logoalt Hacker News

yewenjie • today at 3:14 PM • 40 replies • view on HN

I know this thread will be full of people in denial about full recursive self-improvement / intelligence explosion / superintelligence (the original meaning of the word) within months or a year, so just quoting Scott Aaronson:

"Experience has shown that, even now, there will still be people explaining in patronizing tones why none of this is real and none of it counts. If such people were capable of being impressed by anything that happens in the empirical world, of updating on anything, they would’ve already been impressed and already updated several years ago, long before things had reached the point of an actual Mathocalypse."


Replies

bodge5000 • today at 3:31 PM

My thoughts remain the same since I've been hearing "months or a year" for the past 5 years. Maybe this time truly is different, but unless acceptance of super-intelligence is required to take some action (most people I've seen who hold this view also don't think it should be seriously regulated so I'm not sure what that would be), why not just wait until it happens rather than making more predictions? The only thing I've ever seen come out of these predictions is willing me to "stop resisting", but there's no reason to do that until its here.

If you do want to see something done about it and this isn't just another prediction or a vague warning, more power to you, but otherwise, I'll wait and see

➕ show 9 replies
jrflo • today at 3:22 PM

I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away. In 2025 AI got a gold at the IMO for the first time, a competition for high school kids. Now it's likely solved the biggest problem in math. There's no reason to expect progress to stop here, especially on domains highly amenable to RL.

➕ show 9 replies
dangom • today at 4:12 PM

These are extremely jaw dropping results, but I feel even though it is easy to conflate them with RSI / intelligence explosion, another simpler explanation is that symbolic manipulation is something humans are notoriously bad at, hence we prizing individuals that can rise above and dedicate their lives to it, and call them intelligent.

For computers though, symbolic manipulation is something they are extremely good at, and something that clearly scales well will memory and compute. Therefore I would argue that calling a computer superintelligent would, in analogy to do what we do with humans, require computers to rise above things that are extremely difficult for them. It is of course hard to judge what would be hard for them not being them for the introspection, but it seems the ability to operate outside the realm of symbolic or numeric manipulation would be precisely that.

It stands to reason that if some proofs require chaining 250k pages together, no human would ever be able to put it together. That doesn't mean something that is capable of doing so is necessarily super-intelligent.

Still very impressive, and if true, a wonderful result of the symbiotic interplay between a human and a machine, each at what they are good at.

➕ show 4 replies
FiberBundle • today at 3:36 PM

When Atari games or Go were solved around 10 years ago, nobody was saying that superintelligence was within months just because reinforcement learning worked on a domain that allowed learning by self-play. Now we have a domain that is solvable by reinforcement learning with verifiable rewards. It's very impressive that these models are capable of doing the things they do, but the techniques aren't transferable to the vast majority of other domains. ML research also isn't verifiable in the sense that math is. People who claim it's verifiable, because there's a target that can directly be optimized don't understand why RL works for math. It's that lean is able to massively shrink the search space. This simply cannot be done in other domains in the same way.

➕ show 6 replies
CodeMage • today at 5:31 PM

> I know this thread will be full of people in denial about [my prediction]

You can only be in denial about the past or the present. The future has a chance of being different from your predictions.

Personally, regardless of whether you want to call it "denial" or something else, I'm fervently hoping the AI never achieves superintelligence, but I would settle for it not achieving it until we can do our best to ensure that the superintelligence will be benevolent towards us.

Call me selfish, call me petty, call me a human supremacist, call me anything you want, but I prefer humans being the pinnacle of evolution. I like being able to point to our civilization, flawed as it is, and say "No other species we know of has achieved the same."

Hmorgan12 • today at 3:28 PM

The nature of mathematical proofs requires us to stay vigilant and careful. OpenAI pulled some of the research results from their famous repo due to simple mistakes that where easily spotted by human experts. There have been infamous mass of incorrect Fermi's theorem proofs.

LLM-assisted research is proving to be useful, rewarding and joyful, but we should not treat LLMs as a perfect machine, incapable of making mistakes. Especially in such massive proofs as this one.

I won't say that it "doesn't count". I will simply ask for staying rational.

➕ show 1 reply
xpct • today at 3:45 PM

What's so surprising about this? A lot of these problems can only be solved using a significant amount of capital. At home or work, using a few-hundred dollar token budget, it's mostly a slog and requires handholding for the majority of open-ended problems.

This technology won't be available to the broad public in the scale that it's available to a select group. If it can only be deployed to a limited amount, to most it remains a magical inaccessible box. And that majority should naturally be incentivized to look for cheaper alternatives, and to "keep moving the goalposts", or in simpler terms, looking where they can provide value, of which there is plenty!

➕ show 2 replies
thuo • today at 5:06 PM

I'm one of those people who's skeptical of recursive self-improvement and superintelligence. My rationale for my skepticism is that the algorithms get too much credit for the results. The algorithms ultimately run on hardware built by humans and learn from data generated by humans. In the same vein, I think knowledge work gets too much credit for building things and generating high quality data. The output of knowledge work is only useful if it can be applied to the physical world e.g. a CAD file is useful when the actual thing it describes is made in the real world. Similarly, high quality data comes from several feedback cycles with the physical world. I think people generally underestimate how much of human capability is driven by our sensorimotor skills. I think the idea of recursive self-improvement / superintelligence is just an extension of this attitude.

Humanity as collective does not buy anything, or send the design to the plant to be made. As a collective, humanity transforms the natural world into things, for example, people take fibres from plants, turn them into yarn, weave yarn into fabric, and then turn fabric into clothes. Similarly, humanity as collective does not "buy" chips, people turn quartz, bauxite, copper ore, petroleum, etc into the components of computer hardware.

There's a common adage about how everyone should work as a service person to learn patience and empathy. I think everyone should also work in an assembly line and a mine to learn about where the things they use come from.

snowwrestler • today at 5:35 PM

I don’t understand the technical basis for belief in continuous recursive self-improvement. Frontier models are increasingly impressive at complex math and coding, but the math and code underlying the transformer architecture are not complex. Major advancements in capability have come from scale: more data and more processing power. How does solving the Riemann Hypothesis (or similar achievements) further increase those? Aside from supplying fodder for investment-seeking hype.

➕ show 3 replies
AntiDyatlov • today at 4:03 PM

LLMs are jagged in their capabilities. Opus 4.8 was already an excellent programmer, but it played Pokémon Red far worse than a human child, even if it did, eventually, manage to beat it.

I note that the dazzling events that have happened have all been in coding, cybersec, and math, that is, domains where the correctness of the work can be verified entirely inside the computer. It is unclear whether further progress along these lines leads to crazy scientific breakthroughs too, which is something people are taking as a given when it comes to superintelligence.

cbg0 • today at 3:31 PM

I don't think you need to be in the football team mindset of us vs them, we're all seeing the progress but most of it is still built on top of what humans have put together over decades and not completely new inventions, which is probably what would convince most that we've achieved "superintelligence" or whatever you want to call it.

William_BB • today at 3:57 PM

The pace of progress is incredible, but the end result is hardly unexpected. We're talking about a theoretical domain that's easily verifiable and has no side effects.

You have the input (hypothesis), the output (true/false) and the lean kernel. There exists some path from the input to one of the outputs, you "just" need to find it. You use AI to search this exponential space, occasionally prunning ideas that are dead ends, and throw as many agents as possible to scale it up.

hkalbasi • today at 4:15 PM

In order to keep my mind safe from hype waves, I try to set my expectations periodically and see which of them happened in retrospect, updating my beliefs accordingly. At 2026-09-22 I made a personal 10 statements "prediction" to happen by 2027-03-22, which one of the statements was:

AI will solve exactly one or two millennium problems.

So I will become impressed by something, but only this is not enough. If AI solve three or more problems, I will become impressed and believe that some serious RSI stuff might be happening. If AI couldn't solve any problem (which it seems it did) I would become impressed in the other direction, and believe AI got stuck in a way that I didn't imagined (together with other 9 statements).

There may exist people (like me) who don't believe superintelligence is happening with current set of evidences, but believe it with some stronger evidences.

neuroticnews25 • today at 3:49 PM

Whether AI with it's math achievements is extremely cool, whether it's intelligent in a philosophical sense and whether it's capable of RSI are 3 distinct topics.

svara • today at 4:25 PM

These disagreements are natural and expected if most people are not properly defining their terms. What do you mean by artificial / super intelligence (explosion)?

I think it's a perfectly sane position to acknowledge amazing progress with very large neural networks, expect further progress, and also not buy into the specific mythology around the singularity or extropianism.

These things come bundled together quite often, and that's what I see people (reasonably, imo) resisting.

Doesn't mean there isn't unfair dismissal too, but I feel that's gotten much more rare.

root_axis • today at 3:52 PM

> self-improvement / intelligence explosion / superintelligence

These terms mean different things to different people. What are the particular implications of this that you're suggesting are imminent?

deprave • today at 3:29 PM

I think this is all real and counts. Where I struggle is explaining to non-technical people how their everyday lives are going to change directly or indirectly. I hope these discoveries are the first in a sequence that will culminate in or surpass a cure for cancer and diabetes, cheaper and cleaner energy and water, and other tangible everyday life improvements that influence millions of people.

➕ show 3 replies
codewithcheese • today at 3:24 PM

RL is not scalable to super intelligence. Super intelligence should be able to learn continuosly without verifiable rewards.

➕ show 1 reply
soon_to_be_sb • today at 3:31 PM

I am not impressed because I understand how it works. Not because what things it can or cannot do..

We can start talking when we can train these things without having to beat on its head a million times in a million different ways to teach it one little thing...

➕ show 1 reply
asdfologist • today at 3:29 PM

But it's not clear how well superhuman ability in math generalizes to other domains.

Vegenoid • today at 3:55 PM

This is such a substanceless argument, it’s pure vibes. It just says “Anyone who doesn’t believe in imminent intelligence explosion is an annoying curmudgeon.”

There are very solid arguments about why excellence at proving mathematical theorems does not necessarily imply general intelligence of the sort that people associate with ASI.

➕ show 1 reply
233mhz • today at 3:46 PM

If people kept applying the same progress expectations to cpu frequency since 20001 we'd have 400ghz cpus by now

It's OK to admit we just don't know when or where it'll srop

Also virtually no one argues llms will go away, but a lot of people argue the current companies might, it's completely different.

➕ show 1 reply
geraneum • today at 3:20 PM

> superintelligence within months or a year

Didn’t we already achieve super intelligence?

➕ show 2 replies
ashkankiani • today at 5:31 PM

Neither of us know what the growth curve looks like, so focusing on just calling one side naive is ignoring the naivety of the other side as well. If you want to extrapolate, go ahead, but that doesn't make you some oracle, nor does it make what you're suggesting factual. There are many potential scenarios and we should consider them, sure.

GrollingGrakbor • today at 5:22 PM

You're listing poorly defined words that mean different things to different people and wonder why people disagree with you...? My guess is that a lot of people you perceive this "denial" from are simply realists who are tired of seeing blind hype.

Note that every technology exhibits extreme growth until it... Doesn't. You can't predict the future. Whether any kind of RSI will happen depends on how hard of a problem intelligence is, and the precise point is that no person on earth knows, otherwise we wouldn't need to dedicate an entire research field to it.

Anyone who is sure that X will happen or not happen within Y years is just lying to themselves. For example, if you're so sure RSI will happen, why not invest all the money you have that isn't strictly necessary for your survival into this? Surely it will pay off in Y few years...

And note that I'm not saying it won't happen! I'm saying it's stupid to make a prediction of either case.

➕ show 1 reply
nonameiguess • today at 5:26 PM

I fear this is the result of a massive disconnect based on what you were told to expect. I read Kurzweil when that first came out, and maybe I'm remembering wrong, but I remember the singularity being the rate of technological progress in all domains will be so fast as to be asymptotic to an infinite slope and anything that will ever be possible will happen more or less instantly. Scott Alexander told me in a comments argument on his blog ten years ago that encryption is pointless because an AI will turn Jupiter into computronium and break it. Zvi was bandying about with the "take over its light cone" and "erase all value from the universe" phrases.

If that's the standard, then sorry, I'm not impressed by proving the Riemann hypothesis. If the standard is rapidly improving software for generating other software (which may or may not itself be any better than existing software) and math proofs (that may or may not have any meaningful real-world impact), then sure, I'm impressed. It doesn't mean I expect the daily lived experience of more than a few hundred thousand people globally to change much in the next decade, and most of the ones whose lives will change are software developers and mathematicians.

On that note, I have to imagine you have something different in mind when you talk about the "original meaning" of superintelligence, as you seem to not be talking about the Kurzweil/Bostrom/Yukowsky type of being effectively indistinguishable from God or at least the ancient alien civilizations from the Space Odyssey or Expanse novels, which again, I'm not seeing much of a hint toward because we have very good automated theorem provers. Proving Riemann doesn't mean we can now figure out zero point energy and turn planets into stars.

CarlitosHighway • today at 3:26 PM

Both the deniers and claqueurs are from the same ilk: Silly adults who don't have the intellectual capacity to create or even fully understand the subject, yet possess an almost vulgar self-confidence to constantly opine about its future.

➕ show 1 reply
rhipitr • today at 3:53 PM

people always under emphasize too that the gains of tomorrow aren’t made with the technology/tools of today (I think Kurzweil is where I heard this first). RSI, if landed, seems like that on steroids

igleria • today at 3:32 PM

> people in denial about superintelligence

As long as I get to keep my house...

nozzlegear • today at 5:44 PM

> Experience has shown that, even now, there will still be people explaining in patronizing tones why [...]

This is true no matter which side of the debate you're on. See [1] for an example.

[1] https://news.ycombinator.com/item?id=50575797

wslh • today at 3:22 PM

I would say, seriously, that this is bounded superintelligence. That superintelligence needed top human talent to achieve the goal. It was not a prompt like: "solve the Riemann hypothesis".

Coincidentally, yesterday I did a small presentation on cybersecurity and AI where I made a similar point: scarce top talent in cybersecurity (e.g. Calif) gets superpowers with AI, while AI replaces a lot of the skills of average cybersecurity professionals using simple prompts.

I think the breaking point is when AI can solve complex problems on its own that even top talent can't achieve. IMHO, I've seen some of this in the latest Google models (especially Carbon, and then Argon), where they can do more research independently.

➕ show 1 reply
npn • today at 3:54 PM

No offense, but have your heard about the infinite monkey theorem?

joe_the_user • today at 5:44 PM

I think you can be extraordinarily impressed and believe (or expect) that there significant limits to this technology.

I mean, AI could make very impressive 20 second videos two years ago. Progress has been slow and incremental since; Ten minute video still look far away.

I'd acknowledge is making quick progress on things you can find massive and reliable data on but even in it's good mode it seems to have palpable weaknesses. Someone was claiming Claude has become more and more persuasive over time but all the generators "still sound like slop" to me and many people.

bbor • today at 5:22 PM

Still, I have hope.

It’s not too late to Look Up for you. Yes, I’m talking to you — the person reading this with a vague grimace of disbelief and discomfort and disdain. It’s not too late for you to swallow some pride and reassess. Everyone has to eventually, especially me; why not now, for you?

So please. I’m begging you. LOOK. UP.

➕ show 1 reply
Tanjreeve • today at 3:30 PM

It’s a bait and switch to use a harvard fellow of mathematics solving a problem using AI with “The AI is super intelligent”. People are getting really tired of the constant bait and switches and subterfuge in press releases so unsurprisingly they just start rejecting it by default.

➕ show 2 replies
baggy_trough • today at 4:45 PM

Stochastic parrot truthers in shambles.

caidehen • today at 3:43 PM

I won't believe it until I get a robot to put my dishes into the dish washer, put copper meshes and squeeze spray foam under my siding panels.

dboreham • today at 3:23 PM

We're really using a term invented by Trump yesterday for something we've been working on for 80 years?

➕ show 2 replies
jubilanti • today at 3:30 PM

[flagged]

➕ show 2 replies