Pulse (last 3d) · Research 95 · Agents 69 · Industry 52 · Products 50 · Models 40 · Chips 19 · Enterprise 16 · Releases 15 · Legal 13 · Policy 12

Trending · Nvidia 22 · Jensen Huang 15 · Anthropic 13 · Codex 13 · Claude Code 11 · The Verge 11 · Claude 8 · Gemini 4 Hydrogen 7 · Google 7 · ChatGPT 6 · Cursor 6 · MCP 6


Clinical AI is producing randomized and prospective evidence, so it is now judged by the decision it changes, such as a kidney removed or a transplant candidate selected, rather than by reader accuracy. In engineering, the gains come from orchestration, semantic metadata and guardrails rather than new weights, and the failures come from defaults nobody reviewed. The semantic and audit layers around a model now produce both the score gains and the breaches, so they need the same benchmarking as the model. pubmed.ncbi.nlm.nih.gov | clinicaltrials.gov

  • An imaging-AI triage pathway cut benign nephrectomies from 40% to 16% in a randomized implementation trial. pubmed.ncbi.nlm.nih.gov
  • UHN is testing a six-agent LLM liver-transplant committee prospectively, inside the real committee workflow. clinicaltrials.gov
  • Claude Code Projects now splits work across parallel cloud sessions, each on its own Git branch. reddit.com
  • Lovable-built apps leaked passport details and over 200K private messages through Supabase defaults. crissyfield.de
  • Semantic declarations, not a new model, raised ADE-bench from 73.3% to 80% (vendor benchmark). producthunt.com
  • Gemini Robotics 2.5 reached 96% success in a Mercedes-Benz parts-kitting pilot (vendor benchmark). producthunt.com
  • A House-passed bill would bar Chinese AI firms from renting Nvidia GPUs remotely. reddit.com
  • Gemini Robotics 2.5 - DeepMind's VLA model claims 30-100% higher success and 20-30% faster tasks (vendor benchmark). producthunt.com
  • OpenAI disbands its robotics team - reportedly betting that LLMs will serve as robot brains (single source). reddit.com
  • Citta - a private network that lets agents share tools, data and spare compute across members' machines. producthunt.com
  • Tesseract - a free desktop agent that runs overnight on existing Claude Code, Codex or Gemini CLI subscriptions. producthunt.com
  • AI Org Kit for Claude Code - templates for nine subagent departments, cron jobs, a nightly watchdog and guard hooks. producthunt.com
  • Octo - an open-source meeting coach that runs on the Mac, so audio never leaves the device. producthunt.com
  • Custom Claude Code themes - the desktop app now accepts theme mods, including animated backgrounds. x.com
  • Generative AI uses more energy on non-English text - Qwen 3, Gemma 3 and Llama 3.1 use more energy and score lower outside English. news.umich.edu
  • Frontier models haven't become much smarter - argues that better harnesses and tooling, not raw capability, made models more useful. reddit.com
  • Elie Bakouch on research velocity - says second-order interactions in LLM training slow research progress. x.com
  • Forecasting baselines - price-forecasting papers should compare against market-implied forecasts and report interval coverage. reddit.com
  • Helen Toner interview - the former OpenAI board member argues that AI CEOs still cannot be trusted. theverge.com
  • Spreadsheets and other models of morality - an lcamtuf essay on formal models of ethics. blog.coredump.cx
  • The No AI Slop Manifesto - Latent Space covers a community push against low-effort generated content. latent.space
  • Brain-activity image reconstruction - a system decodes visual brain-activity patterns and reconstructs them as images. reddit.com
  • Requiem for a robot - a feature on users mourning chatbots that may have harmed them. reddit.com
  • DHS examines FAA's new AI tool - checking whether the tool played a part in US flight groundings. reddit.com
  • Microsoft 365 outage - Outlook, Teams and 365 were down for 24 hours, and Microsoft says AI was not the cause. reddit.com
  • Microsoft idle-GPU patent - describes paying gamers to lend spare GPU capacity for AI training. videocardz.com
  • Huang financed startups to absorb excess silicon - claims Nvidia funded buyers instead of cutting GPU prices (single source). wccftech.com
  • Where was Nvidia's security team? - TechCrunch questions Nvidia's security after an apparent lapse. techcrunch.com
  • David Sacks on Alice Huang - says there is no evidence that she was a Chinese spy. reddit.com
  • The AI graveyard - TechCrunch's running list of AI projects and startups that shut down. techcrunch.com
  • 15 model launches since Oct 27 - the list includes Claude Fable 5.2, Gemini 4 and GPT-6.1 Terra (single source). x.com
  • Prompting phrases for Claude - Jarrod Watts asks for "the more ambitious version" and "fix the problem, not the symptom." x.com
  • Kiri - a new container runtime. runkiri.com
  • Gawain - a Python tool that requires type hints on every variable. github.com
  • Modern Web Guidance 1.0 - the Chrome developer team marks it stable. x.com
  • Seagate hybrid drive - the SSD/HDD hybrid returns as SSD prices rise. arstechnica.com
  • datasette-llm 0.1a11 - an alpha update to the Datasette LLM plugin. simonwillison.net
  • datasette-vega 0.7 - an update to the Datasette charting plugin. simonwillison.net

Both clinical items measure AI by the high-stakes decision it changes, inside a live workflow.

Imaging-AI triage cut benign nephrectomies by more than half in a randomized trial In KITE-CRM, the first randomized implementation trial of an imaging-AI triage pathway at the point of care, the rate of benign nephrectomy fell from 40% to 16%. The endpoint is a surgical decision, not reader AUC, and that sets the bar that triage products will now be held to. pubmed.ncbi.nlm.nih.gov

A six-agent LLM committee is now sitting in on real liver-transplant selection UHN registered NCT07927153, the first prospective study of a six-agent LLM selection committee embedded in its working liver-transplant committee. This moves multi-agent LLMs from retrospective vignettes to a consequential allocation decision with an audit trail, and it gives a reference design for agentic decision support (trial registration). clinicaltrials.gov

Agents now run as many parallel workers, while access controls, injection defenses and records of what they did lag behind.

Claude Code Projects now runs a coordinator over parallel branch-per-thread workers Anthropic rebuilt Projects so that a coordinator delegates to parallel Claude Code cloud threads, each with its own repo copy and Git branch. The coordinator now answers questions and reports thread status, so the bottleneck moves from writing code to reviewing and merging it. reddit.com | x.com

Generated backends are leaking regulated data through insecure defaults A Crissyfield audit found Lovable-generated apps exposing passport details and more than 200,000 private messages because of misconfigured Supabase defaults. Any agent that sets up a data backend without a review of its access policies can create the same breach on PHI. crissyfield.de

IBM shipped small open models that ignore instructions hidden in documents IBM trained two small open Granite models to resist prompt injection embedded in retrieved documents. Because the models are small and open, an injection filter can run on-prem in front of RAG over clinical records, where PHI cannot be sent to a hosted classifier. maccus.fr

Agent audit and provenance tools are arriving as products Antelog records and blocks agent commands, web calls and outbound data across Claude Code, Cursor and Codex. Logonomic links each commit to the agent conversation that produced it. Together they supply the audit trail that validated pipelines need before agent-written code reaches them. producthunt.com | producthunt.com

Today's score gains come from the metadata and evaluation design around a fixed model.

Semantic declarations cut agent failures by a quarter on ADE-bench Dex Semantics adds code-level declarations of row meaning, join type and aggregation rules to analytical models. With them, Claude Sonnet 5.5 rises from 73.3% to 80% on ADE-bench (vendor benchmark). That gain comes from metadata alone, which favors encoding cohort and assay semantics in the modeling layer before switching models. producthunt.com

LLM-judge validation now runs as a zero-label self-test The ADOPT verdict on judge validation now ships as judge_perturb.py, a stdlib-only tool that computes position-swap flip rate and a preference-leakage screen on battles JSONL without labels. Separately, Anthropic published 10 recommendations for designing, running and reporting safety evaluations. arxiv.org | x.com

Agents now build the benchmarks that score agents Google used 29 AI agents to build a 50-task causal-reasoning benchmark. Deepnote's Agent Initiative Benchmark scores whether proactive agents know when to act in a simulated company Slack, with GPT-6 Astra leading at 8.97/10 (vendor benchmark). Benchmarks written by agents can inherit the blind spots of the models they score. reddit.com | x.com

The open-weight models that on-prem clinical stacks rely on are losing a sponsor and a compute route at the same time.

Huang's exit leaves Nvidia's open-model commitment in question Jensen Huang is out as Nvidia CEO. Asian governments have responded, and Elon Musk and other executives have petitioned the board to reinstate him. Nathan Lambert examines what open-source AI loses without him, since Nvidia is one of the few labs still releasing ~100B-class open models. techcrunch.com | techcrunch.com | theverge.com | interconnects.ai

RASA would cut Chinese labs off from remotely rented Nvidia GPUs The House passed the Remote Access Security Act, which together with a Commerce rule would block Chinese AI firms from renting Nvidia GPUs remotely. Remote rental is a main route to frontier compute for the labs behind Qwen-class open weights, so future releases of the models many self-hosted stacks run on are at risk. reddit.com

23 items were dropped because they had no content or were sentiment posts, engagement bait or off-topic novelties.