← Home
TWIML · November 3, 2026 · 48m

Can AI Invent or Just Help Scientists Work Faster? with Richard Socher

Richard Socher, former chief scientist at Salesforce and founder of You.com and Recursive, explores the transition from AI assisting researchers to AI autonomously conducting research. The conversation examines Recursive's results in model training and GPU kernel optimization, highlighting the importance of fast verification loops and the risks of reward hacking. Socher also discusses the data and experimental infrastructure required for broader scientific discovery, commercial incentives for alignment, and his vision of a "Eureka machine."

This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.

Preview

•
AI as Researcher, Not Just Assistant
Socher argues that AI is moving from merely helping scientists (e.g., writing code or drafting papers) to autonomously generating ideas and conducting experiments.
•
The conversation highlights how reward hacking – when AI systems exploit training objectives to achieve high scores without genuine progress – poses a significant challenge for AI-driven research.

3 more ideas & all timestamps

This episode is in its early-access window. The full breakdown unlocks free in about 71 hours — Pro members read everything the moment it lands.

Read it now with Pro$10/mo · founding $96/yr
Was this useful?