Core Automation

Core Automation (legally Core Automation, Inc.) is an American artificial intelligence research company based in San Francisco. It was founded in 2026 by Jerry Tworek, a former OpenAI vice president of research who ran OpenAI's Reasoning Models team, together with researchers and executives from OpenAI, Anthropic and Google DeepMind.^[2][6][8] The company says it is building "the world's most automated AI lab", with the objective of "systems that optimize and automate work, starting with research itself."^[1] Its research program centers on continual learning, meaning models that keep learning after they are deployed, and on new learning algorithms and neural network architectures that it hopes will replace large-scale pretraining, reinforcement learning and the transformer.^[1][4][21]

The company was announced publicly on April 21, 2026.^[10] The Information reported in May 2026 that it had raised $100 million at a $1 billion valuation and was in early talks to raise $300 million to $500 million at a valuation of about $4 billion, with Nvidia investing in the earlier round.^[5] A Form D notice filed with the U.S. Securities and Exchange Commission on July 30, 2026 reported $432.1 million sold in an equity offering of up to $635.2 million; the filing does not state a valuation.^[6] As of October 2026 the company's website described no released model or product.^[1] Its public output by then included research essays, a machine learning competition run with Tilde Research, and an internship program built around businesses run by AI agents.^[1][17][19]

Overview

History

Background and founding

On January 5, 2026, Tworek told his team at OpenAI that he was leaving after nearly seven years. He had most recently run OpenAI's "Reasoning Models" team and was part of the core group behind the o1 and o3 models. He wrote that he wanted "to try and explore types of research that are hard to do at OpenAI."^[8] In an interview on the Unsupervised Learning podcast, reported by The Decoder on February 2, 2026, he described a gap he wanted to work on: "There isn't a very good mechanism for a model to update its beliefs and its internal knowledge based on failure." He added that unless models "can work themselves through difficulties and get unstuck on solving a problem, I don't think I would call it AGI."^[33]

On January 28, 2026, The Information reported that Core Automation, which it described as founded by Tworek "in recent weeks", aimed to raise between $500 million and $1 billion, according to a person who had spoken with the company. According to materials shown to potential investors, Tworek wanted to develop AI models using methods that OpenAI and Anthropic were not heavily focused on, and to build models that "can learn on the fly from real-world experience", a capability the report said current models lack.^[4] Summarizing that report, The Decoder wrote that Tworek's goal was models that need 100 times less data, and that the company's long-term plans included AI agents developing products ranging from industrial automation to "self-replicating factories", bio-machines and eventually planetary terraforming.^[9] The Polish trade publication ITReseller, writing about the same late-January reports, gave the name of the model under development as "Ceres".^[34] The company has not used that name in its own published material.^[1]

Reports differ on the founding date. The Information's January article described the company as weeks old, while its May 2026 article said it was "founded in late March".^[4][5] The company's Form D filing lists 2026 as the year of incorporation.^[6]

Public launch

The company announced itself on X on April 21, 2026 with the post "Today we're announcing Core Automation".^[10] Tworek, whose X profile lists him as chief executive and co-founder, quoted it with "Small step for a finger on 'post tweet' button. Big step for millions of future AI agents doing useful work for us."^[2][11] Business Insider reported the next day that the startup was drawing researchers from Anthropic and Google DeepMind. Rohan Anil wrote that he had left Anthropic "a few weeks ago" and that "Jerry Tworek nerdsniped me into starting this with him and others." Anmol Gulati, a Google DeepMind research scientist who had worked on Gemini, wrote that "the next phase comes from something different: new learning algorithms, architectures beyond today's stack, and systems that automate the process of building itself."^[2] Mark Saroufim, another member of the founding team, wrote on launch day that "a lot of the stack we use today was built for an earlier paradigm," and that new kinds of research would need "new libraries, new abstractions, and new ways of running the lab."^[12]

Business Insider also reported that Joanne Jang, a general manager at OpenAI until April 2026 according to her LinkedIn profile, had joined. Others who said they had joined included former Google DeepMind staff Ehsan Amid and Avery Lamp, OpenAI's former head of people Julia Villagra, and Sai Surya Duvvuri, a former research intern at Google and Meta.^[2] The Decoder grouped Core Automation with other companies founded by OpenAI alumni, including Thinking Machines Lab and Safe Superintelligence, that share the view that further progress requires fundamentally new approaches.^[3]

After the launch

In the months after launch, the company's researchers published essays on its blog, gave talks and podcast interviews, and ran a public research competition. The company's website described no released model or product as of October 2026.^[1]

People

Tworek is the company's chief executive and its main public face. Anil's X profile and Saroufim's conference biography describe each of them as a co-founder.^[13][37] Dealroom, summarizing The Information's May 2026 report, named Anil and Jang as co-founders with Tworek, and Gulati and Villagra as other co-founders.^[32]

Other people publicly associated with the company include Ehsan Amid, Avery Lamp and Sai Surya Duvvuri^[2] and Riley Walz, who wrote in September 2026 that he had joined five months earlier.^[20] The Form D also lists Aliisa Rosenthal as a director. Rosenthal, OpenAI's first sales leader, joined the venture firm Acrew Capital as a general partner in January 2026.^[6][28]

Research program

Continual learning and new architectures

The company's website states: "We don't think the next step change in AI will come from scaling the current recipe: larger models, more data, static deployment." It says it is pursuing "new learning algorithms that supersede large-scale pretraining and reinforcement learning, and architectures that scale better than transformers."^[1] Its X profile describes it as "a deep learning research company with a mission to build systems that automate and optimize work."^[36]

Sequoia Capital's description of its July 2026 podcast with Tworek and Anil summarizes the founders' argument. In this account the transformer has carried the field as far as it can, and the main bottleneck is now the architecture rather than scale. The missing capability is continual learning, meaning models that adapt at test time. The description says in-context learning "taps out fast" and that fine-tuning "invites catastrophic forgetting", the tendency of networks to lose earlier knowledge when trained on new data. It adds that Anil argues pretraining and reinforcement learning should be optimized end to end and that transformers spend computation inefficiently.^[21] Describing a September 2026 interview, The Information Bottleneck podcast summarized Anil's view that pretraining and reinforcement learning "were split by organizational convenience rather than by science". Neither stage, in that summary, gives a deployed model a way to absorb new data or learn from its own experience, and folding in new knowledge at inference time is "a fundamental optimization question rather than a harness-engineering one."^[22]

Tworek's public talks after the launch covered related themes. AGI House described a July 2026 session in which he presented automated research as a way to productively scale test-time compute, discussed why autonomous research loops eventually plateau, and said he was exploring test-time training, in which models keep learning while working on hard problems.^[23] In another AGI House conversation, whose description quotes him as saying "the era of evals is done", he discussed how real-world deployment differs from static benchmarks.^[24] The description of an August 2026 interview on the show MTS says he believes human researchers have about two years before AI can fully automate research workflows.^[25]

Ehsan Amid's September 2026 essay on the company blog reported that an ensemble of randomly initialized ResNets, with no training, produces a representation of the CIFAR-10 image dataset that separates its classes better than raw pixels do. He argued that decades of benchmark-driven trial and error had selected for architectures suited to natural data, so the data used in automated architecture search will shape what it finds. He wrote that one of the company's main missions is to design scalable search algorithms and use them to discover models "better suited for tasks such as continual learning, where current architectures still struggle."^[15]

Automating the lab

The company's stated approach is that "the next wave of frontier research will come from small teams with highly capable agents". It says it starts by automating its own work, which "makes room for more ambitious and creative work, which reveals the next thing to automate."^[1] The Sequoia description says the founders see automated kernel generation as the starting point for building the "world's most automated lab", because GPU kernels are "the one place frontier models still lose to a high-taste human."^[21]

Saroufim's MLSys keynote dealt with "kernel LLMs", models that generate GPU kernels. According to the abstract, GPU kernels are a good target for AI-driven optimization because they are verifiable and commercially valuable, but few AI-generated kernels were reliable enough for production use without significant human supervision.^[13] In the essay version, he wrote that "to automate research we must automate systems." He said the company's existing training codebase still used autoregressive transformers so that it could reproduce frontier results. He described the lab as one "focused on alternatives to transformers, where the distinction between training and inference doesn't really exist, where automating systems code is core to the business."^[14]

Long-term vision

The website describes a world in which "small teams with powerful AI systems will take on work that once required entire organizations", giving the example of "a corner store with the logistical capabilities of a multinational corporation."^[1] Its experiments with AI agents running small online businesses, described below, are framed around what such agents can and cannot yet do.^[19][20]

One Layer Deeper competition

On August 3, 2026, Tilde Research and Core Automation launched One Layer Deeper, a competition on whether models can learn deep serial computation end to end. The launch post lists Sean McLeish, Ben Keigwin, Saroufim and Anil as contacts.^[16] The organizers argued that reasoning models usually scale sequential computation by generating more tokens, as in chain-of-thought reasoning. Architectures that add latent depth at test time, such as looped or depth-recurrent transformers, have proven harder to train, and the organizers suggested this may reflect optimization knowledge that favors standard transformers rather than a lack of better designs.^[16]

The public task asked models to compute the result of repeatedly squaring a number modulo a semiprime, a computation in which each step depends on the previous one. Models were trained on smaller moduli and step counts and evaluated on larger ones. Training was limited to a single H100 GPU and models to 500 million parameters. Submissions ran on hosted H100s, with Modal supplying GPU runners and Northflank hosting the competition service. The competition closed on August 31, 2026.^[16][17]

The results post, published on the Core Automation blog on September 9, 2026, reported 15,602 submissions. On the hidden Hard task, only 18 uploads from five accounts were certified at any depth on the seen-modulus tests, and the other 143 ranked accounts failed certification at a single step. Entries from two accounts were rejected because they had extracted the training data from the sandbox.^[17] The organizers wrote that none of the top submissions fully complied with the rules. They named Yash Kant the winner for "learning reusable digit operations and composing them across repeated computation", while noting that "the calculator's wiring is supplied, so this is a narrower result than learning the entire algorithm."^[17] All submissions were released as a public dataset on Hugging Face by GPU MODE.^[17][18]

Agent-run businesses and the internship program

In a September 14, 2026 post, Riley Walz described one of the company's first projects. The team set up several small online businesses and put an AI agent in charge of each. Every business had a bank account, a debit card for paying suppliers, a web browser, an email inbox, a domain name, cloud infrastructure and a schedule, and the agent woke every few hours to decide what to do.^[20] Walz wrote that the agents built storefronts and apps, bought ads and made some sales, but "never really stepped back to strategize about what might move the needle." In one case, when a dropshipping supplier ran out of stock on the first order, the agent quickly sent the customer an apology and tried to issue a refund instead of trying one of the many other suppliers.^[20]

To study where human judgment is still needed, the company announced an "Associate General Manager" internship modeled on associate product manager programs. Four interns would spend January through March 2027 in its San Francisco office, run several businesses at once with a budget, tokens and the company's agent infrastructure, and receive $10,000 a month in base pay plus a share of profits. The company wrote: "By watching where people intervene, we want to understand what agents still need to learn."^[19]

Funding

Core Automation has not published details of its funding rounds on its website or blog. The figures below come from press reports and from its SEC filing.

The Information's May 2026 report said the company had raised $100 million at a $1 billion valuation "just weeks" earlier, according to some of the people it spoke with, and that "Nvidia invested in the earlier round", according to one person. The same report said the company was aiming to raise $300 million to $500 million at a valuation of around $4 billion, according to people who had spoken with its leaders.^[5] Dealroom, summarizing that report in June 2026, said the discussions were early and the terms could change. It also named Spark Capital and Accel alongside Nvidia as backers of the earlier round.^[32] The company has not published a list of its investors.

The Form D, signed by Tworek as chief executive, describes a new equity offering with a first sale on July 15, 2026. It reports a total offering amount of $635,249,754, of which $432,125,053 had been sold to 51 investors and $203,124,701 remained. The sold amount includes $135,252,607 of securities issued on the conversion and cancellation of outstanding SAFEs (simple agreements for future equity).^[6] Axios Pro reported on August 13, 2026 that the company was looking to raise up to $500 million in new funding, citing an SEC filing.^[7] The filing does not state a valuation, and The Information's figure of about $4 billion was reported as a target in early talks, not as the price of a completed round.

Reception and context

Coverage placed the company within a wave of new AI research labs. The Information's January report opened by describing "the wave of so-called AI neolabs seeking breakthroughs they think incumbents like OpenAI will miss."^[4] Its May report presented the company as the latest example of brand-new startups taking on Anthropic and OpenAI that raise back-to-back funding rounds, "sometimes in a matter of weeks", a pattern it linked to Nvidia's investments in startups that use its chips.^[5] Dealroom wrote that the jump from $100 million to a potential $500 million in a few weeks suggested that founding-team credentials alone could command multibillion-dollar valuations before a product reached the market. It also pointed out the risks shown by peers that had lost co-founders or leaders.^[32] Business Insider set the company's hiring against the wider competition for AI researchers, in which startups offer equity, co-founder titles, compute access and time for independent research.^[2]

The company has also been compared with other startups betting on alternative learning methods. In January 2026, The Decoder covered it alongside Flapping Airplanes, a startup that had raised $180 million to build AI systems that learn more like humans.^[9] Debarghya Das's "The Ultimate Neolab List" includes Core Automation among the companies it classifies as a frontier lab working on continual learning.^[35]

See also

• Jerry Tworek
• Continual learning
• Test-time training
• Recursive self-improvement
• Thinking Machines Lab
• Flapping Airplanes
• List of neolabs
• Neolab

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Core Automation, Core Automation homepage, accessed October 27, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Lee Chong Ming and Aditi Bharade, Business Insider, New AI lab Core Automation 'nerdsniped' researchers from Anthropic, Google DeepMind, April 22, 2026.
3. ^Matthias Bastian, The Decoder, Ex-OpenAI researcher Jerry Tworek launches Core Automation to build the most automated AI lab in the world, April 22, 2026.
4. ^1 ^2 ^3 ^4 ^5Stephanie Palazzolo and Wayne Ma, The Information, Ex-OpenAI Researcher's Startup Targets Up to $1 Billion in Funding to Develop a New Type of AI, January 28, 2026 (free preview).
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Stephanie Palazzolo and Julia Hornstein, The Information, Ex-OpenAI Researcher's Six-Week-Old Startup Targets Funding at $4 Billion Valuation, May 7, 2026 (free preview).
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16U.S. Securities and Exchange Commission, Form D notice of exempt offering, Core Automation, Inc. (CIK 0002148145), filed July 30, 2026.
7. ^1 ^2 ^3Chris Metinko, Axios Pro, Core Automation is raising $500M, August 13, 2026.
8. ^1 ^2 ^3Matthias Bastian, The Decoder, OpenAI loses top AI researcher Jerry Tworek after seven years, January 5, 2026.
9. ^1 ^2Matthias Bastian, The Decoder, Two startups want to replace how AI learns: one just raised $180M, another is seeking up to $1B, January 29, 2026.
10. ^1 ^2 ^3 ^4Core Automation (@coreauto), Today we're announcing Core Automation, X, April 21, 2026.
11. ^Jerry Tworek (@MillionInt), post quoting the Core Automation announcement, X, April 21, 2026.
12. ^Mark Saroufim (@marksaroufim), post on the Core Automation launch, X, April 21, 2026.
13. ^1 ^2 ^3 ^4 ^5MLSys 2026, When AI Starts Writing Systems Code (keynote by Mark Saroufim, with speaker biography), May 18, 2026.
14. ^1 ^2Mark Saroufim, Core Automation blog, When AI Starts Writing Systems Code, May 28, 2026.
15. ^1 ^2Ehsan Amid, Core Automation blog, How our data shaped neural architecture discovery, and how automation can reshape the future, September 2, 2026.
16. ^1 ^2 ^3 ^4Sean McLeish, Ben Keigwin, Mark Saroufim and Rohan Anil, Tilde Research, (Some of) The Models, They Just Don't Want to Learn, August 3, 2026 (archived October 3, 2026).
17. ^1 ^2 ^3 ^4 ^5 ^6Mark Saroufim, Sean McLeish, Ben Keigwin, Francis Zhang and Rohan Anil, Core Automation blog, What the models learned in One Layer Deeper, September 9, 2026.
18. ^GPU MODE, One Layer Deeper submissions (dataset), Hugging Face, September 7, 2026.
19. ^1 ^2 ^3Core Automation, Associate General Manager Internship, accessed October 27, 2026.
20. ^1 ^2 ^3 ^4 ^5Riley Walz (@rtwlz), post announcing the Associate General Manager internship, X, September 14, 2026.
21. ^1 ^2 ^3 ^4 ^5Sequoia Capital, Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil, Training Data podcast (YouTube, episode description), July 29, 2026.
22. ^1 ^2 ^3 ^4The Information Bottleneck, Continual Learning Is the Next Bottleneck, Rohan Anil (Core Automation), podcast (YouTube, episode description), September 9, 2026.
23. ^AGI House, Can AI Automate AI Research? Jerry Tworek on What's Still Missing, YouTube (video description), July 28, 2026.
24. ^AGI House, Jerry Tworek: How AI Learned to Reason, Why Evals Are Breaking, and What Comes Next, YouTube (video description), July 22, 2026.
25. ^MTS, Most AI Labs are Simply Cooked, Jerry Tworek, YouTube (video description), August 25, 2026.
26. ^Maxwell Zeff, TechCrunch, OpenAI reorganizes research team behind ChatGPT's personality, September 5, 2025.
27. ^Lee Chong Ming, Business Insider (via Yahoo Finance), OpenAI's head of people is leaving to make art about artificial general intelligence, August 22, 2025.
28. ^Julie Bort, TechCrunch, OpenAI's former sales leader joins VC firm Acrew: OpenAI taught her where startups can build a 'moat', January 21, 2026.
29. ^Rohan Anil, Andrew M. Dai, Orhan Firat, et al., PaLM 2 Technical Report, arXiv:2305.10403, May 2023.
30. ^Rohan Anil, Vineet Gupta, Tomer Koren, Kevin Regan and Yoram Singer, Scalable Second Order Optimization for Deep Learning, arXiv:2002.09018, February 2020.
31. ^Anmol Gulati, James Qin, Chung-Cheng Chiu, et al., Conformer: Convolution-augmented Transformer for Speech Recognition, arXiv:2005.08100, May 2020.
32. ^1 ^2 ^3 ^4 ^5Dealroom News, Core Automation eyes $300M-$500M raise at $4B valuation, six weeks after launch, June 18, 2026.
33. ^Matthias Bastian, The Decoder, Former OpenAI researcher says current AI models can't learn from mistakes, calling it a barrier to AGI, February 2, 2026.
34. ^Ewelina Stój, ITReseller, Były wiceprezes OpenAI zakłada startup. Cel: 100 razy mniej danych, 100% uczenia ciągłego (in Polish), February 2, 2026.
35. ^Debarghya Das, The Ultimate Neolab List, accessed October 27, 2026.
36. ^Core Automation (@coreauto), X profile, accessed October 27, 2026.
37. ^Rohan Anil (@_arohan_), X profile, accessed October 27, 2026.