01MLExpert Is a Generalist Agent Built to Run Machine-Learning Engineering End to End [AI Models & Research]A new arXiv paper outlines a generalist agentic model for autonomous machine-learning engineering — the 'AI for AI' push toward systems that develop and improve other AI systems on their own: data handling, experiment design, training and analysis, with recursive self-improvement framed as the long-term trajectory. It is a research look at what a full-stack ML engineer agent would actually have to hold together. → source
02FDE-Gym Benchmarks AI Agents on Real Customer Delivery, Not Tidy Tickets [AI Models & Research]Most agent benchmarks score well-specified tasks from issue reports; FDE-Gym instead tests the forward-deployed engineer's job — clarifying ambiguous requirements, working in live customer systems and uncovering constraints nobody wrote down. The 'last mile of AI delivery' finally gets its own measurement, and it rewards shipping what a paying customer meant, not what the ticket said. → source
03Researchers Map the Routing Bottleneck Inside Modern Mixture-of-Experts Models [AI Models & Research]Modern MoE architectures mix fine-grained expert segmentation with shared experts, and this analysis works through how routing choices interact with that design — what current routing does well, where its expressiveness breaks down, and what that means for training and serving sparse models. Dense with detail, but a useful technical read. → source
04Meta Quietly Cuts the Terms That Would Have Let It Scrape Publisher Sites for AI Training [AI Tools & Ecosystem]Meta's updated commercial terms declared it could use any content from a business account's domains to 'train and maintain' its products — language legal experts called a permission slip to scrape any publisher with a Meta business account. After a Nieman Lab breakdown and a spokesperson insisting nothing had changed, the provision was removed: the crawling terms, Meta now says, 'only applied to our ads and business products.' → source
05China Puts Model Weights, Agents and Chip Designs on Its Export-Control List [AI Tools & Ecosystem]A draft from Beijing's commerce and science ministries would add 40 technologies to the national export-control list — 23 of them AI, including large language models, agents, chip designs and computing clusters, with model weights and source code restricted above certain performance thresholds. Consultation runs to November 15; API and SaaS access are exempted, but if it passes, China's best weights stop crossing its borders. → source
06Starlink's Call Centers: 20,000 Grok Voice Calls a Day, 70% With No Human [AI Tools & Ecosystem]SpaceX says Starlink support now runs largely on Grok Voice: 20,000 customer calls a day in multiple languages, roughly 70% resolved with no human in the loop, one agent verifying customers, troubleshooting hardware, issuing credits and shipping replacements. Any company can rent the voice at 8 cents a minute — though the headline 70% figure is the company grading its own homework. → source
07FTC Subpoenas OpenAI and Anthropic Over Safety Incidents, Plans Executive Testimony [AI Applications & Industry]The FTC sent civil investigative demands — each more than 50 pages — to OpenAI and Anthropic, seeking documents on security incidents, safeguarding practices, privacy policies and marketing claims in a probe focused on consumer protection. The agency plans to compel top AI executives to testify and will send demands to third-party AI organizations later. OpenAI says it will cooperate. → source
08Schwab and JPMorgan Are Blocking AI Agents From Moving Client Money [AI Applications & Industry]Charles Schwab's CEO says the broker blocks unauthorized AI agents — including Meta's Muse — in the tens every week after clients hand over their passwords: 'We can't control it, the client can't control it, something could go wrong — who's liable?' Schwab will let outside agents view balances through controlled MCP servers and points them at its own assistant; JPMorgan confirmed it is keeping agents out too. Liability, not capability, is the new frontier of agent deployment. → source
09CoreWeave's 2032 Bonds Sink Below the Distress Line — and Musk Says 'Yikes' [AI Applications & Industry]The neocloud poster child's unsecured debt due 2032 traded below 80 cents on the dollar — the level usually read as financial distress — sending the stock down 15.5% in a day. Bloomberg Intelligence now prices roughly 60% odds of default; if CoreWeave, the most leveraged name of the AI buildout, stumbles, its debt-laden peers are likely to wobble with it. → source
10Anthropic's Own Numbers: Junior Technical Hires Fell From 27% to 9% [AI Applications & Industry]In its first employment report, Anthropic says junior-level staff in technical departments fell from 21% to 14% since early 2025 — and junior technical hires from 27% to 9% — while 20% of employees say their job shifted 'from actually doing work to managing Claudes.' Human attention, the company finds, is becoming the productivity bottleneck as Claude's output becomes the deluge everyone else must read and organize. → source
1118 Months and an $8M Fine for the Man Who Faked Billions of AI Music Streams [AI Applications & Industry]Michael Smith — who used hundreds of thousands of AI-generated songs and automated bot listeners to manufacture billions of fake streams and millions in royalties — was sentenced to 18 months in prison and fined over $8 million. Beyond one man's sentence, it is the first real enforcement moment for the platforms still drowning in AI slop on their own metrics. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.