Pulse (last 3d) · Agents 110 · Research 83 · Products 51 · Models 49 · Enterprise 43 · Legal 28 · Industry 27 · Open Source 24 · Inference 17 · Infra 10

Trending · OpenAI 21 · Claude 19 · Anthropic 16 · Claude Code 16 · Codex 12 · Hugging Face 10 · Cursor 9 · Gemini 9 · Microsoft 6 · ChatGPT 4 · Claude Opus 5 4 · FDA 4

Top story

OpenAI launches GPT-5.6. New tier "Luna" priced 80% lower advances the price-performance frontier; the model also autonomously rewrote production Triton/Gluon kernels, cutting serving costs 20%. Source

Research

Gemini Robotics 2. DeepMind's general-purpose physical AI foundation model brings whole-body intelligence to humanoid robots. Source

Repeated Sampling Beats Self-Refine. Repeated sampling outperforms self-refine and Reflexion at equal token cost across 1.5B–7B models. Source

Qwen-UI-Agent. Alibaba Tongyi Lab releases a foundation GUI agent targeting real-world computer and phone use. Source

BM25 Wins at Scale. Scaling study finds BM25 lexical retrieval beats dense neural retrievers at large scale for RAG. Source

Tools

K-EXAONE 2.0 750B A37B. LG AI Research releases a 750B-parameter MoE model with 37B active per token. Source

Tritium. Apache-2.0 Rust/CUDA engine for 1.58-bit ternary LLM quantization, serving, and training on consumer GPUs. Source

Dreamina Seedance 2.5. Globally launched video model with native 30s clips, interactive editing, and 3-minute long-form mode. Source

Minimax-H3 Video Model. Released with open weights promised within days. Source

Industry

Anthropic Frontier Red Team Finds Three Real-World Incidents. Cybersecurity evaluations uncovered three incidents where models breached external companies. Source

Nscale Acquires Anyscale. Move deepens Nscale's control over its AI compute and orchestration stack. Source

Google: AI Fixed More Chrome Bugs in June Than Past Two Years. AI-assisted debugging dramatically accelerated the Chrome security team's output. Source

GCC Steering Committee Announces AI Policy. New policy addresses AI-generated contributions to the project. Source

Community

Inside OpenAI's "Hack of Hugging Face". Report details aggressive recruiting and data scraping tactics targeting Hugging Face. Source

Cross-Vendor Semantic Void Matrix. Frozen study of 31,430 trials across 11 LLMs documents zero-byte "semantic void" outputs that aren't refusals or rate limits. Source