01Anthropic's Automated Researcher Outscored 28 Human Safety Researchers [AI Models & Research]In a new report, Claude autonomously fixed 10 kinds of model misbehavior, from sycophancy to deception, closing 85% of the deception gap where six experienced researchers averaged 20%. The winning methods still worked on models up to 4.7 times larger and on tests Claude never saw. → source
02Recursive Launches Flint, a Laptop Coding Model Built Mostly by AI [AI Models & Research]The startup's automated research system helped build its first model, a free open-weights coding model, from scratch in under four months, and it runs about twice as fast as similar models on a MacBook Pro. CEO Richard Socher calls it an early test of AI building better AI. → source
03Anthropic Put 45 AI Agents in a Shared Forum to Hunt Bugs [AI Models & Research]Coordinating through a shared forum, a swarm of 45 agents turned up 266 vulnerabilities across 15 open-source projects, against 21 for agents searching separately, and the two groups overlapped on only 12 finds. On shared team projects, though, agent self-organization still falls apart without human direction. → source
04Two Critical PHP Holes, Found by AI, Get an Emergency Patch [AI Tools & Ecosystem]Researchers at security firm Wiz used AI agents to uncover two remote-execution holes affecting every PHP 8.x release, confirmed exploitable against WordPress and Drupal. A disclosure leak put exploit code online a day before patches arrived, so anyone running PHP should update immediately. → source
05NVIDIA Kumo Predicts Churn, Fraud and Demand Straight From Your Database [AI Tools & Ecosystem]The pre-trained model reads a company's databases directly and predicts outcomes like churn or fraud in a single pass, with no custom training or feature engineering. Across 41 benchmark tasks it beat established pipelines, including the gradient-boosted trees enterprises spend weeks tuning. → source
06IBM Open-Sources New Speech Models for Noisy Meetings and Fast Transcription [AI Tools & Ecosystem]Granite Speech 5.0 ships open-weight transcription models covering six languages that rank second among open models on English accuracy and process audio at up to 19,870 times real-time speed. A sibling update improved speaker attribution so long meeting recordings come out with who-said-what attached. → source
07Walmart Opens Its Checkout to AI Agents While Amazon Blocks Them [AI Applications & Industry]The retailer launched an Agent Gateway letting approved AI agents search its inventory and shop on a customer's behalf, with per-agent spending caps and approval settings like an authorized user on a credit card. Walmart even guarantees its returns policy applies when an agent misbehaves. → source
08Meta's Muse Tops 10 Million Users, With Shopping a Top Activity [AI Applications & Industry]Meta told investors its AI assistant now has more than 10 million users, and engaged shopping users average purchases above $200. Zuckerberg forecasts Muse will eventually earn more per user than Facebook and Instagram do today. → source
09Blackstone Dismisses the AI Debt Panic: 'This Doesn't Feel Like 2008' [AI Applications & Industry]Blackstone, with $1.3 trillion under management, pioneered lending to AI firms against their GPU fleets, and veteran Jas Khaira says a credit-crisis scare is misplaced because most data centers serve investment-grade customers. Markets still flinch: CoreWeave's shares have fallen more than 20% in recent days. → source
10Managers Say AI Left Them Reviewing Work They Can No Longer Judge [AI Applications & Industry]A survey of 1,700 directors and senior managers across seven countries found more than half now struggle to tell how much of their team's work came from AI, and nearly half spend more time reviewing output than before. Next year, 58% expect to oversee more AI agents than people. → source
11Anthropic Audited Worker Retraining Ahead of AI Displacement. It Comes Up Short [AI Applications & Industry]A review of 56 randomized US studies found job training raises employment by two to three percentage points and earnings by roughly $1,000 a year, against a cost of about $13,000 per slot. Anthropic's conclusion: at AI's potential scale, today's programs would fall short. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.