THE SEED
David’s observation, rebuilt: the entire generative-AI economy grew from a countable handful of freely-posted artifacts — AlexNet (2012), the transformer (2017), GPT-2, CLIP, Whisper, tiktoken, the RLHF data and the constitution (2022). Most carry an MIT or near-free license; most were written by teams of eight to fifty people you can name. The industry is the derivative; the repos are the principal. Here is the seed — what each one was, who posted it, where, why — and where the sowers are now, the philosopher among them.
The Four Natures
the deep seed, the frame, the values turn, the engines
The Thesis
a dozen free commits · the 2017 pivot · the asymmetry
The whole generative-AI economy traces to a countable handful of public artifacts — AlexNet (2012), the transformer (2017), GPT-2 (2019), CLIP & Whisper & tiktoken, the RLHF data and the constitution (2022). Most carry an MIT or near-free license. Most were written by teams of eight to fifty people you can name. The industry is the derivative; the repos are the principal.
Attention Is All You Need (Vaswani et al., Google, 2017) replaced recurrence with attention — and every large model since is a descendant. Eight authors signed it. All eight later left Google; they now run or seeded Cohere, Character.AI, Sakana, Essential AI, Inceptive, NEAR, and lines at OpenAI. The single most consequential git artifact of the era, and its authors scattered.
The input was open, cheap, authored — a git push, an arXiv link, an MIT license. The output was a trillion-dollar industry (NVIDIA alone passed ~$3–4T; the buildout runs to trillions in capex). The value accrued to capital and compute, not to the committers. Vaswani got a citation. Bai got a byline. Nobody on the author lists got a trillion dollars.
The Philosopher
really focusing here — the one who seeded the conscience, not the capability
The seed isn’t only engineers. Amanda Askell is a philosopher — PhD, NYU (2018, Pareto Principles in Infinite Ethics; advisors David Chalmers, Cian Dorr & Shelly Kagan; BPhil, Oxford) — and she sits at the values end of the whole story. She first-authored the 2021 paper that named the frame the field still optimizes toward — helpful · honest · harmless — and she has headed Claude’s character / personality-alignment team since 2021.
In January 2026 she became the primary author of Claude’s Constitution — a ~30,000-word published document (CC0) describing what kind of mind Claude should be. Her job, in her own framing, is to decide how it reasons about right and wrong, how it handles emotional situations, and what personality it presents to the hundreds of millions of people who talk to it. Time 100 AI (2024).
If the engineers seeded the capability, Askell seeded the conscience. The whole industry can compute; she is one of the few people whose job is to decide what it should be — the philosopher’s question, shipped to a planet. Where now: Anthropic, San Francisco — still the one writing the soul.
The Seed — The Born
the artifacts the industry grew from, as ACI .agents (12); the people are credited (cited, not minted) in ‘Where They Are Now’
The Arc
the spark → the pivot → the scaling & the soul → the scattering
AlexNet (Krizhevsky, Sutskever & Hinton, Toronto) crushes ImageNet with a deep net on two GPUs — the moment deep learning becomes inevitable. A couple of files and a paper. Ilya Sutskever is on it; he becomes the through-line of the whole story.
Attention Is All You Need (eight authors, Google). Attention replaces recurrence; the transformer is born. It is the hinge the entire industry turns on — and within a few years all eight authors have left to found or lead their own labs.
Alec Radford leads GPT-2, CLIP, and Whisper at OpenAI — the open engines (‘the genius behind the models’, and he has no PhD). In parallel at Anthropic, Yuntao Bai builds the RLHF/Constitutional-AI usability layer and Amanda Askell writes the values frame. Capability meets conscience; the products land.
The sowers leave. Sutskever → Safe Superintelligence. Radford → independent research. Karpathy → Eureka Labs. The transformer’s eight → Cohere, Sakana, Essential AI, Inceptive, NEAR, Google. And the industry they seeded for free passes into the trillions — the asymmetry, complete.
The Record
the seed commits, and where the sowers are now — cited, public record
The Seed Commits — what · who · where · why · license
the countable handful, dated & attributed
- AlexNet2012 · Toronto · Krizhevsky, Sutskever, Hintonthe spark — deep CNN wins ImageNet on GPUs; deep learning becomes inevitable
- word2vec2013 · Google · Mikolov et al.dense word embeddings — meaning as vectors; the first taste of learned semantics
- Attention Is All You Need2017 · Google · arXiv:1706.03762 · 8 authorsTHE PIVOT — attention replaces recurrence; the transformer, ancestor of every LLM
- gym2016 · OpenAI · custom licensethe RL toolkit that trained a generation of researchers (now Gymnasium)
- GPT-22019 · OpenAI · Radford · custom licensescale + unsupervised pretraining → general capability; ‘too dangerous,’ then released
- CLIP2020 · OpenAI · Radford · MITtext↔image contrastive learning — open-vocabulary vision; the most-built-on vision repo
- Whisper2022 · OpenAI · Radford · MITrobust speech recognition from weak supervision at scale — OpenAI’s most-starred repo (★100k+)
- tiktoken2022 · OpenAI · MITeven the tokenizer is a free repo — the BPE encoder the whole API runs on
- HH-RLHF2022 · Anthropic · Bai et al. · MITthe human-preference data behind RLHF — folded into the-language-of-the-machine
- Constitutional AI2022 · Anthropic · Bai, Askell, Olah · paperharmlessness from AI feedback + a written constitution — folded into constitutional-ai
Where They Are Now — the transformer's eight (all left Google)
the most consequential byline of the era, scattered · public record, 2026
- Ashish VaswaniCEO, Essential AIlead author; earlier co-founded Adept
- Noam ShazeerVP & Gemini co-lead, Googleleft for Character.AI; Google paid a reported ~$2.7B to bring him back (2024)
- Niki Parmarco-founder, Essential AIearlier co-founded Adept (CTO)
- Jakob UszkoreitCEO & co-founder, Inceptivenow applies the transformer to RNA / biology
- Llion Jonesco-founder, Sakana AI (Japan)nature-inspired model architectures
- Aidan GomezCEO & co-founder, Cohereenterprise LLMs; ~$6.8B valuation (2025)
- Łukasz Kaiserresearcher, OpenAIon the GPT reasoning line
- Illia Polosukhinco-founder, NEAR Protocolleft AI for blockchain
Where They Are Now — OpenAI & Anthropic, the sowers
the engine-builders and the seven who split off · public record, 2026
- Ilya Sutskeverfounder & CEO, Safe Superintelligence (SSI)AlexNet 2012 → OpenAI chief scientist → left 2024; the through-line
- Alec Radfordindependent researcher; adviser, Thinking Machines Lablead on GPT/CLIP/Whisper; left OpenAI Dec 2024; no PhD
- Andrej Karpathyfounder, Eureka LabsOpenAI → Tesla → OpenAI → AI-education startup (2024)
- Greg Brockmanpresident, OpenAIco-founder; still there
- Dario Amodeico-founder & CEO, Anthropicled the 2020 split from OpenAI (the ‘seven’)
- Chris Olahco-founder, Anthropicinterpretability — the looking-in line
- Yuntao BaiAnthropiclead author, HH-RLHF & Constitutional AI
- Geoffrey HintonNobel laureate (Physics, 2024); left Google 2023AlexNet supervisor; left to warn about AI risk