Pulse (last 3d) · Agents 77 · Research 56 · Products 45 · Models 34 · Open Source 33 · Industry 26 · Inference 17 · Enterprise 17 · Releases 12 · Legal 12
Trending · OpenAI 14 · Anthropic 13 · Claude 9 · Hugging Face 9 · Claude Code 7 · Claude Opus 5 7 · Gemini 7 · ChatGPT 5 · Fable 5 5 · llama.cpp 5 · DeepSeek 4 · Google 4
Top story
Kimi K3 Goes Open-Weight. Moonshot AI released Kimi K3 (2.8T params, 1M context) as the largest open-weight Chinese model, with weights publicly available alongside inference support on Telnyx and llama.cpp. Kimi K3 Weights · Telnyx Inference · llama.cpp PR
Research
FlowEvo: Compile Successful Agent Traces Into Reusable Executable Skills - Training-free workflow-to-skill compilation loop that turns successful agent traces into persistent skills with contrastive-utility governance. ArXiv
DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data - Learns per-example curation policies for pretraining data mixtures. ArXiv
A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever - A frozen 12B model hits 100% on a verified benchmark with no generation, critiquing static evaluations. HuggingFace
PerceptionBench: Vision Benchmark of 10 Atomic Perceptual Capabilities - Kimi's vision benchmark decomposing failures into 10 atomic perceptual skills mined from 42 existing benchmarks. X Post
Tools
Grok 4.5 - xAI's latest model trained on coding, science, engineering, and math datasets with intelligent and efficient reasoning modes. Product Hunt
Claude Opus 5 - Step-change improvement for long-running agents with gains in coding and professional work. Product Hunt
Microsoft's First Cybersecurity Model + Agentic Security System - Microsoft debuts a dedicated cyber model alongside an agentic cybersecurity system. TechCrunch
Webhound - Budget-controlled research agent returning cited reports after spending a user-specified dollar amount following leads. Product Hunt
whatbroke - Diffs two agent runs showing tool calls, arguments, cost, and output divergence; integrates with CI. Product Hunt
Industry
Nvidia + Microsoft Launch Open Secure AI Alliance Without OpenAI, Google, Anthropic - A new open AI security coalition notably excludes the leading frontier model labs. The Verge
Anthropic's Dario Amodei Responds on Open Weights - Clarifies he doesn't oppose open weights broadly but raises concerns about Chinese AI competition. TechCrunch
Why China is Giving Away Its Best AI Models - The Verge analysis on the strategic rationale behind Chinese labs open-sourcing frontier models. The Verge
$500 RL Fine-Tune of a 9B Model Beats Frontier on Catalog Review - Sparks debate about the economics of large general models vs cheap specialized ones. Hacker News
Community
Private Claude Chats Exposed on Google Search - User-shared Claude chats and Artifacts were inadvertently indexed by Google, exposing private data. Reddit
OpenAI's Hugging Face Breach Reignites Alignment Debate - An OpenAI-related Hugging Face incident reopens questions about alignment and control. TechCrunch
Frontier LLM Bias Benchmark (6 Models, ~20.6K Examples) - Independent evaluation finds broad left-leaning behavior across GPT-5.4, Claude Sonnet 4.6, Opus 4.7, Gemini, and Grok 4.3. Reddit
AI Labs Buying and Destroying Antique Books for Training Data - Investigation reveals labs ingesting rare and sometimes nearly-extinct books, then destroying them. Reddit