01Google's New Agentic Video Understanding Reads Videos Like an Agent — 88% Fewer Tokens [AI Models & Research]DeepMind's new agentic mode for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite lets the model dynamically scan video segments instead of ingesting everything: up to 88% fewer tokens, 66% lower cost, and up to 7% better quality, unlocking sub-second moment retrieval, anomaly detection and precise counting. Available today via the Gemini API and Google AI Studio. → source
02BenchMIRT Audits LLM Benchmarks Question by Question — and Finds They Measure Different Things [AI Models & Research]Ai2 and Hugging Face's BenchMIRT applies item-response theory at the level of individual prompts, revealing that questions inside a single benchmark often measure different abilities — WildJailbreak's harmful prompts load on safety while its benign ones load on general reasoning — so one averaged score can hide what a model was actually tested on. → source
03Verbal Reinforcement Learning: When Plain Language Becomes the Reward Signal [AI Models & Research]A new arXiv survey formalizes Verbal Reinforcement Learning: natural language as the primary feedback channel for improving language agents, conveying intent, preferences and causal structure in forms both humans and models can interpret — RL without the scalar reward, argued in prose. → source
04Your Coding Agent Secretly Ships with an Entire Office Suite — Codex Bundles LibreOffice [AI Tools & Ecosystem]Poking around his cache, Simon Willison found OpenAI's Codex desktop runtime quietly carrying 1.7GB of full Python, Node, git, Poppler — and a headless LibreOffice install, complete with skills that teach the agent when to use them. Coding agents no longer call your tools; they bring their own. → source
05Hugging Face Ships 200+ WebGPU Kernels to Speed Up AI in the Browser [AI Tools & Ecosystem]The WebAI team released @huggingface/kernels, a library for loading optimized WebGPU kernels straight from the Hub, alongside an initial collection of 207 versioned kernel packages — each with shader templates, correctness and benchmark cases. It is the foundation layer for running AI inference at speed inside any browser. → source
06Agent Software Gets a Supply-Chain Check: AIR Raises $50M to Vet Skills and Plugins [AI Tools & Ecosystem]Founded by Unit 8200 veterans and backed by Sequoia and Greenoaks, AIR builds security for the agent software supply chain: discovering agents inside companies, continuously vetting their skills, MCP servers and add-ons, and blocking what fails. It already filters out about 27% of the add-ons it scans. → source
07ChatGPT Health Plugs into Epic's Medical Records — 325 Million Patients, Read-Only [AI Applications & Industry]OpenAI integrated ChatGPT Health with Epic's EHR, which holds data on over 325 million patients, letting clinicians import and summarize notes, labs and medications — read-only, with nothing written back — plus a public-data plugin pulling from PubMed and ClinicalTrials.gov. The company cites 99.1% safe responses across 4,300 physician reviews, even as lawsuits over earlier health advice proceed. → source
08Fable 5.1 Ships: Anthropic Cuts Token Costs and False-Positive Refusals in Flagship Update [AI Applications & Industry]Anthropic shipped twinned updates to its flagship: Fable 5.1, the general-availability version, cuts token cost and reduces false-positive refusals from its safeguards, while Mythos 5.1 remains restricted to vetted partners in cybersecurity and life sciences. The pivot follows months of user complaints about over-restrictive behavior and Fable's struggled adoption. → source
09AfterQuery Becomes Y Combinator's Fastest-Ever Unicorn at $3.2B [AI Applications & Industry]The AI training-data startup went from a $300M valuation in April to $3.2B in five months — the fastest launch-to-unicorn run in YC's history, per partner Gustaf Alströmer. Its edge: not checking whether models answer correctly, but training them to work like doctors and lawyers would, with Nvidia and Legora among customers. → source
10Google's Answer to Canva: You Prompt Instead of Design [AI Applications & Industry]Google launched an AI design tool where the workflow is a prompt rather than a canvas — describe what you want and the system generates the layout. It's Google's most direct challenge yet to Canva's design-on-autopilot territory, and another sign that creation tools are converging on conversation. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.