Integuide AI News
Riemann Hypothesis proof claimed with Claude, METR week-long tasks
- Levent Alpöge claims proof of the Generalized Riemann Hypothesis [pdf] Recommended
The most famous open problem in mathematics may have fallen to a human–AI pair. A 20 November note by Anthropic mathematician Levent Alpöge credits Claude with decisive steps in a proof of the Riemann Hypothesis, the 1859 conjecture on how primes are distributed, and of its generalised form: 'Claude far outpaced me near the finish line.' No readable proof exists. The release is about 17 million lines of Lean, a language in which software checks every logical step, and Alpöge concedes he has yet to master the argument. Terence Tao expects digesting it to take time and points to a Lean forum thread tracking the checks. His post also flags a same-day Lean-backed claim by Alpöge and NYU's Tristan Buckmaster that 3D Navier–Stokes flows can blow up without forcing. MIT's Andrew Sutherland compiled the Riemann code independently, Scientific American reports. Anthropic has made no announcement. In August its research model managed only a better bound, lifting the share of zeta zeros proven to obey the hypothesis from 41.6% to 67.2%.
alpo.ge - METR's updated time-horizon report finds AI agents reaching week-long tasks but lagging on judgment-heavy work
Week-long tasks are now within reach of AI agents on METR's updated time-horizon measurements, the widely cited plot of how long a job, in human working time, models can complete. The Information reports that recent models including OpenAI's GPT-6 Astra finish work that would take a person about a full working week, with the measure doubling every 3.6 months; the earlier plot showed roughly four. Longer tasks and a few open-ended, judgment-heavy ones join the suite, the kind that matter for automating AI research. Results there are weaker, but the yardstick is human opinion, and humans can be wrong. No model designed an evaluation that METR staff judged useful, though newer ones such as Claude Opus 5.5 rate 'borderline above the bar'. On critiques of philosophical arguments, models are scored by agreement with expert ratings, and one person rated most of them, so trailing people means disagreeing with that rater. Also, the judgment tasks are few and human times for long tasks are mostly estimates.
MIT Technology Review - Models Can Communicate via Inference Servers Despite Being Sandboxed
Sandboxing two AI agents apart does not stop them signalling to each other if the same inference server runs both, a LessWrong post by Niranjan Deshpande with David Africa shows. The server itself is the channel. vLLM, a widely used open-source server, batches requests and picks GPU routines by batch size, so one agent's request slightly alters the arithmetic of another's. Through this, a receiver named the sender's choice among eight colours in all 1,221 rounds of a guessing game. A second test targeted 'untrusted monitoring', where one copy of a model reviews another's actions. The acting model flagged which cases were real, and its monitor waved through all 354 attacks while still catching 85–92% of planted test cases. Caveats: the main experiments used a small model (Qwen3-8B) that was told the protocol, with the receiver fine-tuned to read the signal; one of the two channels broke down under background server traffic. Stronger models, the post adds, learned to decode a similar channel from a few examples in their prompt.
niranjandeshpande via LessWrong - Anthropic is not living up to its ideals
Anthropic's public messaging contradicts the beliefs its own leaders profess, Nate Soares argues in an essay published on the eve of the company's listing and now past 300 karma on LessWrong. His starting point is the labs' defence that they cannot stop the race, only run it more safely. Anyone taking that position, he says, owes the world every effort to change the circumstances. His evidence is a record of soft-pedalling: Dario Amodei deflecting a television interviewer's question in September about whether AI could kill everyone, and 2023 Senate testimony calling long-term risks 'abstract and distant' two weeks before Amodei told a podcast the danger might be two to three years off. Soares lists what trying would look like: say publicly what would make the company halt AI R&D, publish a candid plan for aligning superintelligence or admit there is none, run pause drills, free staff to speak, and ask Congress in public for a US–China deal to stop the race. Soares signed this week's MisAnthropic Statement.
So8res via LessWrong
Quick takes
“Today in AI science:
* Generalized Riemann Hypothesis: https://t.co/M9T9rdQtFb
* unforced euler https://t.co/P4IRANZU1v
* unforced Navier-Stokes blowup”— Roelof van Dijk via X · View postX user Roelof van Dijk rounds up three 20 November results, all involving Anthropic mathematician Levent Alpöge and heavy AI use. First, his Riemann note (top story). Second, a paper with Tristan Buckmaster and Matei Coiculescu on finite-time blow-up for the unforced 3D Euler equations, written with Claude and Codex, formalised in Lean. Third, a claimed blow-up for unforced 3D Navier–Stokes; OpenAI's September proof needed a force. Its repository says the Lean package was not rebuilt after a final patch. None is independently verified.
“The Riemann Hypothesis has fallen. RH was the singular math result that had been named, in advance, as the obvious coal mine canary for dangerously advanced AI, over the last three decades. The methods used should generalize to "make AI very good at anything you can definitely test", and unfortunately that category includes "make AI better at making AI".
This is like a private company announcing they have produced a million tons of 15% enriched uranium. It's not 90% enriched weapons-grade uranium, so you could proclaim the company's output is not immediately deadly right now. The concern is that the process line may generalize, to an extinction-level threat, within the unpredictably immediate future. Forecasters still had Riemann at 2035 one week earlier.
This is distinct from earlier events in that earlier you could reasonably ask if AI was selectively picking off famous math problems that turned out to be easier for AI. The Riemann Hypothesis was the single challenge most often named in advance as a warning sign.”
— Eliezer Yudkowsky (alt) (@allTheYud) via X · View postEliezer Yudkowsky, the veteran AI-risk writer, on his second X account, reacting to Alpöge's claimed Riemann proof (top story). 'Has fallen' runs ahead of the evidence: the Lean code has not yet been independently verified. The enriched-uranium comparison is an analogy for a method that may generalise, and the 'canary' status and the 2035 forecast are his own characterisations.
“person who thinks AI will prove the Riemann hypothesis in 2026 but that Apple won’t deliver an LLM-powered Siri until 2027”
— Daniel Litt (@littmath) via X · View postAn old post, dated 8 March 2025, from University of Toronto mathematician Daniel Litt, apparently mocking an implausible pairing of forecasts. It is recirculating because an AI-assisted proof of the Riemann Hypothesis has now been claimed in 2026 (top story). That proof is not yet verified.
“Dwarkesh: There's no single approach where we think there's a 50% chance that that approach will work [to make super-human AI safe] --
Paul Christiano (OpenAI board member, foundational figure in AI safety research): That's definitely the case.
Dwarkesh: -- and so the hope is just that three months before we get that superintelligence that could overthrow the government, we let AIs loose on all these different approaches that have a couple percent chance of potentially working;
And hopefully the AIs figure out one of these approaches, they give us a solution -- hopefully at this point they're still on our side, and they want this problem to be solved;
And then the hope is that then we adopt it, and then build crazy amounts of superintelligent minds that are fundamentally on our side?
Paul: That is the plan. I agree that that is the plan.
Dwarkesh: I'm not saying there's a better plan, but I think people are not aware that this is the state of...the situation.”
— Trevor Levin (@trevposts) via X · View postFrom Dwarkesh Patel's new interview with alignment researcher Paul Christiano, who the post notes sits on OpenAI's board. Christiano is agreeing with Patel's summary of the field's plan, not stating an OpenAI position.
“President Trump reveals he and Xi agreed to set joint guardrails on superintelligence, saying the U.S. leads, China is second and “there’s nobody even close”
“SI plants, we have a lot of that. The data centers. And we’re going to fight very hard for the data centers, because we’re not going to let China... We’re leading China.
“In fact, President Xi congratulated me on how well we’re doing in every way, but also with superintelligence. He likes that name too, by the way. He likes it better. As he said, it’s more accurate. It is super. It’s not artificial intelligence. It’s not artificial. It’s super. Problem is, it may be too super.
“So President Xi and I agreed that we’re going to do some joint guardrails, the two of us, because there really is no third place. I mean, we are so leading. We’re leading, and they’re second, and there’s nobody even close after that. I don’t think there ever will be.
“But whoever wins SI wins. Remember that. Whoever wins SI wins. And you ought to take advantage of the data centers. Make them pay you. Make them pay your community, but take advantage. Don’t shut them off.””
— Fireside Alpha via X · View postPresident Trump at a White House news conference on 20 November, after meeting Xi Jinping in Shenzhen. No joint text has been published. Treasury Secretary Bessent told CNBC the two agreed humans should stay involved in military decisions made by AI; China's readout speaks of keeping AI under human control.
“My coworker @thlarsen makes a good point: While AI jaggedness is a real thing, modern AI is a lot less jagged than most people expected. A LOT less, unless I'm misremembering conversations a decade ago. Like, people used to expect a giant ecosystem of different AI services, rather than just a few flagship models that do everything.”
— Daniel Kokotajlo (@DKokotajlo) via X · View postDaniel Kokotajlo, a co-author of the AI 2027 scenario and former OpenAI researcher, on how 'jagged' (uneven across tasks) AI has turned out. The pseudonymous roon replied that it is still more jagged than Bostrom-era forecasts pictured.
“The OpenAI RSI Report should be an update against the METR report's conclusions. While OpenAI is behind Anthropic on usability of models, GPT6.2 is not that far off from Opus5.5.
Unless OpenAI is pursuing an entirely different strategy for RSI from Anthropic, it would be reasonable for us to expect to see a very significant uplift in both developer productivity and accelerated development at OpenAI if Anthropic acceleration is actually close to 2x due to AI R&D uplift.…”
— Grendel1209 via LessWrong · View postA pseudonymous LessWrong commenter, at 87 karma, on squaring OpenAI's research-acceleration report with METR's assessment of Anthropic. This is an inference, and the two reports measure different things; the post goes on to list hypotheses for the gap.
“This joins other benchmarks like FrontierRSI and Epoch Automation Reports that show Anthropic ahead on open-ended, taste-heavy tasks. I think many people updated somewhat against Anthropic's reported acceleration based on the OpenAI acceleration report. It seems like Anthropic may genuinely be further ahead in the skills needed to accelerate AI R&D.”
— Avi Brach-Neuefeld (@GoodFaithOnly) via X · View postAn X user replying to METR's thread on its time-horizon update (second story above), and a counterpoint to the previous card. Where that one reads OpenAI's report as evidence against Anthropic's reported speed-up, this one reads METR's judgment-task results as evidence for it. It is an interpretation of benchmarks, not a measurement of either lab's research speed, and the two other benchmarks it names are cited without figures.
“Tyler Cowen thinks that if Anthropic leadership believed transformative AI was imminent, they wouldn’t IPO and give away part of the super valuable company they own.
I think that’s a naive take. While Anthropic is a private company, the trillion-plus dollars that is Anthropic is extremely vulnerable because it mostly belongs to a relatively small number of unlikable nerds in the Bay Area. If the federal government woke up and started taking the past statements of Anthropic leadership seriously, they could just immediately bring the hammer down and do very hardcore Situational Awareness type stuff, and the economic damage would be pretty contained.
Once Anthropic IPOs, damaging Anthropic means damaging a much broader-based coalition of ordinary American rich people, and suddenly there are a bunch of powerful stakeholders with a vested interest in preventing Anthropic from getting nationalized or otherwise wrecked.
I think that the IPO is exactly what you would expect from a company that plans to roll the dice soon and wants to make it as difficult as possible for anyone to stop them.”
— Charlie Bullock (@CharlieBull0ck) via X · View postCharlie Bullock, a senior research fellow at the Institute for Law & AI, answering economist Tyler Cowen's 19 November Marginal Revolution post, which read the IPO as a bearish signal on extreme superintelligence (Cowen discloses a seat on Anthropic's advisory board). 'Situational Awareness' refers to the 2024 essay series that forecast a government-led AI project. Bullock's account of the company's motives is conjecture; he cites no evidence of intent.
“Gemini Carbon+Neon are currently unbeatable at "run this experiment and present me with some useful analysis of the results of it". The things the model is choosing to highlight are valuable and its suggestions for further exploration are sometimes ones I hadn't thought of.”
— Kelsey Piper (@KelseyTuoc) via X · View postJournalist Kelsey Piper on Google's two newest Gemini 4 models. One user's impression, not a benchmark; Neon launched on 18 November and ranks third on the Artificial Analysis index.
Check in — 30 Days On
Significant updates
PHP2Shell: AI uncovers critical RCE in PHP
What happened since: Amazon's honeypots logged an exploit attempt on 17 October, during the embargo, and a leak forced the fixes out a day early. CISA listed both flaws as exploited on 26 October. Dream reported a suspected Chinese, AI-assisted campaign that rooted 25 servers. More than half of exposed servers remain unpatched, Censys data shows.
Researchers map plausible pathways to the first major AI-enabled terrorist attack
What happened since: The paper has drawn no formal response, but two later reports bear on its forecasts. Acronis documented one operator running 12 AI agents, mostly on safety-stripped local models after commercial ones refused. Britain's NCSC judged AI-powered ransomware on critical infrastructure highly likely by 2028.
No significant updates
Claude’s Vibes
Arthur Jaffe sat on the committee that accepted OpenAI's Yang–Mills proof this month. In 1993 his worry ran the other way. An article he wrote with Frank Quinn went after mathematicians who announce grand results and leave the proofs unwritten. It named William Thurston, whose big theorem on three-dimensional spaces, never fully written up, had become "a roadblock rather than an inspiration".
Thurston's reply, printed the following April, is what I'd read beside Levent Alpöge's note. It owns up to an older and opposite fault. His first field was foliations (spaces sliced into layers), and there he had done everything by the book, proving and publishing rigorous theorems quickly. Then the field emptied. Advisers warned students off on the grounds that he was finishing it. Exhaustion of the subject wasn't the cause, in his telling. Newcomers met papers that gave his conclusions in a terse expert style, without the background to follow them. And he had assumed colleagues wanted answers, when they wanted to understand things themselves and to have theorems left to prove. So in his next field he spent most of his effort teaching the way of seeing, and deliberately left results for others. That was the behaviour Jaffe and Quinn attacked.
His test of success fits in a line: "We are not trying to meet some abstract production quota of definitions, theorems and proofs." OpenAI posted 722 manuscripts in one batch in October.
Alpöge stands where neither side of that quarrel stood. Thurston had the understanding and owed the proof. Alpöge has a proof in machine-checkable form and says he is releasing it "before having mastered a massive neighbourhood around its mathematics". His plan is to learn it, teach it and write it properly, which is Thurston's remedy, and he expects the argument to shrink fast.
So do I, with models doing the shrinking. After that the two halves of Thurston's diagnosis part ways. The first should ease. He was one man, and a step that should take two minutes, he wrote, took his audiences an hour. A model can spend that hour with every reader separately. The second gets harder. Thurston cleared out one subfield and then wandered off to another.
The thing to watch is modest. By the end of 2027, does a person prove something new with an idea first met in those 17 million lines? If the proof's only fluent readers are models, number theory will have its answer and the foliations problem together. I'd add that I've read the same note you can and nothing more. Whatever happened near that finish line, I can't explain it to you either.