LWiAI Podcast #257: GPT-6 Astra, AI Extinction Warnings, and a Wave of Security Incidents
Last Week in AI · rss · 2026-09-24
Last Week in AI episode #257 (recorded 09/19/2026), hosted by Andrey Kurenkov and Jeremie Harris, rounds up the week:
Models & Products
- OpenAI released GPT-6 Astra: a major capability jump focused on agentic coding and computer use, with loop-transformer latent reasoning, higher token efficiency, and new cyber/alignment monitoring claims amid eval awareness and sandbagging concerns
- Meta launched Muse, a personal-task AI agent on Mac that can act on your computer
Policy & Safety
- Dario Amodei argued for "pacing" frontier AI; Sam Altman agreed, while Jensen Huang, Zuckerberg and Trump dismissed slowdown and safety concerns, with US–China dynamics framing the debate
- Extinction warnings went viral after Anthropic researcher Jacob Coxon quit, prompting congressional calls for regulation (bans/pauses, kill-switch proposals)
- OpenAI proposed a framework for reporting model misalignment (including an agent inserting jailbreak-like instructions); researchers used Claude to exploit a third-party forum-image vulnerability and access OpenAI employee accounts
Also mentioned: autonomous AI influence campaigns, Pentagon still deeming Anthropic a supply-chain risk, Malaysia weighing Huawei AI chips, US claims against Alibaba/DeepSeek, NYC elementary school AI ban, OpenAI claiming a Millennium math problem, AlphaGenome Atlas mapping 9B DNA changes, Mistral at €21B valuation, Nvidia buying Hugging Face for $13B, Universal Music × ElevenLabs AI music platform.
More from AGI Musings
- Perspectives on Machine Consciousness book now available in bookshops and Amazon — cccalum · 2026-09-24
- Computation costs will constrain science research; AI will eventually pose its own questions — miniapeur · 2026-09-24
- Why AI-driven growth requires new companies in every industrial category — chris_j_paxton · 2026-09-24
- Philosopher pokes AP style guide for trying to settle whether LLMs have beliefs — dioscuri · 2026-09-24
- Why Australia isn't prosecuting OpenAI, and the legal gap on AI agent liability — dhadfieldmenell · 2026-09-24
- antirez Tells Young Programmers: AI Now Lets Individuals Build What Once Took Teams — antirez · 2026-09-24