Noeri raises $110M for self-improving AI you own
Noeri, the Stanford AI lab from Yejin Choi and Carlos Guestrin, raised $110M for self-improving AI that companies own. No price yet. Build your evals first.
Noeri, a new AI lab from Stanford professors Yejin Choi and Carlos Guestrin, launched today with $110 million to build self-improving AI that companies own. The pitch is aimed straight at the bill every AI user pays: buy the most expensive model, or accept worse results. Noeri says there is a third option. The claim is big. It is also untested in public, and there is no price.
What actually happened
Bloomberg first reported the round, and The Next Web covered the launch. The facts:
- Money. $110 million led by Menlo Ventures, with Madrona, Kindred Ventures, Wing Venture Capital and NEA. Cisco Investments, Nvidia and Snowflake came in as strategic investors. No valuation and no round stage disclosed.
- Team. Choi and Guestrin are co-CEOs. Choi is a MacArthur Fellow and, per Bloomberg, was most recently a distinguished scientist at Nvidia. Guestrin co-created XGBoost and co-founded Turi, which Apple bought in 2016, then led Apple's central machine learning organization. Headquarters in Palo Alto, office in Zurich.
- Product. The system studies a company's own data and workflows, then changes the setup around a task: the instructions, the tools, which model handles each step, and in some cases the models themselves. Each change comes with an explanation, and a person approves it, like a code review. It works with closed and open-weight models.
- Ownership. Customers keep their data and the workflows and models they build, the company said.
- Proof. One anecdote. Choi told Bloomberg a small Noeri system that learned to use other models beat GPT-5 on one Stanford research team's needs. That is her claim. No public benchmark backs it yet.
No pricing, no general availability date and no named customers. Noeri says it works with partners across industries, including some of its investors.
Why self-improving AI matters for your business
Strip out the lab language and Noeri is selling something we already do by hand for clients: route each step to the cheapest model that passes, tune the prompts and tools, and keep a human on the merge button. That is the right shape. Change proposals with a reason attached, reviewed by a person, beats an agent that quietly rewrites itself.
But "self-improving" only means something against a test. A system can only get better at what you measure. If you have no eval set, a self-improving tool improves against its own idea of good.
So before you sign with Noeri or anyone selling the same promise:
- Write the test first. 50 to 200 real tasks from your business with known good answers. That set is yours, and it travels to any vendor.
- Log cost per task today. If you cannot show the current number, you cannot show savings later.
- Get the ownership clause in writing. "Customers keep the models they build" should mean you can export the weights, prompts and routing rules and run them somewhere else.
- Keep approval on. The review step is the feature. Do not trade it for speed.
Key takeaways
- Noeri launched with $110M led by Menlo Ventures; Nvidia, Cisco and Snowflake invested
- It rewrites prompts, tools, model routing and sometimes models, with a human approving each change
- Customers keep their data, workflows and models, the company says
- No price, availability date, customers or public benchmarks yet
- Build your own eval set and cost-per-task baseline before you buy any self-improving AI
Self-improving needs a scoreboard you own. We build eval sets, model routing and cost tracking on your infrastructure, so any vendor has to beat your numbers, not theirs. See how we build it, or estimate your cost per task.
Sources: The Next Web, Bloomberg via Techmeme.
- #self-improving-ai
- #funding
- #ai-ownership
- #model-routing
- #evals
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
SuperDial: AI voice agents now call AI voice agents
SuperDial's Agent-to-Agent RCM tries EDI, APIs and portals before any phone call, because AI voice agents keep calling each other. Give agents a better door.
Read itSpara Scout AI sales assistant: drafts in, rep hits send
Spara's Scout AI sales assistant watches CRM and buying signals, drafts outreach in the rep's inbox and waits for Send. Free claim, early-access page. Pilot first.
Read it