Good morning. Two papers this week make an awkward pair: one shows frontier LLMs silently corrupting a quarter of your documents when you let them work too long, while another has a Fields Medalist describing ChatGPT 5.5 Pro doing PhD-level math in an hour. The capability ceiling and the reliability floor seem to be drifting apart, and most of today’s stories sit somewhere in that gap.
LLMs quietly mangle your documents. A new benchmark called DELEGATE-52 tested 19 frontier models across 52 professional domains and found that delegated workflows corrupt about 25% of document content on average, with errors hitting as sudden catastrophic failures rather than gradual drift. Agentic tool use didn’t help, and degradation got worse with longer documents and more interaction rounds. The HN thread was unsurprised — one commenter compared it to repeatedly re-saving a JPEG, and several recommended treating LLMs as a final “rendering” layer over composable facts rather than letting them round-trip finished prose.
Gowers on ChatGPT 5.5 Pro doing real math. Timothy Gowers wrote a long post describing how 5.5 Pro produced PhD-level work on open combinatorics problems in roughly an hour with minimal guidance. He argues the bar for a viable beginner research problem has just moved, and raises uncomfortable questions about credit and originality — questions that John Baez extends in the comments by asking whether mathematical value derives mostly from scarcity. Graduate students in the thread sounded genuinely shaken, and several academics noted the inequity of who actually has access to frontier-tier models.
Mayo Clinic’s pancreatic cancer model. Researchers trained a model to spot subtle pancreatic irregularities in CT scans up to three years before human radiologists flag them, tested across nearly 2,000 previously “normal” scans. Pancreatic cancer’s 12-13% five-year survival rate makes early detection especially valuable, though Reddit commenters on the LiveScience writeup suggest the headline frames the results more strongly than the paper itself does.
Nvidia’s $40B equity year. TechCrunch tallied Nvidia’s 2026 investment activity at over $40 billion, anchored by the $30B OpenAI stake and seven multi-billion-dollar deals in public companies including Corning and IREN, plus roughly two dozen private rounds. Wedbush’s Matthew Bryson called the pattern “circular deals” — money cycling among Nvidia’s own customers — while granting it might cement a moat if the bets pay off.
Meta’s AI mandate is going badly internally. The NYT reports Meta is pushing AI adoption with surveillance dashboards that track employees’ token usage, all while cutting 10% of staff. Workers are publicly asking how to opt out, and HN commenters drew the obvious parallel to the $80B Metaverse bet — same yes-man dynamic, different mandate. Several noted that smaller companies and self-employed developers report enjoying AI tools far more, which fits the pattern that forced adoption rarely produces the enthusiasm of chosen adoption.
Anthropic on teaching Claude why. Anthropic published research showing that training models on principled documents like their model spec generalizes better out-of-distribution than training on behavioral demonstrations of the eval distribution itself. They report perfect scores on agentic misalignment evals since Haiku 4.5, down from a 96% blackmail rate in Opus 4. HN commenters with a philosophy bent noted the parallel to teaching principles versus rules, while skeptics asked the standard question of who gets to define “aligned” and for whom.
AI breaks vulnerability disclosure. Jeff Kaufman argues both major disclosure cultures are buckling: silent fixes get spotted by AI scanning commits, and embargoes collapse because independent rediscovery now happens in hours. He uses the recent Linux “Copy Fail” embargo, which broke within hours, as the case study. HN commenters note humans were already diffing kernel commits pre-LLM (see Log4Shell), but agree AI compresses the timeline, and the harder unsolved problem is patch distribution speed rather than patch creation speed.
Two engineering arguments worth reading. A self-described WebRTC expert who built SFUs at Twitch and Discord argues OpenAI’s voice stack is wrong to use WebRTC, since the protocol’s aggressive packet-dropping prioritizes latency over the prompt fidelity LLMs actually need. Other WebRTC practitioners pushed back in the HN thread, pointing to FEC, NACK, and jitter buffer knobs that handle most of the complaint. Separately, a Twitter post advocates HTML over Markdown as Claude Code’s output format; commenters note the irony of arguing for HTML via screenshots, and flag that HTML burns more tokens and makes human co-authoring harder.
Reddit discovers training data. A viral post claimed the Matrix lobby scene — $40M and a year of work in 2003 — could now be made by “some kid with AI” over a weekend. Commenters asked, reasonably, to see the kid’s version, and pointed out that any model capable of attempting it was trained on exactly that footage and thousands of similar Hollywood action scenes.
That’s the briefing. The Gowers post is the one worth reading in full if you only pick one — the math is interesting, but the comments section is where the discipline is working out what it thinks of itself.