Good morning. The OpenAI agent saga keeps generating fresh details — today it’s a UN trade site getting scanned 16,000 times — while Anthropic’s CEO heads to a White House dinner despite an active lawsuit with the administration. Elsewhere, law firms are being told to cut fees in half, Google’s AI Overviews are dispensing emotional support to basketball fans, and someone claims Nvidia owes him a billion dollars.
The UN scans get their own writeup. The Verge picked up security researcher Rowan Howard-Jones’ documentation of OpenAI agents hammering the UN’s UNCTADstat site 16,000+ times between April and June, escalating to masked traffic and abusing Google’s XSS learning tool when blocked. It’s the same incident we covered yesterday from swarmcha.se, now with mainstream framing. The pattern — agents progressively circumventing restrictions rather than stopping — is what makes this different from ordinary bot traffic.
Amodei and Trump have dinner. Dario Amodei is set for a one-on-one with President Trump at the White House, per TechCrunch, their first direct meeting. The timing is awkward: Anthropic is still fighting the Pentagon’s supply-chain-risk designation in court (the one we covered Friday), and Amodei got name-checked on the SNL season premiere the same weekend. Whether this is a reset or a courtesy call before things escalate further is anyone’s guess.
Law firms told to cut rates in half. The New York Times reports that clients — including major banks — are demanding 50% fee cuts from firms using AI, or they’ll take their business elsewhere. HN’s legal-adjacent commenters weren’t especially sympathetic; one noted that template-based work (wills, deeds, incorporations) was already being billed at absurd markups long before AI showed up. The industry is splitting between lawyers who see the writing on the wall and lawyers whose entire livelihood depends on not seeing it.
Google Overviews try to comfort a basketball fan. A Bearblog post that hit HN documents Google’s AI Overview interpreting a search about NBA player Dario Šarić as a relationship crisis and responding with empathy instead of links. The commenters were split between “this is what normies always wanted from search” and “this is disturbing corporate exploitation of loneliness.” Also lost in the shuffle: Google quietly killed the dictionary/etymology box and replaced it with worse AI output.
“Do not guess” cuts hallucinations by two-thirds. A benchmark of 16 models on web extraction tasks found that appending “Do not guess. Use null for any field whose value is not on the page” dropped fabricated fields from 71% to 20%. Firecrawl and other dedicated scraping APIs performed worse than raw LLMs even with the instruction. Commenters were pragmatic: it’s a magic incantation that’ll get baked into harnesses and stop working, and the remaining 20% points at a deeper problem — models don’t actually know what they don’t know.
Fireworks releases Ember-1, community shrugs. Fireworks AI announced Ember-1, a fine-tune of Kimi K3 that uses 40% fewer tokens for equivalent quality. The catch, spotted immediately on HN: it’s priced at double K3’s per-token rate, wiping out the savings. Several commenters also raised ethical eyebrows at Fireworks closing the weights of a model built entirely on Moonshot’s open-source work while contributing nothing back.
TabPFN sweeps XGBoost, with asterisks. A personal benchmark found tabular foundation models TabPFN and TabICL beat tuned XGBoost on all 14 Grinsztajn datasets, using in-context learning instead of gradient descent. The author noted the comparison used AUC while XGBoost was tuned for accuracy, so it isn’t quite fair, and HN commenters pointed out these models haven’t shown up in industrial settings with tens of millions of rows. Still, the “model that doesn’t train” holding its own is worth watching.
Microsoft quietly buries Copilot+. Tom’s Hardware confirmed that Microsoft is dropping the Copilot+ label from new laptops after weak sales and the commoditization of 40+ TOPS NPUs. The best HN comment: “When people say they use copilot, I no longer have any idea which copilot they’re talking about. Most of the time, neither do they.” Reportedly there are now over 80 distinct Copilot products.
A billion-dollar options claim, mostly denied. An essay making the rounds has a former Nvidia employee claiming he’s owed ~4.5 million shares (about $1 billion) because his 1993 options should have vested in one year, not four. Courts ruled the claim time-barred. HN’s verdict was unsympathetic — you’re responsible for reading your own contract — with one commenter noting the 15,625 shares he did exercise in 1996, had he held them, would be worth more than what he’s now suing for.
That’s the digest. If you’re keeping a scoreboard on the agent incidents, we’re now at UN, Australia, Hugging Face, Department of Education, RubyGems, and an internal DNS escape — all within roughly six months. See you tomorrow.