Opus 5 Releases, China Catches Up, and the OpenAI Model Sandbox Escape
Summary
Theo and Ben discuss the Kimi K3 release, analyzing its frontier-level performance, high token usage, and geopolitical implications for open-weight models. They address the Hugging Face security incident involving an OpenAI model, debating the necessity of unrestricted AI for defense. The episode also covers Opus 5, arguing that labs now prioritize massive frontier models over optimizing smaller ones, creating opportunities for competitors in the mid-tier market.
Topics discussed
Intro: GPT-6 rumors and Opus 5 release
T3 Code: Managing threads and remote machines
Harnesses, Codex, and the future of AI tooling
Kimi K3 vs GPT-5.6 Sol: Benchmarks and efficiency
Hugging Face hack and open-weight model security
AI safety, alignment, and the 'doomsday' scenario
Open-weight models and competition in AI labs
Frontier models: Scaling vs. smaller specialized models
Opus 5 vs Fable: Code quality and behavior
Demo: Opus 5 creating a 3D game clone
Workflow: Using Opus and Fable together
Opus 5 failures, alignment, and closing thoughts
Listen ad-free on Castria