core automation

a sourced history of core automation, inc., the san francisco ai lab founded by former openai vp of research jerry tworek, not the industrial-automation firms.

core automation is the san francisco lab jerry tworek founded after seven years running openai's reasoning program — the rubik's-cube hand, codex, o1, o3, gpt-5 — because the research he wanted to do was 'hard to do at openai.' it launched april 21, 2026 with a coordinated 'nerdsniped' cofounder reveal and a reported ~$100m safe round at ~$1b; six weeks later the information had it seeking $300m–$500m at ~$4b, and by july 30 a form d showed $635.2m authorized and $432.1m sold to 51 investors — all before shipping a product, because the product is the lab: 'the world's most automated ai lab,' betting that transformers can't learn continually and a ~20-person team running agents can find what replaces them. this is the record of that bet — the launch, the messy ledger, the competition, the internship, and the gaps.

• history: the company, sourced and dated — the openai years, the january exit, the april launch, the round that kept changing, and the lab in public.
  • the openai years: amsterdam quant finance, then seven years at openai — the rubik's-cube hand, codex, humaneval, and the q*/strawberry line that became o1, o3, and gpt-5.
  • the exit and the founding: the january 5 exit post, the core memory interview, the information naming the company, and a late-march incorporation.
  • the launch: the reported ~$100m safe round at ~$1b, the april 21 launch, and the coordinated 'nerdsniped' cofounder reveal.
  • the round that kept changing: the reported $300m–$500m seek at ~$4b, then the july 30 form d: $635.2m authorized, $432.1m sold to 51 investors, $135.3m of converted safes.
  • the lab in public: the blog corpus, the agi house keynote, the sequoia and mts episodes, one layer deeper, and the agm internship.
• people: tworek and the team he nerdsniped — who they are, what they wrote and said, and which identities stay flagged.
  • the founder: jerry tworek — quant trader, openai vp of research, and the self-authored record at millionintegrals.
  • the cofounders: anil, jang, villagra, gulati, lamp, saroufim, garg, patil — the launch-day reveal and the biographies behind it.
  • the bench: amid, duvvuri, rosenthal, the wider bench, and the inferred backers — the layer around the cofounders.
  • the writing: the self-authored corpus — the homepage, the launch posts, the blog, the reservoir samples.
  • interviews and talks: core memory, training data, the information bottleneck, mts, agi house — where the thesis got said out loud.
  • the open flags: the roster questions that stay open — self-applied titles, the patil name match, the adept attribution, the inferred cap table.
  • the philosophy: continual learning, evals are dead, small teams with capable agents, and the automated lab.
• side quests: the pursuits around the company — the founders' earlier work, the competition and the internship, and the neo-lab ecosystem.
  • the founders' earlier work: the quant years, the shampoo→psgd line, project mariner, openai labs, gpu mode, mosaicml composer, reservoir samples, fysk.
  • the company's quests: the blog corpus, one layer deeper, the agm program, and the pacing letter the lab's employees signed.
  • the ecosystem: tilde research, the alumni ventures feeder, and the neo-lab cohort the company gets grouped with.
• coverage: what the discourse made of core automation — the departure press, the talent-shock launch, the funding ledger, and the thin controversy file.
  • the departure: the verge, the decoder's 'not-so-subtle dig,' and the 'major loss for openai' reads of the january exit.
  • the launch: the nerdsniped coverage wave — bi's signature story, the neo-lab mapping, and a hacker news submission that sat at 1 point.
  • the funding: the $4b seek, the form d reads, and the tracker arithmetic — deal-flow astonishment plus the first skepticism.
  • the founder media: the podcast circuit and the showcase profile — the founders' own channels carrying most of the technical record.
  • the discourse: the skeptics, the deepest secondary analysis, and the one hn thread that actually debated the company.
  • controversies: the unreconciled ledger, the talent-raid frame, the vaporware file, the namesakes, the pacing wrinkle, and the self-reported layer.
  • sentiment eras: sentiment phases: departure intrigue → talent-shock launch → money velocity → substantive evaluation → operationalization.
• the idea: the automated lab as the product — the lineage core automation inherited, what it has and hasn't shown, and what agents do to the argument.
  • lineage: the reasoning program, the optimizer line, the product-and-people bench, the automated-science wave, and the continual-learning problem.
  • what it showed: talent gravity, capital before product, the competition as instrument, and the legible lab.
  • what stayed unproven: no product, continual learning unproven, self-reported automation claims, and the second-hand 'ceres.'
  • the agent-era arguments: the for/against/nuance ledger — contestable exploration, the flywheel, the bitter-lesson trap, the taste problem, the verification gap.
  • successors: openai's own automation program, the neolab cohort, sakana, tilde — competitors, cohort-mates, and the one collaborator.
  • open questions: the questions the next chapter has to answer.
• sources: every source cited in the core automation history, with its date and type.
  • the catalog: the full typed catalog — primary, interview, reporting, archive, community, reference, repo.
• method: how ai agents researched the core automation history, the evidence rules they followed, and what they could not verify.
  • the lanes: the two parallel research lanes and what the formula proved.
  • the evidence rules: the evidence model — confidence, precision, sentiment discipline.
  • known gaps: what the research could not verify — recorded, not hidden.
  • corrections: how to correct the record.

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

Canonical page: https://hraness.com/core-automation