Nathan Lambert

Nathan Lambert × Sebastian Raschka

Anthropic’s distillation warning turns frontier API access into a geopolitical and capability-control issue.Nathan Lambert argues GPU-constrained Chinese labs “obviously should do this”: buying Claude outputs is easier than generating comparable synthetic data internally, while Anthropic labels distributed collection an “attack.”Terms of service mainly let providers terminate access, but renewed enforcement raises the possibility…

Lex Fridman × Nathan Lambert × Sebastian Raschka

The 2026 AI race is becoming plural, with ideas spreading across DeepSeek, Qwen, Kimi, MiniMax, Z.ai, Google, OpenAI and Anthropic while compute, hardware access, culture and distribution determine advantage.Scaling laws now span pre-training, post-training and inference, making coding the clearest monetization wedge and open weights strategic infrastructure; gigawatt-scale Blackwell clusters could support longer RL runs and premium inference, but data rights, benchmark contamination and serving economics remain risks.

Nathan Lambert

RLVR is emerging as a reusable post-training layer for behaviors with checkable outcomes, with its likely moat shifting toward data, infrastructure, environments, and reward design rather than one algorithm.The next bottleneck is long-horizon behavior: proprietary interaction data, search-native tool learning, planning, memory, calibration, and stronger verifiers could improve inference-time scaling, while overoptimization remains a risk as models exploit measured rewards instead of user intent.

Lex Fridman × Dylan Patel × Nathan Lambert

DeepSeek’s V3 and R1 reset the AI cost curve through reinforcement learning, open weights, mixture-of-experts routing, and MLA, not a mythical $5 million frontier model.With roughly 37 billion of 600-plus billion parameters active per token and custom H800 scheduling, efficiency may expand inference demand, while export controls, TSMC, power, and cooling remain decisive constraints.