people
tworek and the team he nerdsniped — who they are, what they wrote and said, and which identities stay flagged.
the founder
- jerry tworek · ceo · @millionintcofounder and ceo — jarosław tworek, executive officer and director on the form d. msc mathematics, university of warsaw; ~5 years in amsterdam quant finance (linkedin: senior quantitative researcher at independent view, jan 2015–dec 2018) after a dropout story he has told on podcasts. openai 2019–jan 2026, ending as vp of research: the rubik's-cube robot hand, the original github copilot model, codex and humaneval, chatgpt plugins and code interpreter, then the q*/strawberry effort that produced o1, o3, and gpt-5's reasoning. self-described 'very cautious ai optimist' at millionintegrals.com.
the cofounders
the launch-day reveal — eight cofounders, most of them announced in their own words. titles are self-applied and inconsistent; the form d is the hard record.
- rohan anil · co-foundercofounder — director on the form d. ~11.5 years at google: systems engineering, then google brain optimization; co-created the shampoo optimizer with vineet gupta and tomer koren, led optimization for palm/gemini pre-training, one of four gemini pre-training leads, distinguished engineer at google deepmind. anthropic pre-training team ~jun 2025 to ~mar/apr 2026 — 'one of the best places to work for a researcher,' until 'jerry tworek nerdsniped me into starting this with him and others.' met tworek at dolores park. ms cs uc san diego; bits pilani.
- joanne jang · co-foundercofounder and mts. stanford bs applied math + ms cs; natural-language work on google assistant, enterprise features at dropbox. openai dec 2021–apr 2026 as gm: product lead on dall·e 2, gpt-4, the chatgpt api, text-to-speech, and embeddings; founding lead of model behavior and steward of the model spec; founded openai labs in september 2025 to prototype human-ai interfaces, reportedly adjacent to the jony ive effort; time100 ai in august 2025. launch-day x bio: 'trying to automate myself @coreautoai.' writes reservoir samples, her own publication.
- julia villagra · co-foundercofounder per dealroom — executive officer on the form d. head of people at hudson river trading, then openai feb 2024–oct 2025: head of hr to vp people to chief people officer. reuters reported she left 'to pursue her personal passion of using art, music and storytelling to help people understand the transition to artificial general intelligence' — how that reconciles with an operating role at a research lab is unreported. her linkedin title at core automation is 'member of the xyz staff' — the company uses playful titles.
- anmol gulati · co-foundercofounder and mts. google deepmind research scientist — tech lead on agents research and computer use (project mariner) in gemini; credits listed: gemini 3.0 multimodal and computer-use, gemini 2.5 code pre-training, gemini 2.0 text representations. earlier senior research engineer at google brain — co-lead of production speech recognition, core team on babelfish/lingvo, one of five recipients of the 2021 brain product & engineering excellence award. one aggregator bio calls him an adept ai cofounder — unverified and likely a conflation; the adept alumnus in the talent list is more plausibly avery lamp.
- avery lamp · co-founder · @averylampcofounder, mts — and the person who submitted the launch to hacker news. software research engineer at google deepmind; earlier mts at adept ai labs and early engineer at mosaicml, where his resume credits the composer training framework and the yahp hyperparameter system; a stealth cofounder in between; google swe intern in 2018. described by fysk as a longtime engineer and lp in the founders you should know community.
- mark saroufim · co-foundercofounder — his own site reads 'cofounder at core automation, pytorch maintainer and cofounder of gpu mode.' path: microsoft to graphcore to meta's pytorch team (torchserve, pytorch/data, benchmarks, ao quantization) to gpu mode, the kernel-engineering community turned company. authors core's systems-code blog posts and led the one layer deeper results writeup.
- kanav garg · co-foundercofounder per linkedin. ex-google deepmind; worked on project mariner — he shared the december 2024 launch calling it 'one of the most stimulating 0->1 projects.' further public background is thin.
- swapnil patil · co-foundercofounder — self-titled 'member of agentic staff.' the best public match is an ex-google software engineer (~2012–2026), cmu phd, ex-microsoft research intern, working on scaled ml systems: gemini/llm serving on gpus, rl performance, large-scale pre-training on google's custom rdma gpu clusters, tpu runtime. the identification rests on a linkedin name match — a name collision is possible, though the systems profile fits the lab's needs; confidence moderate.
the bench
- ehsan amid · teammts since april 2026. ex-google brain/deepmind research scientist (oct 2021–2026), 'core member of the gemini team'; uc santa cruz phd 2020 under manfred warmuth, msc aalto. ml theory, robust learning, optimization — and the author of the september 2 post on neural architecture discovery and automation.
- sai surya duvvuri · teammts — former research intern at google and meta, named in the launch-day coverage.
- aliisa rosenthal · teamdirector on the form d — almost certainly investor-side. openai's first commercial hire and head of sales, june 2022–2025, credited with scaling revenue roughly $10m to $10b and the team from 2 to 300+. joined acrew capital as general partner in january 2026; her board seat suggests an acrew position in the round — inferred, not disclosed.
- the wider bench · teamthe named-but-thin layer: 'riley' (riley@coreauto.com) runs the agm program contact, and linkedin's talent-source stats show hires from humans& (the anthropic-alumni lab), ami — lecun's lab — meta, ut austin, the european commission, and adept. headcount is '20-something' per fysk's august interview, ~21 on linkedin, 14 on one aggregator — with a self-reported plan to reach ~40 by year-end 2026.
- the backers · teamnvidia, spark capital, and accel are the names attached to the reported ~$100m safe round at ~$1b; scribble ventures and threshold ventures appear in secondary trackers; no lead is disclosed for either round. the july form d counts 51 investors but names none of them. sequoia's involvement is inferred from hosting the founders on training data; acrew's is inferred from rosenthal's board seat; alumni ventures is the only investor with its own filing — a $2.1m feeder spv. reported, not confirmed.
the writing
the self-authored record — tworek's millionintegrals deposition, the launch posts, saroufim's systems-code posts, amid's architecture essay, jang's reservoir samples. the lab writes in its own voice.
- 2026million integrals · document, millionintegrals.com · jerry tworekthe self-authored work history — copilot, codex, humaneval, plugins, o1, o3, gpt-5, the chatgpt agent — written in first person, including the typos.
i've conducted all stages of research and trained the model that powered github copilot
i've lead the team that researched training llms to reason with reinforcement learning
- 2026-01-05the exit post · tweet, x · jerry tworekthe exit announcement on x — nearly seven years at openai, leaving for research that's hard to do there. the decoder read it as a 'not-so-subtle dig.'
to try and explore types of research that are hard to do at openai
covered same-day by the verge, the decoder, and the indian press - 2026-04-21the launch posts · tweet, x · jerry tworekthe launch-day pair: the company's 'systems that optimize and automate work, starting with research itself' alongside 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.
- 2026-04-21the cofounder's launch post · tweet, x · anmol gulatigulati's launch-day post — the thesis stated by the cofounder, not the ceo: 'new learning algorithms, architectures beyond today's stack, and systems that automate the process of building itself.'
i've increasingly felt that the current research paradigm — scaling models, data, and static deployment won't get us all the way
- 2026reservoir samples and the labs announcement · essay, joannejang.com · joanne jangjang's own writing — the reservoir samples essays on model personality and human-ai relationships, plus the september 2025 openai labs announcement: 'patterns that move us beyond chat or even agents.'
changing model behavior is an art and a science
- 2026-05-27to automate publishing, first make authoring boring · post, core automation blog · mark saroufimthe first company blog post — 'to automate publishing, first make authoring boring': the lab's own publishing pipeline as a systems warm-up.
- 2026-05-28when ai starts writing systems code · post, core automation blog · mark saroufim'when ai starts writing systems code' — the thesis that automating research means automating the systems layer first, from the pytorch maintainer who cofounded gpu mode.
- 2026-09-09what the models learned in one layer deeper · post, core automation blog · mark saroufim'what the models learned in one layer deeper' — saroufim with mcleish, keigwin, zhang, and anil on the competition results: 15,602 submissions, 18 certified hard uploads, winners wiring around python primitives.
- 2026-09-02the architecture-discovery post · post, core automation blog · ehsan amid'how our data shaped neural architecture discovery… and how automation can reshape the future' — untrained networks already separate cats from dogs; automate the discovery itself.
interviews and talks
- 2025-10-16how gpt-5 thinks · mad podcastthe last interview of the openai years — 'how gpt-5 thinks,' with the bio chapters: poland, dropping out of school, trading. gpt-5 described as roughly 'o3.1.'
- 2026-01-21he left openai to think bigger · core memory ep. 53the exit interview, recorded days after the announcement: big labs have become conservative, deep-learning research at big labs 'is done,' q* became strawberry became o1 — and agi 'likely by 2029' per the decoder's summary.
- 2024-09the o1 interview · fast companythe o1-era press interview — the reasoning program's public face before the exit.
- 2026ex-openai researcher on why he left · unsupervised learning ep. 81the honest-timeline episode: continual learning as necessary for agi, models going 'hopeless' when stuck, robotics in two to three years, and why he left.
- 2026-07-18the agi house deep-dive · agi house auto-research summitthe auto-research summit keynote and interview — evals are dead, humaneval's hand-written origin, o1→o3 as '99% systems, 1% algorithms,' and the open-problems slide eighteen teams built on.
- 2026-07-29building the automated agi lab · sequoia — training datathe flagship episode with tworek and anil, hosted by sonya huang and pat grady: the transformer ceiling, codex compacting after ~20 minutes, fine-tuning as catastrophic forgetting, big labs locked in the coding-agent race, and kernel generation as the first automation target.
- 2026-08-26ai labs are simply cooked · mts podcast'ai labs are simply cooked': three generations of ai labs, a two-year horizon for agents automating research, the taste concession — current agents 'lack deep field understanding and produce low-quality, high-creativity ideas' — plus alignment risks and the 'company brain.'
- 2026the translated long-form · btcc / biggo relaysthe long-form interview relayed through chinese-media translations: the two-year human horizon, the chess-competition analogy for the end state, the 'company brain,' and copying big-tech transformers as 'a losing game.' translation-chain confidence.
- 2026the small-scale eval critique · newsfilter transcript'the problem with testing ai architectures at small scale' — why small-scale evals fail to predict scaled behavior and why rl needed a minimum compute threshold: the reason his ideas couldn't be cheaply pre-proven at openai.
- 2026-09-10continual learning is the next bottleneck · the information bottleneckanil's own episode — 'continual learning is the next bottleneck': the pre-training/rl split as organizational convenience, shampoo's origin story, the qr kernel competition and reward hacking, meeting tworek at dolores park, and the open-source plans.
- 2026from pytorch to gpu mode · podcast ep. 302'from pytorch to gpu mode' — the microsoft→graphcore→meta path and the kernel community that became a company.
the open flags
roster questions the record leaves open — kept visible rather than resolved.
- 2026–the self-assembled rosterthere is no official team page — coreauto.com names nobody. the roster is assembled from linkedin titles, launch posts, the form d's officers and directors, and press; 'co-founder' titles are self-applied and inconsistent — villagra is 'member of the xyz staff,' patil is 'member of agentic staff,' and dealroom upgrades villagra to cofounder.the form d is the only hard record: tworek (executive officer + director, signing as ceo), villagra (executive officer), anil (director), and rosenthal (director). everything else stays labeled.
- 2026the patil questionswapnil patil's identity rests on a linkedin name match — an ex-google ml-systems performance engineer with a cmu phd. the profile fits the lab's needs, but the identification is not cross-confirmed.kept in the roster with the flag attached — moderate confidence, name collision possible.
- 2026the adept attributionone aggregator bio lists anmol gulati as 'cofounder adept ai' — the claim appears nowhere else and looks like a conflation; the adept alumnus in the company's own talent-source list is more plausibly avery lamp.treated as unverified aggregator noise — recorded because aggregators repeat it.
- 2026the inferred cap tablethe investor map is inference: sequoia ran tworek and anil on training data — a common tell — and aggregator copy claims 'backed by sequoia capital,' but no filing or statement confirms it. rosenthal's board seat suggests an acrew position; acrew has disclosed nothing.the form d's 51 investors are unnamed; only alumni ventures is provable, through its own feeder filing.
the philosophy
themes tracked across the writing: continual learning, the automated lab, architecture over scale, evals are dead, small teams, capable agents, big-lab conservatism, systems code and kernels.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.