Zuckerberg's Second Visit

Dwarkesh Patel has quietly become one of the best interviewers in tech. His second conversation with Mark Zuckerberg proves why.

The interview

Dwarkesh's second Zuckerberg episode covers Llama 4, benchmark gaming, the intelligence explosion thesis, DeepSeek, China, export controls, and Meta's AR glasses — and it never feels like a press tour. The key moment: Zuckerberg's claim that AI will write most of Meta's code within 18 months. Patel pushes on what "most" means, and the resulting exchange reveals the gap between a CEO's vision and the engineering reality. The first interview was good; this one is better because Patel has a year of context to draw on and isn't shy about referencing specific claims from last time.

Is progress slowing?

Arvind Narayanan and Sayash Kapoor published Is AI progress slowing down? — examining reports that leading developers hit walls with their next-generation models. Their conclusion is more nuanced than either side wants: declaring model scaling dead is premature, but industry leaders' constant flip-flopping between "scaling is all you need" and "we need new paradigms" exposes how weak their forecasting actually is. The post's sharpest point: if the people building these systems can't predict what works, maybe we should be less confident in anyone's predictions about where this leads.

The tripwire framework

Holden Karnofsky published If-Then Commitments for AI Risk Reduction through the Carnegie Endowment — a framework where AI developers would precommit to specific safety measures triggered by specific capability thresholds. The paper is pragmatic in a way that most AI safety writing isn't: instead of arguing about whether AGI is 5 or 50 years away, it proposes concrete tripwires that work regardless of timeline. A companion piece sketches what those tripwire capabilities might look like.

The thread

A tech CEO making bold claims, researchers questioning the evidence, and a policy thinker building contingency frameworks. The AI conversation is maturing — not because people agree more, but because the disagreements are getting more specific and more productive.