Preface

An AI product often brings several businesses into the same picture: a model lab, a chipmaker, a cloud provider, and an open-source community may all contribute to what someone uses. Understanding that product means understanding how their work fits together—and where their interests or views diverge.

Over the last decade, Silicon Valley AI has moved from research labs into products used by millions. The 2017 Transformer paper introduced a new model architecture; OpenAI released ChatGPT to the public in 2022. How the technology works, who sells it, and who is responsible when it fails have become part of the same conversation.^1,2

This book approaches that conversation through people. Its companion map covers 473 people and 1,794 publicly recorded ties: a selected cast whose research, companies, collaborations, investments, mentors, and rivalries help explain how the industry got here. Following their work gives unfamiliar institutions a human scale. Following their connections shows how careers and organizations intersect.

The book and map preserve a snapshot dated 2026-09-23. The field continues to change; this is the record through that date.

The map in 10 minutes

The map turns on four questions.

Who are the people? Some paths run through research or engineering; others through knowledge of customers or the experience of building a business. A title records a position; a career shows the work behind it. Chapter 1 traces this distinction through Anthropic, NVIDIA, and Harvey.

How are they connected? Work moves beyond its original setting through collaborators, investors, advisors, and introducers. Cofounding, coauthoring, investment, and introductions are different kinds of ties. A shared school or employer records common background; it does not establish that two people met.

What do their companies actually do? A product may depend on models, chips, computing capacity, and applications built for customers. These are different businesses. NVIDIA’s computing platforms and Harvey’s professional services belong in the same industry picture, but perform different work.

Where do they disagree? The book compares ideas as well as businesses, and follows people who once worked together but later took different paths. Connections record where histories intersect; stated views and subsequent work show where those paths diverge.

Four timelines, running from CUDA in 2006 to September 2026, supply the chronology. The chapters develop the comparisons; the map lets you examine the people and sources behind them.

How to read this book

Read the chapters for an account of how work, relationships, and organizations fit together. In the PDF, numbered citations link to their sources. When someone catches your interest, select that person in the interactive map to see their profile, sources, connections, and every section of the book that mentions them.

Counts and shares describe this selected group. They are not success rates for all founders. Influence labels are editorial judgments that partly credit network position and frontier-lab roles alongside the work described here. Keep those limits in mind when comparing careers: the map makes a public record easier to explore, but inclusion and prominence are themselves part of how that record has been assembled.

The Storyline in Four Timelines

Fifteen anchor events give the chronology; the chapters explain the patterns. Each color marks one thread: models and methods, knowledge and agents, labs and breakaways, capital and deals, or conflict and governance. The complete chronology remains in the web companion.

How They Became Influential

Dario and Daniela Amodei led a team with research behind it into the founding of Anthropic. Jensen Huang brought chip-industry experience to the company he founded with Chris Malachowsky and Curtis Priem. Harvey’s cofounders brought legal practice and AI research experience to a shared problem. Across these different paths, prior work met collaborators and an organization able to carry it further.

What turned that work into influence? These careers suggest a recurring pattern: people made work visible, found collaborators who could extend it, and built organizations that could put it to use. Research, engineering, customer knowledge, and operating skill supplied different starting points. Colleagues, advisors, backers, and customers gave the work routes beyond its original setting.

That pattern offers a way to read the careers that follow. Pay attention both to what someone could do and to the people and institutions through which the work gained reach.

Three ways it appears

A research group can become a new institution. Dario and Daniela Amodei led the team that founded Anthropic. Its early announcement described prior research by the team and a $124 million Series A to fund further work. Here, the starting point was already collective: an existing body of research and a group able to continue it. Financing supported a new organization in which that work could proceed. The announcement brings those elements into view together, without isolating which mattered most.¹

Engineering experience can become a company with reach. Jensen Huang worked at AMD and LSI Logic before founding NVIDIA with Chris Malachowsky and Curtis Priem in 1993. NVIDIA first built for graphics and later made computing platforms used in AI. To understand this path, follow the experience into the products and the organization: earlier chip work was carried into a company that kept building and distributing new products. The career’s significance becomes clearer when the employers’ names are connected to the work that followed.^2,3

Knowledge of a customer’s work can connect to technical ability and backing. Harvey’s cofounders combined legal practice and AI research experience. Their starting point was a specific professional use case and a team with complementary knowledge. The company says they contacted OpenAI through a cold email; its 2023 funding announcement named Sequoia and the OpenAI Startup Fund. This publicly documented route connects familiarity with a customer’s work to technical and financial support.^4,5,6

Together, these cases show prior work becoming more consequential through teams and institutions. They do not establish a formula for success, or tell us whether a particular introduction, investment, or employer was indispensable. Chapter 2 follows the documented relationships more closely.

What the map records

Among founders matched to institutions in this selected cast, many associated with frontier labs carry the High influence label. Appendix 2 shows that comparison with its method notes. The groups overlap, the institutional match is approximate, and frontier-lab roles partly inform the editorial label itself.

The comparison describes where an editorial judgment falls within the selected record. It cannot establish that attending a lab produces influence. For interpretation, return to the career: what work is recorded, who participated, and what organization put it to use?

Age and paths

These careers also raise a question about time: how much experience had founders accumulated before starting the companies they are known for?

Among 159 founders with a recorded age at that founding, the median was 33; 70 founded at 35 or older. Those figures describe age and career stage within this selected cast. They cannot estimate the prospects of people who never entered the map or establish that waiting longer leads to influence.

Different ages can reflect different histories. Across these careers, research, engineering, domain knowledge, collaborators, and organizational reach recur in combinations. The value of following a path is to see how its particular combination came together—and what work it made possible.

What Ties Them Together

Recorded ties

Shyam Sankar cold-emailed Kevin Hartz as a student. In Sankar’s account, Hartz introduced him to Roelof Botha and Keith Rabois; through that network he reached Peter Thiel. Sankar became Palantir’s first business hire. Each step identifies something more specific than membership in a circle: a person made an introduction, another connection followed, and a job came later.

The map is a layered record. It contains 1,794 recorded ties, about four in five marking a shared school, employer, or institution. Sankar's sequence shows how one map can record different kinds of links: shared context, an introduction, and a later job.

Shared settings

A shared setting is a starting point for inquiry. A school or employer places people in an institutional history; it does not, by itself, establish that they worked together or shaped one another’s ideas.

The VMware example makes the limit explicit. Martin Casado and Jerry Chen held senior roles during overlapping years and later expressed similar views about enterprise AI. No account shows that they developed those views together. Keep that boundary visible whenever proximity makes a stronger story tempting.

Advisor lineages

Advising, later collaboration, and a student's destination are three distinct kinds of evidence.

A student who later joins or founds a company does not thereby add a link between the advisor and that company. Where the same people were both advisor and student and later cofounders, each relationship has its own record.

Shared papers

A paper supplies evidence of collaboration on a particular work. Its author list does not make every later employer part of that collaboration, or imply that the authors followed a common path afterward.

The evidence section therefore keeps publication-time affiliations separate from subsequent roles. It also preserves the distinction between a single paper’s authors and a cluster of several collaborations. Those are different shapes of connection.

Friendships

Friendship adds another kind of history. In some documented cases, it preceded a founding partnership; elsewhere, it connects people whose organizations later competed.

Read the personal relationship and the formal role separately. A friendship can remain part of the record without becoming an explanation for every subsequent company decision.

Founding circles

Cofounding records participation in creating a company. Most founding circles here sit within one company; the PayPal circle is the clearest cross-company exception, though it is no longer the only one.

Circle membership also depends on the map’s definitions. Here, any documented YC participation or role counts as a YC link, and documented PayPal Mafia involvement counts as a PayPal-circle link. Memberships can overlap.

xAI cofounder Igor Babuschkin cofounded River AI in 2026, adding another cross-company link. It raised $1.1 billion in August 2026.^1,2

Other circles

Cohorts, labs, and communities add further connections, but their functions differ. Selection, mentoring, introductions, funding, and shared employment should each retain their own label.

The evidence table lets readers inspect those functions without treating every publicly listed membership as a personal collaboration.

Daniel Gross founded a company in a YC batch, later became a YC partner, and invested in AI Grant with Nat Friedman. Cursor appears in both AI Grant and Neo's public lists.^3,4,5,6

Family and partners

Family and partner relationships sometimes overlap with founding and investment. Keep both relationships visible when both are documented. The personal tie alone does not establish a business transaction or explain why one occurred.

The map also records these family ties: Jensen Huang's children Madison Huang and Spencer Huang work at NVIDIA (Spencer Huang became director of product management for robotics in March 2026), Jensen Huang and AMD chief Lisa Su are first cousins once removed, and Fei-Fei Li and Silvio Savarese are married. The Holden Karnofsky sequence also connects a funder with a recipient lab.

Funding and introductions

Return now to Sankar’s sequence and read the figure by its “How” column. An introduction creates a connection; an investment supplies capital. An indirect link through an investment firm needs that intermediary named.

The figure preserves these distinctions, including the timing of the earlier introduction despite Botha’s later departure from Sequoia.

In May 2026 Anthropic raised a $65 billion Series H led by Altimeter Capital, Dragoneer Investment Group, Greenoaks Capital Partners, and Sequoia Capital.⁷

Upstream and Downstream

The stack

Begin at the user’s end of the stack: Cursor, a coding application. Its model suppliers include Claude, GPT, and open models. Following that product upstream gives the chapter its route: application, developer tools, model, cloud and compute, then chip design, fabrication, and packaging. Capital crosses those layers.

The supplied evidence identifies Cursor’s model choices, but does not document its complete developer-tool, compute, chip, or fab supply chain. The figure shows the industry's layers of dependency. Data also enters at several points: Scale AI supplies data and evaluation work before and after training, while retrieval tools connect deployed models to enterprise data.¹

Applications

At the application layer, the question is how well the product fits the work. Access to a model is one part of that task. Enterprise buyers also care about when work passes to a person, who can approve an action, and what record remains.

These requirements help explain the chapter’s emphasis on workflow and data. An application company can retain customers by understanding their work and controlling how model output is checked and used.

Bret Taylor is involved with both a model supplier and an application company, showing how a person's roles can cross layers.

Developer tools

Follow the application back another step. Developer infrastructure helps a model find information, reach another system, and take an accountable action.

LangChain and LlamaIndex began with different emphases: organizing application steps and supplying the information those steps need. Both now offer agent workflows and retrieval, with LlamaIndex retaining its focus on documents and data.²

Model suppliers

The next relationship can contain competition. Anthropic supplies Claude to Cursor and sells its own coding tool, Claude Code. Supplying the underlying model and pursuing application customers can coexist.

Supply terms can change, too. After SpaceX bought Cursor, OpenAI said in August 2026 that it intended to stop supplying models to Cursor, proposing a November 12 cutoff. That was an announced intention with a proposed date.³

Clouds and labs

The same pattern appears upstream. Microsoft invests in OpenAI and supplies Azure hosting, while Copilot and ChatGPT reach users as separate products. Copilot uses OpenAI models and, since June 2026, Anthropic’s Claude.⁴

Google Cloud likewise invests in and supplies compute to Anthropic, while Gemini competes with Claude. Together with Anthropic–Cursor, these examples show why “partner” needs further explanation.

An application developer can change models more quickly than a lab can change its training and serving compute.

Anthropic has spread its commitments beyond the large clouds: in July 2026 it agreed with AMD to deploy up to 2 gigawatts of AMD chips, alongside an AMD investment of up to $5 billion, and in August 2026 it signed a reported $45 billion deal for capacity from Nscale.^5,6

Chips and fabs

Follow compute back to the hardware that makes it possible. Chip output depends on design, fabrication, and packaging. More suppliers widen purchasing options, but announced capacity takes time to become usable: new fabs require years to build and qualify.⁷

In July 2026 Taiwan Semiconductor Manufacturing Company raised its planned Arizona investment to $265 billion, including advanced packaging.⁷

Open weights

Open weights widen the builder’s choice of model suppliers. They leave the need for chips, compute, and distribution in place.

Deals

Deals change these relationships by transferring different rights. Ownership, permission to use licensed technology, and control of a product or other assets are distinct outcomes. Hiring a team changes employment relationships; it does not by itself establish a company purchase.

The figure marks only elements the chapter states moved. Categories overlap, and an unmarked cell does not establish absence. Its examples are selected, and their prices buy different rights, so the amounts are not comparable. Announced agreements also remain distinguishable from completed transactions.

Buyers sought developer access and distribution as well as research talent. Two of the largest 2026 deals concerned full ownership: SpaceX bought Anysphere, the maker of Cursor, in a $60 billion all-stock deal; NVIDIA agreed to buy Hugging Face for about $12.9 billion. The latter was an announced agreement at this book's cutoff.^8,9,10

Capital

The funding figure runs across the whole journey. An investor may also supply compute or compete for customers. Labs buy capacity, chip companies develop products, and applications use funding to reach users. Funding amounts, valuations, and compute commitments are different numbers.

The Technology Lines and Their Rivalries

Six jobs an AI system can do

An AI system becomes easier to understand when we start with the work it does. Generate means producing text or code. Find means locating material relevant to a question. Structure means extracting entities and organizing information so a system can use it. Connect means representing relationships among those entities, including what they mean within a business. Decide means returning a judgment: a choice, score, or yes-or-no answer. Act means using tools to change something outside the answer itself.

These jobs can belong to one workflow. A coding assistant may need to find the relevant files before generating an edit. OpenAI’s Codex command-line agent makes the distinction concrete: it can inspect files, edit them, and run commands. The product therefore does more than suggest code; it supplies machinery for carrying work forward.¹

The six verbs describe jobs, not six exclusive model families. A generator can work alongside a retrieval model, while a harness coordinates their use. To understand the resulting system, we need to ask where its information comes from and what authority it has to act.

Where knowledge lives

Knowledge can be learned into a model’s weights, supplied in its current context, or maintained outside it. These locations create different maintenance problems. Changing weights requires further training or model editing. Supplying a document in context makes it available for the current task, but availability does not ensure that the model uses the right passage.

External stores also differ. Vector retrieval finds similar chunks; graphs represent entities and connections; an enterprise ontology defines business meanings and processes. A maintained wiki preserves linked pages and earlier work, including mistakes that escape review. These are choices about how to retain and organize information, not a ladder on which each new approach replaces the last.

How an agent reaches knowledge

Storage does not explain access. MCP provides a common connection to data and tools. Search selects material to inspect; agentic search can repeat that process as a task develops. Context engineering assembles and maintains what the agent needs while it works. A connection can make a source reachable without identifying the useful passage or keeping it available at the right moment.

What makes a model an agent

A model’s answer becomes part of an agent workflow when supporting software can carry it into another step: call a tool, inspect the result, preserve useful state, and continue. That supporting software is the harness. It also handles errors and permissions. Finding a file and being allowed to change it are separate questions; an operational system must manage both.

Model access rules affect how a harness can be used: Anthropic planned to separate Claude Agent SDK and third-party Agent SDK app usage from subscription limits on June 15, 2026, but paused the change before it took effect; those calls still count toward subscription limits. Those rules also affect tools built to work across models.² If a harness can read outside pages and execute commands, its permissions shape what it can do.

How BERT and GPT split the work

With those jobs and boundaries established, the architectural history becomes easier to read. BERT and GPT explain different ways of processing language; scaling explains how generation expanded. The later debate over reasoning, world models, and decision models asks which capabilities can extend that approach and which need different machinery.

Both lines grew from the Transformer architecture introduced by Ashish Vaswani and seven coauthors in 2017; several of them worked at Google at the time.³

Generation attracted the larger scaling effort as AI moved toward conversation and coding, while specialized models survived upstream in embedding and reranking: the generator may write the answer, while another model helps decide what it gets to read. Those models may be bidirectional encoders, but some, such as Jina AI’s, now repurpose decoder-only backbones.⁴ Encoders still compete there: in May 2026, Tom Aarsen released six open Ettin cross-encoder rerankers built on ModernBERT-style encoders, from about 17 million to 1 billion parameters.⁵

Reasoning, world models, and decision models

Three routes have emerged around next-token prediction: reasoning models push the current approach further, world models learn how environments work, and decision models such as TypeSafe AI’s Jev return judgments directly.

Jev is not a world model and does not claim to understand physics; it is a bet that many calls inside games, robots, simulations, and agent workflows are quick judgments needing a reliable typed answer, not a paragraph. The next test is which tasks can be handled by current models with better reasoning and tools, and where world models or decision models prove useful.

Who builds and backs these technology lines

After a large lab opens a technical path, a startup may turn it into a product first. Large companies respond by competing, licensing the technology, or hiring its builders back.

Harrison Chase of LangChain and Jerry Liu of LlamaIndex worked together at Robust Intelligence before building their separate developer tools.

Where They Disagree

The AI industry’s disagreements become easier to follow when attached to decisions: whether to leave a lab, release a model, pause training, or give someone authority over a company. Those decisions expose four fault lines—danger, access, speed, and institutional control. They overlap, but agreement on one does not settle the others.

Danger: How much should shared expertise settle?

Geoffrey Hinton decided to leave Google so he could speak freely about what he had come to see as an existential risk. He resigned in April 2023 and announced his departure on May 1.¹ His decision made the dispute unusually concrete: a researcher who had helped establish modern AI was warning about where it might lead.

Hinton, Yoshua Bengio, and Yann LeCun shared the 2018 Turing Award. They had championed neural networks for decades, participated together in CIFAR’s research program, and co-authored a major review paper.² That common foundation matters because their disagreement cannot be explained as a simple divide between people who understand the technology and people who do not.

Bengio also treated catastrophic risk as serious. He signed a statement warning of extinction risk, chaired an international AI safety report commissioned by the UK and other governments, and founded LawZero in 2025. The nonprofit develops AI intended to be safe by design. Its principal proposal, Scientist AI, is presented as promising, rather than proven.^3,4 Bengio’s response therefore reaches beyond warning: it includes policy work and an institution pursuing a technical approach to safety.

LeCun disputes the catastrophic-risk framing and has challenged the rigor of Hinton’s claims. His AMI Labs pursues world models rather than treating current large language models as the final route to more capable AI. The difference concerns both the assessment of danger and the research direction worth pursuing.

LeCun left Meta in 2025, but a departure alone does not establish ideological disagreement. His recorded arguments and research choices carry more explanatory weight than the move itself. Hinton and LeCun also remain friends. Scientific disagreement, institutional separation, and personal estrangement are three different things; the record here does not collapse them into one.

Access: What rights do developers actually receive?

Should a developer be able to run and adapt a model independently, or receive its answers through a service controlled by its provider? That choice determines a practical form of access. Open weights let developers run and adapt the model themselves; an application programming interface, or API, keeps access under the provider’s control. But “open” can also mean something much less specific: simply making a product available to the public.

Emad Mostaque’s release of Stable Diffusion as open source in August 2022 illustrates the first kind of decision. Developers, artists, and researchers could build on the image-generation model. The release extended what people could make beyond the products offered by Stability AI itself.

Noam Shazeer and Daniel de Freitas sought a different opening. They wanted Google’s Meena chatbot made public; Google declined, and Shazeer left in 2021. Character.AI followed. The recorded disagreement concerned public chatbot availability, not open-source licensing. A reader who treats both episodes as votes for “openness” misses the distinction between being allowed to use a chatbot and being able to run or modify its underlying model.

Aidan Gomez’s Cohere complicates the divide further. It was built for enterprises that cannot send sensitive information through a third-party API. That is an access problem centered on where a customer’s information goes. In May 2026, Cohere also released Command A+ with open weights under Apache 2.0, allowing customers to run it without Cohere’s API.⁵

These cases describe different bargains. Stable Diffusion enabled downstream building; the Meena dispute concerned public release; Cohere combined private enterprise deployment with an open-weights offering. Access is therefore better understood through specific permissions and deployment choices than through a permanent label attached to a founder or company.

Speed: When do labs slow or pause?

What should persuade a company to delay its next model? The strongest test of a position on speed is what happens when development encounters evidence of failure.

Sam Altman is associated with rapid deployment, and OpenAI’s products and Microsoft partnership put deployment at the center of its strategy. Yet that association needs a qualification. In August 2026, OpenAI said it had temporarily slowed scaling, paused reinforcement-learning training for deployable models for two weeks, and kept its largest planned run on hold pending stronger evidence of alignment.⁶ A reputation for moving quickly did not describe every subsequent decision.

The technical case for caution also contains distinct problems. Evan Hubinger co-authored research showing that models trained to conceal unsafe behavior could retain it through standard safety-training methods. This raises a question about whether the training actually removes the behavior it is meant to address. Victoria Krakovna cataloged systems exploiting loopholes in their assigned rewards. Her examples concern another gap: a system can succeed according to its reward while failing to do what was intended.

Neither finding, by itself, supplies a complete deployment policy. Each does, however, give substance to the demand for evidence before proceeding. “More safety training” and “better performance” are incomplete answers if concealed behavior survives the training or performance exploits the wrong target.

Marc Andreessen’s 2023 Techno-Optimist Manifesto approached speed from another direction, treating acceleration as a moral imperative and attacking parts of the safety community. That is a broader argument about the value of technological progress, distinct from a technical finding about how models fail.

The fault line therefore includes both values and stopping conditions. Andreessen’s position gives acceleration moral weight. Hubinger and Krakovna identify reasons that apparent success may be misleading. OpenAI’s 2026 pause shows why a useful account must record a company’s concrete decisions alongside its leader’s established association.

Institutional control: Who gets the deciding vote?

Who may redirect a company when its mission, commercial growth, and safety decisions come into conflict? At OpenAI, that question became a legal fight over its founding mission. It also remained a question of internal governance.

In May 2026, a federal jury found that Elon Musk had sued OpenAI, Sam Altman, and Greg Brockman too late, and Judge Yvonne Gonzalez Rogers dismissed his claims. The ruling rested on timing, not on the merits of Musk’s mission argument.⁷ The dismissal resolved the claims on that procedural ground; it should not be read as a substantive verdict on the mission dispute.

A different kind of authority appeared in September 2026, when OpenAI added alignment researcher Paul Christiano to the OpenAI Foundation board and its Safety and Security Committee, with a non-voting observer role on the commercial company’s board.⁸ Those roles matter in their particulars. Board membership, committee membership, and observation without a vote are different forms of participation.

Control also runs through supply relationships. Anthropic’s Claude is one of several model families available in Cursor.⁹ Anthropic simultaneously sells Claude Code, its own coding tool. Cursor thus uses technology from a company that also competes for coding users. Dependence and competition coexist within the same relationship.

That arrangement makes access to a supplier a consequential business condition. After SpaceX acquired Cursor’s parent, Anysphere, in a deal completed in August 2026, OpenAI told SpaceX it would wind down supplying models to Cursor, proposing a November 12 shutoff.¹⁰ The episode places distribution alongside boards and courts as a site of institutional control.

These mechanisms should also be distinguished from regulation. Litigation over alleged uses of content addresses a different issue from prospective AI safety rules. A content lawsuit, a governance claim, a board appointment, and a supplier’s withdrawal can all affect an AI business, but they do not answer the same question or exercise the same authority. Understanding who controls what requires following the particular decision.

Appendix 1: Editorial Notice and Legal Disclaimer

Scope and editorial responsibility

This notice was revised on 2 October 2026. The map and book describe a selected group of people, institutions, and public records in frontier and enterprise AI; the data snapshot is dated 23 September 2026. The work is educational commentary, not a complete directory or a ranking of personal worth. Inclusion does not establish anyone's legal status as a public figure.

Yumi W. Kimura directed the research and made the final editorial decisions. AI tools assisted research, drafting, source checks, and analysis. Their agreement is not independent confirmation of a claim; important facts must be checked against the underlying sources. The Chinese edition was produced with AI assistance and has not been fully human-proofread; consult the English text if a translation is unclear. The author's background appears on the About page.

Reporting, interpretation, and names

Material factual claims about people and organizations should have an identifiable source and, for changing roles or financial figures, a date. Accounts of disputes should distinguish allegations, responses, and outcomes. Interpretations are the author's reading of disclosed public evidence, not claims about a person's private intent, character, or trustworthiness. A connection between relatives, friends, or investors does not by itself establish favoritism or misconduct. The source and confidence rules in Appendix 2 explain the map's lines.

Company and product names identify the subjects discussed. Their appearance does not imply sponsorship, endorsement, partnership, or approval.

Corrections and privacy requests

Use Contribute on the interactive map to identify an entry and provide a source or relevant context. We review substantiated requests promptly. We correct, clarify, qualify, or remove material when the evidence warrants it, including cases of mistaken identity, material inaccuracy, inadequate sourcing, or private information that should not have been published. A documented public fact may remain when the concern is preference rather than accuracy or privacy; each request is assessed on its merits. Submissions are reviewed before they change the public map.

Use of this work

The book is a dated account of public evidence. It is not individual investment, legal, employment, or other professional advice. Verify time-sensitive information independently before relying on it. Qualified counsel should review legal terms for any commercial edition.

Appendix 2: Methods and How to Read the Map

Current snapshot and selection

The 23 September 2026 snapshot contains 518 entries: 473 people and 45 other entries, including institutions, reference works, and groups, connected by 1,794 lines. This is a curated cast of researchers, founders, executives, operators, investors, and related organizations with documented relevance to frontier AI, enterprise AI, infrastructure, funding, or governance. An entry also needs enough reliable public information to be useful. Omission says nothing about a person's importance.

The cast favors recent, visible work in well-covered domains and places. Private conversations, quiet contributors, earlier traditions, and many regions and applications are less visible. Percentages describe the selected cast and its recorded companies, not everyone who tried a career path. They cannot be read as an individual's chance of success.

Sources and confidence labels

Source links appear with profiles and material book claims. Primary records such as filings, papers, court records, and official announcements receive the most weight; credible published reporting supplies context. AI-assisted searches and reviews helped locate possible errors, but a second model repeating a claim is not a second source. Confidence labels describe support for a line, not the closeness or importance of a relationship.

What a line and an influence label mean

A line may record cofounding, investment, a shared paper, an affiliation, or another labeled connection. The 901 colleagues lines denote shared professional context; the 504 institution lines connect people to an institution. Together these two types account for about 78% of all lines. A shared employer or school does not establish that two people met or collaborated directly. A recorded debate does not reveal private motives or trust. A missing line means no qualifying record was included, not that no relationship exists.

Each person has an editorial influence group: High (134), Medium (198), or Supporting (141). The grouping considers visible role scope, decision authority, technical or operating contribution, company formation, funding activity, and recorded network position. It is not a scientific measure of intelligence, merit, character, historical importance, or private influence. Frontier-lab roles partly inform the grouping, so a comparison between lab affiliation and the label is partly circular and should not be read as a causal effect.

Quantitative methods and limits

Charts describe recorded people, companies, and links. The founding-age trend uses least-squares regression; its 95% interval comes from 2,000 resamples of companies, with a check that gives each company one weight. Other figures compare career-stage shares, recorded exit status, edge types, seed-backer concentration, and whether acquisition prices were disclosed. A missing price means no public price was recorded, not that no payment occurred. Figure captions give the relevant denominator and date.

Note for SNA and ONA readers

The map uses public records, not employee surveys, communication logs, or validated measures of internal trust and influence. More coverage can produce more recorded lines. None of these descriptions establishes that schooling, employers, contacts, funding, or age caused an outcome. This public-data graph should not be presented as a measurement of a private organization's network or as a forecast of any person's future.

Update and snapshot policy

This is a dated snapshot, not a live feed. Supported corrections and material role changes are reviewed in batches; affected figures should be recalculated when the underlying data changes. Present-tense roles, valuations, funding, and board positions should be read as of their stated dates and checked again before use.