01MintAct Unifies UI Grounding, Navigation and Tool Use in One Visual Agent Family [AI Models & Research]A new family of vision-language agents at 2B, 4B and 8B parameters unifies UI grounding, multi-step navigation across mobile, desktop and web, and visual tool use — matching frontier baselines on agentic benchmarks through environment and training-recipe design rather than raw scale. → source
02LeCun's AMI Labs and Fei-Fei Li's World Labs Are Keeping Their World Models Secret [AI Models & Research]The buzziest names in spatial intelligence are also the quietest about products: AMI Labs says "we'll talk when we're ready," and even Physicl, a company supplying their training data, admits it doesn't know what it's feeding. A dark-forest dynamic could shape the field's next commercial race. → source
0373,000 Reddit Posts Reveal What Users Actually Value When Delegating to AI Agents [AI Models & Research]Using Value Sensitive Design, researchers analyzed 73,093 first-person Reddit posts about the OpenClaw agent to map what users prioritize when handing work to autonomous agents — competence, privacy, control and consent — an early manual for agent etiquette written by the users themselves. → source
04APort Vault Replays 4,371 Real Attacks Against AI Agents Handling Your Payments [AI Tools & Ecosystem]A new benchmark replays 4,371 human-written attacks from a public capture-the-flag event against a live payment agent — 14 models from 8 labs, five policy configurations, with and without deterministic guardrails. Agent payment authorization finally gets a proper stress test. → source
05Simon Willison's llm-keys-ui Stops You Pasting API Keys Into Agent Sessions [AI Tools & Ecosystem]Willison's new LLM plugin spins up a local web UI for storing API keys on remote machines, so Codex Remote agents can fetch credentials at runtime without secrets ever landing in a chat session. Stored key values are never displayed back — you can only overwrite them. → source
06Vocci's 6-Gram Titanium Ring Records Your Meetings With a Double-Tap [AI Tools & Ecosystem]The meeting-recorder wars add a form factor: a sub-6-gram titanium ring that double-taps to record up to 8 hours of conversation, with transcripts and AI summaries on your phone. The hardware impresses in testing; the app's AI insights still overshoot their notes. → source
07datasette-explain 0.2.2 Brings Query Plans to Datasette's Stored Queries [AI Tools & Ecosystem]A small release with a useful trick: SQL EXPLAIN plans now work on Datasette's read-only stored-query pages, shipped alongside Willison's upgrade of datasette.simonwillison.net to Datasette 1.0a40 — query validation as you type. → source
08Google Had a Mole Inside the Hacker Gang Behind History's Biggest Supply-Chain Attack [AI Applications & Industry]Google Threat Intelligence revealed a Mandiant analyst had infiltrated TeamPCP's inner circle from near the start — watching the group's unprecedented open-source supply-chain rampage from within, warning breach targets, disrupting exploits and passing identifying details to law enforcement. → source
09'Nobody Is Reading Anything': Life Inside a Company Where Claude Code Ships Everything [AI Applications & Industry]An engineer new to a big company describes full agent-driven development from the inside: specs, code, tests and tickets all made by Claude Code, 12-13 hour days "just to press enter," and nobody on the team reading any of it. The quote is resonating far beyond X. → source
10ScrollEd Turns Any Textbook Into a TikTok-Style Feed [AI Applications & Industry]A TechCrunch Disrupt Battlefield 200 startup turns any PDF or textbook into a vertical feed: scroll for AI-generated video, audio and text, swipe sideways to go deeper, quiz at the end. The bet — if students won't open the textbook, let the textbook meet them in the scroll. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.