Descriptions:
Lex Fridman convenes Sebastian Raschka — author of “Build a Large Language Model from Scratch” and “Build a Reasoning Model from Scratch” — and Nathan Lambert, post-training lead at the Allen Institute for AI and author of the definitive text on Reinforcement Learning from Human Feedback, for a comprehensive review of where artificial intelligence stands in early 2026.
The conversation uses January 2025’s DeepSeek R1 release as its dividing line: the moment an open-weight Chinese model achieved near-frontier reasoning performance at dramatically lower cost, reshaping competitive assumptions industry-wide. From there, Raschka and Lambert examine who is winning the global AI race — contrasting Anthropic’s Claude Opus 4.5 (which has driven extraordinary organic community excitement), Google’s Gemini 3, and OpenAI with the accelerating output from Chinese developers. They dig into the technical substance driving these results: how Reinforcement Learning from Verifiable Rewards (RLVR) works mechanistically, what it actually unlocks versus what is already baked into pretraining weights, illustrated with a striking concrete example of Qwen 3 base jumping from 15% to 50% accuracy on the MATH-500 benchmark in just 50 training steps. This leads to a sharp debate about data contamination in Qwen evaluations and what such rapid gains actually prove.
Scaling laws, the economics of reasoning models with long output contexts, and emerging agent capabilities round out an episode that stands as one of the most technically rigorous and accessible State-of-AI reviews available for 2026.
📺 Source: Lex Fridman
🏷️ Format: Deep Dive
