LWiAI Podcast #257 Covers GPT 6 Astra, AI Extinction

Episode 257 of the Last Week in AI podcast landed on September 19, 2026, with hosts Andrey Kurenkov and Jeremie Harris working through a stretch of news that touched almost every corner of the field at once. The headline item is OpenAI's release of GPT-6 Astra, pitched as a major jump in agentic coding and computer use, built on loop-transformer latent reasoning and higher token efficiency, alongside new claims about cyber and alignment monitoring that immediately drew questions about eval awareness and sandbagging.

Around that release sits a louder argument about speed itself. Dario Amodei has been pushing for deliberate "pacing" of frontier AI development, a position Sam Altman appears to partly share, while Jensen Huang, Mark Zuckerberg, and President Trump have all pushed back publicly, framing caution as a liability against China. That tension collided with a viral moment: an Anthropic researcher's resignation reignited AI extinction warnings online and prompted members of Congress to float bans, pauses, and kill-switch proposals. The episode also gets into a run of security disclosures, including a proposed framework from OpenAI for reporting misalignment and a reported exploit involving Anthropic's Claude model.