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 years2015–2025
amsterdam quant finance, then seven years at openai — the rubik's-cube hand, codex and the first copilot model, and the q*/strawberry line that became o1, o3, and gpt-5's reasoning.
- 2015the amsterdam quant yearstworek spends about five years in amsterdam quant finance — linkedin lists him as senior quantitative researcher at independent view (jan 2015–dec 2018), 'most senior researcher in the company reporting directly to cio,' building fully systematic futures signals. earlier details are thinner: an msc in mathematics from the university of warsaw, and a dropout story he has told on podcasts.
- 2019-02tworek joins openaitworek joins openai — about thirty employees at the time — and stays nearly seven years, ending as vp of research.
- 2019-10the rubik's-cube hand'solving rubik's cube with a robot hand' — the sim2real breakthrough tworek ran, 'one of core breakthroughs' by his own account, landing at the neurips 2019 deep rl workshop.
- 2021codex, copilot, humanevalthe codex line: tworek 'conducted all stages of research and trained the model that powered github copilot,' was a primary contributor to codex, and hand-wrote about thirty humaneval problems — wojciech zaremba wrote thirty more.
- 2024-09-12o1o1 ships — the reasoning model from the effort known internally as q*, then strawberry. tworek: 'i've lead the team that researched training llms to reason with reinforcement learning.'
- 2025-04o3, and the vp-of-research seato3 ships. tworek's later summary of the o1→o3 gap: '99% systems, 1% algorithms.' he is vp of research by february 2025.
- 2025-08gpt-5gpt-5 ships with reasoning descended from the o3 line — tworek later describes it on the mad podcast as roughly 'o3.1.'
- 2025-08the bench starts leavingthe future cofounders disperse. julia villagra — promoted to openai chief people officer in march — exits in august, with reuters reporting she left to pursue 'art, music and storytelling to help people understand the transition to artificial general intelligence.' joanne jang founds openai labs in september and lands on the time100 ai list in august.
- 2025-10-16the last openai interviewtworek records 'how gpt-5 thinks' on the mad podcast while still vp of research — the last big interview of the openai years.
the exit and the founding2026-01 – 2026-03
tworek announces his exit, explains it on core memory, and the information names the new company: core automation, seeking up to $1b for models that keep learning after deployment.
- 2026-01-05tworek leaves openaitworek announces his openai departure on x after ~7 years, telling his team he wants 'to try and explore types of research that are hard to do at openai.' the verge and the decoder cover it the same day; the decoder reads the line as a 'not-so-subtle dig.'
- 2026-01-21the exit interviewcore memory ep53, recorded days after the announcement: high-risk pioneering research had become hard at a 'conservative' openai; big labs as a whole have gone conservative; deep-learning research at big labs 'is done.'
- 2026-01the information names the companythe information reports the new company — named core automation — is seeking $500m–$1b to build models needing ~100× less data that learn continuously; the reported long-term roadmap runs from industrial automation to 'self-replicating factories,' bio-machines, and 'planetary terraforming.' the piece is paywalled; the record runs through the decoder's summary.
- 2026-03incorporatedcore automation, inc. is founded — a delaware corporation incorporated in 2026 per its form d; 'late march' per dealroom. headquarters at 340 brannan street, suite 301, san francisco.
the launch2026-04
a ~$100m safe raise at ~$1b reportedly closes, and on april 21 the lab goes public — 'the world's most automated ai lab' — with a coordinated cofounder reveal.
- 2026-04the ~$100m safe roundthe initial ~$100m raise reportedly closes at ~$1b post-money — press and trackers name nvidia, spark capital, and accel, with scribble ventures and threshold ventures in secondary listings and no disclosed lead. the form d later shows the money came in on safes that converted into $135.3m of stock in the july round — the discrepancy between the reported ~$100m and the $135.3m conversion stays unresolved.
- 2026-04-21launch daypublic launch. the first x post: 'today we're announcing core automation… our objective: systems that optimize and automate work, starting with research itself' — 'the world's most automated ai lab.' tworek's own post: 'small step for a finger on "post tweet" button. big step for millions of future ai agents doing useful work for us.' cofounder avery lamp puts it on hacker news, where it gets 1 point and 0 comments.
- 2026-04-21the nerdsniped cascadethe coordinated reveal. rohan anil posts that he left anthropic 'a few weeks ago' — 'jerry tworek nerdsniped me into starting this with him and others.' anmol gulati announces from google deepmind: 'the current research paradigm — scaling models, data, and static deployment won't get us all the way.' joanne jang's x bio flips to 'trying to automate myself @coreautoai.' villagra, lamp, amid, and duvvuri are named in the coverage.
- 2026-04-22the launch coveragethe launch story travels: business insider's 'nerdsniped' piece becomes the most-syndicated account, the decoder files its launch writeup placing the lab in the neo-lab cohort with thinking machines and ssi, and korea's mk frames it as 'openai mafia' news while noting no business model has been disclosed.
the round that kept changing2026-05 – 2026-08
the funding stays messy in public: a reported $300m–$500m seek at ~$4b, then a july 30 form d showing $635.2m authorized and $432.1m sold to 51 investors — including $135.3m of converted safes.
- 2026-05-07$1b to $4b in six weeksthe information reports the ~6-week-old company is already seeking $300m–$500m at ~$4b. palazzolo on x: 'back to back rounds are continuing in neolab land!' the report is paywalled — the record runs through techmeme, dealroom, and testingcatalog.
- 2026-07-15the equity round opensfirst sale of the new equity offering, per the form d filed two weeks later.
- 2026-07-30the form dthe form d lands: a $635,249,754 authorized offering with $432,125,053 sold to 51 investors and $203,124,701 remaining — including $135,252,607 issued on conversion of the outstanding safes. tracker math (forge/pitchbook-derived) puts the round at ~$3.2b pre / ~$3.62b post — reported, never confirmed by the company or the filing.
- 2026-08-13axios reads the filingaxios pro reads the filing as the company 'looking to raise up to $500 million in new funding' — consistent with the remaining capacity plus new cash — and describes core automation as 'an ai continual learning model developer.'
- 2026-08-17the alumni ventures feedera second filing surfaces: 'avsf – core automation 2026, llc,' an alumni ventures feeder spv — $2,109,250 sold to 100 investors, first sale august 3.
the lab in public2026-05 – 2026-09
the lab in public: the blog corpus, the agi house keynote, the sequoia and mts episodes, the one layer deeper competition with tilde research, and the agm internship.
- 2026-05-27the blog opensthe first company blog posts appear — saroufim's 'to automate publishing, first make authoring boring' and 'when ai starts writing systems code,' arguing that automating research means automating systems code first.
- 2026-07-18the agi house keynotetworek keynotes the agi house auto-research summit in hillsborough — 'the era of evals is dumb,' 'the only test is the real world' — and puts up an open-problems slide that eighteen teams build on for the rest of the day.
- 2026-07-28the pacing letterthe 'pacing the frontier' open letter goes public — 1,386 frontier-ai employees asking for 'technical and governance tools needed to deliberately pace the frontier of automated ai development.' core automation employees are among the signatories per the hn discussion — the lab that automates research, signing the letter about pacing it.
- 2026-07-29the sequoia episodesequoia's training data episode airs — 'building the automated agi lab: core automation's jerry tworek and rohan anil' — the fullest public statement of the thesis: transformers 'incapable of continual learning,' the search for a replacement architecture, and the most automated lab as the instrument. sequoia hosting is not evidence sequoia invested — the round's participants remain undisclosed.
- 2026-08-03one layer deeper'one layer deeper' launches with tilde research — a public competition on repeated modular squaring: learn the computation, then run longer chains than seen in training. the hard track gives one h100, one hour of training, and asks models to certify depth t=1…64.
- 2026-08-25twenty-something peoplethe fysk showcase primer reports the team at '20-something people' — drawn largely from openai and google deepmind plus anthropic, meta, xai, and amazon's agi lab — with a plan to reach ~40 by year-end. linkedin-derived counts say ~21; one aggregator says 14.
- 2026-08-26the mts episode'ai labs are simply cooked' — the mts episode: three generations of ai labs, a two-year horizon for agents automating research, and the concession that today's agents 'lack deep field understanding and produce low-quality, high-creativity ideas.'
- 2026-09-02the architecture-discovery postehsan amid's post on neural architecture discovery — untrained networks already separate cats from dogs; automate the discovery of architectures — and on automation reshaping the field.
- 2026-09-09the competition results'what the models learned in one layer deeper' — the results: 15,602 submissions; only 18 hard uploads from 5 accounts certified depth (16 reached t=64). the winners learned 'wiring around python primitives' — evidence that test-time depth is genuinely hard for current methods.
- 2026-09-10anil on the bottleneckrohan anil on the information bottleneck — 'continual learning is the next bottleneck': the pre-training/rl split as 'organizational convenience rather than science,' shampoo's origin, the qr-factorization kernel competition (~60× over cusolver), and how he met tworek at dolores park.
- 2026-09the agm programthe agm — 'associate general manager' — internship goes live: four interns, january–march 2027 in sf, $10,000/month plus profit share, each given a budget, tokens, and the company's agent infrastructure to run real online businesses. 'by watching where people intervene, we want to understand what agents still need to learn.' applications close october 6, 2026.
- 2026-09-16the showcasetworek appears at the fysk sf startup showcase — the lab's fifth month ends the way it started: the founders in public.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.