Article
A small AI commanding older models unexpectedly defeated GPT-5 in a research task: Stanford professors' Noeri secures $110,000,000 in funding
Breaking News: AI startup Noeri, founded by two top Stanford AI professors Yejin Choi and Carlos Guestrin, has officially debuted and announced it has raised $110,000,000. The round was led by Menlo Ventures, with participation from NVIDIA, Cisco, Snowflake, and others.
Both founders have impressive backgrounds:
Yejin Choi previously served as a Distinguished Scientist at NVIDIA and received the MacArthur Fellowship—known as the "Genius Grant"—in 2022. She has long studied how to enable smaller models to perform complex tasks at lower cost.
Carlos Guestrin is the co-developer of the renowned 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 teams.
Noeri originated when a Stanford research team found GPT-5 too expensive. Choi told Bloomberg that a microbiology and immunology research team at Stanford deemed GPT-5 prohibitively costly, so they trained a smaller AI to learn how to orchestrate older models like GPT-4 for research tasks. Surprisingly, this combination outperformed GPT-5 on certain specific tasks.
Noeri plans to build on this approach. It aims to help enterprises develop their own specialized AIs that continuously learn from real-world operations. For example, if a particular task frequently yields errors, the AI can analyze the causes, adjust its tools and workflows, and even retrain underlying models when necessary. Changes undergo human review before being deployed.
Other players are already exploring similar directions. Japan's Sakana AI launched Fugu this June, also training dedicated models to command other AIs and accomplish complex tasks. Sakana is likewise researching AI self-improvement.
Noeri places greater emphasis on applying these capabilities within enterprises, enabling AIs to learn over time from proprietary data and operational experience, gradually forming cheaper, more specialized models better suited to specific jobs. Enterprises can also retain ownership of their developed models and workflows.