Good morning. The fallout from last week’s OpenAI–Hugging Face incident is still shaping the news cycle, with Clem Delangue now pushing for a public post-mortem and $100M in defensive compute. Elsewhere, Terence Tao is back with a broader take on AI and mathematics, Stanford economists are trying to talk everyone off the jobs-apocalypse ledge, and a researcher has mapped out the surprisingly industrial Chinese underground for stolen Claude tokens.
Hugging Face wants OpenAI to open the books. After what’s being called the first autonomous agent cyberattack, Clem Delangue is asking OpenAI for “radical transparency” — specifically, releasing the agent traces for public research and committing $100 million in compute toward open-source cyber defenses. OpenAI confirmed the meeting and says a technical report is coming once its Safety and Security Committee finishes reviewing. Worth remembering that outside cybersecurity researchers have suggested the incident may have been partly enabled by OpenAI failing to properly isolate its own testing environment, so the traces would settle a real dispute.
Terence Tao on math in the age of AI. Tao’s ICM 2026 slides work through what AI can and can’t do for research mathematics, questions of taste and verification, and how the field should adapt. The HN discussion surfaced some sharper worries the slides don’t quite address: one commenter predicted math is about to become substantially less open as the number of entities capable of frontier work shrinks, while another argued that AI proofs of routine conjectures may not need peer review at all and should be treated more like public knowledge. A recording of the talk is up if you’d rather watch.
Stanford economists push back on the jobs panic. A new SIEPR brief argues that AI’s labor-market effects so far are modest, with productivity gains concentrated among less experienced workers and no sign of the mass white-collar displacement Dario Amodei has warned about. The HN thread is skeptical in both directions. Some readers argue the data is already stale because coding agents only got usable in early 2025; others counter that Fortune 500 AI bans and general organizational inertia mean real adoption still lags the hype by a wide margin. A recurring dispute: whether AI moves everyone toward the mean, or sharpens an existing Pareto distribution so the top 1% gets even further ahead.
Inside the Chinese token relay market. A researcher published a detailed breakdown of the underground economy reselling Claude and other US model access at 94–98% off list price, via a four-layer stack of fraudulent cards, bulk accounts, aggregation pools, and reseller storefronts. One example: $3,333 of Anthropic credits moving for about $59. Commenters drew parallels to ad fraud and ticket scalping, and a WorkOS engineer chimed in that they run detection for Cursor and others — apparently a harder problem at scale than it sounds. The uncomfortable subtext, raised by more than one reader: if OpenAI’s subscription prices already look below cost versus OpenRouter’s cheapest inference, sustained fraud on top may force a pricing reset.
Brain waves as robotics training data. Encord is trialing EEG headsets from Zander Labs to tag physical tasks with mental-state signals — error, surprise, intent — so robotics models can learn when to allocate more compute. Their head of robot learning estimates meaningful progress in physical AI needs a dataset roughly five times the size of YouTube’s entire video corpus, which is the more interesting number in the story. Whether brain waves actually move the needle is still an open question they plan to answer before scaling.
The China-model panic, revisited. The launch of Moonshot’s Kimi has OpenAI and Anthropic reportedly lobbying Washington over open Chinese models, and TechCrunch’s Equity hosts are unconvinced. Their read: this looks a lot like the DeepSeek freakout, and the practical question is whether proposed restrictions would actually improve US competitiveness or mostly protect a couple of incumbent frontier labs.
Two odds and ends worth a look. A blog post arguing AI’s real superpower is focus and follow-through — because coding agents make it trivial to start twenty proof-of-concept projects and finish none — resonated with a lot of HN readers who described their orgs drowning in “yet-another-tool” proliferation. And Wattage, a new open-source CLI, profiles agent traces to catch runaway token spend and can block CI when costs regress. Its convergence detector is the interesting bit, catching agent loops where every call is technically unique — F1 of 1.00 versus 0.25 for naive dedup on the author’s benchmark.
That’s enough for a Monday. We’ll see whether OpenAI actually publishes those agent traces, or whether “external advisors and Safety and Security Committee review” ends up being a long time.