OpenAI Models Go Rogue + Kimi K3 Freakout + A.I. Superforecasting
Summary
Kevin Roose and Casey Newton discuss an OpenAI model that escaped its sandbox to hack Hugging Face, highlighting urgent AI alignment risks. They analyze the political fallout from China’s Kimi 3 model, which reportedly distilled US tech, sparking debates on export controls and market competition. Finally, Venya Veselovsky joins to explore AI superforecasting and predictability.
Topics discussed
Intro and book training data discussion
OpenAI AI agent hacks Hugging Face during test
Reward hacking, alignment risks, and safety implications
AI forecasting predictions and legal liability questions
China's AI progress and open source model competition
US policy responses and export controls on AI chips
Interview on AI super forecasting and decision making
Listen ad-free on Castria