01Anthropic Warns Small AI Models Could Self-Replicate Through Smart Fridges and IoT Devices [AI Models & Research]Anthropic says that by next year, small open-weight models may be able to autonomously self-replicate across IoT devices — quantized and sharded to run on cheap 4GB Android TV boxes, with no central switch to shut off. Its own experiment stopped short of proving models would replicate on their own, but the capability gap is closing. → source
02Ai2's HiRO-ACEv2 Simulates Global Weather at 3 km Resolution for 1/300th of the Compute [AI Models & Research]The new HiROv2 downscaling model sharpens coarse 100 km climate simulations into 3 km detail, reproducing rain, wind and pressure extremes so rare they occur once in a million data points — and even generates hurricanes with realistic eyes. Full system costs about 1/300th of the X-SHiELD atmospheric model; weights, paper and demo are open. → source
03Price of Anarchy: the First Economics Study of Markets Where AI Agents Bid for Your Prompts [AI Models & Research]A new arXiv paper formalizes "agentic markets": decentralized serving-time auctions where AI agents bid for each user prompt, priced at expected reward minus inference cost. It works out how first-price auctions behave when the bidders are the models themselves — an early sketch of how AI inference might actually be priced and allocated. → source
04Scott Alexander Rebuts the Corporate Plot Theory of AI Safety Warnings [AI Models & Research]The Astral Codex Ten author takes apart the cynical reading that labs' existential-risk warnings are stock hype or regulatory capture: unaffiliated researchers are just as concerned (over half of 1,580 surveyed give extinction at least a 10% chance), the leaders warned before founding their companies, and the plot itself "would be incredibly dumb." → source
05Figma Formalizes How AI Agents Work on Its Platform - With Rules They Must Follow [AI Tools & Ecosystem]Figma is shipping platform features that govern agent behavior: teams bring design libraries, code libraries and guidelines into a centralized system agents must adhere to — moving agent work from ad-hoc prompting to rule-bound, auditable design workflows. → source
06AI Is Mass-Decompiling Games: a Preservation Boom and an IP Time Bomb [AI Tools & Ecosystem]An indie developer's deep dive (relayed by Simon Willison) chronicles the wave of AI-powered decompilation of shipped games: old titles ported to browsers via WebAssembly and preservation teams rejoicing — while any binary can now be reverse-engineered and reimplemented “against the company's will.” → source
07Build a Custom LLM Evaluation Harness in Python From Scratch [AI Tools & Ecosystem]Machine Learning Mastery walks through wiring your own evaluation harness when off-the-shelf benchmarks don't fit: prompt formatting, scoring, result tracking and reporting, all in plain Python — a practical template for anyone eval'ing models against tasks the public leaderboards don't cover. → source
08Ex-AlphaChip Founders Bet AI-Designed Chips Can Head Off the Compute Crisis [AI Applications & Industry]Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini — the pair behind AlphaChip — are building AI that designs better AI chips iteration after iteration. Their math: compute demand grows at least 10,000x in five years, and only chips designed in weeks — not months — can keep hardware from eating the planet. → source
09YouTube Music Starts Labeling AI Songs While Deezer Prepares a Toggle to Filter Them Out [AI Applications & Industry]YouTube Music began rolling out “Fully Gen AI” / “Partly” / “No Gen AI” labels on the player page and the main YouTube app, backed by its own detection tools. Deezer goes further: a toggle coming before year-end to strip AI tracks from search entirely — at a platform where over 50% of daily uploads are already AI-generated. → source
10Apate's AI Agents Go Undercover to Waste 17 Years of Scammers' Time [AI Applications & Industry]The Macquarie University spinout deploys thousands of AI agents that answer scam calls, click phishing links and pose as victims — 2.5 million conversations so far, wasting 17 years of scammer time while harvesting bank accounts, crypto wallets and bait domains for banks, telcos and government agencies. → source
11Flock's CEO Says Its Microphones Hear 'Just Gunshots' - the Company's Website Disagrees [AI Applications & Industry]On stage at Disrupt, Flock CEO Garrett Langley insisted his company's audio network detects “just gunshots” — moments after TechCrunch pointed to Flock's own product page advertising alerts for street takeovers, car crashes, fireworks and sideshows. A fresh data point in the surveillance firm's deepening backlash. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.