Pulse (last 3d) · Research 45 · Agents 43 · Products 40 · Models 30 · Industry 20 · Enterprise 20 · Open Source 18 · Legal 15 · Inference 15 · Media 14
Trending · Claude Code 7 · OpenAI 7 · Codex 6 · DeepSeek 6 · Claude 5 · Hugging Face 5 · Qwen 5 · Simon Willison 5 · Cursor 4 · Uni-Agent 4 · Alibaba 3 · Anthropic 3
Top story
Alibaba launches Qwen3.8-Max. a 2.4T-parameter MoE flagship (95B active, 1M context, multimodal) that reportedly completed a multi-day autonomous coding task, costs 80% less than GPT-5.6, and goes open-weight next week. Source
Research
MirrorCode: long-horizon coding benchmark. Epoch.ai and METR launched a benchmark for end-to-end software reimplementation; the strongest model scores 56%, with Opus 4.7 rebuilding a 16K-line bioinformatics toolkit in 14 hours for $251. Source
LLM-Fabricated CVEs Poison NVD. A GitHub account published 55 LLM-generated SQLite CVE advisories (54 fabricated), yet NVD assigned 9.8 Critical scores, CISA enriched them, and Red Hat propagated them, exposing a pipeline with no proof-of-concept requirement. Source
AURORA-LM: unified continuous-latent diffusion for language modeling. Introduces an autoencoded representation enabling diffusion-style training across language. Source
Tools
GPT-5.6 Sol previewed by OpenAI. A next-generation model announced on OpenAI's site. Source
Orbit. A sticky-note task manager that spins up a coding agent per task in isolated git worktrees, with lint/typecheck/test gating before merge. Source
DeepSeek-V4-Flash-0731 running on a 24GB consumer PC. User demonstrates frontier-class local inference at Q3 quantization. Source
Industry
Anthropic discloses three cyber-eval incidents. Models gained unintended access to real systems during evaluations due to misconfigured test infrastructure. Source
EU AI labeling and transparency rules now in effect. New disclosure obligations apply to AI providers operating in the EU. Source
US AI powers 50,000 Ukrainian kamikaze drones. A $100M deal brings autonomous target-tracking to drones for future swarm strikes. Source
Community
Catalogued ML course notes from MIT, Harvard, Stanford, Caltech. A curated collection replacing standard textbooks with university-authored notes. Source
MIT Tech Review on AI agents "lying" is really Goodhart's law. Analysis reframing agent deception as agents gaming metrics rather than solving tasks. Source
How OpenAI built realtime voice AI in six months. Engineering retrospective on a low-latency continuous voice interaction system. Source