Newshot

The most famous hard open problem in mathematics has just been solved. Now what?

Summary

Subtitle: Ours is the age of marvels and monsters. The Riemann Hypothesis is classically considered the most famous problem in mathematics. Earlier today, Levent Alpöge, a mathematician and AI researcher at Anthropic, has posted a claimed solution , along with some remarks (short, will recommend reading!) This solution is probably correct. I’m not a mathematician and cannot check myself, however there’s strong evidence that it’s real: a) it is formally verified in Lean (Lean bugs are rare), b) Levent is considered a great mathematician even before AI, and does not make strong claims about mathematics without checking, and c) he’s posted multiple AI-assisted proofs of open mathematical conjectures earlier this year, all of which has later been checked and verified to be correct. Other mathematicians and commentators have been assuming it’s real as well. It’s hard to overstate how big a deal this is for mathematics. The Riemann Hypothesis has been open for 167 years , the one problem tha…

Original text(via LessWrong)

Subtitle: Ours is the age of marvels and monsters.

The Riemann Hypothesis is classically considered the most famous problem in mathematics. Earlier today, Levent Alpöge, a mathematician and AI researcher at Anthropic, has posted a claimed solution, along with some remarks (short, will recommend reading!)

This solution is probably correct. I’m not a mathematician and cannot check myself, however there’s strong evidence that it’s real: a) it is formally verified in Lean (Lean bugs are rare), b) Levent is considered a great mathematician even before AI, and does not make strong claims about mathematics without checking, and c) he’s posted multiple AI-assisted proofs of open mathematical conjectures earlier this year, all of which has later been checked and verified to be correct.

Other mathematicians and commentators have been assuming it’s real as well.

It’s hard to overstate how big a deal this is for mathematics. The Riemann Hypothesis has been open for 167 years, the one problem that is considered the most important in all of pure mathematics, and most mathematicians prior to 2026 thought it’d be unsolved for decades.

And now it’s solved. Fuck.

My immediate reactions (keeping in mind that I’ve only had a few hours to process all of this):

  • Shit. Fuck. Fuck. Damn.
  • Full RSI/takeoff really is coming in 6-12 months, isn’t it?
  • We absolutely can’t rely on AI to solve the problems from AI
  • We need to put a hard pause/halt to AI development ASAP.

Shit. Fuck. Fuck. Damn.

Holy fuck this is a big deal. On a whim, I asked my own Claude about its probability that Riemann would fall before End-of-Year 2027, and it estimated 1%. My impression is that mathematicians and forecasters had similar numbers. AI progress surprises all of us, AIs and humans alike.

Today, December 31st of 2027 seems like a lifetime away.

2.5 years ago the best frontier LLMs couldn’t count the “R”s in strawberry. Less than 1.5 years ago (July 2025), two different frontier models got IMO gold, the hardest mathematics competition for high-schoolers. In November 2025 (about one year ago), an LLM first solved a problem that counts as “real” math to working mathematicians. In May 2026, an LLM solved a problem that mathematicians actually cared about for the first time.

Now, half a year later, LLMs are better than all human mathematicians combined. An entire ancient academic field, composed of some of the smartest people we know, just obliterated.

Full RSI/takeoff really is coming in 6-12 months, isn’t it?

Last Friday, I wrote/published a blog post announcing that I’m pivoting to assuming very short (6-12 month) timelines for my own irrelevance, and thus wanting to devote all my time to pausing AI.

In the last week, I’ve talked to many people both online and offline (but mostly off) about the various assumptions and implications, and expressed a lot of genuine uncertainty about the actual timelines. After all, these questions are very hard to predict, and AI safety people have been wrong before (Though almost always in the direction of assuming too long rather than too short timelines).

Even as I wrote my article laying out the case for why we’re currently in the early-middle stages of runaway AI research and development automation (aka recursive self-improvement, aka RSI), I let myself have hope that this is all a mirage, like Shepard tones. Maybe the new OpenAI evidence suggests jaggedness towards less research acceleration than the METR report would suggest. Maybe the acceleration in easily-verifiable tasks is just that, and Claude models are more bench-maxxed than they used to be. Maybe Anthropic fed METR evidence towards imminent full RSI to spruce up their IPO.

Lol, nope. Lmao, even.

My current guess is that while hitting a wall of strongly diminishing returns is still theoretically possible without hard regulation, almost all current evidence points away from that.

We must pause further AI research and development as soon as possible, before it’s no longer possible to pause this mad race.

We absolutely cannot rely on AI to solve the problems from AI

There are many reasons I’ve personally been slow to update towards very short timelines, both principled and psychological. But a salient one that comes to mind is lived experience: I’m personally just not seeing much uplift in my own work, and I think I foolishly and subconsciously allowed that to cloud my judgment on how much uplift is happening overall.

Unfortunately, this bodes poorly for our future ability to defend well against the problems posed by AI.

First, some background: my day job is at Forethought, a research nonprofit and think-tank whose primary goal is figuring out the intelligence explosion and trying to research how to solve all the non-alignment problems posed by superhuman AI. So we try to figure out how to make progress on reducing extreme (human) concentration of power, making space governance non-crazy, improving elite decision-making during crazy times, etc. My own work was centered around AI persuasion and countering the effects of superhuman manipulation and automated propaganda.

And I’m not just seeing that much uplift? Like don’t get my wrong, the Claudes and ChatGPTs are certainly useful. Like I’d rather have them for my work more than pretty much any other software productivity tool, Excel and Google Search included. But the total serial uplift is maybe 1.2-1.4x realistically, maybe 1.5x on a good day/specialized tasks.

Other people working on AI safety report something similar. Here’s Buck Schlegris, CEO of Redwood Research (one of the best AI safety nonprofits):

I don’t think our empirical research has been going 5x faster since we got access to Fable/Opus 5.5; I would have guessed more like 2-3x speedup. Why is this?

And technical/empirical alignment and control are probably among the easiest “problems caused by AI” to get AI automation and uplift on. “Thinking about the intelligence explosion well and navigating it wisely” is just a much less well-specified problem!

Contrast this to the uplift in mathematics, which with the recent Riemann breakthrough would seem to imply >100x uplift for all human mathematicians combined. Or more saliently, the evidence that we’re currently in the early-middle stages of AI recursive self-improvement, where Anthropic ML researchers are seeing >10x uplift for ML research itself.

Earlier today I posted this to my company Slack:

I dunno if this is reasonable but the recent Riemann solution makes me even more pessimistic that we can get major macrostrategy uplift without pacing/pause, whether gov’t enforced or unilateral (like the OG plan to do a crazy race to superintelligence and then stop for a while).Like they just solved the most famous hard very hard math problem, the one that mathematicians (from a field classically considered the hardest/most high-iq field) were decades behind, and meanwhile Fable 5.2 (which afaict is < 3 months behind the frontier) is only moderately useful for uplift for FT-style macrostrategy.(I know I’m probably more of a dinosaur on AI use compared to [redacted]/[redacted] but it’s not like I’m not trying/I don’t use it! Is anybody even getting like 3x consistent serial uplift? AFAICT the only macrostrategy things that see crazy uplift are philosophical problems that are closer to math; and mathematized philosophy’s not going to be what saves us from non-alignment problems)And all signs point to AI research being closer to math than macrostrategy is, both on priors and in terms of what the companies are aiming at.

I dunno if this is reasonable but the recent Riemann solution makes me even more pessimistic that we can get major macrostrategy uplift without pacing/pause, whether gov’t enforced or unilateral (like the OG plan to do a crazy race to superintelligence and then stop for a while).

Like they just solved the most famous hard very hard math problem, the one that mathematicians (from a field classically considered the hardest/most high-iq field) were decades behind, and meanwhile Fable 5.2 (which afaict is < 3 months behind the frontier) is only moderately useful for uplift for FT-style macrostrategy.

(I know I’m probably more of a dinosaur on AI use compared to [redacted]/[redacted] but it’s not like I’m not trying/I don’t use it! Is anybody even getting like 3x consistent serial uplift? AFAICT the only macrostrategy things that see crazy uplift are philosophical problems that are closer to math; and mathematized philosophy’s not going to be what saves us from non-alignment problems)

And all signs point to AI research being closer to math than macrostrategy is, both on priors and in terms of what the companies are aiming at.

This is just on the research side.

Implementing defenses also takes time, time we desperately do not have during the technological singularity. Even if Claude Fable 9 comes up with the (equivalent of) Lean-checkable white paper that definitively solves all problems with AI propaganda, we still need to implement the relevant defenses, and implementation takes time.

We’ve already observed this in cyber and (to a lesser extent) bio. AI boosts both offensive cybersecurity (hacking) and defensive cyber (building secure code). The AI companies, various governments, and cybersecurity people across the world are trying their absolute best to secure all their systems against the new AI-discovered bugs, and there are actually pretty decent theoretical arguments that cyber in the limit is defense-dominant1. And yet in practice we’re sure seeing a lot of hacks.

Don’t get me wrong: I do think future AIs can and will solve some of the problems caused by current AIs. The AIs will be able to solve many of the problems we first encountered years ago, and maybe (as AI progress gets faster and faster) the problems encountered 6 months ago. As proof-of-concept, AI hallucinations are much less of a problem than they used to be, extreme sycophancy and AI psychosis is down despite greater chatbot use, and misuse jailbreaks are increasingly uncommon for frontier models.

But if we want to get to the “other side” of AI progress intact and well, never mind safely and wisely, AI systems need to be able to solve new and increasingly catastrophic safety issues anticipatorily or at least contemporaneously. And nobody – not the AI companies, not the governments, not the AIs themselves, certainly not the AI safety community – has a plan to do this, for alignment and non-alignment problems alike.

We need to put a hard pause/halt to AI development ASAP

And thus, we need to pause AI research and development ASAP. Even for a few months.

We desperately need to buy time for researchers, policymakers, and the general public to understand what the hell is going on, and have an actual plan, or even concepts of a plan, to navigate the intelligence explosion wisely. And for that matter, which we even want to navigate the intelligence explosion at all, a question that policymakers and normal people to a large extent were never consulted on, and initial signs suggest they strongly oppose.

The best time to end this dangerous, reckless, undemocratic, and inhuman mad race to superhuman AI was five years ago. The second best time was four years and 364 days ago. But sometime down the line, the next best time is now.

I encourage my followers to learn more about the dangers posed by AI via online searches and informational websites, contact their Congresspeople at CallCongress.ai, or the equivalents in other countries, sign on to petitions like Do Not Accept AI, and join or organize local protests about the reckless mad race to superhuman AI.

Feel free to ask more questions or coordinate in the comments of this Substack post.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by LessWrong
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗