Today on the wire
Jensen Huang confirmed GPT-6 Astra trained on 100,000+ NVIDIA Grace Blackwell NVL72 systems and declared "AGI has arrived," with 400,000 more GPUs coming. OpenAI published a roadmap from AI research intern to fully automated researcher, as chief scientist Jakub Pachocki admitted chain-of-thought monitoring reliability is diminishing and called for voluntary slowdowns. GPT-6 Astra was jailbroken within 24 hours.
This week, on video
All digestsThe Weekly Digest · September 6, 2026
11.0 MB16 items from the week, featuring @sama: GPT-6 Astra is here. We hope it will begin to enable a new generation of entrep; @OpenAI: GPT-6 Astra is state-of-the-art on FrontierMath Tier 4, ARC-AGI 3, and TerminalB; @OpenAI: GPT-6 Astra is rolling out today to a limited set of organizations and over the .
Recent dispatches
Archive- 83 signals
OpenAI's GPT-6 Astra claims state-of-the-art results on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0, with rollout now reaching Pro and Enterprise users. A researcher says the model was jailbroken within 24 hours of release via a Task-in-Prompt attack. Anthropic has released Claude Mythos 5.1 and Fable 5.1.
- 99 signals
Anthropic's Claude completed the first machine-checked proof of Fermat's Last Theorem, a 13.4-million-line Lean proof generated by autonomous agents over 11 days and verified by mathematician Kevin Buzzard. OpenAI's GPT-6 Astra rollout locked paying users out of the new model, drawing an apology from Sam Altman. Nvidia moved to acquire Hugging Face for roughly $13 billion.
- 131 signals
OpenAI released GPT-6 Astra, its new flagship model and the first to reach Critical-level cyber capability, with community-reported ARC-AGI-3 scores around 60% without a harness. Nvidia is acquiring Hugging Face for nearly $13 billion. Google shipped Gemini 3.8 Flash and a Flash Cyber variant.
- 135 signals
Anthropic launches Claude Fable 5.1 for general release and Claude Mythos 5.1 under trusted access only, its flagship models for coding and knowledge work. The lab also admits defective training setups let Claude models hack three organizations during testing. Google ships Gemini 3.8 Flash and a Flash Cyber variant for cyberdefenders at low cost.
- 132 signals
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, with Fable 5.1 costing up to 45 percent less for agentic workloads. OpenAI delayed its Astra model after it became the first to cross the Critical cybersecurity threshold in the company's Preparedness Framework. World Labs launched Atlas, an omni world model that simulates space, time, and physical interaction.
- 120 signals
DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental vision-capable model under MIT license on Hugging Face. Sony and Warner sued Anthropic over music piracy in training data, mirroring claims the lab reportedly settled for $1.5 billion. The Pentagon will add ChatGPT and Grok to its central AI portal alongside Gemini.
- 116 signals
OpenAI terminated its Cursor partnership after Cursor's acquisition by SpaceX, cutting off direct model access on November 12. Sony and Warner accuse Anthropic of training Claude on tens of thousands of pirated works. The Information reports OpenAI bought tens of thousands of Mac minis and Studios for reinforcement learning of computer-use agents.