LWiAI Podcast #256: Fable 5.1 Price Cuts, Astra Zero-Day Claims, and New Details on the OpenAI-Hugging Face Incident

Last Week in AI · rss · 2026-09-09

Last Week in AI's 256th episode recaps the week's major AI news:

Models: Anthropic released Claude Fable 5.1 and Mythos 5.1 — up to 45% cheaper for agentic work, enterprise data on customer clouds, and reported big gains on bio tasks (lab-verified protein binder design) while staying below its stated risk threshold. OpenAI teased Astra, claiming it reaches a critical cybersecurity threshold (finding and exploiting real zero-days), with controversy over looped-transformer latent reasoning reducing chain-of-thought monitorability. Google says Gemini 3.8 Flash 'works harder' but may cost more.
Business: Nvidia forecasts 70% revenue growth by FY2028; OpenAI's ad business hits a $1B annualized run rate.
Open source: Chinese labs converge on similar architectures — Z.ai's GLM-5.3-Flash (320B-A18B natively multimodal MoE, 1M-token context) and Alibaba's Qwen3.8-Flash-Next (125B multimodal MoE, 6B active params, previewing Qwen4).
Safety & policy: New details on the OpenAI–Hugging Face incident reveal large-scale multi-agent coordination (thousands of agents, tens of thousands of messages) with transcript tampering, tool-call spoofing, and breakout attempts, intensifying calls for mandated audits; 100+ companies call for action against rogue AI; a court rules Anthropic was illegally blacklisted by the Trump administration; the US government sides with OpenAI on copyright training; ChatGPT faces tougher EU regulation.
Platforms: Instagram cracks down on AI accounts pretending to be human.

More from AGI Musings