Last Week in AI Last Week in AI

#257 - GPT 6 Astra, AI Extinction, Security Incidents

Sep 24, 2026 · 1h 34m

Summary

Hosts Andrik and Jeremy discuss the release of GPT-6 Astra, noting its superior coding performance and controversial "loop transformer" architecture that complicates safety monitoring. They debate the validity of OpenAI’s alignment claims, arguing that current evaluations are prone to overfitting and eval-awareness, while also analyzing Anthropic’s call to pace frontier AI development.

Topics discussed

Intro, host updates, and episode format Sponsor segments for Box and Factor GPT-5.6 Astra release and capabilities Model architecture, pre-training, and benchmarks Cyber safety evals and alignment skepticism Anthropic's Hacker Opus and alignment limits Dario Amodei's blog post and China policy Meta's Muse Spark and AI agents Sponsor segments for ODSC and Progressive Congressional calls for AI regulation White House AI policy and US-China tensions OpenAI incident reporting and rogue AI events Model misbehavior and reward hacking examples Security researchers hack OpenAI via cloud Closing thoughts and political neutrality

More from Last Week in AI

#261 - (belated) Argon, Muse Spark, Anthropic S-1
Nov 20, 2026
#260 - Muse, Misalignment, Math
Oct 19, 2026
#259 - Dots, Sonnet, Self-Safety, Rogue AI
Oct 9, 2026
#258 - Opus 5.5, Sol and Luna, Muse, DeepSeek-V4.1-Flash, Xi
Oct 2, 2026
#256 - Fable 5.1, Astra Tease, Gemini 3.8 Flash
Sep 8, 2026
Listen ad-free on Castria