Public dossier section · @ben

people

This is @ben's profile revision 1. The original date and editorial confidence below are attributed to its publisher, not independent verification.

Original credited author: Hraness

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

Back to the full profile · Markdown version · Original document

Published text

core automation: people

tworek and the team he nerdsniped — who they are, what they wrote and said, and which identities stay flagged.

the people

jerry tworek (ceo, @millionint): cofounder 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. (sources: million integrals — tworeks homepage, jerry tworek — linkedin, jerry tworek speaker bio (warsaw.ai), core automation, inc. — form d (sec), openai loses top ai researcher jerry tworek after seven years (the decoder), he left openai to think bigger — core memory ep. 53)
rohan anil (co-founder): cofounder — 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. (sources: core automation, inc. — form d (sec), continual learning is the next bottleneck — rohan anil (the information bottleneck), rohan anil bio (teahose), rohan anil — google scholar, rohan anil — theorg profile, new ai lab core automation nerdsniped researchers from anthropic, google deepmind (business insider))
joanne jang (co-founder): cofounder 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. (sources: joanne jang — linkedin, joannejang.com, joanne jang speaker bio (womenpm), joanne jang profile (prometheus), joanne jang profile (forbes), new ai lab core automation nerdsniped researchers from anthropic, google deepmind (business insider))
julia villagra (co-founder): cofounder 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. (sources: core automation, inc. — form d (sec), julia villagra — linkedin, openai leadership updates, march 2025, openai chief people officer julia villagra leaving (business insider), core automation eyes $300m–$500m raise at $4b valuation (dealroom note))
anmol gulati (co-founder): cofounder 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. (sources: anmol gulati — linkedin, anmol gulati — google scholar, anmol gulatis launch-day post (texxr mirror), x bios mirror (digg), new ai lab core automation nerdsniped researchers from anthropic, google deepmind (business insider))
avery lamp (co-founder, @averylamp): cofounder, 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. (sources: avery lamp — linkedin, avery lamp — resume, your showcase primer: core automation (fysk), core automation launch submission by averylamp, the nerdsniped launch story, yahoo syndication)
mark saroufim (co-founder): cofounder — 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. (sources: marksaroufim.com, mark saroufim — github, from pytorch to gpu mode — saroufim interview, mark saroufim — linkedin, the core automation blog)
kanav garg (co-founder): cofounder 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. (sources: kanav garg — linkedin, core automation eyes $300m–$500m raise at $4b valuation (dealroom note))
swapnil patil (co-founder): cofounder — 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. (sources: swapnil patil — linkedin (likely match), x bios mirror (digg))
ehsan amid (team): mts 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. (sources: ehsan amids research site, the core automation blog, the nerdsniped launch story, yahoo syndication)
sai surya duvvuri (team): mts — former research intern at google and meta, named in the launch-day coverage. (sources: new ai lab core automation nerdsniped researchers from anthropic, google deepmind (business insider), the nerdsniped launch story, yahoo syndication)
aliisa rosenthal (team): director 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. (sources: core automation, inc. — form d (sec), openais former head of sales is entering vc (fortune), aliisa rosenthal — acrew capital, aliisa rosenthal — linkedin)
the wider bench (team): the 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. (sources: associate general manager program (coreauto.com), core automation — linkedin company page, your showcase primer: core automation (fysk), core automation company profile (silicon valley investclub — conflates the 2001 namesake))
the backers (team): nvidia, 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. (sources: core automation company profile (sacra), core automation funding profile (ai market watch), core automation, inc. — form d (sec), avsf – core automation 2026, llc — form d (sec), aliisa rosenthal — acrew capital, training data episode recap (kazuha))

writings

2026 — million integrals (document, tworek, millionintegrals.com): the self-authored work history — copilot, codex, humaneval, plugins, o1, o3, gpt-5, the chatgpt agent — written in first person, including the typos. source

"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"

2025-10-16 — how gpt-5 thinks (podcast, tworek, mad podcast): the 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.' source
2026-01-05 — the exit post (tweet, tworek, x): the 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"

2026-01-21 — he left openai to think bigger (podcast, tworek, core memory ep. 53): the 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. source
2026-04-21 — the launch posts (tweet, tworek, x): the launch-day pair: the company's 'systems that optimize and automate work, starting with research itself' alongside tworek's own post. source

"small step for a finger on 'post tweet' button. big step for millions of future ai agents doing useful work for us."

2024-09 — the o1 interview (interview, tworek, fast company): the o1-era press interview — the reasoning program's public face before the exit. source

"what we managed to train our models to do is this very natural way of reasoning… it looks a little bit more human."

2026 — ex-openai researcher on why he left (podcast, tworek, unsupervised learning ep. 81): the honest-timeline episode: continual learning as necessary for agi, models going 'hopeless' when stuck, robotics in two to three years, and why he left. source
2026-07-18 — the agi house deep-dive (talk, tworek, agi house auto-research summit): the 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. source

"the era of evals is dumb" "the only test is the real world" "there is no way to outsource your understanding"

2026-07-29 — building the automated agi lab (podcast, tworek, sequoia — training data): the 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. source

"core automation is a lab created to build models that continuously learn from deployment. we believe that transformers are incapable of continual learning." "our quest is to find that new architecture, find that transformer replacement. and we want to build the most automated lab to do it." "the first step to replacing transformers is appreciating deeply how far they were able to carry us."

2026-08-26 — ai labs are simply cooked (podcast, tworek, 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.' source
2026 — the translated long-form (interview, tworek, btcc / biggo relays): the 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. source

"my rough estimate is that human researchers will continue to play a significant role in ai research discoveries for at least two more years" "if we can compress the time for an experiment from one month to one day, that's a 30x acceleration"

2026 — the small-scale eval critique (interview, tworek, 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. source
2026-09-10 — continual learning is the next bottleneck (podcast, anil, the information bottleneck): anil'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. source

"organizational convenience rather than…science"

2026-04-21 — the cofounder's launch post (tweet, gulati, x): gulati'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.' source

"i've increasingly felt that the current research paradigm — scaling models, data, and static deployment won't get us all the way"

2026 — reservoir samples and the labs announcement (essay, jang, joannejang.com): jang'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.' source

"changing model behavior is an art and a science"

2026-05-27 — to automate publishing, first make authoring boring (post, saroufim, core automation blog): the first company blog post — 'to automate publishing, first make authoring boring': the lab's own publishing pipeline as a systems warm-up. source
2026-05-28 — when ai starts writing systems code (post, saroufim, core automation blog): '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. source
2026-09-09 — what the models learned in one layer deeper (post, saroufim, core automation blog): '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. source
2026 — from pytorch to gpu mode (podcast, saroufim, podcast ep. 302): 'from pytorch to gpu mode' — the microsoft→graphcore→meta path and the kernel community that became a company. source
2026-09-02 — the architecture-discovery post (post, amid, core automation blog): '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. source

controversies

2026– — the self-assembled roster: there 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. reaction: 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. (sources: core automation, inc. — form d (sec), coreauto.com — company site, core automation eyes $300m–$500m raise at $4b valuation (dealroom note), core automation — linkedin company page)
2026 — the patil question: swapnil 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. reaction: kept in the roster with the flag attached — moderate confidence, name collision possible. (sources: swapnil patil — linkedin (likely match))
2026 — the adept attribution: one 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. reaction: treated as unverified aggregator noise — recorded because aggregators repeat it. (sources: x bios mirror (digg), core automation — linkedin company page)
2026 — the inferred cap table: the 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. reaction: the form d's 51 investors are unnamed; only alumni ventures is provable, through its own feeder filing. (sources: building the automated agi lab — tworek and anil (sequoia training data), training data episode recap (kazuha), core automation, inc. — form d (sec), avsf – core automation 2026, llc — form d (sec), aliisa rosenthal — acrew capital)

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/people

Original source occurrences

auto-research is moving from benchmarks to real-world loops (agi house) · agihousecommunity.beehiiv.com · 2026-07-23 (day) · original ID: src-ci-agihouse-beehiiv · Other public citations of this URL
"there is no way to outsource your understanding"; verification-bottleneck framing
translated tworek interview (biggo relay) · finance.biggo.com · 2026 (year) · original ID: src-ci-biggo-interview · Other public citations of this URL
same interview; o1/o3 compute-allocation backstory
translated tworek interview (btcc relay) · btcc.com · 2026 (year) · original ID: src-ci-btcc-interview · Other public citations of this URL
chinese-media interview relay; "third-generation ai lab"; 30x experiment compression; "company brain"; chess analogy
core automation eyes $300m–$500m raise at $4b valuation (dealroom note) · app.dealroom.co · 2026-05-08 (day) · original ID: src-ci-dealroom-4b · Other public citations of this URL
$300–500m at ~$4b; $100m at $1b closed weeks prior; investors nvidia/spark/accel; cofounders incl. villagra; "isn't close to releasing" a product
core automation launch submission by averylamp · hacker news · 2026-04-21 (day) · original ID: src-ci-hn-launch · Other public citations of this URL · avery lamp
submission by cofounder averylamp; 1 point, 0 comments
continual learning is the next bottleneck — rohan anil (the information bottleneck) · the-information-bottleneck.com · 2026-09-10 (day) · original ID: src-ci-infobottleneck-anil · Other public citations of this URL
anil interview; pretrain/rl split as "organizational convenience"; kernel-competition + hiring segments; dolores park origin
training data episode recap (kazuha) · kazuha.ai · 2026-07 (month) · original ID: src-ci-kazuha-sequoia · Other public citations of this URL
episode summary; claims "backed by sequoia capital" (unconfirmed)
core automation — linkedin company page · linkedin · 2026-09 (month) · original ID: src-ci-linkedin · Other public citations of this URL
~21 employees; key-executive list; sf hq; founded 2026
marksaroufim.com · marksaroufim.com · 2026-09 (month) · original ID: src-ci-marksaroufim · Other public citations of this URL · mark saroufim
saroufim: core automation cofounder, pytorch maintainer, gpu mode cofounder
ai labs are simply cooked — jerry tworek (mts) · snipd · 2026-08-26 (day) · original ID: src-ci-mts-podcast · Other public citations of this URL
"ai labs are simply cooked"; three-generations model; 2-year horizon; agents' taste deficit; "company brain"
the problem with testing ai architectures at small scale (newsfilter transcript) · newsfilter.io · 2026 (year) · original ID: src-ci-newsfilter-transcript · Other public citations of this URL
small-scale eval critique; rl needs compute threshold
core automation company profile (sacra) · sacra · 2026-09 (month) · original ID: src-ci-sacra · Other public citations of this URL
funding recap; monetization speculation; paradigm-timing risk analysis
building the automated agi lab — tworek and anil (sequoia training data) · youtube · 2026-07-29 (day) · original ID: src-ci-sequoia-podcast · Other public citations of this URL · sonya huang and pat grady
tworek+anil; transformer-ceiling thesis; codex ~20-min compaction; kernel-generation automation; hosts huang & grady
openais vp researcher jerry tworek resigns (times of india) · timesofindia.indiatimes.com · 2026-01 (month) · original ID: src-ci-timesofindia · Other public citations of this URL
exit memo verbatim; chinchilla/rl-on-robots history
openais research lead is departing after nearly seven years (the verge) · the verge · 2026-01-05 (day) · original ID: src-ci-verge-departure · Other public citations of this URL
tworek exit; memo/x post quoted
the nerdsniped launch story, yahoo syndication · finance.yahoo.com · 2026-04-22 (day) · original ID: src-ci-yahoo-nerdsniped · Other public citations of this URL
bi syndication; adds ehsan amid, avery lamp, julia villagra, sai surya duvvuri
auto-research summit recap (agi house) · blog.agihouse.org · 2026-07 (month) · original ID: src-hp-agihouse-recap · Other public citations of this URL
auto-research summit recap; open-problems slide; 18 teams
jerry tworek deep-dive (agi house) · blog.agihouse.org · 2026-07-29 (day) · original ID: src-hp-agihouse-tworek · Other public citations of this URL
deep-dive on tworek keynote/interview — evals, o1→o3, continual learning, founding
core automation funding profile (ai market watch) · ai-market-watch.com · 2026 (year) · original ID: src-hp-aimarketwatch · Other public citations of this URL
~$532m total claim; jul ~$432m at ~$3.6b post (pitchbook/forge-derived); scribble/threshold also listed
ehsan amids research site · sites.google.com · 2026-09-16 (day) · original ID: src-hp-amid-site · Other public citations of this URL · ehsan amid
amid bio: ucsc phd 2020 (warmuth), aalto msc, google brain oct 2021–2026 gemini core, joined core apr 2026
rohan anil bio (teahose) · teahose.com · 2026 (year) · original ID: src-hp-anil-teahose · Other public citations of this URL
anil bio: distinguished engineer deepmind, 1-of-4 gemini pretraining leads, shampoo, anthropic
new ai lab core automation nerdsniped researchers from anthropic, google deepmind (business insider) · business insider · 2026-04-22 (day) · original ID: src-hp-bi-nerdsniped · Other public citations of this URL
launch-day roster: anil "nerdsniped" quote; gulati quote; jang, villagra, lamp, amid, duvvuri named
openai chief people officer julia villagra leaving (business insider) · business insider · 2025-08 (month) · original ID: src-hp-bi-villagra · Other public citations of this URL
villagra's openai exit; reuters' "art, music and storytelling…agi" line
training data episode summary (biggo finance) · finance.biggo.com · 2026-07-29 (day) · original ID: src-hp-biggo-summary · Other public citations of this URL
episode summary incl. the "first step to replacing transformers" tweet
associate general manager program (coreauto.com) · coreauto.com · 2026-09-16 (day) · original ID: src-hp-coreauto-agm · Other public citations of this URL
associate general manager internship page — jan–mar 2027, $10k/mo, apply by oct 6, 2026
the core automation blog · coreauto.com · 2026-05-27 (day) · original ID: src-hp-coreauto-blog-data · Other public citations of this URL · core automation
company blog corpus: five 2026 posts — saroufim on authoring and systems code, the one layer deeper launch and results, amid on neural architecture discovery
coreauto.com — company site · coreauto.com · 2026-09-16 (day) · original ID: src-hp-coreauto-site · Other public citations of this URL
company homepage — "world's most automated ai lab," research/approach/vision/team copy
he left openai to think bigger — core memory ep. 53 · corememory.com · 2026-01-21 (day) · original ID: src-hp-corememory-ep53 · Other public citations of this URL · ashlee vance
exit interview; q*/strawberry/o1 lineage; big labs "conservative"
openai loses top ai researcher jerry tworek after seven years (the decoder) · the-decoder.com · 2026-01-05 (day) · original ID: src-hp-decoder-departure · Other public citations of this URL
tworek exits openai; quote "explore types of research that are hard to do at openai"
google caught up because openai stumbled (the decoder) · the-decoder.com · 2026-01 (month) · original ID: src-hp-decoder-google · Other public citations of this URL
core memory recap: agi ~2029; continual learning "one of the last important building blocks"; google catch-up framing
x bios mirror (digg) · digg · 2026 (year) · original ID: src-hp-digg-bios · Other public citations of this URL
mirror of x bios (tworek ceo bio; jang "trying to automate my work @coreautoai"; gulati incl. unverified "cofounder adept ai")
jerry tworek on o1 reasoning (fast company) · fast company · 2024-09 (month) · original ID: src-hp-fastcompany-o1 · Other public citations of this URL
o1-era tworek quotes on reasoning
joanne jang profile (forbes) · forbes · undated (unknown) · original ID: src-hp-forbes-jang · Other public citations of this URL
forbes/30u30-style profile quotes
your showcase primer: core automation (fysk) · newsletter.foundersysk.com · 2026-08-25 (day) · original ID: src-hp-fysk-primer · Other public citations of this URL
"20-something people," ~40 by year-end; lamp = fysk lp; sept 16 showcase
kanav garg — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-garg-linkedin · Other public citations of this URL
co-founder; ex-deepmind (project mariner post, dec 2024)
anmol gulati — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-gulati-linkedin · Other public citations of this URL
deepmind roles: mariner/agents tl; gemini 3.0/2.5/2.0 credits; google brain speech
anmol gulati — google scholar · scholar.google.com · 2026-09-16 (day) · original ID: src-hp-gulati-scholar · Other public citations of this URL
gulati publications
anmol gulatis launch-day post (texxr mirror) · texxr.com · 2026-04-22 (day) · original ID: src-hp-gulati-x · Other public citations of this URL · anmol gulati
mirror of launch-day post
joanne jang — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-jang-linkedin · Other public citations of this URL
openai gm dec 2021–apr 2026; labs/model-behavior history; time100 post
joannejang.com · joannejang.com · 2026-09-16 (day) · original ID: src-hp-jang-site · Other public citations of this URL · joanne jang
personal site — stanford cs/applied math
avery lamp — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-lamp-linkedin · Other public citations of this URL
deepmind swe-research; adept mts; stealth co-founder
avery lamp — resume · averylamp.me · 2026-09-16 (day) · original ID: src-hp-lamp-resume · Other public citations of this URL · avery lamp
mosaicml composer/yahp; google intern 2018
how gpt-5 thinks — mad podcast, october 2025 · youtube · 2025-10-16 (day) · original ID: src-hp-mad-podcast · Other public citations of this URL
"how gpt-5 thinks" — bio chapters (poland, dropping out, trading); "gpt-5 ≈ o3.1"
sean mcleish — personal site · mcleish7.github.io · 2026-09-16 (day) · original ID: src-hp-mcleish-site · Other public citations of this URL
sean mcleish (co-author, sep 9 results post) — tilde research intern
million integrals — tworeks homepage · millionintegrals.com · 2026-09-16 (day) · original ID: src-hp-millionintegrals · Other public citations of this URL · jerry tworek
tworek's self-authored work history (copilot, codex, o1, o3, gpt-5, chatgpt agent)
openai leadership updates, march 2025 · openai · 2025-03 (month) · original ID: src-hp-openai-leadership · Other public citations of this URL
openai announcement of villagra's cpo promotion
swapnil patil — linkedin (likely match) · linkedin · 2026-09-16 (day) · original ID: src-hp-patil-linkedin · Other public citations of this URL
likely match: ex-google ml-systems perf (gemini-on-gpu, rdma clusters, tpu runtime); cmu phd
joanne jang profile (prometheus) · prometheusroot.com · 2026 (year) · original ID: src-hp-prometheus-jang · Other public citations of this URL
model spec stewardship; oai labs; reservoir samples; jony ive adjacency (reported)
aliisa rosenthal — acrew capital · acrewcapital.com · 2026-09-16 (day) · original ID: src-hp-rosenthal-acrew · Other public citations of this URL
acrew gp bio
openais former head of sales is entering vc (fortune) · fortune · 2026-01-22 (day) · original ID: src-hp-rosenthal-fortune · Other public citations of this URL
rosenthal → acrew gp; openai first commercial hire jun 2022
aliisa rosenthal — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-rosenthal-linkedin · Other public citations of this URL
"$10m→$10b" revenue scaling claims (self-reported)
mark saroufim — github · github · 2026-09-16 (day) · original ID: src-hp-saroufim-github · Other public citations of this URL
github profile — @coreautom/@pytorch/@gpu-mode
mark saroufim — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-saroufim-linkedin · Other public citations of this URL
gpu mode/pytorch work list
from pytorch to gpu mode — saroufim interview · snipd · 2026 (year) · original ID: src-hp-saroufim-podcast · Other public citations of this URL
"from pytorch to gpu mode" — microsoft→graphcore→pytorch→gpu mode path
rohan anil — google scholar · scholar.google.com · 2026-09-16 (day) · original ID: src-hp-scholar-anil · Other public citations of this URL
anil publications (gemini, palm, etc.)
avsf – core automation 2026, llc — form d (sec) · sec · 2026-08-17 (day) · original ID: src-hp-sec-formd-avsf · Other public citations of this URL
"avsf – core automation 2026, llc" — alumni ventures feeder spv, $2,109,250, 100 investors, first sale 2026-08-03
core automation, inc. — form d (sec) · sec · 2026-07-30 (day) · original ID: src-hp-sec-formd-coreauto · Other public citations of this URL
form d: $635,249,754 offering; $432,125,053 sold; $135,252,607 safe conversion; 51 investors; first sale 2026-07-15; officers/directors; 340 brannan st sf
core automation company profile (silicon valley investclub — conflates the 2001 namesake) · siliconvalleyinvestclub.com · 2026-07 (month) · original ID: src-hp-svinvestclub · Other public citations of this URL
caplight/forge-sourced table: apr seed $100m @ $1b post; jul $432.125m @ $3.62b post
building the automated agi lab — timestamped transcript (teahose) · teahose.com · 2026-07-29 (day) · original ID: src-hp-teahose-transcript · Other public citations of this URL · sonya huang and pat grady
timestamped verbatim quotes from the episode
core automations first x post (texxr mirror) · texxr.com · 2026-04-21 (day) · original ID: src-hp-texxr-coreauto · Other public citations of this URL · core automation
mirror of company's first x post ("today we're announcing core automation…")
tworeks launch post (texxr mirror) · texxr.com · 2026-04-21 (day) · original ID: src-hp-texxr-tworek · Other public citations of this URL · jerry tworek
tworek's launch tweet ("small step for a finger on 'post tweet' button…")
rohan anil — theorg profile · theorg.com · 2026 (year) · original ID: src-hp-theorg-anil · Other public citations of this URL
anthropic mts since jun 2025; ucsd ms; bits pilani
building the automated agi lab — tworek and anil (getpodcast record) · getpodcast.com · 2026-07-29 (day) · original ID: src-hp-training-data · Other public citations of this URL · sonya huang and pat grady
sequoia training data ep w/ tworek & anil
jerry tworek — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-tworek-linkedin · Other public citations of this URL
openai roles/dates; independent view (amsterdam) 2015–2018
ex-openai researcher on why he left — unsupervised learning ep. 81 · unsupervised-learning.simplecast.com · 2026 (year) · original ID: src-hp-unsupervised-learning · Other public citations of this URL
tworek on scaling limits, continual learning, why he left
julia villagra — linkedin · linkedin · 2026-09-16 (day) · original ID: src-hp-villagra-linkedin · Other public citations of this URL
openai cpo jun 2024–oct 2025 (per li text); hrt head of people; "member of the xyz staff"
jerry tworek speaker bio (warsaw.ai) · warsaw.ai · 2020 (year) · original ID: src-hp-warsawai · Other public citations of this URL
bio: msc mathematics univ. of warsaw; "first five years of his career in the hedge fund industry"
joanne jang speaker bio (womenpm) · womenpm.org · undated (unknown) · original ID: src-hp-womenpm-jang · Other public citations of this URL
bio: google assistant nlu, dropbox; ms cs + bs applied math stanford; joined openai 2021

Structured research records

jerry tworek · core-automation-people

Original location · source-8de928b054e68b0e9907 · source-25cfbaa387cc68d9c9be · source-151613bef2589429f7a1 · source-1ebcf9332be7769ac171 · source-bd2ad8edfd1e1375ba4a · source-333f666b0fccca3d6903

Original structured fields

{
  "handle": "@millionint",
  "id": "tworek",
  "name": "jerry tworek",
  "note": "cofounder 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.",
  "role": "ceo",
  "sourceIds": [
    "src-hp-millionintegrals",
    "src-hp-tworek-linkedin",
    "src-hp-warsawai",
    "src-hp-sec-formd-coreauto",
    "src-hp-decoder-departure",
    "src-hp-corememory-ep53"
  ]
}

rohan anil · core-automation-people

Original location · source-1ebcf9332be7769ac171 · source-a3b7895376914a87589b · source-61732b7ebccb89c28970 · source-b827a04cf94ad65ec1f8 · source-fd11b57639406bbbeada · source-fdb69ed0a36be5776aaf

Original structured fields

{
  "handle": null,
  "id": "anil",
  "name": "rohan anil",
  "note": "cofounder — 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-sec-formd-coreauto",
    "src-ci-infobottleneck-anil",
    "src-hp-anil-teahose",
    "src-hp-scholar-anil",
    "src-hp-theorg-anil",
    "src-hp-bi-nerdsniped"
  ]
}

joanne jang · core-automation-people

Original location · source-043117ab23b38f0eee88 · source-dc48be8f83d31389aa83 · source-ba4be762f58beb1f2669 · source-b5c1d30a07a38f828a5f · source-20d4054b5c21495368ea · source-fdb69ed0a36be5776aaf

Original structured fields

{
  "handle": null,
  "id": "jang",
  "name": "joanne jang",
  "note": "cofounder 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-jang-linkedin",
    "src-hp-jang-site",
    "src-hp-womenpm-jang",
    "src-hp-prometheus-jang",
    "src-hp-forbes-jang",
    "src-hp-bi-nerdsniped"
  ]
}

julia villagra · core-automation-people

Original location · source-1ebcf9332be7769ac171 · source-71300673c6428de26449 · source-436ebd0b8ac6f5a679c9 · source-c054436dabb4b0e49fc5 · source-a15b1429f33e57a2d599

Original structured fields

{
  "handle": null,
  "id": "villagra",
  "name": "julia villagra",
  "note": "cofounder 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-sec-formd-coreauto",
    "src-hp-villagra-linkedin",
    "src-hp-openai-leadership",
    "src-hp-bi-villagra",
    "src-ci-dealroom-4b"
  ]
}

anmol gulati · core-automation-people

Original location · source-5678020f38da84e7abe6 · source-a113db3192cb433ab60b · source-b4ce13eb7a1b18d2eb4d · source-a73eb42cb9cfd24f8b9f · source-fdb69ed0a36be5776aaf

Original structured fields

{
  "handle": null,
  "id": "gulati",
  "name": "anmol gulati",
  "note": "cofounder 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-gulati-linkedin",
    "src-hp-gulati-scholar",
    "src-hp-gulati-x",
    "src-hp-digg-bios",
    "src-hp-bi-nerdsniped"
  ]
}

avery lamp · core-automation-people

Original location · source-faafa35b6ee0731debdb · source-ea467af8b1655e907e76 · source-d2983547bdc9fbe8d6a3 · source-2d59fd16262cbbb66af8 · source-096287f02ff129f28b9d

Original structured fields

{
  "handle": "@averylamp",
  "id": "lamp",
  "name": "avery lamp",
  "note": "cofounder, 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-lamp-linkedin",
    "src-hp-lamp-resume",
    "src-hp-fysk-primer",
    "src-ci-hn-launch",
    "src-ci-yahoo-nerdsniped"
  ]
}

mark saroufim · core-automation-people

Original location · source-dd8c9743bfd56ba656d7 · source-67dc87109153fe8dd9f9 · source-73871eca2ea73150f07f · source-d5ecdad014481f8752c8 · source-b6609bc25d69b16aaa6b

Original structured fields

{
  "handle": null,
  "id": "saroufim",
  "name": "mark saroufim",
  "note": "cofounder — 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-ci-marksaroufim",
    "src-hp-saroufim-github",
    "src-hp-saroufim-podcast",
    "src-hp-saroufim-linkedin",
    "src-hp-coreauto-blog-data"
  ]
}

kanav garg · core-automation-people

Original location · source-b066261ed766d0f814b7 · source-a15b1429f33e57a2d599

Original structured fields

{
  "handle": null,
  "id": "garg",
  "name": "kanav garg",
  "note": "cofounder 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-garg-linkedin",
    "src-ci-dealroom-4b"
  ]
}

swapnil patil · core-automation-people

Original location · source-0672852c748308031626 · source-a73eb42cb9cfd24f8b9f

Original structured fields

{
  "handle": null,
  "id": "patil",
  "name": "swapnil patil",
  "note": "cofounder — 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.",
  "role": "co-founder",
  "sourceIds": [
    "src-hp-patil-linkedin",
    "src-hp-digg-bios"
  ]
}

ehsan amid · core-automation-people

Original location · source-4f7781d80eb21eff67de · source-b6609bc25d69b16aaa6b · source-096287f02ff129f28b9d

Original structured fields

{
  "handle": null,
  "id": "amid",
  "name": "ehsan amid",
  "note": "mts 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.",
  "role": "team",
  "sourceIds": [
    "src-hp-amid-site",
    "src-hp-coreauto-blog-data",
    "src-ci-yahoo-nerdsniped"
  ]
}

sai surya duvvuri · core-automation-people

Original location · source-fdb69ed0a36be5776aaf · source-096287f02ff129f28b9d

Original structured fields

{
  "handle": null,
  "id": "duvvuri",
  "name": "sai surya duvvuri",
  "note": "mts — former research intern at google and meta, named in the launch-day coverage.",
  "role": "team",
  "sourceIds": [
    "src-hp-bi-nerdsniped",
    "src-ci-yahoo-nerdsniped"
  ]
}

aliisa rosenthal · core-automation-people

Original location · source-1ebcf9332be7769ac171 · source-99e07cd452943a444c1e · source-f104bcdcefdd6f467acb · source-7cfc0dcc7b6f4c5325ec

Original structured fields

{
  "handle": null,
  "id": "rosenthal",
  "name": "aliisa rosenthal",
  "note": "director 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.",
  "role": "team",
  "sourceIds": [
    "src-hp-sec-formd-coreauto",
    "src-hp-rosenthal-fortune",
    "src-hp-rosenthal-acrew",
    "src-hp-rosenthal-linkedin"
  ]
}

the wider bench · core-automation-people

Original location · source-466b261c17d90367bdfe · source-a43a6672c482a60bec6c · source-d2983547bdc9fbe8d6a3 · source-425914e251a310e9baa9

Original structured fields

{
  "handle": null,
  "id": "bench",
  "name": "the wider bench",
  "note": "the 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.",
  "role": "team",
  "sourceIds": [
    "src-hp-coreauto-agm",
    "src-ci-linkedin",
    "src-hp-fysk-primer",
    "src-hp-svinvestclub"
  ]
}

the backers · core-automation-people

Original location · source-c800d245b22a6234598a · source-5ab6913e09ab1028784d · source-1ebcf9332be7769ac171 · source-99f92fa5fd2f37485555 · source-f104bcdcefdd6f467acb · source-4d44ba49a1861a456536

Original structured fields

{
  "handle": null,
  "id": "backers",
  "name": "the backers",
  "note": "nvidia, 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.",
  "role": "team",
  "sourceIds": [
    "src-ci-sacra",
    "src-hp-aimarketwatch",
    "src-hp-sec-formd-coreauto",
    "src-hp-sec-formd-avsf",
    "src-hp-rosenthal-acrew",
    "src-ci-kazuha-sequoia"
  ]
}

continual learning · core-automation-themes
Original location
Original structured fields
```
{
  "text": "continual learning"
}
```
the automated lab · core-automation-themes
Original location
Original structured fields
```
{
  "text": "the automated lab"
}
```
architecture over scale · core-automation-themes
Original location
Original structured fields
```
{
  "text": "architecture over scale"
}
```
evals are dead · core-automation-themes
Original location
Original structured fields
```
{
  "text": "evals are dead"
}
```
small teams, capable agents · core-automation-themes
Original location
Original structured fields
```
{
  "text": "small teams, capable agents"
}
```
big-lab conservatism · core-automation-themes
Original location
Original structured fields
```
{
  "text": "big-lab conservatism"
}
```
systems code and kernels · core-automation-themes
Original location
Original structured fields
```
{
  "text": "systems code and kernels"
}
```

million integrals · core-automation-writings · 2026 (year)

Original location · source-8de928b054e68b0e9907

Original structured fields

{
  "date": "2026",
  "id": "wrt-millionintegrals",
  "kind": "document",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-millionintegrals",
      "text": "i've conducted all stages of research and trained the model that powered github copilot"
    },
    {
      "sourceId": "src-hp-millionintegrals",
      "text": "i've lead the team that researched training llms to reason with reinforcement learning"
    }
  ],
  "sourceIds": [
    "src-hp-millionintegrals"
  ],
  "summary": "the self-authored work history — copilot, codex, humaneval, plugins, o1, o3, gpt-5, the chatgpt agent — written in first person, including the typos.",
  "themes": [
    "architecture over scale",
    "big-lab conservatism"
  ],
  "title": "million integrals",
  "traction": null,
  "url": "https://millionintegrals.com/",
  "venue": "millionintegrals.com"
}

how gpt-5 thinks · core-automation-writings · 2025-10-16 (day)

Original location · source-87c9a20b6ad4f98d43dc

Original structured fields

{
  "date": "2025-10-16",
  "id": "wrt-mad-podcast",
  "kind": "podcast",
  "person": "tworek",
  "quotes": [],
  "sourceIds": [
    "src-hp-mad-podcast"
  ],
  "summary": "the 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.'",
  "themes": [
    "architecture over scale"
  ],
  "title": "how gpt-5 thinks",
  "traction": null,
  "url": "https://www.youtube.com/watch?v=RqWIvvv3SnQ",
  "venue": "mad podcast"
}

the exit post · core-automation-writings · 2026-01-05 (day)

Original location · source-bd2ad8edfd1e1375ba4a · source-2ff44485c491e4a4a770 · source-62830dab964b96919baa

Original structured fields

{
  "date": "2026-01-05",
  "id": "wrt-exit-post",
  "kind": "tweet",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-decoder-departure",
      "text": "to try and explore types of research that are hard to do at openai"
    }
  ],
  "sourceIds": [
    "src-hp-decoder-departure",
    "src-ci-verge-departure",
    "src-ci-timesofindia"
  ],
  "summary": "the 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.'",
  "themes": [
    "big-lab conservatism"
  ],
  "title": "the exit post",
  "traction": "covered same-day by the verge, the decoder, and the indian press",
  "url": null,
  "venue": "x"
}

he left openai to think bigger · core-automation-writings · 2026-01-21 (day)

Original location · source-333f666b0fccca3d6903 · source-7405e0b34edfba397dec

Original structured fields

{
  "date": "2026-01-21",
  "id": "wrt-corememory",
  "kind": "podcast",
  "person": "tworek",
  "quotes": [],
  "sourceIds": [
    "src-hp-corememory-ep53",
    "src-hp-decoder-google"
  ],
  "summary": "the 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.",
  "themes": [
    "big-lab conservatism",
    "continual learning"
  ],
  "title": "he left openai to think bigger",
  "traction": null,
  "url": "https://www.corememory.com/p/he-left-openai-jerry-tworek",
  "venue": "core memory ep. 53"
}

the launch posts · core-automation-writings · 2026-04-21 (day)

Original location · source-e8c5fec49ac7294f4776 · source-6d86cef8b7d059006a8d

Original structured fields

{
  "date": "2026-04-21",
  "id": "wrt-launch-post",
  "kind": "tweet",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-texxr-tworek",
      "text": "small step for a finger on 'post tweet' button. big step for millions of future ai agents doing useful work for us."
    }
  ],
  "sourceIds": [
    "src-hp-texxr-tworek",
    "src-hp-texxr-coreauto"
  ],
  "summary": "the launch-day pair: the company's 'systems that optimize and automate work, starting with research itself' alongside tworek's own post.",
  "themes": [
    "the automated lab",
    "small teams, capable agents"
  ],
  "title": "the launch posts",
  "traction": null,
  "url": "https://texxr.com/handle/millionint",
  "venue": "x"
}

the o1 interview · core-automation-writings · 2024-09 (month)

Original location · source-6769d5c0758441fb6f7a

Original structured fields

{
  "date": "2024-09",
  "id": "wrt-fastcompany-o1",
  "kind": "interview",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-fastcompany-o1",
      "text": "what we managed to train our models to do is this very natural way of reasoning… it looks a little bit more human."
    }
  ],
  "sourceIds": [
    "src-hp-fastcompany-o1"
  ],
  "summary": "the o1-era press interview — the reasoning program's public face before the exit.",
  "themes": [
    "architecture over scale"
  ],
  "title": "the o1 interview",
  "traction": null,
  "url": "https://www.fastcompany.com/91246222/openai-researcher-jerry-tworek-human-brain-o1-models",
  "venue": "fast company"
}

ex-openai researcher on why he left · core-automation-writings · 2026 (year)

Original location · source-be5dae4a31d8ab20ba5a

Original structured fields

{
  "date": "2026",
  "id": "wrt-unsupervised-learning",
  "kind": "podcast",
  "person": "tworek",
  "quotes": [],
  "sourceIds": [
    "src-hp-unsupervised-learning"
  ],
  "summary": "the honest-timeline episode: continual learning as necessary for agi, models going 'hopeless' when stuck, robotics in two to three years, and why he left.",
  "themes": [
    "continual learning",
    "big-lab conservatism"
  ],
  "title": "ex-openai researcher on why he left",
  "traction": null,
  "url": "https://unsupervised-learning.simplecast.com/episodes/ep-81-ex-openai-researcher-on-why-he-left-his-honest-agi-timeline-the-limits-of-scaling-rl-v8n6gexx-fZAknHYv",
  "venue": "unsupervised learning ep. 81"
}

the agi house deep-dive · core-automation-writings · 2026-07-18 (day)

Original location · source-031e6807adae79efb2d3 · source-2ab09564c60328f89fa5 · source-174997f87efc94eec2a8

Original structured fields

{
  "date": "2026-07-18",
  "id": "wrt-agihouse",
  "kind": "talk",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-agihouse-tworek",
      "text": "the era of evals is dumb"
    },
    {
      "sourceId": "src-hp-agihouse-tworek",
      "text": "the only test is the real world"
    },
    {
      "sourceId": "src-ci-agihouse-beehiiv",
      "text": "there is no way to outsource your understanding"
    }
  ],
  "sourceIds": [
    "src-hp-agihouse-tworek",
    "src-hp-agihouse-recap",
    "src-ci-agihouse-beehiiv"
  ],
  "summary": "the 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.",
  "themes": [
    "evals are dead",
    "the automated lab",
    "continual learning"
  ],
  "title": "the agi house deep-dive",
  "traction": "18 teams worked his open-problems slide at the summit",
  "url": "https://blog.agihouse.org/posts/jerry-tworek-deep-dive",
  "venue": "agi house auto-research summit"
}

building the automated agi lab · core-automation-writings · 2026-07-29 (day)

Original location · source-d31ac7a38bdf5dfb0e64 · source-6f10a8ea42b7d0de1151 · source-bedbcfbe402d8788c612 · source-14dd8994f0135ba55018

Original structured fields

{
  "date": "2026-07-29",
  "id": "wrt-training-data",
  "kind": "podcast",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-hp-teahose-transcript",
      "text": "core automation is a lab created to build models that continuously learn from deployment. we believe that transformers are incapable of continual learning."
    },
    {
      "sourceId": "src-hp-teahose-transcript",
      "text": "our quest is to find that new architecture, find that transformer replacement. and we want to build the most automated lab to do it."
    },
    {
      "sourceId": "src-hp-teahose-transcript",
      "text": "the first step to replacing transformers is appreciating deeply how far they were able to carry us."
    }
  ],
  "sourceIds": [
    "src-ci-sequoia-podcast",
    "src-hp-training-data",
    "src-hp-teahose-transcript",
    "src-hp-biggo-summary"
  ],
  "summary": "the 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.",
  "themes": [
    "continual learning",
    "the automated lab",
    "architecture over scale",
    "big-lab conservatism"
  ],
  "title": "building the automated agi lab",
  "traction": "the hn submission of the video drew 2 points, 0 comments",
  "url": "https://www.youtube.com/watch?v=2RJiaf0SY8s",
  "venue": "sequoia — training data"
}

ai labs are simply cooked · core-automation-writings · 2026-08-26 (day)

Original location · source-b61f38c141b9a164a9e4

Original structured fields

{
  "date": "2026-08-26",
  "id": "wrt-mts",
  "kind": "podcast",
  "person": "tworek",
  "quotes": [],
  "sourceIds": [
    "src-ci-mts-podcast"
  ],
  "summary": "'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.'",
  "themes": [
    "the automated lab",
    "small teams, capable agents"
  ],
  "title": "ai labs are simply cooked",
  "traction": null,
  "url": "https://share.snipd.com/episode/28a168a3-3d27-431e-8808-849d05d9fc34",
  "venue": "mts podcast"
}

the translated long-form · core-automation-writings · 2026 (year)

Original location · source-361b2040d2360d8af452 · source-84159fb04dc6497f7498

Original structured fields

{
  "date": "2026",
  "id": "wrt-translated-interview",
  "kind": "interview",
  "person": "tworek",
  "quotes": [
    {
      "sourceId": "src-ci-btcc-interview",
      "text": "my rough estimate is that human researchers will continue to play a significant role in ai research discoveries for at least two more years"
    },
    {
      "sourceId": "src-ci-btcc-interview",
      "text": "if we can compress the time for an experiment from one month to one day, that's a 30x acceleration"
    }
  ],
  "sourceIds": [
    "src-ci-btcc-interview",
    "src-ci-biggo-interview"
  ],
  "summary": "the 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.",
  "themes": [
    "the automated lab",
    "small teams, capable agents"
  ],
  "title": "the translated long-form",
  "traction": null,
  "url": "https://www.btcc.com/en-AU/news/deep-article/11448",
  "venue": "btcc / biggo relays"
}

the small-scale eval critique · core-automation-writings · 2026 (year)

Original location · source-961cfb82fe701429df61

Original structured fields

{
  "date": "2026",
  "id": "wrt-newsfilter",
  "kind": "interview",
  "person": "tworek",
  "quotes": [],
  "sourceIds": [
    "src-ci-newsfilter-transcript"
  ],
  "summary": "'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.",
  "themes": [
    "evals are dead",
    "architecture over scale"
  ],
  "title": "the small-scale eval critique",
  "traction": null,
  "url": "https://newsfilter.io/interviews/the-problem-with-testing-ai-architectures-at-small-scale-jerry-tworek-core-autom-0f49567b-479c-5bbb-87f9-22c4fc3e4c7d",
  "venue": "newsfilter transcript"
}

continual learning is the next bottleneck · core-automation-writings · 2026-09-10 (day)

Original location · source-a3b7895376914a87589b · source-61732b7ebccb89c28970

Original structured fields

{
  "date": "2026-09-10",
  "id": "wrt-infobottleneck",
  "kind": "podcast",
  "person": "anil",
  "quotes": [
    {
      "sourceId": "src-ci-infobottleneck-anil",
      "text": "organizational convenience rather than…science"
    }
  ],
  "sourceIds": [
    "src-ci-infobottleneck-anil",
    "src-hp-anil-teahose"
  ],
  "summary": "anil'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.",
  "themes": [
    "continual learning",
    "systems code and kernels",
    "architecture over scale"
  ],
  "title": "continual learning is the next bottleneck",
  "traction": null,
  "url": "https://www.the-information-bottleneck.com/p/continual-learning-is-the-next-bottleneck-2fc",
  "venue": "the information bottleneck"
}

the cofounder's launch post · core-automation-writings · 2026-04-21 (day)

Original location · source-fdb69ed0a36be5776aaf · source-b4ce13eb7a1b18d2eb4d

Original structured fields

{
  "date": "2026-04-21",
  "id": "wrt-gulati-launch",
  "kind": "tweet",
  "person": "gulati",
  "quotes": [
    {
      "sourceId": "src-hp-bi-nerdsniped",
      "text": "i've increasingly felt that the current research paradigm — scaling models, data, and static deployment won't get us all the way"
    }
  ],
  "sourceIds": [
    "src-hp-gulati-x",
    "src-hp-bi-nerdsniped"
  ],
  "summary": "gulati'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.'",
  "themes": [
    "architecture over scale",
    "the automated lab"
  ],
  "title": "the cofounder's launch post",
  "traction": null,
  "url": "https://texxr.com/handle/anmol01gulati",
  "venue": "x"
}

reservoir samples and the labs announcement · core-automation-writings · 2026 (year)

Original location · source-20d4054b5c21495368ea · source-b5c1d30a07a38f828a5f · source-dc48be8f83d31389aa83 · source-043117ab23b38f0eee88

Original structured fields

{
  "date": "2026",
  "id": "wrt-reservoir",
  "kind": "essay",
  "person": "jang",
  "quotes": [
    {
      "sourceId": "src-hp-forbes-jang",
      "text": "changing model behavior is an art and a science"
    }
  ],
  "sourceIds": [
    "src-hp-prometheus-jang",
    "src-hp-jang-site",
    "src-hp-jang-linkedin"
  ],
  "summary": "jang'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.'",
  "themes": [
    "small teams, capable agents"
  ],
  "title": "reservoir samples and the labs announcement",
  "traction": null,
  "url": "https://joannejang.com/",
  "venue": "joannejang.com"
}

to automate publishing, first make authoring boring · core-automation-writings · 2026-05-27 (day)

Original location · source-b6609bc25d69b16aaa6b

Original structured fields

{
  "date": "2026-05-27",
  "id": "wrt-authoring-boring",
  "kind": "post",
  "person": "saroufim",
  "quotes": [],
  "sourceIds": [
    "src-hp-coreauto-blog-data"
  ],
  "summary": "the first company blog post — 'to automate publishing, first make authoring boring': the lab's own publishing pipeline as a systems warm-up.",
  "themes": [
    "the automated lab",
    "systems code and kernels"
  ],
  "title": "to automate publishing, first make authoring boring",
  "traction": null,
  "url": "https://www.coreauto.com/blog",
  "venue": "core automation blog"
}

when ai starts writing systems code · core-automation-writings · 2026-05-28 (day)

Original location · source-b6609bc25d69b16aaa6b

Original structured fields

{
  "date": "2026-05-28",
  "id": "wrt-systems-code",
  "kind": "post",
  "person": "saroufim",
  "quotes": [],
  "sourceIds": [
    "src-hp-coreauto-blog-data"
  ],
  "summary": "'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.",
  "themes": [
    "the automated lab",
    "systems code and kernels"
  ],
  "title": "when ai starts writing systems code",
  "traction": null,
  "url": "https://www.coreauto.com/blog",
  "venue": "core automation blog"
}

what the models learned in one layer deeper · core-automation-writings · 2026-09-09 (day)

Original location · source-b6609bc25d69b16aaa6b · source-c69ac40920ebde955b9a

Original structured fields

{
  "date": "2026-09-09",
  "id": "wrt-old-results",
  "kind": "post",
  "person": "saroufim",
  "quotes": [],
  "sourceIds": [
    "src-hp-coreauto-blog-data",
    "src-hp-mcleish-site"
  ],
  "summary": "'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.",
  "themes": [
    "evals are dead",
    "systems code and kernels"
  ],
  "title": "what the models learned in one layer deeper",
  "traction": null,
  "url": "https://www.coreauto.com/blog",
  "venue": "core automation blog"
}

from pytorch to gpu mode · core-automation-writings · 2026 (year)

Original location · source-73871eca2ea73150f07f · source-dd8c9743bfd56ba656d7

Original structured fields

{
  "date": "2026",
  "id": "wrt-gpumode-podcast",
  "kind": "podcast",
  "person": "saroufim",
  "quotes": [],
  "sourceIds": [
    "src-hp-saroufim-podcast",
    "src-ci-marksaroufim"
  ],
  "summary": "'from pytorch to gpu mode' — the microsoft→graphcore→meta path and the kernel community that became a company.",
  "themes": [
    "systems code and kernels"
  ],
  "title": "from pytorch to gpu mode",
  "traction": null,
  "url": "https://share.snipd.com/episode/c8577ba1-ca13-42ac-841f-5aceef0ca3af",
  "venue": "podcast ep. 302"
}

the architecture-discovery post · core-automation-writings · 2026-09-02 (day)

Original location · source-b6609bc25d69b16aaa6b · source-4f7781d80eb21eff67de

Original structured fields

{
  "date": "2026-09-02",
  "id": "wrt-amid-post",
  "kind": "post",
  "person": "amid",
  "quotes": [],
  "sourceIds": [
    "src-hp-coreauto-blog-data",
    "src-hp-amid-site"
  ],
  "summary": "'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.",
  "themes": [
    "the automated lab",
    "architecture over scale"
  ],
  "title": "the architecture-discovery post",
  "traction": null,
  "url": "https://www.coreauto.com/blog",
  "venue": "core automation blog"
}

the self-assembled roster · people-controversies · 2026– (approximate)

Original location · source-1ebcf9332be7769ac171 · source-440a6bec8c8cc37b4d74 · source-a15b1429f33e57a2d599 · source-a43a6672c482a60bec6c

Original structured fields

{
  "date": "2026–",
  "detail": "there 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.",
  "id": "cty-roster-assembly",
  "reaction": "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.",
  "sourceIds": [
    "src-hp-sec-formd-coreauto",
    "src-hp-coreauto-site",
    "src-ci-dealroom-4b",
    "src-ci-linkedin"
  ],
  "title": "the self-assembled roster"
}

the patil question · people-controversies · 2026 (year)

Original location · source-0672852c748308031626

Original structured fields

{
  "date": "2026",
  "detail": "swapnil 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.",
  "id": "cty-patil-identity",
  "reaction": "kept in the roster with the flag attached — moderate confidence, name collision possible.",
  "sourceIds": [
    "src-hp-patil-linkedin"
  ],
  "title": "the patil question"
}

the adept attribution · people-controversies · 2026 (year)

Original location · source-a73eb42cb9cfd24f8b9f · source-a43a6672c482a60bec6c

Original structured fields

{
  "date": "2026",
  "detail": "one 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.",
  "id": "cty-gulati-adept",
  "reaction": "treated as unverified aggregator noise — recorded because aggregators repeat it.",
  "sourceIds": [
    "src-hp-digg-bios",
    "src-ci-linkedin"
  ],
  "title": "the adept attribution"
}

the inferred cap table · people-controversies · 2026 (year)

Original location · source-6f10a8ea42b7d0de1151 · source-4d44ba49a1861a456536 · source-1ebcf9332be7769ac171 · source-99f92fa5fd2f37485555 · source-f104bcdcefdd6f467acb

Original structured fields

{
  "date": "2026",
  "detail": "the 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.",
  "id": "cty-investor-inference",
  "reaction": "the form d's 51 investors are unnamed; only alumni ventures is provable, through its own feeder filing.",
  "sourceIds": [
    "src-ci-sequoia-podcast",
    "src-ci-kazuha-sequoia",
    "src-hp-sec-formd-coreauto",
    "src-hp-sec-formd-avsf",
    "src-hp-rosenthal-acrew"
  ],
  "title": "the inferred cap table"
}