Wink Pings

Two Stanford Professors Raise $110 Million to Let AI Get Stronger on Its Own Inside Companies

Noeri comes out of stealth and announces $110 million in funding. What Yejin Choi and Carlos Guestrin want to build is an AI system that can continuously learn in real work, build its own tools, and select models by problem. A single superintelligence is their opposite.

Noeri came out of stealth today and announced $110 million in funding. Menlo Ventures led the round, with participation from Madrona, NVIDIA, NEA, Kindred Ventures, Wing VC, AIX Ventures, Principal VC, Cisco, Snowflake, and others. The company also announced its product and research direction: enable enterprises to build specialized AI using their own data and experience, and let it keep getting stronger in production environments.

Noeri发布

Yejin Choi and Carlos Guestrin are co-founders. In a tweet, Yejin framed the question directly: should the future of AI be monolithic or plural? Her answer leans toward the latter. The name Noeri is explained as 'the plural of intelligence,' or 'many intelligences.' The website vision states it more concretely: if the future ends up with only one superintelligence owned by a few people, its blind spots will become the world's blind spots.

The two founders have serious backgrounds. Yejin is a professor of computer science at Stanford, a senior fellow at HAI, has led research at Ai2 and NVIDIA, is a MacArthur Fellow, and is a TIME100 AI honoree. Menlo's Deedy added in a tweet that she is one of about 30 computer scientists who have received that fellowship. Carlos is a Stanford professor, director of the Stanford AI Lab, co-creator of XGBoost, and a member of the National Academy of Engineering. Noeri is his seventh startup; his previous company, Turi, was acquired by Apple in 2016, after which he led machine learning and AI at Apple. Both have h-indices above 100.

Noeri官网首页

The product page breaks self-improvement into four things. Learning on the job: every task failure and correction becomes a training signal, requiring very little labeled data. Building its own tools: it modifies prompts as well as skills, tools, and orchestration, and creates new tools when existing ones are insufficient. Training its own models: train models on the enterprise's own work so that each task runs on a more suitable model. Tracking returns: quality and cost are measured using the enterprise's own traffic and metrics, not public leaderboards. Yejin's tweet also mentions choosing a right-sized model for each problem, at a cost that can be a fraction of frontier models.

在岗学习

The research page lists four words: Create, Own, Evolve, Discover. In plain terms: let the system optimize itself, generate data and curricula, and build tools; use distillation and lightweight updates to master domains with little data and compute; continue learning after deployment without forgetting; and use search, reasoning, and multi-model teams to propose, test, and correct answers. The team roster includes Prithviraj Ammanabrolu, Danqi Chen, Andreas Krause, and others, with offices in Palo Alto and Zurich.

Yejin Choi
Carlos Guestrin

Among community reactions, Snowflake's official account said that a plural-AI future will require enterprises to be able to run many models and agents on their own data. Wall Street Mav's comment was more direct: companies should own AI, not rent the same model everyone else uses.

There are also doubts. Lawrence W. Abrams argued that Noeri's pitch sounds like entering a crowded full-stack inference market, and that the ultimate asset is a post-trained open-weight model owned by the enterprise and deployed on-prem or on a cloud VM, with the issue being the asset's useful life; he also suggested that a single LLM with a continuous RL harness acting as a 'contract worker' might be cheaper. This skepticism deserves to be put on the table. What Noeri needs to prove is that the gains from continuous learning can cover the costs of maintaining multiple models, multiple tools, and continuous training.

Menlo partner Deedy's launch tweet: https://x.com/deedydas/status/2122381278016360805
Noeri website: https://www.noeri.ai/

发布时间: 2026-11-17 05:27