Stanford Professors Launch Noeri, an AI Lab Backed by $110M from Nvidia
Most AI models stop learning once training ends, so companies face a choice: overpay for a top-tier model or accept worse results from a cheaper one. Noeri, a lab founded by Stanford professors Yejin Choi and Carlos Guestrin, raised $110M to change that. Its system studies a client's data and reconfigures the task itself, while a human approves the changes the way they would in a code review. For teams paying for expensive APIs, this is a path to savings.
AI-processed from TNW; edited by Hamidun News
Almost everyone who has deployed AI in production has faced the same fork: use the strongest model and pay for every request, or save money and accept noticeably worse results. The new lab Noeri offers a third option. It was founded this year by Stanford professors Yejin Choi and Carlos Guestrin, and it raised $110M at launch.
Menlo Ventures led the round. Madrona Venture Group, Kindred Ventures, Wing Venture Capital and NEA also took part. Cisco Investments, Nvidia and Snowflake came in as strategic investors. Bloomberg was the first to report the deal. Noeri has not disclosed its valuation.
How Noeri wants to teach a system to learn on the job
The core idea is simple. Most models stop acquiring new knowledge once training is complete, while Noeri's systems, according to the company, keep improving as they are used.
The mechanics work like this. The AI studies a specific company's data and workflows. It then changes not only the model itself but everything around it: instructions, the set of tools, the choice of model for each step and, if it pays off, the models themselves. Picture a call center. The system may notice that a small, cheap model handles simple inquiries without trouble, and that the expensive model should be called only for complex disputes. It will rewrite the prompts and redistribute the steps.
Every change comes with an explanation, and a human approves it, as in a code review. This is an important detail: automatic optimization without oversight worries the people responsible for reliability. Here, an engineer can see exactly what changed and why, and can reject the edit.
The system works with both closed models and models with open weights. Data, workflows and the resulting models remain the client's property.
A small model that manages large ones
The idea partly grew out of work with a team of Stanford microbiologists and immunologists. They needed to use OpenAI's models more cheaply. Choi told Bloomberg that their small AI learned to use other models and, on that team's tasks, outperformed GPT-5. It did, however, call GPT-4 rather than GPT-5: according to her, Stanford has too little money to run the newest model all the time.
This is still the founder's account, not a published benchmark. The messages contain no verifiable figures on how much the system saves money or improves quality.
The second part of Noeri's plans concerns science. The company wants AI to work alongside scientists: proposing hypotheses, running experiments and learning from the results. Example tasks include new molecules, proteins and complex technical systems. In the company's view, small specialized models make that kind of search affordable.
Who stands behind the project and what it means for practitioners
Yejin Choi is a MacArthur fellow, a former professor at the University of Washington and senior director at the Allen Institute for AI. Bloomberg writes that she was also a distinguished scientist at Nvidia. Carlos Guestrin founded Turi, a machine learning platform that Apple acquired in 2016, and later led Apple's central machine learning and AI organization. He is also one of the creators of the open-source XGBoost library.
Noeri is headquartered in Palo Alto, with an office in Zurich, and the company is hiring researchers and engineers in both cities. It already works with partners across several industries, including some of its investors.
Guestrin explained the company's position to Bloomberg: both founders are concerned about centralization, where a single company owns everything. Hence the emphasis on compatibility with third-party models and on the client keeping control.
For developers and product managers, the takeaway is this. Choosing a model is increasingly part of the system itself rather than a one-time decision. It is already worth measuring quality and cost at each step of a workflow, keeping prompts and tools under version control, and checking where an expensive model can be replaced with a cheaper one. If Noeri delivers on its promises, such optimization will become automatic, with humans keeping the final say. For now, the company has money, a strong team and a pitch. Measurable results for customers are still to come.
- Nov 21, 2026Nvidia doubled profit, but shares fell 2% over the search for a new CEO
- Nov 21, 2026Nous Research fine-tuned China's Qwen3.8, and propaganda in its answers dropped from 22% to 1%
- Nov 21, 2026Samsung Wants to Buy Modine, a Data Center Cooling Supplier, for Nearly $6 Billion
- Nov 21, 2026Reuters: China's $3.2B AI Data Center Buildout Relies on Nvidia Chips as Huang Steps Down
Key Questions Answered
How Noeri wants to teach a system to learn on the job?
The core idea is simple. Most models stop acquiring new knowledge once training is complete, while Noeri's systems, according to the company, keep improving as they are used. The mechanics work like this. The AI studies a specific company's data and workflows. It then changes not only the model itself but everything around it: instructions, the set of tools, the choice of model for each step and, if it pays off, the models themselves. Picture a call center.
What are the key details on A small model that manages large ones?
The idea partly grew out of work with a team of Stanford microbiologists and immunologists. They needed to use OpenAI's models more cheaply. Choi told Bloomberg that their small AI learned to use other models and, on that team's tasks, outperformed GPT-5. It did, however, call GPT-4 rather than GPT-5: according to her, Stanford has too little money to run the newest model all the time. This is still the founder's account, not a published benchmark.
Who stands behind the project and what it means for practitioners?
Yejin Choi is a MacArthur fellow, a former professor at the University of Washington and senior director at the Allen Institute for AI. Bloomberg writes that she was also a distinguished scientist at Nvidia. Carlos Guestrin founded Turi, a machine learning platform that Apple acquired in 2016, and later led Apple's central machine learning and AI organization. He is also one of the creators of the open-source XGBoost library.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.