AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

AI Is Crossing the Frontier of Human Knowledge - Kevin Weil on AI for Science

· The a16z Show

Kevin Weil, formerly Chief Product Officer and VP of Science at OpenAI, joins the a16z Show to make the case that AI's most consequential impact won't be productivity but scientific discovery. He describes leading the newly formed "OpenAI for Science" group, which is focused on using frontier models to solve problems beyond the frontier of human knowledge—pointing to roughly 10-12 open mathematics problems solved in January alone, mostly by GPT-5.2, as early evidence. Weil argues we are in the "glimmer phase" for AI in science, where models can almost do things that previously seemed impossible, and predicts a rapid progression similar to coding capabilities over the past 18 months. Weil outlines a vision in which AI accelerates breakthroughs in math, physics, materials science, medicine, and fusion—essentially bringing "the science of 2050 to 2030." The approach combines in-silico reasoning with experimental validation through robotic labs operating on RL feedback loops, where models design experiments, run simulations, send refined parameters to scalable physical labs, and iterate on results. He emphasizes that frontier reasoning must extend from minutes to hours, days, and months for these harder problems. On product and the broader ecosystem, Weil discusses recent launches including Prism (an AI-native scientific writing tool) and the apps platform for distribution. He highlights Open Claw as a surprising, emergent example of AI agents collaborating, built in just three days on Codex. For founders, he advises using ensembles of models rather than single-prompt approaches, notes that enterprise/B2B is leading this wave (in contrast to past tech shifts) because models already perform economically valuable work and offset inference costs. He is bullish on startups broadly because OpenAI itself cannot anticipate all the use cases unlocked by emergent model capabilities.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.