see than product demos suggest.

Read full profile ↓

Microsoft, Salesforce, ServiceNow, and Workday fight for policy authority inside core workflows. Databricks and Snowflake fight for governed retrieval and data control. LangChain, CrewAI, n8n, Composio, and Zapier fight for orchestration and execution reliability. Glean, Neo4j, Stardog, TigerGraph, and Alation fight for context quality and traceability.

The least solved control point is escalation intelligence. Not what a model can say, but what the organization should decide, who signs, who is informed, and who carries the consequence. That is where organizational-network products can shift from optional analytics to operational infrastructure.

The 12 People Who Bridge Everything

Every model in this book runs on silicon someone else built. The most important AI story of the decade is not who fine-tuned the best transformer, it is who got allocated H100s, who designed the next generation, and who fabricated them. Research on industrial bottlenecks consistently shows that the layer no one argues about in public is often the layer that matters most; in AI, that layer is the chip supply chain. This chapter covers the people who own it.

---

Lisa Su, The Comeback Engineer Who Made AMD a Buyer

Lisa Su was born November 7, 1969 in Tainan, Taiwan, and moved to New York as a child. She earned her BS, MS, and PhD in electrical engineering from MIT, where her dissertation work pioneered silicon-on-insulator transistors. After roles at Texas Instruments, IBM (where she co-led the Cell processor team that powered the PlayStation 3), and Freescale, she joined AMD in 2012 and became CEO in 2014, when the company was widely considered weeks from insolvency (Fast Company). Under her leadership AMD's stock rose more than 50x and the company acquired Xilinx (completed February 2022, approximately $49B per AMD IR) and Pensando, then in 2024 began shipping Instinct MI300X GPUs to Microsoft, Meta, and Oracle as the first credible alternative to NVIDIA's H100.

Read full profile ↓

She is Jensen Huang's first cousin once removed (full family-tree detail in Chapter 1). She received the Semiconductor Industry Association's Robert Noyce Award in 2021 and was named TIME's CEO of the Year in 2024.

Sources: Wikipedia · AMD leadership

Morris Chang, The Father of the Foundry Model

Morris Chang was born July 10, 1931 in Ningbo, China, fled the civil war as a teenager, earned mechanical engineering degrees from MIT (BS, MS) and an electrical engineering PhD from Stanford, and spent 25 years at Texas Instruments rising to Group VP. In 1985 the Taiwanese government invited him to lead what would become Taiwan Semiconductor Manufacturing Company. He invented the "pure-play foundry" model: TSMC would make no chips of its own, only fabricate chips designed by others. That decision is the reason NVIDIA, AMD, Apple, Qualcomm, and every modern fabless chip company can exist. He retired in 2018 and is treated as a national figure in Taiwan; his biography is required reading at TSMC.

Read full profile ↓

Sources: Wikipedia · TSMC company profile · Computer History Museum Fellow

C.C. Wei, The Engineer Holding the Single Point of Failure

Che-Chia "C.C." Wei was born 1953, joined TSMC in 1998 from Singapore-based Chartered Semiconductor, and became CEO in 2018 and chairman in 2024 succeeding Mark Liu. TSMC fabricates roughly 70% of the global foundry market and an even higher share of leading-edge AI chips (Focus Taiwan), every H100, every Blackwell, every M-series Apple silicon, every Strix Halo AMD chip. The geopolitical risk concentrated in one company on one island is the most discussed dependency in the AI infrastructure stack. C.C. Wei has overseen the $40B Arizona expansion (Phoenix Fab 21) and the planned Kumamoto and Dresden fabs as diversification responses, though the most advanced nodes remain in Hsinchu and Tainan.

Read full profile ↓

Sources: Wikipedia · TSMC executive bio

Andrew Feldman, The Wafer-Scale Bet

Andrew Feldman co-founded Cerebras Systems in 2015 with Gary Lauterbach, Sean Lie, JP Fricker, and Michael James, fresh off the SeaMicro acquisition by AMD in 2012. Cerebras built the Wafer-Scale Engine, the largest single chip ever manufactured, 850,000 cores etched onto an entire 12-inch silicon wafer rather than diced into hundreds of small dies. Cerebras IPO'd on NASDAQ as CBRS in May 2026 at a projected valuation of $35B-$49B (Markets Insider), then partnered with G42 (UAE) on the Condor Galaxy AI supercomputer network. Feldman's pitch is that wafer-scale inference is faster and cheaper than GPU clusters for any workload that fits, and that NVIDIA's networking bottleneck is permanent.

Read full profile ↓

Founded Cerebras Systems (Post-corporate · 2015). Andrew Feldman co-founded SeaMicro in 2007, which was acquired by AMD in 2012. He then served as Corporate Vice President and General Manager at AMD from 2012 to 2014 before co-founding Cerebras Systems in 2015. Exit: Cerebras Systems is preparing for an IPO on Nasdaq under the ticker CBRS, with pricing expected around May 13, 2026, and a projected valuation of approximately $35 billion to $48.8 billion. (Cerebras - Wikipedia · Andrew Feldman On Generating Over $1B From Selling His Startups And Now Raising $200M To Accelerate Deep Learning - Alejandro Cremades)

Sources: Wikipedia · Crunchbase

Jonathan Ross, The Original TPU Engineer Who Built Groq

Jonathan Ross is the engineer who, while an intern at Google in 2013, designed the first prototype of what became the Tensor Processing Unit (TPU), the chip family that gave Google in-house silicon for neural network training. He left Google in 2016 and founded Groq in 2017 with the thesis that inference, not training, is where the long-term silicon margin sits. Groq's Language Processing Unit (LPU) runs Llama, Mixtral, and other open models at output speeds that competitors cannot match on identical models. Groq raised $640M Series D in August 2024 at a $2.8B valuation led by BlackRock with participation from Cisco, Samsung Catalyst, and Type One.

Read full profile ↓

Founded Groq (Mid-career · 2016). Jonathan Ross left Google in 2016 after nearly five years, where he designed and implemented the core elements of the original Tensor Processing Unit (TPU) as a 20% project. Exit: Groq has raised significant funding, including a $640 million Series D round in August 2024 at a $2.8 billion valuation, and a $750 million Series E round in September 2025 at a $6.9 billion valuation. The company also entered a licensing agreement with Nvidia in December 2025 (CNBC), but said it would continue to operate independently.

Sources: Wikipedia · Crunchbase

Hock Tan, The Quiet Broadcom Strategy Behind Anthropic Silicon

Hock Tan was born 1952 in Penang, Malaysia, earned mechanical engineering degrees at MIT and an MBA at Harvard, and built Avago Technologies into the consolidator that bought Broadcom in 2016 and kept the latter's name. Under Tan, Broadcom became the second-largest semiconductor company by market cap, partly on Apple, Cisco, and Google networking sockets, partly on the custom-silicon "XPU" partnerships disclosed publicly with Google (TPU manufacturing), Meta (MTIA), ByteDance, and most recently Anthropic. The Anthropic deal, reported in November 2025, places Broadcom at the center of the second-source AI training chip supply that hyperscalers and frontier labs are quietly assembling to reduce NVIDIA dependence.

Read full profile ↓

Founded Pacven Investment (Mid-career · 1988 · age 37). After earning his MBA in 1979, Hock Tan held finance roles at General Motors and PepsiCo, and served as a managing director at Hume Industries from 1983, before co-founding Pacven Investment in 1988. (Hock Tan - Wikipedia · Who is Hock Tan, Broadcom CEO and new Meta Board Member? | Technology Magazine)

Sources: Wikipedia · Broadcom leadership · The Information (Nov 2025)

The conventional story is that openness is simply the generous choice. The harder version is that it is also a strategic one: when you cannot out-resource the largest labs on the planet, releasing your weights turns every developer in the world into a distribution channel. A parallel set of labs, fine-tuners, agents, and inference providers grew up around that insight. This chapter covers the people who built it, and what it actually cost them to give the work away.

Arthur Mensch, The Polite French Counterweight

Arthur Mensch is a graduate of École Normale Supérieure and INRIA, did his PhD on optimization with applications to brain imaging, and joined DeepMind in 2020 where he was a lead author on the Chinchilla scaling laws paper. He co-founded Mistral AI in April 2023 with Guillaume Lample and Timothée Lacroix, both Meta AI researchers from the LLaMA team. Mistral 7B, released in September 2023 under Apache 2.0, was the first open-weight model to outperform Llama 2 13B on most benchmarks and became the foundation for hundreds of fine-tunes. As of 2024, Mistral had raised over $1B at multi-billion-dollar valuations and shipped Mistral Large, Codestral, and the Mixtral mixture-of-experts series, while remaining the most prominent European frontier lab.

Read full profile ↓

Founded Mistral AI (Mid-career · 2023 · age 31). Arthur Mensch co-founded Mistral AI in April 2023 after working for nearly three years as a researcher at Google DeepMind, where he focused on large language models and multimodal systems. Prior to DeepMind, he completed a PhD and postdoctoral research in machine learning. Exit: Mistral AI is still an active, privately held company. As of September 2025, it was valued at over $14 billion after securing significant funding rounds. The CEO, Arthur Mensch, has stated that the company is not for sale. (Mistral AI - European AI Innovation · Arthur Mensch - Wikipedia)

Sources: Wikipedia · Mistral AI · Mistral about

Clément Delangue, The Hub That Hosts the Open Camp

Clément Delangue co-founded Hugging Face in 2016 with Julien Chaumond and Thomas Wolf, originally as a chatbot for teens. The company pivoted to an open-source ML library in 2018 with the release of `transformers`, the Python package that became the default interface for using and fine-tuning every major language model. Hugging Face today is the GitHub of machine learning: 1M+ models hosted, 10M+ datasets, and the central distribution channel for every open-weight release from Meta, Mistral, DeepSeek, Alibaba Qwen, Nous Research, and Stability AI. The company raised $235M Series D in 2023 at a $4.5B valuation with backing from Google, NVIDIA, Salesforce, and Sequoia. Delangue's public message is consistent: AI must be open or it will calcify into a regulated oligopoly.

Read full profile ↓

Founded Hugging Face (A few years out · 2016 · age 27). Clément Delangue graduated with a Master in Management degree from ESCP Business School in 2012. Before co-founding Hugging Face in 2016, he held product roles at machine learning startup Moodstocks (2011-2012) and social listening startup Mention (2013-2014), and also co-founded VideoNot.es (2010-2011). Exit: Hugging Face is still running and has not had an exit. The company reached a valuation of $4.5 billion in 2023 after raising $235 million in Series D funding. (Hugging Face revenue, valuation & funding | Sacra · Hugging Face Hits $4B Valuation After Salesforce Ventures-Led Round - Crunchbase News)

Sources: Wikipedia · Hugging Face · Hugging Face profile · Crunchbase

The most consequential open-weight frontier release of this period came from outside the US lab pipeline: DeepSeek, profiled in Chapter 3 alongside the other frontier labs.

"teknium", The Most-Downloaded Fine-Tuner on the Internet

"teknium" is the pseudonymous co-founder and head of post-training at Nous Research, one of the most-followed independent AI labs on X. He spent 2023 fine-tuning open-weight models on consumer GPUs and publishing them on Hugging Face under the OpenHermes brand; his models cumulatively crossed 70M downloads. Nous Research formalized around teknium, Karan Pratap Singh, and Jeffrey Quesnelle (emozilla) in 2023, raised a $5.2M seed in mid-2024 led by Distributed Global, then a $50M Series A in mid-2025 valuing the lab around $1B. Nous ships the Hermes model series and Hermes Agent, a runtime for tool-using, locally-runnable agents that is one of the most popular open alternatives to closed agent platforms.

Sources: Nous Research · Hugging Face profile

Jeffrey Quesnelle ("emozilla"), The Architect Behind Hermes Agent

Jeffrey Quesnelle, who posts as "emozilla", is Nous Research's co-founder and CTO. He is best known for the YaRN context-extension method (NeurIPS 2024), which is now standard in long-context open models, and as the technical architect behind Hermes Agent, the open-source agent runtime that lets users run tool-using AI agents on local hardware against Nous Hermes or other open models. In a publicly visible AMA in April 2026 he framed Hermes Agent as "the local-first alternative to ChatGPT's agent mode," and that framing has put Nous in direct conceptual conflict with the closed-agent platforms covered earlier in this book.

Sources: Nous Research · YaRN paper

The LLaMA Team at Meta

Meta's LLaMA series, shipped under the GenAI organization led by Yann LeCun's research direction and the engineering leadership of Ahmad Al-Dahle (and previously Joelle Pineau before her move to Cohere covered earlier in this book), is the single most-downloaded family of open-weight models, with LLaMA 2, LLaMA 3, and LLaMA 4 collectively used as the base for thousands of derivatives. Mark Zuckerberg has publicly committed Meta to open-weight releases through at least the LLaMA 5 generation, citing the same reasoning Linux founders cited in the 1990s: when the underlying technology is too important to leave in the hands of one vendor, the answer is to release it.

Sources: Meta LLaMA

Chapter 7 · Conflicts & Power Struggles

In AI, the public fights are where power gets renegotiated. This chapter follows the explicit conflicts: lawsuits, board crises, public resignations, market battles, and partnership disputes. The pattern-synthesis chapter that follows steps back from those disputes to ask what the network as a whole keeps producing.

Conflict Timelines: Shared Ties Before the Public Fights

Public fights in AI tend to look like sudden ruptures in the press. The map says otherwise. Most of the major conflicts in this chapter sit on top of long shared histories: co-founders, mentor-student pairs, board allies, joint authors. The chart below puts the strongest of those conflicts on a single timeline. Green marks prior shared work; orange marks the documented divergence point; red marks the moment the dispute went public.

The pattern is consistent across the five rows. Musk and Altman were OpenAI co-founders for three years before any public disagreement. LeCun and Hinton shared three decades of work, two industry roles, and a Turing Award before the 2023 split. The November 2023 board crisis was four days from firing to reinstatement, but the underlying tensions trace back to the for-profit restructuring four years earlier. The Anthropic walkout was preceded by four years of Amodei working inside OpenAI. The Andreessen exchange runs through a16z portfolio companies he had been backing for nearly a decade. Public fights in this dataset are not the absence of prior relationship. They are the consequence of one, played out in front of an audience.

Elon Musk vs. OpenAI (Lawsuit, 2024-2025)

In February 2024, Elon Musk filed suit against OpenAI and Sam Altman in California state court, alleging OpenAI had breached its founding nonprofit mission by commercializing through Microsoft. Reuters reported on the original filing. He voluntarily dismissed the lawsuit in June 2024, one day before a scheduled hearing on OpenAI's motion to dismiss, without explanation. In August 2024, his legal team publicly stated an intent to refile in federal court with expanded allegations, as CNBC reported. The status of any federal case as of this writing should be verified against current court records before any specific claim about it is acted on. Active litigation as of 2026-05-16 Legal observers have noted that the lawsuit, whatever its personal dimensions, forced a public reckoning with a genuinely unsettled question: what does a nonprofit AI mission actually require when commercial partnerships enter the picture? (For Elon Musk's full profile, SpaceX, Tesla, xAI, see Chapter 1.)

Yann LeCun vs. Geoffrey Hinton, The Public Split Between Two Long-Time Collaborators

Geoffrey Hinton was Yann LeCun's mentor and champion, hired him at Bell Labs in the 1980s, recommended him for his first academic position. For thirty years they were allied members of the same intellectual project. When Geoffrey Hinton resigned from Google in May 2023 (MIT Technology Review) and began warning about AI existential risk, Yann LeCun disagreed publicly, arguing that his longtime mentor was not applying sufficiently rigorous scientific standards to claims about risk. The exchange marked the moment a thirty-year intellectual alliance became a thirty-year debate that the rest of the field now navigates.

The OpenAI Board Crisis (November 2023)

Documented in full in Part II. Post-crisis feuds of note:

• Helen Toner vs. Sam Altman: In May 2024, Helen Toner gave a detailed account on the TED AI podcast directly contradicting several of Sam's public statements about the reasons for his firing. Sam's allies pushed back publicly. The account of "why Sam was fired" remains genuinely contested between parties who were present.

• Ilya Sutskever's limited public explanation: Ilya has not given a detailed public account of his reasoning during the events. His subsequent SSI founding was read by some publicly quoted observers as consistent with previously reported concerns about OpenAI's direction (The Globe and Mail).

Emad Mostaque, Stability AI's Turbulent Period

Emad Mostaque's decision to release Stable Diffusion as open source in August 2022 was one of the most significant democratizations of generative AI to date, enabling millions of developers, artists, and researchers to build on a frontier image-generation model without gatekeeper API access. He founded Stability AI in 2020 and raised $101M at a $1B valuation in October 2022. He resigned as CEO in March 2024 amid Forbes reporting on operational dysfunction and concerns about financial management at the company; multiple senior executives departed and Getty Images filed a copyright suit over watermarked training images (The Guardian). Active litigation as of 2026-05-16 What makes this a power struggle rather than a private misfortune is that the founder's departure, the senior team's exits, and the investor pressure that drove both were governance mechanics, not personal ones. He has since founded Intelligent Internet, a decentralized-AI effort, and continues to speak publicly. His record is contested: the open-source release reshaped a generation of tooling; the company's operations during his tenure drew sustained criticism. Wikipedia

Read full profile ↓

Founded Stability AI (Post-corporate · 2019 · age 36). Before co-founding Stability AI, Emad Mostaque spent over a decade as a hedge fund manager, specializing in crude oil trading and advising governments. He has publicly cited family motivations for moving into AI. Exit: Emad Mostaque resigned as CEO and from the board of Stability AI on March 23, 2024, to pursue decentralized AI. The company has not had an exit and continues to operate under new leadership while facing financial and market challenges. (Stability AI - Wikipedia · Emad Mostaque - Wikipedia)

Marc Andreessen's Techno-Optimist Manifesto vs. the AI Safety Community

Marc Andreessen's October 2023 "Techno-Optimist Manifesto" (5,000 words, published on a16z's website) called effective altruism "our enemy," described AI doomers as adversaries, and positioned accelerationism as a moral imperative. It provoked detailed rebuttals from AI safety researchers and direct criticism from EA leaders including William MacAskill (Oxford philosopher whose professional community overlaps substantially with Anthropic's Amanda Askell), opening a sustained public argument about whether accelerationism and safety-mindedness are genuinely incompatible worldviews or competing emphases within a shared goal. The most-read and most-argued-over document produced by a venture capitalist in the last decade. (Marc Andreessen co-created Mosaic, co-founded Netscape, and built a16z into a $42B+ fund, fuller profile in Chapter 7.)

LeCun, Hinton, and Bengio on X: An Ongoing Public Debate

The Yann LeCun-Hinton-Bengio public argument has become a recurring genre of AI Twitter drama that plays out multiple times per year. All three are active on X. Their disagreement is genuine, substantive, and consequential, it shapes how governments, journalists, and the public understand AI risk. Junior researchers report feeling pressured to pick sides. The three people who co-invented deep learning cannot agree on whether the technology poses existential risk. That the founders of deep learning disagree so openly about its consequences is either the most important scientific disagreement of our time or a demonstration that even earned authority does not resolve genuinely hard empirical questions, and either way the rest of the field has to do the work of thinking it through. If you are running a company that deploys these systems, the debate does not pause while the godfathers sort it out. You make a call, you build the guardrails you believe in, and you update when evidence arrives. That is not a lesser version of the intellectual work. It is a different kind.

The Model Race: ChatGPT, Claude, Gemini, and Grok

ChatGPT vs. Claude vs. Gemini vs. Grok is the consumer AI contest that will shape how billions of people interact with intelligence. As of December 2024, ChatGPT had over 300 million monthly active users and first-mover advantage. Claude is gaining share in enterprise and developer markets through superior reasoning and safety characteristics. Gemini has Google's distribution, embedded in Search, Android, and Workspace. Grok has Elon Musk's X platform and a willingness to say things the other models won't. The consumer AI market is not winner-take-all; it is becoming a competition in which each model serves a different kind of user and earns trust in a different way. The surprising implication is that having multiple strong models may produce better results for users than a single dominant model would, because each competitor is forced to improve a different part of what intelligence means in practice.

Perplexity vs. Google Search, the most asymmetric competition in technology. Perplexity processes approximately 100 million queries per month. Google processes 8.5 billion per day. Yet Perplexity's answer-engine model, which synthesizes sources into direct responses rather than returning links, represents an existential threat to Google's ad-supported search model. The New York Times, BBC, and Dow Jones have all sued Perplexity for scraping their content to generate answers without driving traffic back to publishers, as Forbes reported. The lawsuits are the clearest indicator that Perplexity has built something publishers consider genuinely threatening to their survival.

The pattern is that AI conflict moves in layers. The model wars shape who controls capability, and the enterprise agent wars show how that contest reaches buyers and distribution.

The Enterprise Agent Race: Accountability, Budgets, and Risk

At 7:12 a.m., in a glass conference room, a COO asks the question now shaping software budgets: if an agent approves the wrong contract, who owns the mistake?

That is the enterprise war in one line. Marc Benioff says Agentforce should sit where revenue teams already work. Bill McDermott says ServiceNow should sit where operational handoffs already break. Satya Nadella says Copilot should ride inside Microsoft surfaces employees never leave. Aneel Bhusri says Workday holds the HR and finance records that agents cannot hallucinate around.

Then specialists exploit the gaps. Bret Taylor and Clay Bavor built Sierra around customer-experience agents and scaled it fast. Arvind Jain built Glean around enterprise retrieval quality. Aaron Levie repositioned Box as a governed AI content system. Winston Weinberg focused Harvey on legal workflows where citation mistakes become liability.

So the market keeps splitting, then recombining. Suites win distribution, specialists win accuracy, buyers run both until one stack can prove lower risk at higher speed.

Mutual Fit, Overlap, and Replacement Risk

Most market maps look scientific and say very little. They rank benchmark scores and avoid the buying process where competition is decided.

Read full profile ↓

Use a harder filter. Same buyer, same quarter, same workflow slot.

By that filter, Microsoft Copilot, ChatGPT Enterprise, and Gemini for Workspace collide directly in daily knowledge work. Agentforce and ServiceNow collide in ticket-to-approval workflows. Databricks and Snowflake collide in governed retrieval for enterprise agents. LangChain, CrewAI, n8n, Composio, and Zapier collide on orchestration reliability, authentication, retries, and action logging.

The surprising part is that overlap and fit can both be true. A company buys LangChain and still buys Copilot. A company buys ServiceNow and still buys Agentforce. Replacement usually comes later, when one layer quietly absorbs the neighboring layer.

Useful filter: same buyer, same workflow slot, integration depth, switching cost, and policy friction. High scores across all five means you are looking at a real head-to-head fight.

The Code Editor Race: Cursor, Windsurf, and Google Distribution

Cursor (~$50B) vs. Windsurf (Google talent licensing + Cognition acquisition) is the clearest head-to-head contest in AI code editors. Cursor, built by four MIT students with no marketing budget, has become the default AI-native code editor for a generation of developers. The Windsurf deal resolved in a three-way split: OpenAI's LOI collapsed, Google licensed the core team (~$2.4B), and Cognition acquired the remaining IP and staff (see the full deal breakdown in Chapter 5.5). Google is integrating the Windsurf team into its developer platform. The competition is asymmetric: Cursor has developer love and organic adoption; Google has distribution and integration with Cloud, Android Studio, and the broader Google developer platform.

Anthropic ships Claude Code while Claude powers Cursor, the supplier-competitor paradox. Anthropic is simultaneously the model provider that makes Cursor possible and a competitor building its own AI coding tool (Claude Code). This is the classic platform tension: AWS competing with companies that run on AWS, Google competing with companies that advertise on Google. The question is whether Anthropic will use its position as model provider to advantage its own coding tools, or whether the market is large enough that both can win.

The Video Generation Race: Pika, Runway, and Sora

Pika vs. Runway vs. Sora, three companies competing for Hollywood budgets and creator wallets. Pika (founded by former Stanford AI PhD students Demi Guo and Chenlin Meng) had raised a total of $135M as of June 2024 (per sacra.com). Runway (founded by Cristobal Valenzuela) pioneered video generation with Gen-2 and Gen-3. OpenAI's Sora entered the market with the ChatGPT brand behind it. The competition is not just technical; it is a question of whether AI-generated video becomes a creator tool, a Hollywood pipeline, or a consumer product. Each path requires a different sales strategy.

Llion Jones, co-author of "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture, has said publicly that he is "absolutely sick of transformers" and is building something new at Sakana AI ($2.65B valuation). Jones is the only Transformer co-author who has explicitly committed to replacing the architecture he helped create. If Sakana's research produces a post-Transformer framework, it would be the most consequential act of creative destruction by an inventor since, it is difficult to find a precedent. There is something generative about this: the people most qualified to replace a framework are usually the ones who built it and can see exactly where it breaks.

Helen Toner, The Paper That Became a Focal Point

Helen Toner is a researcher at Georgetown's Center for Security and Emerging Technology (CSET), where her work on AI policy and US-China technology competition has informed U.S. government strategy. She served on OpenAI's board from 2021 to 2023. Wikipedia

Read full profile ↓

Co-authored an October 2023 academic paper that included a comparison some observers (and reportedly some inside OpenAI) read as a criticism of OpenAI's safety practices while she sat on its board. What made this hard was not that a board member had a dissenting view. Board members are supposed to have dissenting views. The hard part was that the original board compact, a small group of researchers governing a nonprofit with a safety mission, had been stress-tested by a company that grew faster than any governance structure built for it. The board action in November 2023 was one of the most consequential corporate governance episodes in AI to date. Toner has spoken publicly about her perspective since. The Verge on Toner's account

Chapter 8 · Hidden Ties & Pattern Synthesis

In high-uncertainty markets, personal trust becomes operating infrastructure. The patterns in this chapter are not gossip; they are the connective tissue between the labs, funds, and boards covered in the earlier chapters, and they help explain decisions that the formal org charts cannot.

The AI power map is held together less by org charts than by repeat ties: shared formations, marriages and friendships across competing firms, one philanthropist who funded both labs, and recurring patterns of school, immigration, and project shape. Those ties help explain why the conflicts of the previous chapter keep rearranging the same cast rather than replacing it.

The Eight Canonical Motifs of Power Transfer

Career paths in AI are not random. Eight repeating motifs account for the vast majority of power transfers in this dataset, and they show up again and again across the patterns documented in the rest of this chapter:

The “safety exit” trigger is unique to OpenAI. No other company in the dataset has produced a comparable exodus pattern. When researchers leave OpenAI, they tend to found safety-oriented competitors. When researchers leave Google, they tend to found commercial labs. When researchers leave Meta, they tend to go to Europe.

Different relationship types elevate different people

When you split the network by relationship type, the key connectors change. That appears to matter because trust, access, and talent transfer travel through different kinds of relationships, not through one generic web of ties.

Cofounder Clusters: Who Has Built With Whom

One cofounder circle in this cast stands out for spanning multiple companies and multiple decades. Beyond that, most cofounder ties stay inside the company they were formed for. There are 82 documented cofounding ties between 78 people in this dataset, grouped into 24 distinct circles. The single cross-company circle, plus the two largest single-company founding teams, are shown below.

The Family Map: Personal Ties at the Top

8 of 29 tier-1 people in this dataset have a documented family tie to someone else in the cast (within-cast share 28%, n=29, 95% CI ~14-48%). The AI industry's most powerful institutions are, in several cases, literally run by families. Read as a composition pattern in this curated cast, not as a base rate for the broader industry.

The VMware Heritage: Two Competing AI VCs With Strikingly Similar Frameworks

Martin Casado (a16z General Partner) and Jerry Chen (Greylock Partner) are among the most prominent enterprise AI investors in Silicon Valley. They publish competing frameworks for the same thesis: Martin Casado calls it "AI Moats"; Jerry Chen calls it "Systems of Intelligence." Both argue that frontier models commoditize quickly and the durable enterprise advantage sits at the proprietary data, workflow integration, and domain ontology layers above the model. The two frameworks propose a strikingly similar architecture for durable enterprise value. One plausible explanation, consistent with public VMware employment records and AllThingsD's reporting on the Nicira acquisition, is that both authors spent formative years in the same VMware product organization.

Martin Casado joined VMware as a Fellow and CTO of Networking and Security in 2012, after VMware acquired his company Nicira for $1.26B (AllThingsD). Jerry Chen was VMware's VP of Product Management and Marketing for cloud and application services for the same overlapping years. VMware in this era, building Cloud Foundry, launching vCloud, navigating the shift from virtualization to cloud, was one of the most intellectually serious product organizations in enterprise software. It published whitepapers. It ran technical conferences. It took frameworks seriously as competitive tools. Both Martin Casado and Jerry Chen absorbed that habit of treating frameworks as competitive tools and now deploy it in venture capital. The competitor-VCs-with-identical-theses story is actually a story about a shared intellectual formation at a single company a decade earlier.

The Sequoia-Greylock Marriage: Pat Grady × Sarah Guo

The Greylock connection is layered further: Sarah Guo's former Greylock colleagues include Jerry Chen (who co-invested with Jack Altman in LlamaIndex's seed round) and Reid Hoffman (LinkedIn co-founder, OpenAI board founding member). The Pat Grady-Guo marriage places two investors within overlapping professional networks that also include Jerry Chen and Reid Hoffman: Pat Grady (Sequoia), Sarah Guo (Conviction), Jerry Chen (Greylock), Reid Hoffman (Greylock / OpenAI founding board). In network terms, the marriage links otherwise separate professional circles in a way that may matter for access and visibility.

The Hackathon That Accidentally Invented Both LLM Frameworks

As Harrison Chase has described publicly (LangChain blog), in October 2022, Robust Intelligence, an AI security startup founded by Harvard professor Yaron Singer, held an internal hackathon. Two of its machine learning engineers, Harrison Chase and Jerry Liu, participated. Harrison Chase built what he called LangChain. Jerry Liu built what he called GPT Index (later renamed LlamaIndex). Both released their projects open-source almost immediately afterward. Within months, LangChain had millions of downloads and LlamaIndex had hundreds of thousands. Together, they became the de facto LLM application stack for the entire developer community, influencing how hundreds of millions of dollars in enterprise AI was built.

Neither man was founding a company. Neither had VC backing. Neither had spoken to investors beforehand. The two most important LLM application frameworks in the world were created by the same two-person team, at the same employer, during the same weekend hackathon, independently, without coordination, and both went on to raise tens of millions in venture capital. Jerry Liu's seed round was led by Jerry Chen at Greylock and included Jack Altman (Sam's brother, Lattice founder) as an angel investor (Greylock). Harrison Chase's seed was led by Benchmark (LangChain). Both companies are now unicorns.

Sierra's 18-Year Friendship Origin

As reported by Forbes, soon after Bret Taylor's resignation from Salesforce became effective in late January 2023, Clay Bavor called him. They arranged to have lunch at a Mediterranean restaurant in Palo Alto. They talked for hours over herbal tea. By the end of the meal, they had decided to start a company together. That company became Sierra (Forbes).

What makes this origin story striking is the eighteen-year gap it spans. Taylor had hired Bavor at Google in approximately 2005, when Taylor was working on Google Maps. Bavor stayed at Google for almost two decades, through Gmail, Google Drive, Google Daydream VR, while Taylor went to FriendFeed, Facebook, Quip, and Salesforce. They had remained friends across that entire span. When Taylor walked out of his second co-CEO role in the Valley, Bavor was the person who called within hours. Sierra raised $1B in its first institutional round, $175M in its second, and $950M in its third at a $15.8B valuation (TechCrunch). A friendship maintained across eighteen years and multiple employers turned a single lunch into one of the most valuable AI startups on earth. There is no standard playbook for that path. The thing that made Sierra possible was not the credentials, both men had those in abundance. The pattern is consistent with a trust advantage: prior history reportedly let the founders move quickly, and Sierra closed its $1B first institutional round within months. Whether that speed was attributable to the prior trust is a plausible hypothesis, not a documented mechanism.

Open Philanthropy's Lab Funding Thread

Open Philanthropy, with Holden Karnofsky in a senior decision-making role, was an early funder of both OpenAI (before the commercial pivot) and Anthropic (founded 2021). The same philanthropic fund directed money to the two most important frontier AI labs in the world, which are now intense commercial competitors. Holden Karnofsky subsequently married Daniela Amodei, President of Anthropic. As of January 2025, he joined Anthropic's staff (Wikipedia).

The sequence is notable: Open Philanthropy supported OpenAI and later Anthropic while Karnofsky held a senior role there; he later married Anthropic President Daniela Amodei and then joined Anthropic's staff. A person who helped lead philanthropic grantmaking in AI safety later became an employee of one recipient organization. None of the individual decisions in that sequence are scandalous on their own. The sequence raises governance and disclosure questions that readers may reasonably note, and it developed over time through a series of separate, publicly known steps.

Palantir's Employee #13 and the PayPal Thread

As Sankar has described in public interviews, Shyam Sankar got into Palantir via Kevin Hartz. As a student, Shyam Sankar cold-emailed Hartz (co-founder of Xoom, later EventBrite), who introduced him to the PayPal Mafia network: Roelof Botha (now Sequoia managing partner), Keith Rabois, and, through that network, Peter Thiel. Shyam Sankar joined Palantir in March 2006 as employee #13, the company's first business hire. The chain: student → cold email → Xoom co-founder → PayPal Mafia → Peter Thiel's AI company → 18 years later, Palantir CTO and one of the company's billionaires through stock appreciation. As of June 2025, Shyam Sankar was granted a direct commission as Lieutenant Colonel in the U.S. Army Reserve's Executive Innovation Corps (Wikipedia). The PayPal thread runs from Peter Thiel's network directly to the Pentagon's defense AI strategy.

Benioff's Critique of Microsoft's OpenAI Dependency

At Davos in January 2025, Salesforce CEO Marc Benioff said publicly that Microsoft's relationship with OpenAI "cracked when it hired Mustafa Suleyman," as reported by the Economic Times. His framing: that Satya Nadella bringing in a CEO-level AI leader (Mustafa Suleyman) inside Microsoft while continuing to depend on OpenAI for AI capabilities creates an apparent tension within Microsoft's AI strategy. Marc Benioff contrasted this with Salesforce, which he described as operating from a single CEO structure and a single AI product (Agentforce).

One context the Davos remarks did not address: in October 2025, Salesforce announced that Agentforce 360 integrates directly with ChatGPT (Salesforce), an integration that creates its own meaningful Salesforce-to-OpenAI dependency. Readers can weigh the Davos critique alongside that subsequent partnership announcement. The "cracked relationship" observation about Microsoft is consistent with separately reported tensions; the comparative independence framing for Salesforce is harder to square with the October 2025 partnership.

Bret Taylor and Mark Zuckerberg: A Publicly Stated Friendship

As reported by Forbes, when Taylor left Facebook as CTO in 2012 to found Quip, he called Zuckerberg "my mentor and one of my closest friends" (Forbes). Taylor has subsequently been: CTO at Facebook (2010-2012), co-CEO at Salesforce (2021-2023), chairman of OpenAI's board (2023-present), and CEO of Sierra (2023-present). He has served on OpenAI's board during a period when OpenAI and Meta AI are in intense competition. He runs Sierra, which competes with Meta's enterprise AI products. The author has not identified public remarks by Taylor criticizing Zuckerberg or Meta. Zuckerberg has not publicly addressed the competition framing. The publicly described friendship has persisted alongside roles at organizations that now compete in AI.

Leopold Aschenbrenner: From Columbia Valedictorian to OpenAI Departure to AI Macro Investor

Leopold Aschenbrenner graduated as Columbia's valedictorian in May 2021 at age 19. By June 2022 he was a grantmaker at the FTX Future Fund, Sam Bankman-Fried's EA-aligned philanthropic arm. He resigned shortly before FTX collapsed in November 2022. He joined OpenAI's Superalignment team in 2023. OpenAI dismissed him in April 2024, characterizing the dismissal as a security policy violation tied to a document he had shared externally (The Information); Aschenbrenner has publicly disputed the characterization, attributing the dismissal to internal politics around a security memo he wrote about Chinese intelligence threats. He published "Situational Awareness", a 165-page AGI essay, which went viral. He raised a hedge fund (Situational Awareness LP) backed by Patrick and John Collison (Stripe), Nat Friedman, and Daniel Gross. As of its most recently aggregated 13F filings (via whalewisdom.com), the fund reported a U.S.-equity portfolio concentrated in electricity, semiconductors, and compute infrastructure. 13F filings disclose long U.S.-equity holdings only, not total assets under management or fund-level returns; figures circulating online about total AUM or first-year performance have not been independently verified for this work. He was approximately 23 as of mid-2025.

The Salesforce-ServiceNow Territorial Dispute

As CNBC reported, in July 2024, Marc Benioff appeared on Jim Cramer's Mad Money alongside Workday CEO Carl Eschenbach to announce a joint partnership. During the appearance, he described Salesforce as "McDonald's" and ServiceNow as "Wienerschnitzel", implying ServiceNow is a regional specialty while Salesforce is the global default. McDermott responded days later at a conference: "Have we really gotten that far under their skin that they're doing this type of a thing?" He called the comments "unhinged." The underlying territorial dispute: Salesforce was releasing an IT service management product to compete on ServiceNow's home turf; ServiceNow was pushing into CRM to compete on Salesforce's home turf. Two enterprise software companies had been quietly competing over vertical territory for years. The McDonald's/Wienerschnitzel exchange was when the dispute went public. The AI agent layer has made this collision inevitable: Agentforce and ServiceNow's AI agents do overlapping things for overlapping customers, and the product distinction is collapsing faster than the companies can reposition themselves.

Extended Roster · 34 more from this chapter

Lukas Biewald, GM Weights & Biases / CoreWeave (b. 1981). American AI entrepreneur. Started career at Yahoo! GitHub | Sources: Wikipedia Weights and Biases Lukas Biewald site

Joelle Pineau, Chief AI Officer (b. 1974). One of Canada's most prominent AI scientists. PhD from CMU supervised by Sébastien Thrun; career at McGill, then Mila / Université de Montréal. | Sources: Wikipedia Mila Quebec AI Institute TechCrunch

Jason Wei, Researcher. Dartmouth undergrad; Google Brain researcher under Quoc Le. Co-invented chain-of-thought (CoT) prompting, the technique that taught LLMs to reason step-by-step (NeurIPS 2022, one of the most-cited AI papers ever). GitHub | Sources: Jason Wei site arXiv Tech Times Google Scholar

Jack Rae, Distinguished Scientist. British AI researcher. Led pre-training for Google DeepMind's Gemini. | Sources: Google Scholar Cognitive Revolution

Bhavin Shah, SVP & GM Moveworks and AI, ServiceNow. Founded Moveworks in 2016 to bring AI-powered enterprise support to large organizations; ServiceNow acquired Moveworks for $2.85B in 2025 (announced March, closed December). Prior to Moveworks, founded Refresh.io (acquired by LinkedIn 2015) and led projects at LeapFrog. | Sources: TechCrunch Moveworks ServiceNow

Athena Karp, SVP AI Strategy, Workday. Founded HiredScore in 2012, building AI-powered talent orchestration software for responsible/explainable HR AI. Workday acquired HiredScore in 2024; she now leads Workday's AI strategy as SVP and continues as GM of HiredScore. | Sources: Workday Blog Georgetown JCW

Nick Mehta, Special Advisor / Board Member, Gainsight (b. 1977). Defined and led the customer success category for 12+ years as Gainsight CEO, scaling past $100M ARR and 20,000+ customers before stepping down to Board Director & Special Advisor. Previously CEO of LiveOffice (acquired by Symantec). | Sources: F5 The Software Report PubMatic IR

Francesca Rossi, IBM Fellow · Global Leader for Responsible AI (b. 1962). IBM Fellow and Global Leader for Responsible AI and AI Governance. Past president of AAAI (Association for the Advancement of AI). | Sources: Wikipedia IBM Research AI for Good (ITU)

Howie Xu, Chief AI & Innovation Officer, Gen · Stanford Lecturer. Howie Xu is Chief AI & Innovation Officer at Gen Digital and a guest lecturer at Stanford GSB since 2018. He founded VMware's networking unit during its rise to a $40B market cap, was founder/CEO of TrustPath (acquired by Zscaler), served as SVP of AI at Palo Alto Networks and VP of AI/ML at Zscaler, and was previously an EIR at Greylock Partners. | Sources: Gen Digital Greylock News Substack

Lucas Beyer, Researcher. Belgian/Swiss computer vision researcher. Co-invented the Vision Transformer (ViT), 'An Image is Worth 16x16 Words' (ICLR 2021), showing transformers match CNNs at scale on vision tasks. | Sources: Lucas Beyer site Google Scholar GitHub

Lynn Wu, Associate Professor, Wharton · Director of Embodied AI. Associate Professor at Wharton studying how AI and technology transform work, organizations, and innovation. Director of Embodied AI research. | Sources: Wharton OID Google Scholar MIT IDE Wharton Executive Education

Megh Gautam, Building the Agentic Enterprise, Box. Building the agentic enterprise at Box. Former CPO at Crunchbase, previously at Dropbox, Twilio, and Microsoft. | Sources: about.crunchbase.com Stanford Entrepreneurs Modern CTO

Adam Evans, Former EVP & GM AI Platform / Agentforce, Salesforce. Adam Evans leads Salesforce's AI Platform and Agentforce. A serial founder, he co-founded RelateIQ (acquired by Salesforce in 2014 for $392M) and Airkit.ai (acquired by Salesforce in 2023), and earlier founded the health division at Palantir. | Sources: Salesforce Time TechCrunch Salesforce Ben

Ajeya Cotra, Senior Researcher, Open Philanthropy. Ajeya Cotra is a senior researcher at Open Philanthropy focused on risk assessment for loss of control from advanced AI and transformative AI timelines. She authored the influential 'Biological Anchors' framework for forecasting transformative AI and co-writes the Planned Obsolescence Substack on AI futurism with Kelsey Piper. | Sources: Planned Obsolescence 80,000 Hours EA Forum

Charles Isbell, Chancellor, UIUC (b. 1968). Charles Lee Isbell Jr. is the 11th Chancellor of the University of Illinois Urbana-Champaign since August 2025. | Sources: Wikipedia Georgia Tech faculty Illinois Chancellor Georgia Tech

Cynthia Rudin, Lehrman Distinguished Professor of CS, Duke (b. 1976). Cynthia Rudin is the Lehrman Distinguished Professor at Duke with appointments in Computer Science, ECE, Statistical Science, Mathematics, and Biostatistics, and directs the Interpretable Machine Learning Lab. She is the leading advocate for inherently interpretable ML over black-box explainability, particularly for high-stakes decisions like criminal justice and healthcare. | Sources: Duke CS profiles Wikipedia Google Scholar Duke SCAI

Hila Lifshitz-Assaf, Professor, Warwick Business School / Faculty Harvard LISH. Co-author of the landmark 'Jagged Frontier' field study (with Dell'Acqua, Mollick et al.) showing how AI tools reshape consultant performance at BCG. Heads the Artificial Intelligence Innovation Network (AiiN) which links industry leaders and academic researchers studying real-world human-machine collaboration. | Sources: Hila Lifshitz site Warwick Business School Harvard Lish Google Scholar

Hyung Won Chung, Researcher. MIT PhD. Co-invented instruction fine-tuning at Google Brain, the FLAN-PaLM papers (2022) that showed fine-tuning on 1800+ tasks dramatically improves LLM generalization. | Sources: HW Chung site arXiv Google Scholar

Iyad Rahwan, Director Center for Humans & Machines, Max Planck (b. 1978). Syrian-Australian scientist born in Aleppo who established the field of Machine Behavior with Manuel Cebrian and Nick Obradovich. Founding director (since 2019) of the Center for Humans & Machines in Berlin; previously associate professor at MIT Media Lab (2015-2020) where he led the Scalable Cooperation Group. | Sources: Wikipedia Max Planck Institute Iyad Rahwan site Google Scholar

Karim Lakhani, Dorothy Hintze Professor / Director LISH, Harvard Business School. Karim R. Lakhani is the Dorothy & Michael Hintze Professor at HBS, founder and co-director of the Laboratory for Innovation Science at Harvard (LISH), and principal investigator of the NASA Tournament Laboratory. | Sources: Harvard Business School Wikipedia

Karl Sims, Computer Graphics & Artificial Life Researcher (b. 1962). Pioneer of artificial life simulations, his "Evolved Virtual Creatures" (1994) remains a landmark in the field. MIT Media Lab alum. | Sources: Wikipedia Karl Sims site MacArthur Foundation MIT Media Lab Karl Sims site

Kelly Trindel, Head of Responsible AI, Workday. Dr. Kelly Trindel leads Responsible AI at Workday, governing AI ethics and governance for the company's AI-powered HR and finance platform. | Sources: Workday Blog Workday Life blog PLI

Madison Huang, Senior Director, NVIDIA Omniverse & Physical AI (b. 1990). Jensen Huang's daughter. Started as an intern at NVIDIA and worked her way to Senior Director leading Omniverse and Physical AI, the simulation platform that underpins robotics and autonomous systems. | Sources: baike.baidu.com Storyboard18 Seoul Economy Daily Korea Herald UPI

Manuela Veloso, Head AI Research JP Morgan / Professor Emeritus CMU (b. 1957). Manuela Veloso, born in Lisbon, Portugal, is Head of AI Research at JPMorgan Chase, where she founded the firm's AI research lab in 2018, and Herbert A. Simon University Professor Emerita at Carnegie Mellon's School of Computer Science. | Sources: Wikipedia CMU CS CMU The Alan Turing Institute

Mike Booth, VP Product Marketing, Workday. Vice President of Experiential Product Marketing at Workday since April 2025, leading AI-powered product demonstrations and storytelling. Previously held leadership roles at Salesforce (Global VP Solution Strategy, VP Solutions Engineering), Yext, and Skedulo (SVP Customer Solutions), with deep experience in CRM/ERP solution engineering. | Sources: Workday

Milind Tambe, Gordon McKay Professor / Director AI for Social Good, Harvard. Gordon McKay Professor of Computer Science at Harvard (since 2019) and concurrently Principal Scientist and Director of 'AI for Social Good' at Google Research. Previously a Professor at USC where he co-founded the Center for AI in Society (CAIS). | Sources: Wikipedia Harvard Teamcore Harvard Kennedy School Harvard SEAS

Mingxi Wu, Co-Founder & SVP Engineering, TigerGraph. Co-founded TigerGraph (originally GraphSQL) in 2012 and led engineering for eight years before being promoted to CEO in June 2023. Co-designed the GSQL query language with Alin Deutsch in 2015. | Sources: Crunchbase Datanami Wikipedia

Philipp Herzig, CTO & Chief AI Officer, SAP. Computer scientist who became SAP's Chief Technology Officer in January 2025, after serving as Chief AI Officer leading SAP's AI agenda across its ERP, HR, and supply chain products. Holds a PhD from Technische Universität Dresden with research and patents in event streaming and gamification. | Sources: SAP Technology Magazine CIO

Rebecca Hinds, Head of Work AI Institute, Glean. PhD researcher and bestselling author of 'Your Best Meeting Ever.' Heads the Work AI Institute at Glean, studying how AI transforms workplace collaboration. Previously led workplace research partnerships with Stanford and Asana. | Sources: Glean Rebecca Hinds site HR Policy

Sinan Aral, David Austin Professor / Director MIT IDE. Sinan Aral is the David Austin Professor at MIT and Director of the MIT Initiative on the Digital Economy (IDE), succeeding Erik Brynjolfsson in 2020. His research on social networks, social media, and digital strategy includes landmark work on the spread of false news, and he authored 'The Hype Machine.' He is also a founding partner at Manifest Capital and Milemark Capital and has received the Microsoft Faculty Fellowship, NSF CAREER Award, and MIT Sloan's Jamieson Award for Teaching Excellence. | Sources: MIT Sloan MIT IDE Sinan Aral site

Tsedal Neeley, Naylor Fitzhugh Professor / Chair HBS AI Academy. Tsedal Neeley is the Naylor Fitzhugh Professor of Business Administration at Harvard Business School, where she serves as Senior Associate Dean and Chair of the MBA program and is founding Chair of the HBS AI Academy. She studies digital transformation, remote/global work, and cultural change, and authored the award-winning book 'The Digital Mindset' which introduced the '30% rule' for thriving in an AI-driven era. | Sources: Harvard Business School Harvard Business School Tsedal Neeley site HBR

Patterns of similarity

A handful of schools educated most of the cast. Across the 420 people profiled in this book (as of May 2026), Stanford appears in 77 biographies (by documented education field), Harvard in 49, MIT in 45, UC Berkeley (counting undergraduate and graduate degrees combined) in 28, Carnegie Mellon in 22, and Oxford in 21. After that the distribution flattens fast, Wharton 14, IIT 12, Princeton 11, Yale 11, Cambridge 9, but the headline is that you can predict roughly one in five of these biographies just by guessing "Stanford."

Three sub-patterns sit inside that education concentration. The first is the academic family, the parent who was already a professor or researcher, often the parent who handed over the first programming book. Sam Altman's mother is a dermatologist and his father worked in real estate, but the more telling case is Demis Hassabis, whose chess-prodigy childhood routed straight through Cambridge and University College London before DeepMind. Yann LeCun went from the Sorbonne to Bell Labs to NYU; his children grew up in the same kind of research-household he had reproduced. The pattern is hereditary in the soft sense: the dinner-table conversation was already about ideas, papers, citations.

The second sub-pattern is the immigrant pipeline, almost entirely through India, Israel, the UK, and China. Sundar Pichai (Madurai → IIT Kharagpur → Stanford → Wharton → Google) is the canonical version, but the structure repeats: an elite undergraduate in the home country, a graduate program in the US, a first job at Google or Microsoft or Bell Labs, then a founder role a decade later. Arvind Jain (IIT Delhi → Glean) and Amit Zavery (Gujarat → Oracle → Google Cloud → ServiceNow) sit on this line. So does the Israeli variant, Unit 8200 instead of IIT, Tel Aviv University instead of Stanford, but the same shape ending in a Silicon Valley founding.

The third sub-pattern is the dropout from an elite school, which is the most cinematic and the rarest. Sam Altman left Stanford. Michael Truell of Cursor left MIT. Scott Wu of Cognition is a competitive-programming prodigy who skipped the standard graduate-school step. But almost every "dropout" in the cast dropped out of a school that 99% of the population could not get into in the first place, the dropout is a sorting mark, not a counter-mark.

The hobby data is the quietest finding. Reading, hiking, and cooking dominate. Chess shows up four times, which is a lot for a hobby that requires explicit study; tennis and running cluster among the operators and investors more than the researchers. Sailing appears three times and is almost entirely a partner-at-a-VC-firm tell. None of this is causal. It is the texture of the cohort.

The opposite case

Some names break the pattern hard enough that the name itself is the surprise.

Eliezer Yudkowsky left school in middle school and never returned. He has no undergraduate degree, no graduate degree, no institutional research affiliation in the traditional sense, and yet MIRI and LessWrong shaped the way a substantial fraction of the cast thinks about alignment risk. The intellectual lineage runs through blog posts, not citations.

Aravind Srinivas missed the IIT computer-science cutoff by a hair and ended up in electrical engineering at IIT Madras instead, then enrolled at Berkeley for a PhD in computer science and walked out before finishing to start Perplexity. He is the cast's clearest case of a near-miss on the canonical India pipeline turning into a founder anyway.

Liang Wenfeng built DeepSeek not out of a research lab but out of High-Flyer, a quantitative hedge fund he had already run for years. The implication, that a trading firm with enough GPUs can produce a frontier model, is among the most disruptive structural facts of 2025 and 2026, and it came from outside the lab pipeline entirely.

teknium has no documented legal identity, no university, no LinkedIn. They fine-tune open-weight models on consumer GPUs, run Nous Research as a public collective, and ship Hermes releases that compete with corporate labs on specific benchmarks. The cast contains exactly one pseudonymous principal at this level of influence.

Chamath Palihapitiya arrived in Canada as a Sri Lankan refugee, worked at Burger King at fourteen, did not collect a PhD, and routed into the cast through AOL and early Facebook rather than through a research lineage. He is one of the few capital allocators in the book whose biography reads more like an immigrant-labor story than a credentialing story.

Andrej Karpathy does have a Stanford PhD under Fei-Fei Li, so he is not credential-light. The pattern-break is what he did with it: he turned into a public teacher, hand-coding GPT from scratch on YouTube to an audience of millions, when nearly everyone else at his level retreated into private research. The Slovak-family Czechoslovakia escape narrative is the second half, he is the rare frontier-lab alumnus whose origin story he tells himself in public.

Geoffrey Hinton went through English public school, Cambridge, and an Edinburgh PhD, so on paper he is the most pattern-conforming person in the cast. The opposite case is his stance: he resisted the field's commercial absorption for decades, left Google publicly to speak about risk, and refused to ride the Turing-Test wave on its own terms. His mirror-life is the one he chose not to live.

Projects, by category

Grouping the cast by project rather than by biography produces a different map. The category boundaries are softer than they were a year ago, but they are still the cleanest way to walk the landscape.

Foundation model labs. The lab tier still anchors the cast even as it stops being the only place frontier work happens. OpenAI (Sam Altman, Greg Brockman, Ilya Sutskever) · Anthropic (Dario Amodei with Daniela Amodei and the founding seven) · Google DeepMind (Demis Hassabis, Jeff Dean, Koray Kavukcuoglu) · xAI (Elon Musk) · Mistral AI (Arthur Mensch, Guillaume Lample, Timothée Lacroix) · Cohere (Aidan Gomez, Ivan Zhang, Nick Frosst) · DeepSeek (Liang Wenfeng) · Meta GenAI (under Yann LeCun's research direction).

Vector and graph databases. The retrieval layer underneath every agent stack. Pinecone (Edo Liberty, Ash Ashutosh as CEO) · Chroma (Jeff Huber) · Neo4j (Emil Eifrem) · TigerGraph (Yu Xu) · Weaviate (Bob van Luijt, Etienne Dilocker).

Agent frameworks and orchestration. The plumbing layer that turns a model into a workflow. LangChain (Harrison Chase) · LlamaIndex (Jerry Liu) · CrewAI (Joao Moura) · Composio (Shrey Batra) · Sierra (Bret Taylor, Clay Bavor) · Hermes Agent / Nous Research (teknium, Jeffrey Quesnelle).

Coding agents and IDE-native AI. Probably the fastest-moving category by revenue. Cursor / Anysphere (Michael Truell) · Cognition / Devin (Scott Wu) · GitHub Copilot (Thomas Dohmke) · Replit (Amjad Masad, Haya Odeh) · Codeium / Windsurf (Varun Mohan).

Hardware and silicon. The constraint layer everything else negotiates with. NVIDIA (Jensen Huang) · AMD (Lisa Su) · TSMC (Morris Chang, C.C. Wei) · Cerebras (Andrew Feldman) · Groq (Jonathan Ross) · Broadcom (Hock Tan, including the Anthropic XPU partnership).

Open-source and open-weights camp. Often defined more by their release practices than by their geography. Mistral · LLaMA team at Meta · Hugging Face (Clément Delangue) · DeepSeek · Nous Research.

Enterprise AI courts. The incumbents bolting agentic features onto installed bases. Salesforce / Agentforce (Marc Benioff, Clara Shih, Joe Inzerillo) · ServiceNow / Now Assist (Bill McDermott, Amit Zavery) · Workday (Aneel Bhusri, Gabe Monroy) · Microsoft AI / Copilot (Satya Nadella, Mustafa Suleyman) · Glean (Arvind Jain).

Behavioral and organizational context layer. The newest category in the book, and one of the least populated. BehaviorGraph (Yumi Kimura), a runtime context layer for who-should-act decisions inside enterprise AI · Atlassian Teamwork Graph (Tamar Yehoshua), the same problem from inside the Atlassian suite · Lattice (Jack Altman), adjacent on the people-data side.

The category boundaries above were sharper twelve months ago than they are now. Model labs are turning into agent platforms (OpenAI shipping Operator, Anthropic shipping Claude Code as a first-party product). Hardware vendors are training their own models to validate their silicon. Vector databases are bundling agent frameworks. Enterprise suites are shipping the same retrieval-plus-orchestration stack the indie agent companies built. What no one fully anticipated is that the blur between categories is not evidence the market is maturing; it is evidence the market is still trying to figure out what it is. By the next edition of this book, half of the headers above will need to be redrawn, and the mirror lives will mirror each other one click closer.

Chapter 9 · The Outsider Playbook: How People Enter the AI Power Network

The previous chapters mapped the mechanics: who has power, how power moves, where the network bends. This chapter asks the practical question that follows from the data: if these are the mechanics, what can a non-insider do with them?

This chapter is for the reader who looked at the profiles in this book and thought: I am not in this network, and I am not 25. The data from the previous chapters suggests that entry is harder but far from impossible. Of the 70 founders in this dataset who entered AI after age 35, 35 are classified as outsiders by this book's elite-school rule (Stanford, MIT, Harvard, Berkeley, CMU, Oxford, Cambridge; Caltech, Princeton, and Dartmouth are not on the list, which is an editorial choice that puts Dario Amodei, Fei-Fei Li, Clay Bavor, and Mira Murati on the outsider side). All 35 sit at tier-3 influence or higher. If Caltech, Princeton, and Dartmouth were added to the elite list, the outsider count would drop to 31 and the share at T3+ would remain 100%. The headline is robust to that specific definitional choice, but it is sensitive to whatever list you pick. The median founding age for outsiders in this dataset is 43. The oldest successful outsider founded a company at 69.

The Six Entry Paths

Not all entry paths are equal. The data reveals six categories with sharply different success rates:

The Paths That Don't Work

In this dataset, failure usually means trying to enter AI without an anchor strong enough to substitute for years inside the field.

These failure paths are conservative: most outright failures are not in this book by definition, because the book covers people who became influential, so the real failure rates outside this dataset are higher.

If You Start Here: 10-Year Trajectories

Based on the career arcs in this dataset, here is what a 10-year trajectory looks like from different starting points.

What the Successful Outsiders Have in Common

Three patterns recur across the 19 outsiders who reached the top three tiers:

The long apprenticeship. 77% of successful outsiders spent 10-20+ years in a corporate role before founding. The corporate career was not wasted time. It was the apprenticeship that produced the domain expertise, the customer relationships, or the infrastructure knowledge that became the founding advantage.

Infrastructure knowledge beats brand. The strongest combination is infrastructure expertise plus domain depth. Ash Ashutosh went from HP Storage VP to founding Actifio (acquired by Google) to CEO of Pinecone. Edo Liberty went from Amazon AI Labs Director to founding the canonical vector database. The infrastructure does not need a brand. If you are the person who built the foundational piece the world depends on, you are automatically in the network.

No age ceiling. The data shows no upper bound. Geoffrey Hinton co-founded DNNresearch at 65. Eric Schmidt launched Schmidt Sciences at 69. Yann LeCun founded AMI Labs at 65. The founders over 60 in this dataset all have tier-1 results. Late entry is hard. Late entry at the top is not impossible if the credibility indicator is strong enough.

The Nonlinear Bet: Illia Polosukhin and Parallel Infrastructure

One pattern in this dataset is the insider who deliberately leaves the obvious path. Illia Polosukhin co-authored "Attention Is All You Need" in 2017, then co-founded NEAR Protocol, a layer-1 blockchain network that has raised over $500M and processes billions of transactions.

He is one of very few people who can credibly claim to have co-invented both modern AI and a major blockchain. Most researchers pick a lane. Polosukhin treated both as serious bets on how the next decade of infrastructure gets built. Most people in this dataset who tried to run two separate companies in parallel did not sustain both. Polosukhin is one of the few captured cases where parallel bets appear to have compounded, plausibly because both bets share the same underlying thesis: that open, decentralized infrastructure is what durable AI products eventually need to run on. The lesson for outsiders is narrower than it looks: parallel bets work when the bets share a thesis, not when they share only your attention.

How Each Tier Jump Works in This Dataset

T4/T5 to T3: descriptive associations in the sample. PhD holders are overrepresented at T3+ relative to T4/T5 in this dataset (43% vs 17% of subgroup, raw counts; direction of causality unknown, PhD is also confounded with age, field, role type, and inclusion probability). Co-founder edges to already-connected founders also appear more often among T3+ cases than T4/T5 cases. Public teaching is more common at T3 than at T4/T5 in this sample (26% vs ~7% with current keyword detection). Whether teaching causes upward mobility or prominence causes teaching is unknown, read it as a behaviour observed more often in the higher tier, not a prescription.

T3 to T2: The network jump. Here the composition pattern shifts. The proportion of individuals with public teaching experience is lower in T2 than in T3 in this sample (T2 share is markedly lower than 26%); whether this reflects declining utility, a compositional role shift, or reverse causality is unknown. What the cross-sectional data does show clearly is that the T3 to T2 gap is dominated by who you know: 59% of T2 people are connected to T1 vs 26% of T3 (+33 percentage points). Having worked at OpenAI (+17pp) or Meta (+11pp) also matters. This jump cannot be made from a keyboard. It requires being in rooms with decision-makers.

T2 to T1: The closed club. 96% of T1 people are connected to other T1 people. Average degree jumps from 6.9 to 10.8. At the top, prior network ties matter at least as much as technical credentials. You reach T1 when T1 already knows you.

What This Analysis Does Not Capture

Scope bias. This book analyzes the core AI industry: frontier labs, foundation model companies, the investors and infrastructure builders directly shaping how AI is built and deployed. The tier rankings reflect influence within that specific world. They do not cover vertical AI (healthcare AI, legal AI, climate AI), the broader investment industry, enterprise software outside AI, government, academia outside AI research, or any other domain. A person who is T5 in this dataset might be the most influential person in their own field. The tiers, the entry paths, and the success rates above are skewed toward the angle this book examines. Readers building careers in vertical AI, applied ML, AI policy, or adjacent industries should read these patterns as directional signals, not as a map of their own terrain.

Family and personal ties. This dataset tracks professional relationships: co-founding, co-authoring, investment, mentorship, and employment. It does not capture private or family-based pathways into the network. In practice, personal relationships with people already in the network, having a parent or sibling who is already in the network, or having children who enter the field can all create access and opportunity that this analysis cannot measure. Several people in this book are married to each other, and at least three of the most consequential co-founding pairs are siblings. Family ties are real infrastructure in this industry, and this dataset undercounts them. The playbook above is for people building from outside those channels. If you have family connections to the network, the timeline compresses significantly, but this data cannot tell you by how much.

Same Age, Different Starting Points

At the same founding age, the outcomes diverge dramatically based on starting position. These are real people in this dataset who all founded their companies at the same age.

How Fast Does Each Background Reach the Top?

The timeline to influence depends heavily on your starting position. Here is the median founding age by background type for people who reached the top two tiers.

If you are reading this and trying to figure out what to do with the patterns above, see If You Start Here: 10-Year Trajectories in Chapter 9.

A Note for Younger Readers

For founders under 30, the dataset shows a different pathway. Young founders with an elite credential reached the top two tiers 58% of the time; without that credential, the rate dropped to 33%, still meaningful but materially lower. The observed early-career path that produces the highest conversion has four components: a strong lab, a high-output advisor (the Hinton, Abbeel, or Fei-Fei Li of a given generation), one foundational publication, and co-founders the researcher already trusts. The median founding age in this dataset is 33, not 22. The "founders are college dropouts" narrative is the exception in the data, not the rule.

The rest of this section pulls seven patterns out of the 420 profiles that are specifically useful if you are 22 to 35, tech or non-tech, and do not start from inside the network. They are the counter-images to the founder mythology that dominates the press cycle.

Seven patterns hiding in 420 careers

The single line, if you are 22 to 35 and not from inside. You do not need to be American, you do not need to have gone to Stanford or MIT, you do not need to have started a company by 22, and you do not need a CS degree. The shortest documented path to tier-1 conversion is: get hired by a frontier lab or a serious infrastructure company, work five to ten years, leave with a network and a narrow thesis, and ship something other people in the network can recognize as deep. The map in this book is not a tournament bracket. It is a directory of doors.

Appendix

Appendix 1 · Editorial Notice & Legal Disclaimer

Last reviewed: 2026-05-16. This notice follows standard editorial-and-legal disclaimer practice for educational and journalistic works on public figures. Qualified counsel should review it before commercial publication; it is not a substitute for advice of counsel.

A note from the author

This book was directed by me and co-written by a few of my AI agents.

I designed the project, set the research direction, defined the chapter structure, chose what belonged and what did not, set the editorial standards, designed and reviewed the analytical methods, shaped the network maps, interpreted the structural findings, and made the final strategic argument.

AI agents, including Claude, Codex, Gemini, and GPT, helped draft biographies, profile cards, fact boxes, translations, and research summaries. They also supported source checks, consistency reviews, and methodology audits.

Most people in this book are not people I know personally. The factual base comes from public sources, supported by AI-assisted research and fact-checking. This approach can reveal patterns at scale, but it can also miss context or inherit gaps in public records. The book should be read as human-directed analysis of public evidence, produced with AI assistance, not as a traditional line-authored biography.

The project began as a private research tool to understand who trained whom, who funded whom, which institutions produced which spinouts, and which career paths appeared to lead to influence in modern AI. After the network and ML analysis surfaced useful patterns, I published a version for readers to inspect.

This is a living work. If you find a factual error, missing context, or mischaracterization, contact the author directly. Corrections are reviewed in good faith.

About the Chinese edition: A Chinese translation is available. It was produced entirely by AI using GPT and Claude, and has not been fully human-proofread. Please refer to the English original if anything seems unclear.

Purpose, audience, and scope

This project uses public data to show what successful AI researchers, founders, operators, and investors often share: institutions, relationships, career paths, funding patterns, and trust networks.

The dataset includes 420 people across the core AI industry: top-tier AI researchers, influential founders, key operators and executives, and the investors and institutions shaping frontier outcomes. The map shows who trained, funded, hired, co-founded, competed, or split; the book analyzes recurring career and funding patterns.

The pattern is talent pipelines, lab diaspora, capital flow, and outsider entry paths. The goal is to help entrants, mid-career operators, investors, and journalists see career structure, not rank people.

Scope note: this project focuses on the core enterprise and frontier AI power network, not every person or company in AI globally. It is educational and journalistic: not a directory, ranking, financial product, or professional-advice substitute.

Sources and methodology

Sources are weighted roughly in this order: primary documents (SEC and similar regulatory filings, court documents, official press releases); major business press (NYT, WSJ, Bloomberg, FT, Forbes); specialist publications (The Information, TechCrunch, VentureBeat, Stratechery); interviews and podcasts (Acquired, All-In, Possible, No Priors); company pages and public bios; and Wikipedia cross-checked against primary sources where it appears.

Data assembly and verification used Anthropic's Claude and OpenAI's Codex, with Google's Gemini 2.5 Flash with Google Search grounding for independent source-tier checks on contested or volatile claims. Multi-agent audits cover sourcing, tone, lawsuit and negative-claim red flags, and fix-plan peer review. Each new source URL is fetched to confirm claim support. Every map connection carries evidence_type and confidence, defined in Appendix · How to Read the Map. Claims were validated against external links to primary sources based on the best available evidence at the snapshot date.

What this work claims and what it does not

Despite this verification process, the work is provided "as is" and without any warranty of any kind, express or implied. Information about living people, private companies, and active business relationships changes constantly. Numbers in this work (valuations, headcounts, funding rounds, ages, marital and family status, role titles) reflect the best available public source at the time of writing and may be outdated by the time you read them. The author and contributors do not warrant the accuracy, completeness, timeliness, or fitness for any particular purpose of any information in this work.

Known biases of this dataset

Four structural biases shape what this dataset can and cannot say. They come from cast selection, observable relationships, scope, and denominator limits. Read every percentage, centrality score, and "X% reached Y" claim through these filters.

1. Recency bias (the data captures the modern AI industry, not the history of the field). Of the 194 people in the cast with a recorded founding year, 46% founded a company between 2010 and 2019, and another 30% founded between 2020 and 2025. Only 4 founders are recorded for the entire 1980s. Earlier institutions, expert systems, statistical NLP, and long-running academic programs appear mostly through later-visible alumni:

• Claims about "who shaped modern AI" reflect the modern AI industry, not the field's full history. Many AI-winter figures (1980s-2000s) are absent because they did not become visible in the post-2010 industry build-out.
• Cohort patterns (e.g., "the 2010s decade produced 46% of cast founders") are within-cast patterns; they reflect the founders the book covers, not the actual rate of AI company formation in the broader economy.
• The 2020s cohort (n=59 through mid-2025) is partial; its size will keep growing.

2. Public-visibility bias (the map shows documented ties only). Every node and edge rests on a public source: press release, regulatory filing, interview, news article, public profile, paper, podcast, or court document. The graph maps the public record, not the full social structure. What is missing:

• Off-the-record investors and angel networks. Some important capital relationships never appear in press releases. The investor edges (93 edges) are a lower bound.
• Private advisors, informal mentorships, and undocumented introductions. The advisor and mentor edges (49 combined) cover the public record; real mentor networks are wider.
• Friendships and personal relationships outside public events. The map captures marriages, siblings, and on-the-record friendships. Social ties are mostly invisible unless one person has described them publicly.
• Family ties that have not been reported. The 21 family/sibling edges are documented family relationships. Cousins, in-laws, and step-family relationships likely exceed 21.
• Private networks inside T3+ leaders' companies. The people working at these T3+ leaders' companies are themselves accomplished operators; each carries their own private network, and that layer doesn't appear on this map.

The practical consequence: "X is a bridge node" means X has the most documented ties to the cast. That usually tracks real brokerage, but it also reflects public coverage. Sam Altman appears as the network's highest-betweenness node partly because he sits near the center of the documented public record about AI.

3. Domain bias (the cast is core AI and enterprise AI). The 420 people were selected for relevance to the core AI industry (frontier labs, foundation models, infrastructure, AI-focused capital) and enterprise AI (AI products and platforms for businesses). That scope undersamples adjacent areas:

• Vertical AI. Healthcare AI, legal AI, defense AI, financial AI, climate AI, biology and drug-discovery AI. Leaders appear only when they are also active in the core AI conversation.
• Applied AI outside the lab pipeline. Robotics, autonomous vehicles, embedded ML, edge AI, and AI applied to physical industries are represented thinly.
• Non-US, non-China AI networks. European AI is partially covered (Mistral, DeepMind UK roots, ENS Paris alumni); Indian, Israeli, Latin American, African, and Southeast Asian AI networks are minimally represented.
• Academic AI outside the industry pipeline. Faculty who do important AI research but have not crossed into industry founding, advising, or board work appear less often.

Read every "the AI industry does X" claim as "the slice of the AI industry this book maps does X." The slice is narrow; the book is not a directory of the field.

4. Percentages are within-cast shares, not success rates. Every percentage in this book ("43% of T3+ are PhD holders," "80% of captured outsider tooling founders reached T1-T3," "76% of the dataset is in the Bay Area") is computed across the curated cast. Observable core-AI prominence is rare, and the true denominator (everyone who attempted a given path) is not knowable from public data:

• A within-cast share is a directional finding, not a probability of success. "80% of captured outsider tooling founders reached T1-T3" describes included people; it does not mean an outsider building AI tooling has an 80% chance of reaching the top tiers.
• The denominator of attempts is missing. For every captured outsider who reached T3+, an unknown number tried a similar path and did not become public. They are not counted.
• Strategies can self-defeat at scale. Some entry paths show up disproportionately partly because few people pursued them at the time. If many more people copy the same playbook, the competitive baseline rises and marginal advantage shrinks. These patterns describe captured conditions, not crowded-playbook outcomes.

How to use these caveats in practice. When citing a percentage, centrality ranking, "who connects whom" claim, or "this kind of person succeeds" claim, name the slice: "In the cast of 420 publicly visible core-AI and enterprise-AI figures captured in this dataset, ...". The methodology audit log (notes/ML_AUDITING.md) uses this framing.

Public figures and fair commentary

The individuals discussed are public figures or have entered the public conversation about AI through company affiliation, investment activity, public statements, regulatory testimony, or commercial roles. Biographical and commercial information comes from public sources. Reports on disputes, feuds, lawsuits, or controversies rest on court filings, on-the-record statements, and published reporting. Commentary is the author's good-faith assessment.

Nothing in this work is intended to defame, harass, or deliberately misrepresent anyone. Factual claims are sourced. Characterizations, judgments, and comparisons are framed as the author's view.

Trademarks and third-party names

All company names, product names, and trademarks referenced here belong to their owners. Their use is for identification, reference, and commentary. No endorsement, sponsorship, partnership, or approval is implied.

Corrections, contributions, and the standard for removal

This is a living document. The author's editorial posture follows mainstream journalism practices and Wikipedia's Biographies of Living Persons policy: factual corrections are welcomed and applied promptly; accurate, sourced content about a public figure's public activities is not removed on request.

Factual corrections. If you are mentioned and believe a factual statement is incorrect, out of date, or misattributed (wrong birth year, wrong job title, wrong company, misquoted statement, broken family fact, outdated valuation, and so on), contact the author (see Contact below). Substantiated corrections will be reviewed and applied; where appropriate, the correction and date will be noted.

Characterization disputes. Characterizations of documented public events (feuds, disputes, business decisions, public statements) are the author's good-faith assessment based on cited evidence. The author will consider a written response or clarification and, if it materially changes the public record, add it or update the framing. Accurate sourced reporting will not be removed by preference.

Removal requests. Removal will be considered against a narrow standard. The author will consider removing material that (a) describes genuinely private information that should not have been included and that is not material to the subject's public role; (b) misidentifies the subject (wrong person); (c) is defamatory under applicable law; or (d) violates applicable law. Requests outside those categories, including requests to remove accurate, publicly sourced content about a public figure's public activities, are not generally granted. This standard supports the work's educational and journalistic purpose and matches mainstream business journalism and collaborative reference works covering public figures.

Contribute panel Planned / in progress The Contribute panel supports four submission types: suggest someone missing, correct a fact, add a connection we do not have, or join the map yourself. The live submissions backend, "claim node" flow, and verified-contributor badge are planned features and not yet shipped.

No professional advice

Nothing in this work is investment advice, legal advice, medical advice, employment advice, or any other professional counsel. Do not make business, financial, legal, or personal decisions from it without independent verification and appropriate professionals.

Limitation of liability

To the maximum extent permitted by applicable law, neither the author, the contributors, the host platform, nor any associated parties shall be liable for any direct, indirect, incidental, special, consequential, exemplary, or punitive damages arising out of or in connection with the use of, or inability to use, this work or the information contained in it, whether based on warranty, contract, tort, or any other legal theory, and whether or not the author has been advised of the possibility of such damages.

Contact

Corrections, takedown requests, and content questions: contact Yumi Kimura via LinkedIn (LinkedIn). Use direct contact until Contribute launches.

Appendix 2 · How to Read the Map (General Methods)

The map is a graph: people, companies, labs, and movements are nodes; relationships are edges. Each edge carries a strength tier derived from type and evidence_type, so classification is reproducible from raw data.

The strength tiers

Tier	Includes	Examples
Strong	Co-founders, current direct reports, spouses and family, PhD advisor ↔ student, mentor relationships, active board seats, three or more co-authored papers.	Dario ↔ Daniela Amodei (siblings, co-founded Anthropic); Hinton → Sutskever (PhD advisor); Sam Altman ↔ Greg Brockman (co-founded OpenAI).
Medium	Single co-authorship, funding round without board seat, conference co-panel, public mutual citation, sequential employment at one firm, documented public feud or tension.	Vaswani ↔ Gomez (Transformer paper co-authors, 2017); a16z's check into a single company round; Musk ↔ Altman (documented feud).
Weak	Same school different cohort, same company at different times with no overlap, alumni networks without direct contact, generic "colleagues" without overlap detail.	Two Stanford CS alumni who graduated a decade apart; two ex-Google researchers who never overlapped.
Latent	Co-location only: same hometown, same neighborhood, shared movement identifier (EA, e/acc), same conference circuit. A structural marker, not a conventional relationship.	Two SF Hayes Valley residents who both identify as e/acc; two founders who grew up in Pittsburgh.
Affiliation	A person ↔ institution edge: currently works at, previously worked at, founded, leads, or sits on the board of. It is a membership fact, not an interpersonal tie: "who is at this company / lab / fund," not "how does X know Y."	Sam Altman → OpenAI (CEO); Dario Amodei → Anthropic (CEO, co-founder); Jeff Dean → Google DeepMind (Chief Scientist).

A note on "affiliation"

Affiliation is the only tier that does not describe a person-to-person relationship. It connects a person to an institution (company, lab, fund, university, movement). About half of all edges are affiliation, which is why big labs look dense even when their people may not know each other.

Read affiliation as scaffolding (who is where) and the four interpersonal tiers as relationships (who knows whom, how directly).

Why the tiers matter

Strong ties carry trust and high-bandwidth information; they show how influence moves through small circles. Medium and weak ties, in Granovetter's 1973 paper "The Strength of Weak Ties," carry new information across clusters. Latent edges matter only in aggregate: a shared movement label or city may point to shared context, not direct contact.

Three rules of thumb: (1) one strong tie between clusters suggests coordination; (2) bundles of medium ties without a strong tie suggest a situational, reversible relationship; (3) latent markers prompt closer review, not proof of a relationship. Several stacked latent markers are worth a second look, but still need public evidence.

What the data can prove and what it cannot

This map is a public-record reconstruction built from public sources, AI-assisted research, and human editorial review. Its limits follow from that scope:

• Some edges are assumptions, not directly verified facts. Example: overlapping at the same lab can imply likely contact, but it is not a documented relationship.
• The map is incomplete by construction. Low-profile people, unwritten relationships, off-the-record investors, unreported family ties, and friendships outside public events are under-represented.
• The strength tier reflects available evidence, not the full relationship. A real strong tie may appear as medium or weak because it was never reported; a strong tag may reflect public co-authorship without a deeper relationship.
• Confidence (verified or inferred) and strength are tracked for every connection. Read them together: a strong inferred edge is a hypothesis worth checking, not settled fact.

Claim, contribute, correct

Because the map is incomplete and partly inferred, it is a living document, not a frozen reference. Readers can participate four ways:

• Claim your node. If you appear in the map, you can claim your node, update your bio, fix your role or links, and flag wrong or missing relationships. Verified claimants get a "verified" badge.
• Correct a fact. Contact the author directly with a factual correction and source (the Contribute panel is planned but not yet live). Substantiated corrections are reviewed and applied; the correction and date may be noted.
• Add a connection. If a relationship is missing (co-founder, mentor, spouse, investor, advisor), submit it with a public source or first-person account for review.
• Suggest someone missing. Propose an important person or institution with a one-line reason and sources, especially if they appear in the book's prose but do not yet have a node.

For policy on corrections, removal, and public figures, see the Appendix · Editorial Notice & Legal Disclaimer immediately above.

How this analysis was built

The findings come from one dataset (464 nodes including 420 people, with 1,510 relationship and institution edges in the current snapshot) and standard network methods. The pipeline is local and reproducible; every conclusion comes from the dataset, not the model. The two diagrams summarize the pipeline and methods.

A note for SNA and ONA readers

Edge label semantics: evidence_type and confidence

Every edge carries two fields beyond strength: evidence_type and confidence. Strength says how weight-bearing the relationship is. Evidence and confidence say how the relationship was established and how certain the data is. Read the three fields together.

The confidence field has four values. They describe evidence status, not importance:

Value	What it means	Edge count	When you see it
verified_public	Confirmed by at least one primary or major-outlet public source (SEC filing, court record, press release, NYT/WSJ/Bloomberg/FT, official company page, Wikipedia cross-checked against a primary source).	987 of 1,510 (65%)	A co-founder relationship in an official filing; a funding round announced by press release and covered by a major outlet.
inferred	A reasonable conclusion from two or more independent public sources that do not directly state the relationship.	414 of 1,510 (27%)	Two researchers overlapped at the same lab for two documented years; an advisor appears in board documents without a formal announcement.
book_inference	An editorial inference from the analysis phase, usually for contextual connections implied by the narrative but not yet traced to a specific public source.	108 of 1,510 (7%)	A latent or weak edge from shared movement affiliation (e.g., EA or e/acc), or a context connection from biography rather than a stated relationship.
reported	Sourced to one public news report or published statement, without independent corroboration.	1 of 1,510 (<1%)	A connection sourced to a single interview or profile piece that has not appeared elsewhere.

How to read confidence in practice. The main split is verified_public (treat as fact, subject to snapshot date) versus inferred and book_inference (treat as hypotheses if citing the relationship). The 28% inferred category is normal for a public-record knowledge graph; private advisory relationships and informal mentorships often leave no primary-source trail. For fact-checking, confirm whether the edge is verified_public or inferred. Raw data: data/edges.json.

The evidence_type field goes one level deeper. It records the source class behind the edge. The taxonomy clusters into a few families:

• Institutional record (academic record, education record, employment record, career history, publication record, published paper): official university pages, company announcements, arXiv, or CVs. The largest single category, covering 456 edges.
• Company and org record (company founding, same company, same firm, same firm leadership announcement): official company pages, press releases, or LinkedIn-equivalent public career records. Covers 395 edges.
• Shared affiliation (e.g., "shared affiliation: google," "shared affiliation: openai"): person-to-institution edges showing a documented employer. Covers 250 edges. These are affiliation edges, not interpersonal relationships; see "A note on affiliation."
• Interview or biography (interview/biography, bio mention, author attestation): long-form interviews, published biographies, or first-person accounts. Covers 118 edges, often for mentor-student or early-career relationships.
• Financial or funding record (funding announcement, investment deal, financial record): press releases, SEC filings, or major-outlet funding coverage. Covers 83 edges.
• Shared school (e.g., "shared school: stanford," "shared school: tsinghua"): co-alumni tags. Used only for latent and weak edges; shared alumni status is context, not a documented relationship.

Other evidence_type values are case-specific citations for notable events (e.g., "CNBC, NYT reporting on Nov 2023 board vote," "SoftBank $41B investment, Stargate JV"). A specific label points to a traceable public record.

Inclusion rules: who is in the 420-person cast and why

The cast covers 420 people plus 44 institutional nodes, for 464 total. That editorial choice shapes every pattern the book reports. Here is how the cast was selected and what that means.

The four inclusion criteria. A person enters the cast if they meet at least one of these:

1. Frontier-lab founding or senior leadership. Co-founders, CEOs, CTOs, presidents, chief scientists, and comparable leaders at OpenAI, Anthropic, Google DeepMind, xAI, SSI, Mistral, and other frontier-tier labs.
2. Outsized documented influence on the industry's direction. Foundational researchers, investors, executives, and people who shape funding rounds, hiring decisions, governance structures, or multiple independent narratives in the book.
3. Network relevance. People who connect otherwise-disconnected clusters. Bridge figures at the junction of capital, talent, and institutional networks qualify on this criterion alone.
4. Enterprise and infrastructure coverage. Key executives, founders, and investors in compute, cloud, tooling, and enterprise distribution entered the cast when their work intersects the frontier-lab narrative.

What is deliberately not in the cast. These categories are underrepresented or absent:

• Vertical AI specialists (healthcare AI, legal AI, defense AI, climate AI) unless they also appear in the core industry conversation captured by this book's sources.
• Academic AI researchers without industry crossover. Faculty who publish important work but have not founded companies, joined lab leadership, or appeared in the cast's funding or governance networks.
• Pre-2010 AI figures from expert-systems and statistical-NLP waves who did not re-emerge in the modern deep-learning build-out. They appear through students and intellectual descendants, not as primary nodes.
• Non-US and non-China networks outside a small set of exceptions (Mistral, DeepMind's UK roots, and a few European research clusters). This is a documented gap, not a judgment about importance.
• People who asked not to be included, where the editorial standard for removal was met (see Appendix 1 corrections and removal policy).

Why 420 and not more. Below T3, public-record coverage thins: fewer press releases, formal bios, and sourced relationship claims. More people would raise node count but lower average source quality and edge density. The 420 number is where the quality/noise tradeoff fit the book's analytical goals.

For implications on percentages and rankings, see "Known biases of this dataset" in Appendix 1. Domain bias, recency bias, and the within-cast-percentage caveat (Bias 4) flow from cast selection.

How to read a tier number

The five influence tiers (T1 through T5) appear in claims like "80% of captured outsider founders reached T3+" or "the network's T1 cluster is 29 people." The full scoring criteria, 4-dimension rubric, and distribution table (T1: 29, T2: 105, T3: 179, T4: 72, T5: 35) live in Part 1's "How the five tiers are scored" section. This section explains how to read a tier number.

What a tier number reflects. Tier scores one observable thing: current institutional role scope inside core AI, as visible from the public record. The four dimensions (institutional leverage, capital and decision authority, technical or operating contribution, and network position) are proxies for role weight. T1 means load-bearing for the industry's direction. T5 means connector or contextual figure rather than primary actor.

What a tier number does not reflect. Tier is not intelligence, talent, character, or long-term historical significance. It excludes private influence and adjacent fields where a person might rank higher. It changes over time: a T3 person who founds a lab, raises a large round, and publishes a foundational paper can move to T1; a T1 person who steps back can drift toward T3 or T4. Tier is a role snapshot, not a person judgment.

The construct caveat. Tier assignment is editorial judgment. There is no published inter-rater reliability study for this taxonomy. Two editors can reasonably disagree on T2 versus T3. The book uses tier as a rough grouping variable (e.g., what share of T2+ people have a PhD). It is useful for coarse patterns, not individual ranking. Every association using tier is descriptive within the cast, not causal.

The public map consolidates five tiers into three (High, Medium, Supporting) to reduce ranking overread. The five-tier system remains in the data; readers see the three-tier sidebar. See Part 1's influence pyramid for the mapping.

Update and snapshot policy

The dataset is a snapshot, not a live feed. This matters for any claim about a number, role title, or relationship status.

What "snapshot date" means. The data reflects public sources through approximately May 2026, when the most recent audit pass completed. Present-tense claims ("X is CEO of Y," "the fund has $Z under management," "A and B are co-founders of C") describe the public record then. Roles change, people leave, and rounds close; the snapshot does not update retroactively.

For volatile claims (valuations, funding rounds, board membership, role titles), the book uses dated attribution: "as of [month year], [source] reported that..." An undated present-tense company or role claim should be read as snapshot-date, not necessarily current.

When the dataset gets refreshed. Updates happen in batches. Triggers are: (1) a significant new cast-member development that changes a foundational claim (a merger, lab founding, major dispute going public); (2) a reader correction that passes Appendix 1's standard; or (3) a periodic audit of source quality, broken links, and inferred edges across much of the dataset. Git history records the update log, with each commit tied to its change and source.

How corrections flow in. Factual errors (wrong role title, institution, date, relationship) are corrected when sourced. Characterization disputes are considered but do not automatically change the text. Removal requests follow Appendix 1's narrow standard. Every correction goes through the same multi-agent verification loop as the original data: a source is fetched and confirmed before the change is applied. Unverified corrections are not accepted, regardless of sender.

What changes and what does not change across updates. The node list, edge list, and biographical details can change. Analytical claims (percentages, centrality scores, motif findings) are recalculated only when the data changes enough to shift the conclusion. Minor corrections (misspelled name, updated job title) do not trigger full re-analysis; major corrections (mis-attributed co-founder relationship, wrong tier for a prominent figure) do. If a recalculation changes a finding, the book notes it with the revision date.

For review history and corrections through the current edition, see the notes and git history. For corrections and removal requests, see Appendix 1.

Appendix 3 · Full Roster

This appendix is a compact index of the 420 people in the map. For full bios, relationship context, and search, use the interactive graph and click any person.

Why these people?

These 420 people were initially selected through an AI-assisted research process that started from known frontier AI labs, major AI infrastructure companies, enterprise AI platforms, top AI investors, foundational researchers, and organizational-context experts, then expanded outward through public relationships such as cofounding, employment, funding, board roles, advisor ties, academic lineages, coauthorship, acquisitions, and documented family or trust relationships.

The roster is a reference layer, not a chapter. Inclusion follows four selection rules, applied together:

1. Decision authority in modern AI. The person makes calls on training, deployment, funding, hiring, or governance at a frontier lab, hyperscaler, model company, infrastructure firm, enterprise application, or AI-focused fund.
2. Documented public role. The role is verifiable from public sources such as company pages, regulatory filings, peer-reviewed papers, mainstream press, public board records, or public research profiles.
3. Network position matters. The person sits at the intersection of two or more of: a lab, a fund, a university, a board, a paper, or a founding team.
4. Editorial relevance. The person is referenced in at least one chapter, or is a documented relationship to someone who is.

The roster is a snapshot. Coverage skews toward the people the book argues about, and the interactive graph remains the canonical surface for searching and exploring connections.

Appendix 4 · Schools & Universities

This index lists the 13 universities that appear as institution nodes on the map, ordered by affiliated-node count (descending). Each entry shows the total count and the primary sub-field or talent channel the institution represents in this network. Use it to find which school produced the most people in a given role type, or to cross-reference the talent-flow diagram in Part 4. Two people sharing an institution affiliation did not necessarily overlap in time or know each other.

Counts: affiliated-node totals come from institution-type edges in the underlying graph. Affiliations span undergraduate, graduate, postdoc, and faculty roles.

Appendix 5 · Company & Product Guide

This index covers the companies, labs, and products referenced in the book, grouped by layer: frontier labs, compute, enterprise SaaS/workflow, data and tooling infrastructure, specialized AI, and capital. Each entry shows the number of named individuals tracked in this dataset (if any), a one-line description, and why the company is relevant to this network. Use it to quickly orient on any organization mentioned in the text.

Created and edited by Yumi W. Kimura