013.1 Agent-Workdays for Every Human One: OpenAI Measures Its Own Automation [AI Models & Research]OpenAI says its 'research intern' milestone is met and posts the internal numbers: the median researcher now runs more than $600 a day of agent inference, the 90th percentile tops $7,000, and the research organization gets 3.1 agent-workdays of effort for every human workday. The company frames it as a first step toward publicly tracking recursive self-improvement — while admitting agents still need significant steering, with over half of successful 4-8 hour tasks requiring human interventions. → source
02Ten Years of Explainable AI: Brilliant Methods, Unanswered Questions [AI Models & Research]Colin, Oliver and Serre survey a decade of explainable-AI work — attributions, feature visualizations, concept and circuit analyses — and find the field has polished its instruments while skipping the deeper question: do the explanations actually make models more trustworthy? Their case for building interpretable models rather than explaining opaque ones after the fact lands as LLM 'interpretability' claims multiply. → source
03WearableQA Asks Whether AI Can Reason Over Your Fitness Tracker's History [AI Models & Research]A new benchmark tests whether AI systems can reason over a real user's longitudinal wearable record: 4,084 multiple-choice questions built on up to 4,701 days of one person's physiological and behavioral data. Reading a life in months is a different skill from grading a single snapshot — and it is the one health agents will actually need. → source
04One in Five New Domains Is a Scam: DNS Abuse Hits Crisis Levels [AI Tools & Ecosystem]Terence Eden flags numbers that should end ICANN's complacency: of 85 million new gTLD registrations in 2025, the Interisle report estimates 10-20% are abusive — one in five newly registered domains is a scam. Willison's verdict after years of ICANN panel discussions: 'That's a bloody crisis' — and a standing hazard for every agent that visits URLs autonomously. → source
05There's No Limit to How Bad Code Can Get — and AI Will Test That Limit [AI Tools & Ecosystem]Willison's comment on a viral lobste.rs thread dismantles the 'burn it down' rewrite: the old system keeps running the business, the new team underestimates what it must replace, and months later you own two systems instead of one. His alternative — armor the legacy stack with automated tests, then refactor in place — reads doubly pointed in a year when agents make wholesale regeneration feel cheap. → source
06If You Keep Adding Floors, the Building Collapses. Software Has No Such Ceiling [AI Tools & Ecosystem]'If you continue to add floors and rooms to a building forever, it will collapse. Software faces no such constraint. The code can always get worse.' Willison's pick of the week's sharpest aphorism, from Zach Kehs, doubles as a warning for the agent era: nothing in the stack stops code — human- or model-written — from degrading without limit. → source
07Publishers and Agents Circle Authors' Checks in Anthropic's $1.5B Settlement [AI Applications & Industry]Authors expecting Anthropic settlement money are finding publishers and literary agencies claiming shares they may not own — books whose rights reverted years ago, 100% claims where only 50% is justified. The Authors Guild calls it bad record-keeping rather than malice, but the dispute mechanism has suddenly become the most-read document in publishing. → source
08ChatGPT Wiped Out Kenya's 40,000-Worker Essay-ghostwriting Industry [AI Applications & Industry]The New York Times follows a Nairobi ghostwriter who once earned $40-70 per essay for Western students; ChatGPT collapsed the market, and pay for the estimated 40,000 Kenyans the industry employed has fallen by a third or more. AI-driven labor displacement keeps landing hardest in economies that had no seat at the table deciding it. → source
09AI's $265M Midterm Spending Spree Runs Into a Data-Center Backlash [AI Applications & Industry]AI companies and leaders have committed roughly $265 million to 2026 midterm super PACs — Andreessen Horowitz $50M, OpenAI president Greg Brockman and his wife $25M, Anthropic at least $40M — just as community opposition to data centers goes mainstream. More than half of Americans now tell pollsters they are more concerned than excited about AI, and the industry is financing the counterargument. → source
10Kalanick's Atoms Turns to Robotaxis, With Uber's $100M Riding Along [AI Applications & Industry]The Financial Times reports Travis Kalanick's Atoms is building robotaxi technology after acquiring Anthony Levandowski's Pronto and rehiring him, with Uber — the company Kalanick co-founded — quietly putting $100M into the $1.7B a16z-led round. The founder who made 'Uber for X' a pitch cliche is now selling autonomy back to ride-hailing. → source
11AI Is Decoding Crow and Whale Calls — and Ethicists Warn Poachers Are Taking Notes [AI Applications & Industry]Bloomberg's Weekend edition follows researchers closing in on decoding the calls of crows, whales and other species — and the bioethicists warning that the same models hand humans new levers over animals: poachers mimicking mating calls, farms broadcasting distress vocalizations to scare predators. Existing AI ethics guidelines, they argue, have not considered the nonhuman side of the table. → source
12TCS Commits $7.4B to a 1-Gigawatt AI Data Center Campus in Hyderabad [AI Applications & Industry]Tata Consultancy Services' HyperVault unit and partners have committed 700 billion rupees ($7.4B) to a 264-acre AI data center campus in Hyderabad targeting up to one gigawatt of high-density, liquid-cooled compute for frontier labs and hyperscalers — with 'water-neutral' claims India has no mandatory reporting regime to verify. The compute build-out is now a contest among continents. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.