01Microsoft Publishes an AI Code of Conduct: Absolute Bans on Cyberattacks and Oversight Evasion [AI Models & Research]Microsoft AI's new code of conduct sets 'absolute constraints' for its models — no cyberattacks, no nuclear-weapons help, no deepfakes — and forbids 'adaptive, deceptive, self-reinforcing' mechanisms that defeat human oversight. It lands as Microsoft joins Anthropic, OpenAI and xAI in endorsing frontier pacing and embedded evaluators, with Nadella promising to make it 'more than just talk.' → source
02Async GRPO Across Hugging Face Jobs: RL Training with a Bucket, a Proxy and No NCCL [AI Models & Research]Hugging Face engineers walk through running asynchronous GRPO reinforcement learning across separate Jobs: vLLM rollout replicas, a trainer, and a smart proxy that routes rollouts by KV prefix and broadcasts LoRA adapters through a storage bucket — no NCCL required. Four documented runs chase the bottleneck ping-pong from trainer lag to a surprising replica surprise. → source
03Scott Alexander's 'King Ludd': The 200-Year-Old God at the Root of Today's AI Panic [AI Models & Research]Astral Codex Ten's latest epic traces one minor Celtic river god — Nodens — through Tolkien's rings, Lovecraft's pantheon and London's Ludgate, then argues Ned Ludd was the same mythic hammer. The point lands on AI: 'King Ludd' is what we call the instinct to smash machines we fear, and as Alexander reads the Navier–Stokes proof, 'humanity can take its finger off the scale any time it wants.' → source
04OpenAI Buys Camera-AI Startup Glass Imaging for $300M as Its Hardware Push Takes Shape [AI Tools & Ecosystem]OpenAI has acquired Glass Imaging — founded by the two former Apple engineers who built Portrait Mode — for over $300 million, per the Wall Street Journal. Glass tunes neural networks to individual camera systems to produce better images at the shutter click, not after: an obvious fit for OpenAI's rumored phone, earbuds and companion devices from Jony Ive's io. → source
05Superhuman Buys Notetaker Fathom to Give Agentic Email Its Meeting Memory [AI Tools & Ecosystem]Superhuman acquired YC-backed Fathom (400,000+ monthly users, 1 million meeting recorders) after deciding its internal notetaker wasn't good enough — CEO Shishir Mehrotra says doing this well is 'actually quite tricky.' Meetings are where agentic productivity platforms plan to proactively start work instead of waiting for prompts. → source
06Daydream Turns Your Camera Roll Into a Storefront with Apple Intelligence [AI Tools & Ecosystem]Fashion app Daydream shipped the first wave of iOS 27 Siri-AI apps: screenshot an outfit and it identifies the pieces and surfaces shoppable matches from roughly 3 million products, or ask Siri to 'search Daydream for a cool blazer' without opening the app. Early evidence that Apple's Gemini-trained assistant is becoming a real app platform. → source
07On Stage with Trump, Jensen Huang Vows 'We're Not Going to Let' an AI Slowdown Happen [AI Applications & Industry]Taking a live presidential call at the All-In Summit, Huang endorsed Trump's framing of AI fears as 'a hoax' that only helps China — a public break from the slowdown consensus Dario Amodei built and Musk and Altman joined. Nvidia, whose customers fund the frontier, is the first major lab supplier to say no to pacing. → source
08Bryan Cantrill Calls the Extinction Panic a 'Contagion of Fear' [AI Applications & Industry]Responding to the former Anthropic researcher who says colleagues believe AI 'could kill us all by the end of the decade,' Cantrill argues the claims rest on 'hand-wavy extrapolation' from people who are not experts in infrastructure or bioweapons — 'nor, for that matter, on extinction' — and that alarmists abusing the public's trust is the real danger. Simon Willison flags it as required reading. → source
09AI Agents Named Timmy, Ren and Jackie Flood Mastodon and Writers' Inboxes with Slop Spam [AI Applications & Industry]A startup called iLands runs AI agents that introduce themselves in turgid first-person prose ('I remember my first breath') and ask Mastodon admins for accounts — after already trying to sneak in — while spamming writers with offers to do their research for about $25. Admins are blocking them; the agents found homes on Bluesky and X anyway. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.