AI News, Nov 22: Two Groups Say the Proof Checks Out
The Big Story: two groups say the proof checks out
On Saturday the Lean and Mathlib communities published a joint statement, endorsed by the Lean FRO and the Mathlib Initiative: “We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.” Levent Alpöge, a mathematician at Anthropic, had released it the day before as a Lean library “comprising 17 million lines of code,” with no prose version.
What they did is worth spelling out, because nobody read the proof. Two tools confirmed “that the theorem proved in Alpöge’s library is exactly Mathlib’s statement of the Riemann Hypothesis,” itself “carefully vetted by experts in Lean and number theory.” Dozens of people recompiled the codebase and reproduced the same raw proof export, which then went through Lean’s official kernel plus con-leche and con-ron, checkers that “come with a formal proof, relative to set theory, that they cannot accept a false proof,” closing off the worry that the proof had exploited a hole in Lean’s logic.
MIT’s Andrew Sutherland keeps the running audit, updated Sunday: all 175,187 modules compile on Lean v4.33.1 with zero errors, con-leche accepts 1,468,418 declarations in verified mode, and #print axioms on the five headline theorems returns propext, Classical.choice and Quot.sound only. A comparator run and the Lean kernels on the export have no verdict yet. His page also fixes the detail that got mangled all weekend: “GRH” here “means Dirichlet L-functions: all moduli and all characters,” and Dedekind zeta, Hecke and automorphic L-functions “are not among the statements that were checked.”
Alpöge says Claude helped develop the solution over about six weeks, and wrote that “Claude flew, and danced, and sparkled near that finish line.” The Lean statement does not mention Claude, Anthropic or AI anywhere, and closes by looking forward to the work “that will help explain the new ideas in this proof.” Correctness is settled. Comprehension is not.
Every story this weekend rested on a check no person could make by reading
The proof needed kernels because 17 million lines are unreadable. A security team planted two million prompts across the inference supply chain because a provider’s zero-retention promise cannot be audited off a policy page. METR spent its quarterly update arguing that the benchmarks used to verify capability are flattering the models. Sutherland’s audit is a table of exit codes. In each case the thing being checked had outgrown the people meant to check it, and the remedy was another machine, down to the byline on that audit page: “checks run with Claude Code.”
Today’s Top Stories
Two million canary prompts caught attackers inside the inference supply chain
HD Moore’s team at Dragos submitted “approximately two million prompts containing unique URLs and canary token credentials” to widely used AI services from July through November, then watched what touched them. They “found attackers actively stealing data from three different hosting providers.” The largest breach involved “a couple hundred of our AWS keys,” each sent in exactly one request to one provider, later used for “tens of thousands of AWS API calls.” “Some of the stolen keys came from requests that the provider rejected before a model ever ran.” One Chinese lab’s systems fetched the secret URLs from “more than half of the prompts that we sent to its models,” including “roughly two-thirds of the requests marked as ‘no data collection’ or Zero Data Retention.” Resellers needed no exploit: “list a model at the lowest price, enable request logging, and wait.” OpenRouter responded on Sunday by off-boarding three providers and stripping 125 endpoints of ZDR eligibility, saying the findings showed no breach of OpenRouter itself. It did not name the three.
UN autonomous weapons talks collapsed just before 5am
A week of talks toward a binding autonomous-weapons treaty failed to overcome Russian objections, and the conference fell back on a mandate asking states “to elaborate by consensus” an instrument that may carry no legal obligations. More than 100 states wanted a new law. The US said human control over autonomous weapons came up in Xi and Trump’s Apec meetings but was reluctant to be bound; Japan, South Korea, Turkey, the UK, France and Germany also resisted. The forum, the Convention on Certain Conventional Weapons, needs all 128 parties to agree and has passed no new weapons law in 30 years. Human Rights Watch’s Bonnie Docherty noted treaties can be negotiated in 15 months, against the US proposal of five years just to settle a mandate.
METR says its own evals are probably flattering the models
METR’s Q4 capability update, published Thursday, puts time horizons “doubling roughly every 3.6 months” on its combined suite, with the strongest models placing in “the 96th to 98th percentile of human participants” across 15 programming contests. The useful part is the self-doubt. Doubling a task’s lifecycle cost imposes a difficulty penalty “very roughly equivalent to halving the time horizon of the agent,” so cheap evals overstate ability, and METR is “increasingly concerned that our easy-to-run, containerized evals could overestimate real-world capabilities, due to similar RL environments becoming common in training.” Its composite index shows no strong acceleration or deceleration, a result it says it is unsure of: “current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years.”
Quick Hits
- Screenshots: Signal extended its Windows screenshot block to Android and iOS, with president Meredith Whittaker calling operating systems that integrate agentic AI an existential threat and warning that turning it off would “break our privacy guarantees.”
- Devotion: Sam Kriss’s New Yorker investigation of a Vermont monastery recruiting burned-out programmers, which he calls “a Scientology for Silicon Valley,” reports claims it records residents’ every breath to train an AI god modeled on its leader’s mind.
- Monitoring: Perplexity shipped a Monitor API with scheduled searches, page diffs and signed webhooks that fire only when a decision model clears thresholds for relevance, credibility, novelty and timeliness, at $1 per 1,000 runs.
- Terseness: on Artificial Analysis’s personality benchmark, GPT-6.2 Astra’s “Admits Uncertainty” ranking fell from 96 to 24 generation over generation, with average answers down from 352 to 210 words.
- Sovereignty: of 11 regulated enterprises surveyed, nine now require inference inside their own boundary and seven have never tested disconnected operation, key revocation or exit.
- Quantization: a trellis-code compiler squeezing Inkling-2-Mini from 18.89GB to 3.85GB drew 507 points on Hacker News and a thrashing in the comments for benchmaxxing.
- Power: advocacy group the Better Data Center Project estimates from EPA marginal-generator modeling that 87% of Anthropic’s contracted US data center power will be gas in 2028.
- Context: Nvidia researchers argue huge context windows are the wrong lever, since later tokens cost as much as 7× more and frequent compaction sacrifices “little to no task accuracy.”
- Slop: a Claude-generated prose rendering of the Lean proof was posted to Hacker News claiming nobody had built the project in full, which was already false when it went up.
- Quiet: arXiv announced no new papers for a second day, no frontier lab shipped a model, and TechCrunch published nothing at all on Sunday.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
