Nathan Lambert and Florian Brand discuss the current state of open AI models following the Kimi K3 release. They cover the performance gap between open and closed models, why Chinese labs are producing competitive models (capital efficiency, focused teams, increasing compute access), and a roundup of Chinese providers including Kimi, GLM, Qwen, DeepSeek, and MiniMax. The conversation addresses the US open-model ecosystem with players like Thinking Machines, Poolside, and Nvidia. A significant portion debates distillation — specifically pushing back on Ben Thompson's claim that distillation becomes more impactful during RL training, arguing instead that SFT distillation has limited impact and RL-stage distillation is prohibitively expensive. They also discuss cybersecurity implications of potentially banning Chinese open-weight models in the US, and close with predictions on which labs will reach frontier status by year-end.
