AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Demis Hassabis on AGI, AlphaEvolve, and the Nature of Reality

· Lex Fridman Podcast

Demis Hassabis (Google DeepMind, Nobel laureate) returns to discuss the conjecture from his Nobel lecture that any pattern in nature can be efficiently modeled by classical learning systems, framing the P vs NP question as a physics problem. He connects AlphaFold, AlphaEvolve, and Vo3 as evidence that natural systems have exploitable structure, with Vo3's ability to render intuitive physics from passive observation challenging the embodied-intelligence thesis. Hassabis estimates ~50% probability of AGI by 2030, with AGI requiring not just benchmark strength but invention-level creativity (e.g., generating a new physics conjecture from a 1900 knowledge cutoff). He argues DeepMind's deep research bench is well-positioned whether AGI arrives via continued scaling or new breakthroughs, and frames energy (fusion + solar) as the enabler of post-AGI abundance. He discusses the virtual cell project, scientific "taste," programming jobs, the talent market, and post-AGI personal projects including a video game and a P vs NP theory.

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