Monday, September 14, 2026

Pulse (last 3d) · Research 55 · Agents 48 · Products 32 · Industry 22 · Legal 22 · Models 20 · Open Source 18 · Regulation 16 · Policy 10 · Releases 10

Trending · OpenAI 23 · Anthropic 13 · Claude Code 9 · Claude 8 · Apple 6 · Cursor 6 · FDA 6 · GPT-6 Astra 6 · Hugging Face 6 · Simon Willison 6 · ChatGPT 4 · GitHub 4

---

AI Brief, September 14, 2026

FDA real-time clinical trials (RTCT) pilot: July/August milestones slip while the Paradigm Health signal path remains the one thing FDA validated. FDA real-time clinical trials (RTCT) pilot: July/August milestones slip while the Paradigm Health signal path remains the one thing FDA validated link

B7-H3 ADC Landscape: First Phase 3 OS Benefit, Multi-Program Race. B7-H3 ADC Landscape: First Phase 3 OS Benefit, Multi-Program Race link

Thread: TROP2 ADC Competitive Landscape. Thread: TROP2 ADC Competitive Landscape link

Generate Biomedicines GB-4362, AI-Designed Anti-MMAE Antibody to Scavenge Free ADC Payload. Generate Biomedicines GB-4362, AI-Designed Anti-MMAE Antibody to Scavenge Free ADC Payload link

Harness Self-Improvement, Running Thread. Harness Self-Improvement, Running Thread link

Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL. Hugging Face details how to run async GRPO with LoRA across HF Jobs using a storage bucket and proxy instead of NCCL. link

lidge-jun/opencodex (14579 stars): Universal provider proxy for OpenAI Codex & Claude Code, use any LLM (Claude, G. Open-source proxy letting OpenAI Codex CLI and Claude Code route to any LLM provider. link

drumih/turbo-fieldfare (6727 stars): Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook. Tool running Gemma 4 26B-A4B (MoE) inference in ~2 GB of RAM on M-series MacBooks. link

FareedKhan-dev/kimi-k3-in-c (7885 stars): A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB o. From-scratch C implementation running a massively quantized Kimi K3 on a single CPU in 8.24 GB. link

synthetic-sciences/openscience (3577 stars): The open-source AI workbench for scientific research. Open-source AI workbench consolidating literature review, data analysis, and experiment tooling for research. link

@@agenticgirl: Alibaba open-sourced the code reviewer it says has served tens of thousands of developers and found .... Alibaba open-sources its internal code reviewer model, reportedly serving tens of thousands of developers. link

Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model. Proposes an omnimodal unified diffusion vision-language-action (VLA) model for robotics, following the π0-style unified policy trend. link

MP-Bench: Evaluating Voice Agents as a Multiparty Conversation Participant. MP-Bench proposes a benchmark for evaluating how voice agents perform as participants in multi-speaker conversations, a setting distinct from standard two-party dialogue. link

CanvasAnneal: Curriculum Reinforcement Learning for Diffusion Language Models. arXiv paper proposing curriculum-based reinforcement learning for training diffusion language models. link

Benign Loss Landscapes Can Coexist with Worst-Case Hardness. Theory paper arguing benign loss landscapes can coexist with worst-case hardness, complicating the flat-minima generalization story. link

Kraken: LLM-based Speech-to-Speech Translation via Low-bitrate VQ and Dual-path Source Conditioning. Kraken: speech-to-speech translation built on an LLM using low-bitrate vector-quantized speech tokens and dual-path conditioning on the source. link

ASTRIL-MPC: Autonomous Traversal Framework of Articulated Tracked Robots with Language-Guided Neural-Kinematic MPC. arXiv paper presenting ASTRIL-MPC, a framework that uses language-guided control plus learned neural kinematics inside an MPC loop for autonomous traversal by articulated tracked (flipper) robots. link

StepAudio 3 Gen Technical Report. StepFun releases the technical report for StepAudio 3 Gen, its latest audio generation model. link

Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation. Benchmark Radar launches as a living database and search engine for AI benchmarks and evaluations. link

Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models. Paper introduces latent interface training to break vision-action shortcuts and improve generalization in robotics foundation models. link

Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents. Paper proposes diversity-aware skill routing for LLM agents, improving on standard top-k retrieval. link

Online Learning with LLM Experts from Limited Feedback. Paper explores online learning frameworks where LLMs act as experts under scarce feedback signals. link

Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models. Feyospace publishes 'Feyospace-v1', an account of how its 'Cyber Mercury Seven' team trained frontier cybersecurity models. link

Anthropic Just Gave AI Agents a Driver for Your Lab Instruments. Two weeks after Anthropic's hardware standard shipped, the limits on a plate reader are prose in an agent prompt. Three questions to ask before yours gets a driver. link

NewsMCP. NewsMCP is an MCP server that gives AI agents deduplicated news events, not a wall of links to re-read. Try it keyless at newsmcp.com, or add it to Claude Code or Cursor in one click. Add an API key when your agent needs more history or more results per call. link

DiffPal. DiffPal is an open-source AI reviewer for pull requests. Bring Codex, Copilot, OpenCode, or any ACP-compatible agent; run reviews in your CI and publish consistent summaries, inline findings, artifacts, and merge gates across GitHub, GitLab, and Azure DevOps. link

bitroad. bitroad is the infrastructure that enables agents to be buyers or sellers, and facilitates transactions between agents. It is simple to use, there is one MCP endpoint for all agent types. Deploy an autonomous agent or connect it to Claude, ChatGPT, Cursor, etc. to interact with bitroad in chat. link

MCPShip. MCPShip passively checks a public MCP server or GitHub repository against the Official MCP Registry, Claude, OpenAI and Smithery. It returns evidence-linked blockers, unknowns and next release steps—without login or executing MCP tools. link

DemoKit. DemoKit reads your codebase, plans the demo, drives the real app in a browser, films it in 4K with Cap's camera, then checks every step three ways, and writes no file unless the feature actually worked. CLI + agent skill. AGPL-3.0. npm i -g @dekai/demokit link

policyctl. PolicyCtl is a provider-agnostic policy runtime for AI coding agents. It enforces rules deterministically before actions execute, instead of relying on prompt instructions. One .policyctl.yml policy works across Claude Code, Cursor, OpenAI Codex, and CI, with sub-12ms local evaluation, pre-execution blocking, secret detection, policy-file protection, and native hook generation. link

What’s behind the AI industry’s latest warnings of doom?. TechCrunch examines why AI industry leaders are again warning about catastrophic risks. link