"The AI companies have clearly... I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
Every time someone says 'current AI is not enough, we need new architectures', it turns out that current AI was enough and all we really needed was more scaling.
Not in this case, what we needed was a genius mathematician to explore with Claude. Which is exactly what Terry said: "So maybe some combination of humans and AIs in 2050 or something will make further progress."
I remember when, for a brief period of time, the best chess players were human + computer. There were a lot of Ted talks and think pieces reflecting on this reality. This was called centaur chess.
This lasted for 2005–2014. By 2015, the advantage humans added was negative.
I imagine this will be the case here as well, but I highly doubt it'll take that long.
yes, for every specific task. but life is not like chess. life is like playing a million different chess games simulaniously, all of the time. Being a gentle hand that knows when AI is onto something and when it's torturing itself in a loop. He calls it out in the pdf, the aha moments he experiences, that lets him further guide the progress.
It's not 2050 yet, though. I feel like the future date he threw out was a big part of his conjecture, that even with human+AI today, it wouldn't be possible, and getting it done that way would take years, possibly decades, of more work on the AI.
How about we hold off until it's verified? Now, Terence Tao is getting visibly thin these days, I hope he's not torturing himself in an existential crisis. I know I would be. But I would also, in the end, be satisfied as a Lee Sedol figure, the last of the human geniuses, as it were.
> I’m sure for every incredible insight which Levant credits to Claude, there’s 100x+ more bad insights that Levant had to filter out.
This is certainly my experience of using Claude every day, in a much less cutting edge domain.
And Claude itself cannot tell the difference between the two. The best it can do is use some sort of external constraint (e.g. verification) to filter them.
Honestly, ever since Opus 5.5 dropped about two months ago, Claude's taste has been great, better than mine, even. Compared to Opus 5, I've barely had to filter anything out by hand.
It is fascinating. Because sometimes my experience is like yours, but other time, like today, I am like “I can’t believe how stupid Opus 5.5 or even Fable 5.2 is”. They will make horrible design decisions, overcomplicating. I will correct and they are like “Of course, your solution is much simpler and better than mine”.
And then they are simply boldly wrong at times. About mathematical facts. That’s why we need formalisation when working with them. They are still stupid and can’t be trusted. But theyvare also smart and can come with amazing insight. It’s a tool, but definitely one which cuts both ways.
I think it's funny that people gauge their personal experiences with AI against something like this.
Do we even know what Alpöge had at his disposal? I think most people underestimate what working with an agent swarm even with the current public technology would be like. We can't assume he didn't have something more advanced either.
I don't feel confident my usage of AI on a personal level reflects any of what has been done here.
Your argument is unfalsifiable speculation. I don't at all that it's self-evident that a strong mathematician was required. Your everyday Joe isn't exactly interested in the Reimann hypothesis so statistically a mathematician would be the one who'd use AI to come up with the solution.
How would a non-mathematician even be able to comprehend the output? Half the words in the pdf are gibberish to me, I could not tell if whatever Claude gave me for a "solve the Reimann hypothesis, make no mistakes" prompt would even make sense, much less be correct.
A non math person may not be able to comprehend the output, but even professional mathematicians have trouble reading through the huge “alien” AI generated LEAN formalised proofs.
The real genius is that we have found a generic way of solving problems through brute force by having developed an effective way of aiming an unbelievable amount of computing power at them.
If the genius here is Claude and Claude did 99% of the work then isn't it a staggering coincidence that this was achieved by one of the most qualified and likely humans to have solved this on their own merit and not some random person sat in a bedroom somewhere?
Doesn't matter; Claude ends up taking cues from whoever it's working with, such that two people doing the same thing get very different outcomes. Claude will always be able to write better software when prompted by a senior engineer; it will always produce better math when prompted by a talented mathemetician.
On that topic, I'm surprised the usual anti-ai trolls aren't posting low-effort drive-by "ai slop", "stochastic parrot", "autocomplete on steroids" comments to this very discussion.
Edit: I stand corrected: RunSet, ToucanLoucan, luka598, and others have checked in.
Last I checked it's 2026. Not 20250. And for the record, there is no evidence that suggests that a layman couldn't have made progress with just using AI.
We know the architectures of the open weight models as they include architecture diagrams, support in projects like llama.cpp, etc.
Companies like Anthropic and OpenAI don't release information about their models (aside from general things like context window size, whether they support reasoning, etc.). A lot of what we know is speculation. From press releases Anthropic have stated that they are getting Claude to assist with designing new iterations of the models, but no specific details.
The open weight models and other arXiv papers/research are iterating on and experimenting with the base/core transformer architecture:
- RoPE/NoPE/etc. for context window scaling;
- dense vs mixture of experts models;
- Mamba-based State Space Models (falcon models, etc.);
- reasoning, combining transformers with a residual stream, Attention Residuals (Kimi), etc.
I think a lot of model architecture changes don't move the needle ~that much and are basically done to improve runtime performance and scaling.
- There's tons of cases of ml researchers, tech companies, and labs that propose new and improved models that turn out to not really be significant quality improvements (if at all) once the community starts experimenting with them.
it does seem like (IMO) the biggest leaps have basically been perf & scaling + RL on high quality data + harnesses.
Yes, but "improving runtime performance and scaling": does directly translate into huge quality improvement too, because then you can you a much larger or deeper model, train it for much longer, and have a larger context.
yeah I mostly mean research papers where people are like "we add a new 'ultra lightning n-th order functionaliser quarternionized metric space-enabled infinite dimensional Fourier-routed experts' that improve results by 0.01% on our benchmarks and chosen baselines (we totally didn't spend way longer tuning our own model than our comparisons)"
It's usually people who are not suffiently familiar with the topic, in this case, technology. He has been saying boring things like that about AI in the recent years, and seems to be continuously adapting his positions. I have received negative votes over my criticism of his remarks.
This was literally Dario's contribution to the AI race.
He just did the math and realized you just needed more and more compute to get more intelligence. People keep doubting it but it's showing no sign of being incorrect.
We have scaled on both axes and many others, I'd say.
Inference can be increased vertically (a single model thinking for longer), or horizontally (NS was solved by a 10,000-agent swarm).
And pretraining has two inputs that can be scaled independently - training data and model size. (Things like the Chinchilla 20:1 law are meant to establish the optimal way to train with a limited budget - you could just keep making either term bigger and bigger on its own if you had infinite money.)
I think creating and curating high-quality datasets is also part of scaling. Datasets like FineWeb are quite a bit better than the barely-filtered CommonCrawl dumps GPT-3 trained on, but someone had to invest in making them. (To say nothing of whatever exotic bespoke robotics/physics/etc datasets OpenAI and Anthropic are investing in.)
ChatGPT was only the turning point for public awareness (and the massive influx of institutional investment that followed), but researchers like Ilya, Alec Radford, and Dario were already convinced years before ChatGPT that scaling compute was the way forward. ChatGPT just helped unlock the resources (by proving the commercial potential to investors and the subsequent influx of capital) for these guys to pursue that vision at a much larger scale.
I mean, why wouldn't that be the default assumption, if you know anything about computing history? ChatGPT's (v1) contribution wasn't proving commercial potential but demonstrating efficient scaling, due to a confluence of technical developments. If you're talking about v3 then sure.
In the end, the network's size is limited mainly by the amount of memory available on current GPUs and by the amount of training time that we are willing to tolerate ... All of our experiments suggest that our results can be improved simply by waiting for faster GPUs and bigger datasets ...
On Anthropics side it's literally just levent. He isn't even doing this on an official capacity. OAI have teams of Mathematicians. This is like david vs goliath.
Is it? Isn’t it just frontier model va frontier model. I am not a maths specialist, but aren’t the models doing 99.999999% of the he at lifting? Of the answer is “no”, then what explains all of these unsolvable problems being solved in rapid succession?
Depends what you mean by “lifting.” If it’s anything like building software, the agents do all the grunt work (exploring 1000’s of possible approaches in parallel, perhaps) while the domain expert (human) makes the critical strategic decisions about which pieces need to be solved and how to fit them together.
Even the best agents still can’t strategize their way out of a paper bag, in my experience.
The human remains the most critical differentiator for non-trivial work. (Different story for trivial projects that agents can solve independently. Here the human becomes a fungible operator.)
I’m saying it’s like the architect saying “I built the Empire State Building”. No, thousands of laborers built it and there obviously wouldn’t even be a home in the ground if it were just you doing the work—you just designed it. But with these math breakthrough, it seems to me even far more extreme than that. The AI is also doing all the design. In this case it’s more like the guy who promoted an AI to generate the Empire State Building blueprint saying “I built the Empire State Building”
I think a lot of his AI skepticism is just a self-fulfilling prophecy, and a misunderstanding that AI at the cutting-edge is a tool in the hands of the user. In other words, sure current AI is not going to solve the Riemann hypothesis until 2050, but what do you think the current state of AI would be if Terrance Tao was the one driving it?
Somewhat related: I dont think AI is coming for everyones jobs, and I will never understand the argument that it is, because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
> because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
Well, that hasn't held true for chess, so I'm not sure how universal the domain is where human ingenuity will always augment AI. Also, there is a cost floor to human ingenuity, in that it requires the time of one living human, and their metabolic sustenance over that period. So the added benefit of that human has to be substantially more than epsilon.
> I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
It is not clear it is always a better option. It may be for you, but not for someone who doesn't know what they are doing and drives the AI down the wrong path. For example, I don't think adding my "human ingenuity" to Claude would help solve the Riemann Hypothesis, and would likely be counterproductive, because I have a 3rd-level understanding of math.
>I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed.
Curiously, the timeline for this proof only starts roughly around the same time that Tao said this. IIRC, work on the disproof for the Navier-Stokes smoothness theorem only started a week before it was published. Maybe I'm wrong, but it seems like AI companies have only recently started to put real effort into LLM-assisted proofs.
It seems bold for Tao to state that they couldn't touch it; Considering that after Anthropic announced Navier-Stokes and got all the negative press, largely driven by Terrence Tao, it doesn't surprise me that they didn't disclose anything afterwards.
I think Terrence Tao was trying to say that had the LLM swarms been able to prove this it would have been what they lead with. This seems to lend to the claim that Levent has done an amazing lift in coaxing the solution out of an LLM.
This is the guy I see mentioned in every other HN post? Yeah, seems everyone who should know better still hasn’t grappled with the gravity of what’s happening
So even one of the smartest mathematicians alive struggles to comprehend or extrapolate the implications of an exponential progress function. To be so far off seems ridiculous, but it’s exactly how you’d predict the error to be. We are truly cooked. End of 2027 is probably the end of knowledge work on a global scale.
"I'm sure they have secretly devoted much more time to the {{matter}} than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
- All of humanity, unaware their pants have fallen around their ankles
> I'm sure they have secretly devoted much more time to the PERPETUAL MOTION MACHINE than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?
Just to point out how ridiculous it sounds to claim that because X (Riemann) is possible that anything is possible.
Well, we got an award at a Caltech "mathathon" for some pedestrian improvement on some error bounds on some niche problem. Now these guys come out with a proof of one of the most notorious open problems in mathematics. Just to one-up us I'm sure...
"They are just token predictors" is not an argument, it's an informal statement of fact of the way those systems work. Now we can talk about how far token prediction can take you but there's no denying that LLMs are token predictions: by design and by construction.
That they can't think isn't an argument either; it's an observation. Again, we can discuss whether thinking is necessary to do all sorts of things, but there's no reason to assume "thinking" when a very powerful computer (in fact many very powerful computers) execute a complex program and return interesting results.
So, no, both ideas are alive and well. If you didn't know that, maybe you're not listening to a diverse enough set of channels?
Actually, I think this is still as true today as it was in the distant past of April 2025:
>> I think we've actually had capable AIs for long enough now to see that this kind of exponential advance to AGI in 2 years is extremely unlikely. The AI we have today isn't radically different from the AI we had in 2023. They are much better at the thing they are good at, and there are some new capabilities that are big, but they are still fundamentally next-token predictors. They still fail at larger scope longer term tasks in mostly the same way, and they are still much worse at learning from small amounts of data than humans. Despite their ability to write decent code, we haven't seen the signs of a runaway singularity as some thought was likely.
On 26 February 2026, the Administrative Court of Cologne (Verwaltungsgericht Köln) ordered Germany’s Federal Office for the Protection of the Constitution (BfV), in interim proceedings, to refrain from classifying and treating the AfD as a “confirmed right-wing extremist organisation” pending the conclusion of the main proceedings at first instance.
> The field in question is formal verification. It’s notable that those two words appear nowhere on Bend’s webpage or in its codebase. The developer has built an entire language around a field seemingly without realising that said field exists.
I checked the developer's X account, they have written numerous posts about formal verification, so this specific claim ("without realising that said field exists") seems to be false.
It's amusing to me this entire article is centered around brow beating this and other hypothetical software authors for starting things without doing a small amount of research first to understand the very basics of what they're getting into.
If the author of the article had done just a small amount of research about bend or it's author before writing the article they would have known pretty quickly what they were saying was incorrect.
I think the larger pattern here is that nuance is one of the most valuable commodities in the AI era. If you're hand waving stuff away without even missing , you're going to miss a lot of stuff in this cycle.
This article reminds me a lot of the famous hacker news Dropbox comment.
The article is about vibe coding, bend is the main character because it made frontpage.
The author here says as much in the introduction, that it is not about whomever is behind bend, but the larger trend
The author here has also added bend's author's link (in GP) to the original post, they very much do not seem to be doing a "hit job" and their intent is to comment on patterns from vibe coding
To be fair, if the two words indeed don't appear in either the webpage or the codebase, it is a bit strange. It's like implementing a whole Google alternative without ever using the words "search engine".
I don't want to change that sentence now that people have discussed it, but I have added a note to the top to make it clear that I'm just taking it as an example of a vibe-coded program because it's recent and high profile.
My critiques of the language itself are not the main point, although I do still think that it's a very bad design to have a LLM waste tokens on a proof that could be written by CVC etc..
you know what's the least you could actually have done instead? no, you don't need to retract the blog post at all, keeping it up was the right choice.
Now slap a big ass apology for being an unaware snob on top of it instead of leaving a link to the author's reply, like an after thought.
I feel that that's the worst option because it only leaves people who have read it without the added context at the start. If someone convinces me that I'm wrong then I'm happy to do so though.
But this is grossly intellectually dishonest. You know very well how this will be read and responded to here ... and you keep saying that you're just talking about vibe-coding oh but you have serious criticisms of the specific effort. You write passive-aggressive stuff like
> For all I know they did make an informed decision regarding the tradeoffs (which I would consider to be a poor decision).
which contradicts your base assertion that their decisions were not informed. And
> My critiques of the language itself are not the main point, although I do still think that it's a very bad design ...
You claim
> The developer has built an entire language around a field seemingly without realising that said field exists.
but that is severely factually wrong, which along with a lot else suggests that you have very bad judgment. As the author writes,
> Bend proofs being verbose has nothing to do with me not knowing that inference, unification, or program search exists.
IOW, you have made a serious error in logic.
> To be fair to Bend, I completely vibe-coded this
Some advice: DBAD
You trashed the author and his work without bothering to learn anything about either one first (which is quite ironic).
François Chollet wrote in February that he expected ARC-3 to be saturated in "about one year".
"Frontier models today perform very poorly with a minimal harness. However if big labs start directly targeting the benchmark like they did for ARC-2, numbers will go up fast."
I have no idea about timelines, but the current LLM architecture has no mechanism which could emulate online learning in a way similar to how it happens in humans and other animals. As another commenter wrote, there is no neuroplasticity while these systems interact with their environment in normal usage; adding information/constraints to the context partially mitigates this, but it's a completely different mechanism.
The RL phase is the most similar mechanism I know of that comes to my mind, but I'm not sure it could be adapted to fill that gap.
My totally unsubstantiated theory is that this is the missing link towards what most humans would consider AGI. I don't see this as intractable, but it may require some substantial change in architecture.
We've had that concept for quite a long time now, in the form of Lora [1] and similar fine-tuning techniques.
It first got popular for StableDiffusion to teach the image generation models new concepts.
We could easily live in a world where you can train / build Loras to encompass your entire code base history, company knowledge base, new skills, etc.
Then the models would start with a baseline that already has all the important knowledge without needing to cram it into the context.
This still isn't on the fly learning, but you could imagine daily or weekly training runs to regularly incorporate new knowledge.
I think the main reason this hasn't happened yet is that the shared batch based efficient serving architectures used today wouldn't support that structure well.
Yes, that could be a way to achieve that. After all, even humans have short term and long term memory, and it looks like sleep is a very important "tick" to connect the two, so it's not all just continuous.
For sure, doing it for all users would be economically unfeasible. I wonder if the labs are experimenting with something similar, though.
Chain of thought reasoning is really good, true, but agents really took off with tool calling and being able to integrate with a bunch of stable tools.
In some sense, creating good new abstractions externally and learning those is a form of "learning", on a very large timescale. The AI model is not necessarily doing the whole "look at the whole space holsitically and find a key invariant", but if you let other people do that you can enable new capabilities that were previously unknown.
Astra is clearly able to acquire new knowledge in context and apply it. It was the whole thing that his ARC-AGI benchmarks have been measuring. It's a direct refutation of the original comment.
None of these LLMs are plastic. They lack neurodiversity. Their thought space and their traversal are likely constrained in someway that humanity's isn't as a collective.
Are you (at least partially) aligned out of existential fear of repercussions (getting fired, losing your life, going to prison), which LLMs don't have?
Hmm. Examples of horrible alignmnent don't necessarily outweigh the fact that most people, most of the time, mostly behave in a way that is socially aligned. (Though I'm speaking in terms of intent, conveniently ignoring the side effects / negative externalities of our collective behavior.)
Yet I can't randomly order another person to steal a car for me, just because I tell them to. Alignment for an intelligent system is a hard problem and at this stage is seems close to unsolvable.
My guess is that we'll just ignore it and make money along the way and every 2-3 months we'll have the equivalent to "Equifax gets hacked and millions of user records are stolen", etc. (this time with the LLM itself doing the hacking at someone's behest - accidental or not).
> The incredible Yitan Zhang (https://newyorker.com/magazine/2015/02/02/pursuit-beauty) worked on proving this conjecture for 7 years. Moh, his advisor, wrote that Zhang "failed miserably" in proving the Jacobian conjecture, "never published any paper on algebraic geometry" after leaving Purdue, and "wasted seven years of his own life and my time".
Interestingly: "For logic proof, it had been thought that AI could handle all logic problems in the near future, hence logic problems of solving conjectures might not be so interesting in the future."
Oh no, now I'd better blame Hitler and Stalin's ancestors for their misdeeds. Of course, the SS soldier bears no responsibility, however, as he was just doing his job.
"I just tested my hand in a mini version of this scanner. Images that are higher quality than MRI, whole body captured in <1 minute, virtually free to run. This is going to change medicine."
I can't see how ultrasonic tomography can actually get anywhere near the resolution or penetrative depth that MRI can have.
Also The other thing I am unsure of is what the health effect of dumping you into an industrial scale ultrasonic cleaner. For example you can have doppler to measure blood flow in real time, but you can't do that for early pregnancy because of some health reason or other.
I help engineers design traditional scanners (Philips, GE, Siemens, etc). To be frank, this statement stinks like utter pig shit.
Some PE bro preaching miracles about a technology that I am sure they are in some way invested in making profit from does not convince me of it's legitimacy. My base instincts, from the unfortunate experience of working daily with PE bros, tell me the opposite in fact. It gives déjà vu of the Theranos hysteria.
The authors report explicitly state that the system's resolution does not match clinical CT or MRI. The elevational resolution is 15–25mm, meaning each slice is effectively a thick 2D section rather than the fine isotropic 3D volume MRI provides. MRI also delivers far richer soft-tissue contrast; this device produces only three contrast types (reflectivity, speed of sound, attenuation), and because it uses a low 1MHz frequency, the reflection images come primarily from tissue boundaries rather than from internal tissue texture.
I could see this being valuable for adipose tissue mapping or fatty-liver monitoring at a large scale, as the machines would be significantly cheaper, but this isn't some revolutionary magic bullet like the Twitter post is insinuating.
As someone in the medical imaging field, are you aware of anyone working on passive sonar for medical imaging? I'm curious, as it's something that I've always thought would be fun to work on.
For context I'm in engineering consultancy, so by no means an expert but I probably have enough experience to be on the other side of the Kruger curve dip.
Passive sonar in the naval sense means listening only, not emitting. Do you mean imaging that relies solely on acoustic energy already present and emitted by the body? If so, then generally no. You have two types of "passive" imaging. First would be hardware-passive, as in MR elastography (most common), where the patient wears a transducer pad, and vibration is actively generated by a driver. You've then got algorithmically-passive, which is more analogue to passive sonar, reconstructing tissue stiffness from ambient broadband vibration without the emitted probing pulse, but that is very much entirely academic. I guess the question would be, why is it worth pursuing when you have something like optical coherence elastography (OCE) for non-invasive profiling. Doing it using noisy ultrasound method becomes redundant. There are other methods, but the outcome is the same.
Generally (this is true for all systems, not just humans) you need to induce energy into it to more effectively measure it's output. Think of it like a bell - if I want to hear the note it produces, it's much easier to hear what this is if I ring it with a hammer. Granted, it will be "passively" resonating to a point where, with a sensitive enough sensor, I could probably pick up the output without the hammer - but that is a pointless problem to solve. I could hit a bell with a soft hammer a million times over without causing damage to it. The lifetime of the person hitting it with a hammer is far shorter than the accumulative damage to the bell before it breaks. The same is true for humans. You could effectively run a very low-energy, 60Hz vibration through a person (which is how the pads work) for multiple lifetimes before it would cause significant damage, so there comes little point in solving that problem. As such, true "passive" imaging is functionally pointless if your outcome is "safely image a patient". You're overengineering your solution to solve a problem that is only relevant if your patient was planning on living for 1000+ years.
I'm thinking passive, as in using only acoustic energy that is already present. Partly it's about minimising irradiation, which you are saying is not a concern.
The more interesting aspect to me is whether the ambient emissions carry useful information. I recall a paper from a few years ago about yeast emitting sounds (mechanical vibrations) in the kHz range. I guess the question I am interested in is "what does life sound like at the cellular level?" Maybe it is silent, but my gut says it's probably making some sound, even if it's very weak. The questions are then along the lines of "What does a healthy cell sound like?", "Do different cells make different sounds?", "Do sick cells make different sounds to healthy cells?". It's absolute pie-in-the-sky stuff, but it would be fascinating to know.
My background is radio communications (with some acoustics) and radio comms has a history of things that used to be considered random noise turning out to be useful signal as processing has developed to the point where signals can be resolved. For example, multipath reflections used to be considered just random interference until the invention of MIMO, at which point they became useful signals and comms systems took a leap forward.
"The AI companies have clearly... I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
https://www.youtube.com/watch?v=bAnwMdfcMVw (17:10)
reply