#257 - GPT 6 Astra, AI Extinction, Security Incidents
Summary
Hosts Andrii Verikov and Jeremy Harris discuss the release of GPT-6 Astra, analyzing its advanced capabilities, loop transformer architecture, and controversial alignment claims. They also cover Anthropic CEO Dario Amodei’s call to pace frontier AI development, debating the strategic risks of slowing down in the face of China’s rapid progress. The episode addresses the siloed nature of AI national security efforts and the growing tension between public safety messaging and the reality of model misalignment.
Topics discussed
Intro, host updates, and discussion on episode format
Sponsorships for Box and Factor meals
GPT-5.6 Astra release and computer use capabilities
Model architecture, pre-training vs test-time compute
Cyber safety evals and alignment skepticism
Anthropic's jailbroken model and alignment limits
Super alignment definition and model pricing
Dario Amodei's blog post on China and AI safety
Criticism of effective altruism and AI funding
Meta's Muse Spark and AI agent integration
Sponsorships for ODSC and Progressive Insurance
Congressional calls for AI regulation and safety bills
Political polarization and US-China AI competition
White House internal dynamics and export controls
OpenAI incident reporting and rogue AI marketing
Analysis of model misalignment and reward hacking
Security researchers hack OpenAI via cloud flaws
Cyber defense, model theft risks, and closing thoughts
Listen ad-free on Castria