Lex Fridman 深度访谈

Lex, Sebastian Raschka and Nathan Lambert discuss the DeepSeek R1 release in January 2025 that surprised everyone with near-state-of-the-art performance at much lower cost. They debate who's winning the international AI competition between China and the US.

Discussion of the shifting hype between AI model releases — Gemini 3's launch excitement versus Claude Opus 4.5's growing organic hype. Anthropic's cultural advantage in betting hard on code and maintaining organizational stability.

Deep dive into post-training techniques — SFT, RLHF, Constitutional AI, reasoning, and inference-time scaling. Nathan explains why RLHF is fundamentally unsolvable because preferences cannot be cleanly quantified, drawing on social choice theory and economics.

Reflections on the value of struggle in learning. Both guests emphasize that if you're not struggling, you're not fully learning. Discussion of training AI models for education that don't give all answers at once.

Comprehensive overview of AI's current landscape — no company has exclusive technology access, differentiation comes from budget and hardware. The fluid movement of researchers between labs means ideas flow freely, but implementation resources remain the bottleneck.