Pulse (last 3d) · Agents 92 · Research 79 · Models 48 · Products 31 · Enterprise 29 · Open Source 28 · Releases 24 · Industry 21 · Legal 18 · Media 17
Trending · OpenAI 11 · Claude 10 · Inkling 10 · Kimi K3 6 · Thinking Machines 6 · Anthropic 4 · Claude Code 4 · Gemini 4 · xAI 4 · AGENTRY FIND 3 · ChatGPT 3 · Codex 3
Top story
Thinking Machines Lab ships Inkling. Mira Murati's Thinking Machines Lab released Inkling, a 975B-total / 41B-active multimodal MoE with open weights (BF16 + NVFP4), vLLM/SGLang/Unsloth integration, and a companion post-training framework called Tinker; it tops the U.S. open-weight leaderboard at 41 on the Artificial Analysis Intelligence Index. HuggingFace
Research
Gemini Diffusion: Block-Parallel Denoising at 1-2k tok/s. Google DeepMind brings discrete-token diffusion to production scale, hitting 1-2k tokens/second via block-parallel denoising. Source
Two Camps of Reasoning Post-Training: Checkable or Not. Reasoning post-training splits on the reward signal: one camp rewards a checkable answer, the other manufactures reward for everything else. Source
ATHENA-R1, Treatment Reasoning Agent Trained with RL Scientific Feedback over 212 Biomedical Tools. A multi-institution effort (Harvard Medical School, Oxford, Mount Sinai, Clalit, Zitnik lab) builds an RL-trained clinical reasoning agent over 212 biomedical tools. arXiv
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget. Enables RL training past 2M-token contexts under a fixed GPU budget. HuggingFace
Tools
Kimi K3. Moonshot AI's 2.8T open-weight model with native vision and 1M context, ranking #3 on the Intelligence Index ahead of Opus 4.8. ProductHunt
Bonsai 27B: Frontier Reasoning at 1.7 Bits a Weight. A 27B reasoning model with natively binary/ternary weights that runs a full agentic RAG loop on a laptop or phone. Source
xAI open-sources Grok Build coding agent harness and TUI. Release includes CLI tools to build a robust agent harness. Source
ARA, Decision-Replay Infrastructure for ML Models. Records every model decision with its exact producing state for queryable incident debugging and label-leakage-free training data joins; free Community Edition runs on-prem. ProductHunt
Industry
Moonshot's Kimi K3 sends AI and semiconductor stocks into a tailspin. China's largest open-weight model revives DeepSeek-era fears about the economics of U.S. AI infrastructure spending. Source
Databricks hits $188B valuation. New funding round cements Databricks' position as a major AI data-platform incumbent. TechCrunch
Why the first GPU financiers are turning to inference chips in a $400M deal. A $400M deal signals the shift from GPU financing to specialized inference-chip investment. TechCrunch
Patreon stops asking AI bots not to scrape, and starts blocking them. Patreon moves from robots.txt requests to active infrastructure-level blocking of AI scrapers. TechCrunch
Community
I analyzed 60 documented AI coding-agent failures, 47% were critical, and the root cause usually wasn't the model. CVE-style review finds 47% critical and flags confidence miscalibration as the dominant root cause. Reddit
Kimi K3 landed third on the Intelligence Index. Detailed benchmark write-up suggesting open-weight parity with closed frontier models; weights reportedly drop July 27. Reddit
Xi Jinping calls for more open-source AI: 'China is ready to be more open'. China's president publicly endorses greater openness at the World AI Conference. Business Insider
Bonsai 27B runs locally on an iPhone, a 27B model in 3.9GB. Practitioner thread on running a 27B ternary model on-device in under 4GB. Reddit