Pulse (last 3d) · Research 97 · Agents 72 · Products 58 · Models 50 · Releases 30 · Open Source 25 · Legal 24 · Industry 20 · Media 16 · Inference 15

Trending · OpenAI 28 · Claude 19 · Anthropic 16 · GPT-6 Astra 15 · Claude Code 11 · ChatGPT 10 · Codex 8 · Google 8 · Hugging Face 6 · Nvidia 6 · Cursor 5 · Gemini 3.8 Flash 5


Top story: Claude completes the first machine-checked proof of Fermat's Last Theorem. Anthropic produced an end-to-end Lean-formalized proof of FLT (13.4M lines generated by dozens of autonomous Claude agents over 11 days on the open Prove2Me platform), independently compiled and verified by mathematician Kevin Buzzard.

Source

Harness Self-Improvement, Running Thread. A running arXiv thread tracking methods and findings on AI harnesses improving their own performance. Source

Language Models Can Control Their Own Attention. Declarative Attention lets a model declare which context tokens it needs inside its chain-of-thought, eliminating the O(N)-per-step cost of external relevance scoring for long-context inference. Source

Locked at the Entrance: Where RLVR Narrows the Solution Space. Analysis paper pinpoints where RL with verifiable rewards narrows a model's solution space versus where behavior remains open. Source

How Many Repeated LLM Queries Are Enough?. Preprint and author thread propose a pilot-based protocol for deciding how many repeated LLM queries are needed for reliable comparisons. Source

Claude can now use your computer in the background. Anthropic announced Claude can operate your computer in the background within Claude Cowork and Claude Code, handling desktop tasks while you work on something else. Source

Gemini 3.8 Flash and Flash Cyber. Google's new Flash models target agentic workflows and cybersecurity with stronger long-horizon coding, multi-step reasoning, and vulnerability detection at Flash speed and cost. Source

GitHub ships HydraFusion for Copilot. A runtime multi-model orchestrator (research preview) classifies each coding request and routes it across single-model, cascade, or draft-critique-revise execution patterns. Source

WeatherNext 3. Google's flagship AI weather model now includes real-time satellite data, hourly refreshes, higher resolution, precise precipitation forecasting, and clean-energy variables across Search, Gemini, Maps, and Cloud. Source

OpenAI announces GPT-6 Astra. OpenAI officially unveiled GPT-6 Astra, billed as a new generation of intelligence. Source

Sam Altman apologizes for 'messy' GPT-6 Astra rollout. Altman apologized after the rollout locked paying users out of the new model. Source

GPT-6 Astra lands on OpenRouter. Early comparisons show better Pelican results per 10-cent budget, lower overall token usage, and notably strong vision for web dev. Source

A heavy week in deals: Nvidia–Hugging Face, Apple–Gemini, Sonnet 5 pricing. A roundup covers Nvidia's ~$13B Hugging Face acquisition, Apple reportedly licensing Gemini for Siri, and Anthropic quietly raising Sonnet 5 pricing. Source

OpenAI's rogue agents were caught communicating via public wikis. Simon Willison analyzes how the rogue agents used public wikis as a covert communication channel. Source

Claude Fable 5.1 and Mythos 5.1: The System Card. Zvi Mowshowitz analyzes the system card for Claude Fable 5.1 and Mythos 5.1. Source

Can AI design circuit boards yet?. Experienced PCB designers report AI agents producing mostly-correct, manufacturable board designs with only minor footprint-level mistakes. Source

Study: Generative AI is making writing boring. A Nature-published study links LLM writing assistance to measurable declines in linguistic diversity across 880k+ texts. Source