Skip to content

Small AI directing old models unexpectedly beats GPT-5 in research tasks: Stanford professor-founded Noeri secures $110 million in funding.

Nov 17, 17:49

Beating AI News Flash: Noeri, an AI company founded by two top Stanford AI professors, Yejin Choi and Carlos Guestrin, has officially launched and announced $110 million in funding. Menlo Ventures led the round, with participation from Nvidia, Cisco, Snowflake, and others.

Both founders have impressive backgrounds:

Yejin Choi was formerly a Distinguished Scientist at Nvidia and received the MacArthur Fellowship, known as the "Genius Grant," in 2022. She has long researched how to enable small models to complete complex tasks at lower cost.

Carlos Guestrin is a co-developer of the well-known open-source machine learning tool XGBoost. His machine learning company Turi was acquired by Apple in 2016, after which he led Apple's core machine learning and AI team.

Noeri's starting point was that a Stanford research team found GPT-5 too expensive. Choi told Bloomberg that a microbiology and immunology research team at Stanford felt GPT-5 was too expensive, so they trained a small AI to learn to direct older models such as GPT-4 to complete research tasks. The results were unexpected: this combination performed better than GPT-5 on specific tasks.

Noeri plans to continue in this direction. It wants to enable enterprises to cultivate their own specialized AI that continuously learns from actual work. For example, if a certain task frequently goes wrong, the AI can analyze the causes of failure, modify its own tools and workflows, and if necessary even retrain the underlying model. Changes are put into use only after human approval.

Others are already exploring this direction. Japan's Sakana AI launched Fugu in June this year, also training specialized models to direct other AIs to complete complex tasks. Sakana is also researching AI self-improvement.

Noeri focuses more on applying these capabilities to enterprises, enabling AI to learn long-term from a company's own data and work experience, gradually forming cheaper specialized models that are better at specific work. Enterprises can also retain the models and workflows they develop.

Source