#228 - GPT 5.2, Scaling Agents, Weird Generalization
Summary
This episode covers GPT-5.2’s benchmark performance and enterprise focus, Runway’s new world models, and Tencent’s Hunyuan 2.0. It also discusses Disney’s $1B OpenAI investment, NVIDIA chip export tensions, and Unconventional AI’s massive seed round. Finally, the hosts analyze Claude’s internal “soul” document and new research on scaling agent systems.
Topics discussed
Sponsors: Box and SilverMirror
Intro and episode overview
OpenAI GPT-5.2 release and benchmarks
Runway Gen-4 and robotics models
US-China chip export controls and tariffs
China's domestic semiconductor efforts
OpenAI, Disney, and copyright issues
Unconventional Labs and analog AI chips
Slack revenue and Anthropic valuation
FAQS leaderboard and fact-checking
Anthropic's Claude 4.5 and identity
Multi-agent systems and scaling laws
Video-based world models for robotics
Self-play and synthetic data training
LLM reasoning and belief entrenchment
Reinforcement learning and training stability
DeepMind UK and US AI regulation
Emergent capabilities and safety risks
Compute saturation and AGI timelines
AI security institutes and chip smuggling
Listen ad-free on Castria