Pulse (last 3d) · Research 71 · Agents 59 · Products 52 · Industry 44 · Models 24 · Chips 20 · Policy 13 · Infra 12 · Enterprise 12 · Open Source 10
Trending · Nvidia 21 · Claude Code 17 · Jensen Huang 13 · Codex 12 · Anthropic 9 · Claude 8 · Cursor 7 · TechCrunch 7 · MCP 5 · OpenAI 5 · ChatGPT 4 · FDA 4
Top story
Clinical AI now has randomized and prospective outcome evidence: a randomized trial cut benign nephrectomies from 40% to 16%, and Toronto General is testing a six-agent LLM committee on live transplant decisions. At the same time, FDA is building the regulatory infrastructure, with a master-file pathway for validated models and a PDUFA VIII letter that makes AI a named accountability item. Clinical models are now judged on a trial registration and a reusable regulatory dossier rather than an AUROC, so validation work should be planned to produce both. pubmed.ncbi.nlm.nih.gov | fda.gov
The Shortlist
- In a randomized trial, imaging-AI triage cut benign nephrectomies from 40% to 16%. pubmed.ncbi.nlm.nih.gov
- Toronto General is testing a six-agent LLM committee on real liver-transplant selection decisions. clinicaltrials.gov
- FDA's model master file could let many submissions cite one validated model. fda.gov
- PDUFA VIII names AI as an FDA accountability item through 2032. fda.gov
- An open 0.8B CPU model extracts entities from a schema written as a prompt, with no retraining. producthunt.com
- One LLM edited another model's weights and raised its math score from 7.9% to 48.1%. reddit.com
- No ClinicalTrials.gov registration backs Isomorphic's Phase 2 claim. clinicaltrials.gov
Also Noted
- Qwen4-Max - an MoE model with native vision and 1M context, priced below Qwen3.8-Max. x.com
- LongCat-2.5 - Meituan's open 1.6T agent model has a deploy guide only for Huawei NPUs. reddit.com
- Nvidia chips in China - export-banned chips power China's top models, and Congress moves to tighten rules after a spy scandal. reddit.com | theverge.com
- US AI national champions - per the WSJ, Nvidia, Amazon, Microsoft, OpenAI and Anthropic will be named to win overseas business (single source). x.com
- EUROPA - the EU's open model project targets nearly 1 trillion parameters. maccus.fr
- Anthropic efficiency slide - an IPO roadshow slide reportedly shows superexponential gains in pretraining compute efficiency (single source). x.com
- Unbundling vision encoding - Meta FAIR and NYU trained 200+ MLLMs: encoder pretraining is optional, but vision-only parameters are essential. i.redd.it
- Emergent biology in a molecular-interaction foundation model - the paper reports emergent biological capabilities (preprint, unreplicated). huggingface.co
- OpenAI science - a post on progress toward a lattice Standard Model, plus a new dedicated science blog. openai.com | openai.com
- Agentic quantum circuit discovery - agents search for fault-tolerant quantum circuits. arxiv.org
- Training-free early exit - a linear probe folded into the
</think>output row, so the edited weights run anywhere. reddit.com - MMFlash - faster speculative decoding for vision-language models. huggingface.co
- Quantization divergence viewer - compares two quantizations of one model and shows the token where they diverge. v.redd.it
- llama.cpp on Gemma4 - a chart of llama.cpp throughput gains over time. i.redd.it
- TokenEconomizer - Microsoft's black-box cost control for coding agents, learned from experience, at 6,751 stars. github.com
- No Second Run - asks whether recursive self-improvement stops at frontier pretraining; an 801-30B-A3B base model is coming. reddit.com
- Benchmark trust poll - 56 practitioners rank Terminal-Bench 4, Terminal-Bench Science, FrontierCode and FrontierSWE as most trusted. x.com | x.com
- Benchmarks as vendor entry - argues that a public benchmark frontier models fail is a data vendor's fastest route into labs. x.com
- Perit Marketplace - sells owned data and expert-built RL tasks to frontier labs, with $11M revenue in 4.5 months. producthunt.com
- AI reviewers favor appendices - LLM paper reviewers are biased toward content placed in appendices. huggingface.co
- Verification transport - tests whether verifier judgments still hold once agent trajectories are abstracted. arxiv.org
- Mech interp for agent failures - a case study on web agents hallucinating page elements. arxiv.org
- GH-AWTrace - a large dataset of execution logs from GitHub agentic workflows. huggingface.co
- TradeGauntlet - a benchmark for compliance controls on autonomous trading agents. arxiv.org
- MEM-SQL - adds a memory mechanism to LLM text-to-SQL. huggingface.co
- Fake-brand hallucination check - ChatGPT and Google AI Mode invented nothing in 18 answers about 9 fictional brands. reddit.com
- ARCHER - moves small models from imitating rationales to executable reasoning. arxiv.org
- Interactive program induction - language-model priors guide interactive program synthesis. arxiv.org
- Step-aware decoding for MDLMs - reduces hallucination in masked diffusion language models. huggingface.co
- Sim-trained reasoning robots - reasoning models trained in simulation operate real robots natively. arxiv.org
- Gamer - a generalist vision-language-action model for long-horizon, real-time game agents. huggingface.co
- Agentic robotics - a perspective paper on robotic systems built on foundation models. huggingface.co
- lucidml - a 630M Apache-2.0 world model generates a playable open world frame by frame. producthunt.com
- Multiplayer world-model game - a world model on each player's own GPU generates that player's view. vor.tech
- JVT - a simple pixel-space video diffusion transformer. huggingface.co
- Disen-Forcing - separates semantic anchoring from motion in autoregressive video diffusion. huggingface.co
- mischief 1 - a Wan 2.2 add-on that claims 5x faster image-to-video (vendor benchmark). x.com
- WaferTetris - freeform workload mapping for wafer-scale accelerators. arxiv.org
- Bharat Semi - an Indian AI sensor chip startup raises a $24M seed round. techcrunch.com
- FETTUCCINE - fast, efficient brain-to-text decoding on mobile devices. arxiv.org
- sEMG keystroke decoding - muscle activity, language and typing behavior all shape live decoding accuracy. arxiv.org
- BAL - bidirectional autoregression in latent space for learning human movement representations. arxiv.org
- eDIAR - training-free streaming diarization for open speaker sets and long meetings. huggingface.co
- StreamEmbed - real-time indexing from pixels for screenshot retrieval. arxiv.org
- Divergences between near-equal distributions - analyzes numerical instability when computing divergences between nearly equal distributions. arxiv.org
- Elliptic sampling - optimal interpolant schedules for Gaussian targets and beyond. arxiv.org
- Conic classification heads - asymptotic margins and feature confinement under softmax cross-entropy. arxiv.org
- Contrastive learning with score matching - combines contrastive latent learning with denoising score matching. arxiv.org
- Score-based neural ODEs - applied to modeling reaction-diffusion systems. arxiv.org
- Imitation-bootstrapped MARL - imitation learning bootstraps multi-agent reinforcement learning. arxiv.org
- Alignment by socialization - models how alignment tradeoffs are learned under selection. arxiv.org
- JFinBench - Sakana's Japanese finance benchmark, built from the latest certification exams and refreshed continuously. huggingface.co
- Stanford HAI on jagged gains - why AI's gains are uneven across tasks and sectors. reddit.com
- Code Buster - an open-source repository review CLI with no source uploads or model calls. producthunt.com
- Polylane - reviews pull requests for security flaws, tracing attacker-controlled input to the vulnerable line. producthunt.com
- MCP Cursor Guard - signed cursors bound to their filters stop silent MCP pagination replay failures; MIT. producthunt.com
- recursiva - an AGPL issue tracker that gives agents scoped tokens and one inbox for their questions. producthunt.com
- Mercury.sh - MIT-licensed tool that escalates from one model's answer to a configurable model panel. producthunt.com
- Pixel Office - an open-source local workspace for managing agents with roles, memory and permissions. producthunt.com
- Papiro - a self-hosted OCR document archive that idles at 5 MB RAM and includes an MCP server. producthunt.com
- Qwen3.8-27B Heretic merge - a community merge that adds vision and abliterated weights for local agents. reddit.com
- Airflow 3 ML tutorial - an interactive orchestration tutorial in English and Spanish that updates Made With ML's lesson. reddit.com
- Social Sim'27 CFP - an AAAI-27 workshop on LLM social simulation, with papers due Nov. 20. reddit.com
- South Korea exam cheating - an AI-assisted exam cheating scandal has caused a national uproar. reddit.com
- Tor 0.4.9.17 - a rapid-response DoS-defense release shipped during a sustained DDoS attack. forum.torproject.org
Clinical AI evidence is moving from accuracy claims to registered, prospective outcomes
Each item replaces a retrospective accuracy number with a trial registration, a randomized outcome or a registry check.
Imaging-AI triage cut benign nephrectomies from 40% to 16% in a randomized trial KITE-CRM, a Shanghai-led point-of-care randomized implementation trial, sent cystic renal masses through an imaging-AI triage pathway and cut benign nephrectomy rates from 40% to 16% compared with standard reporting. It measures a surgical outcome rather than reader accuracy, which is the deployment-level evidence payers and regulators have lacked for imaging AI. pubmed.ncbi.nlm.nih.gov
Toronto General is testing a six-agent LLM committee on live transplant selection University Health Network registered NCT07927153, a pre-post study of an "AI Selection Committee" with agents for hepatology, surgery, cardiology, anesthesiology, psychiatry and social work, advising real liver-transplant candidacy decisions. It is the first prospective test of agentic AI in organ allocation, and its endpoints will become the template for evaluating multi-agent clinical consensus (registration, no results). clinicaltrials.gov
Apple opens a US health-AI study that can enroll 200,000 people Apple launched an app-based US study of how people use AI for health, with capacity for up to 200,000 participants. Prospective usage data at that scale would give a consumer platform a real-world evidence base that no single health system can match (press coverage). reddit.com | theverge.com | techcrunch.com
No trial registration backs Isomorphic's "first AI-designed drug to Phase 2" claim A ClinicalTrials.gov query for Isomorphic returns zero registered studies, and no first-in-human Phase 1 record exists either. Registries are now the cheapest audit of AI drug-discovery claims, and this claim fails it. clinicaltrials.gov
FDA is building the infrastructure to treat models as reusable, accountable regulatory objects
Both documents give AI a formal place in FDA's machinery, one as a filing mechanism and one as a funded obligation.
FDA already has a master-file path for validated models, and QMIN could extend it beyond generics FDA runs a Drug Master File-style mechanism for filing a validated model once and referencing it as a regulatory object, so far used in generic-drug review. Extending it through the QMIN program would let a vendor qualify a model once for many applications to cite, the closest thing yet to a registry for clinical models. fda.gov
PDUFA VIII makes AI a named FDA accountability item and ties user fees to Phase 1 geography The PDUFA VIII goals letter commits FDA to AI deliverables it will be measured against, and it ties prescription drug user fees to where Phase 1 data are generated. Both terms run through 2032, which puts explicit cost and accountability lines on AI-assisted review and on offshore early-phase data in sponsor planning. fda.gov
Agents now rewrite their own harnesses, memories and weights, and verification is the bottleneck
Every self-modification result today depends on whether the change can be judged before it is adopted.
Harness self-improvement is now an RL objective ModHarness-RL treats the agent harness as a learnable program that evolves during agentic RL, so harness improvement is trained instead of hand-tuned (preprint, unreplicated). A second paper adopts self-generated updates only when evidence supports them, the kind of gate a regulated deployment of self-improving agents would require. arxiv.org | huggingface.co | arxiv.org
Agent memory fails on behavior, not recall In a new failure case, an agent's self-written notes for its next attempt told it to repeat the move that lost the game. A recall benchmark would score those notes as accurate memory even as performance got worse, which is the gap behavioral-gain evaluation is meant to measure. reddit.com | arxiv.org
One LLM edited another model's weights and raised its math score from 7.9% to 48.1% An LLM modified another model's parameters directly and lifted its math score more than sixfold (single source). WeightBench, released the same day, benchmarks LLMs as weight-space programmers, which turns the technique from a stunt into a measured capability. reddit.com | arxiv.org
Narrow models and local inference now fit inside the PHI boundary
Tasks that used to need an external API now run on a CPU or a single consumer GPU.
A 0.8B open-weight extractor takes its entity schema as a plain-English prompt ReDefiNER, built on Qwen3.5-0.8B, takes entity types, definitions and inclusion rules in a prompt of up to about 10,000 tokens and returns exact spans with a type and a confidence score. It ships as a 0.77 GB int8 file under Apache 2.0 that runs offline on a CPU, so clinical NER schemas change without retraining or PHI leaving the network (vendor claims). producthunt.com
Air-gapped moderation with PII detection costs a flat $299 Argus 1 returns true or false on 12 categories of English text, including PII and self-harm, and runs air-gapped from one Docker command. It costs $299 once, or $0.40 per 1,000 hosted requests, so patient-facing chat can screen input without a third-party API (vendor claims). producthunt.com
Consumer hardware now serves 27B dense and large MoE models at usable speed Basalt 0.4.0 claims up to 770 tok/s on Qwen 27B on one RTX 5090, and BigMoeOnEdge runs any MoE GGUF on 8 to 16 GB cards and Macs faster than Strata and colibri (vendor benchmarks). Serving capable open models on-prem no longer requires datacenter GPUs. v.redd.it | reddit.com
Dropped
13 items: rumors, opinion posts, a demo video, a running thread with no update since August, or no discernible content.