Scientific progress is heavily influenced by how many bodies you can throw at a problem.
The more experiments you can run, with more variety and angles the more data you can get, the higher the likelihood of a breakthrough.
Several huge scientist are famous not because they are geniuses, but because they are great fundraisers and can have 20/30/50 bodies to throw at problems every year.
This is true in virtually any experimental field.
If LLMs can be de facto another body then scientific progress is going to sky rocket.
Robots also tend to be more precise than humans and could possibly lead to better replication.
But given that LLMs cannot interact with the real world I don't see that happening anytime soon.
What can be said about scientists and bodies is interesting but ultimately irrelevant.
Edit: I'd add that various LLMs/neural-nets have turned out to be great tools for research. I simply find the scientist-equivalent position problematic.
https://arxiv.org/abs/2509.06503
They set up scoreable computational science problems and do search over solutions.
What type of interaction do you envision? Could a non-domain-expert, but somewhat trained person provide a bridge? If the LLM comes up with the big ideas and tells a human technical assistant to execute (put the vial here, run the 3D printer with this file, put the object there, drive in a screw), would that help? But dexterous robots are getting more and more advanced, see CoRL demos right now.
No, because the bottleneck isn't the thinking but running experiments.
I worked in solar research, assembling a cell to test implied 40 different steps and from beginning to testing it was around 4 to 5 days.
This means that in one year working full time I will realistically run 40ish different experiments. Many of those will need to be done multiple times, and when you have 40 different steps that can go wrong and kill your efficiency this further compounds.
Thus realistically are running 5 to 10 different experiments (or better, a handful plus their variations).
At no point in this process you're like "yeah, if only LLMs could provide ideas", it's just not true, you get millions of ideas, time and bodies are the limit.
Another delay point is getting collaborators schedules to align for meetings on progress or potential directions.
Placing the results in context takes some time but not so much as you might guess if you are constantly reading and writing sourced paragraphs and skeleton papers needing only results plopped in when they are ready and some exposition in the discussion section.
Writing the code might be the fastest step in the process already.
Pair LLMs with machines and robotics and you are getting closer
Yes they can...VLAs exist.
Jumper's work is the poster child of AI success in science; this isn't about a new domain being revolutionized by it.
I will throw out an idea I've been thinking about recently about a far less ambitious idea, but related: Amber (MD package) provides Force Field names and partial charges for a number of small organic molecules in their GeoStd set. I believe these come from its Antechamber program. Would it be possible to infer useful FF name and Partial charge for arbitrary organic molecules using AI instead, trained on the GeoStd set data?
No it wouldn't. I've seen anti-AI people try to make this sort of argument repeatedly and it doesn't make any sense.
It's an attempt to smuggle in an ad-hominem. It's relying on the fact that people who hate AI also hate people who work in AI.
Carter sold his peanut farm to avoid conflicts of interest, Trump launched a coin pump & dump on his first day.
If a Nobel Prize winner works for a corporation, that should be disclosed (the original title contained "Nobel Prize Laureate" instead of "DeepMind Director").
But I suppose that in the current age where everyone just wants to get rich these courtesies no longer matter.
https://research.nvidia.com/labs/dbr/blog/illustrated-evo2/
It's nice to see more and more labs using ai for drug discovery, something truly net positive for society.
The cynists will comment that I've just been sucked in by the PR. However, I know this team and have been using these techniques for other problems. I know they are so close to a computationally-assisted proof of counterexample that it is virtually inevitable at this point. If they don't do it, I'm pretty sure I could take a handful of people and a few years and do it myself. Mostly a lot of interval arithmetic with a final application of Schauder that remains; tedious and time-consuming, but not overly challenging compared to the parts already done.
https://x.com/DeryaTR_/status/1972115494787338484
>...noticed an email from one of my PhD students sent more than eight years ago, outlining a highly complex immune cell experiment that would run for several weeks and asking me to make corrections
>...Incredibly, GPT-5 Pro would have been as good as, if not better than, me at making these corrections, interpretations, analyses, and follow-up experiment suggestions! The experiment would also have yielded better results thanks to more precise planning...
Maybe the era of AI speeding things is upon us. Maybe not so long till AIs are helping make better AIs?
Maybe its utopia, maybe its akin to an Eliezer Yudkowsky prediction, who knows. Regardless of the specific outcome, its a huge gamble with effectively unlimited risk.
Next jump given by AI (not LLMs specifically, I mean “machine learned systems” in general) is navigation. Even with large amounts of speed some problems are still impractically large, we are using AI to better explore that space, by navigating it smarter, rather than just speeding through it combinatorially.
Still at the top of the benchmarks of integer optimization by huge margin are the traditional usual suspects. Same in constraint programming and SAT.
Not published just yet are experiments for finding solutions to mathematical problems traditionally found with SAT solvers, at much larger scale than was previously possible.
Uh, ok, I didn't claim that. At All.
Deep Learning machine learned features have definitely helped us (meaning my company) over hand engineered features, allowing us to navigate our problem space significantly faster
Three points jumped out
1) "really when you look at these machine learning breakthroughs they're probably fewer people than you imagine"
In a world of idiots, few people can do great things.
2) External benchmarks forced people upstream to improve
We need more of these.
3) "the third of these ingredients research was worth a hundredfold of the first of these ingredients data."
Available data is 0 for most things.
> Available data is 0 for most things.
I would argue that we need an effective alternative to benchmarks entirely given how hard they are to obtain in scientific disciplines. Classical statistics has gone very far by getting a lot out of limited datasets, and train-test splits are absolutely unnecessary there.
Even if this is a super high bar, I think more papers in ML for science should strive to be truly interdisciplinary and include an actual science advancement... Not just "we modify X and get some improvement on a benchmark dataset that may or may not be representative of the problems scientists could actually encounter." The ultimate goal of "ml for science" is science, not really to improve ML methods imo
- Build a complex intractable mathematical model (here, Navier-Stokes)
- Approximate it with a function approximator (here, a Physics Informed Neural Network)
- Use the some property of function approximator to search for more solutions to the original model (here, using Gauss-Newton)
In a sense, this is actually just the process of model-based science anyway: use a model for the physical world and exploit the mathematics of the model for real-world effects.
This is very very good work, but this heritage goes back to polynomial approximation even from Taylor series, and has been the foundation of engineering for literal centuries. Throughout history, the approximator keeps getting better and better and hungrier and hungrier for data (Taylor series, Chebyshev + other orthogonal bases for polynomials, neural networks, RNNs, LSTMs, PINNs, <the future>).
You didn't say anything to the contrary, and neither did the original video, but it's very different than what some other people are talking about in this thread ("run an LLM in a loop to do science the way a person does it"). Maybe I'm just ranting at the overloading of the term AI to mean "anything on a GPU".
I also wouldn't say this is entirely "classical". Old, yes, but still unfamiliar and controversial to a surprising number of people. But I get your point :-).
That's a strong claim. Is it based on more than the linked work on some model problems from fluid mechanics?
I will say that I dread the discourse if it works out, since I don't believe enough people will understand that using a PINN to get new solutions of differential equations has substantially no similarity to asking ChatGPT (or AlphaProof etc) for a proof of a conjecture. And there'll be a lot of people trying to hide the difference.
PINNs are different in concept, yes, but clearly no less important, so the additional attention will be appreciated. Asking LLMs for proofs is a different vein of research, often involving Lean. It is much further behind, but still making ground.
> In three space dimensions and time, given an initial velocity field, there exists a vector velocity and a scalar pressure field, which are both smooth and globally defined, that solve the Navier–Stokes equations.
When working toward a problem of this magnitude, it is natural to release papers stepwise to report progress toward the solution. Perelman did the same for the Poincare conjecture. Folks knew the problem was near a solution once the monotonicity proof of the W functional came out.
You can just ignore them. I see a lot of science-literate folks try to meet the anti-science folks as if they're on equal footing and it's almost always a waste of time. Imagine if every time you talked about biology you had to try to address the young earth creationists in the room and try to pre-rebut their concerns.
Before commenting I did check out your profile and saw that you weren't AI hater. My intention was the opposite -- I was trying to point out that some arguments don't survive on logic, but on emotional appeal. When enough smart people repeat something, it can slip past our skepticism. But of course my comment was terse and you had no way of knowing that.
In my view, among all people who make claims about what is on the forefront of science or knowledge, the most credible claims come from people whose research is on the forefront of science or knowledge. That's why the claims of experts are considered more credible in their fields than non-experts, even when the non-experts are experts in nearby fields.
So, applying that general rule to this particular case, we would expect the people to most credibly talk about the application of AI to science to be people who know the most about the application of AI to science.
There are other scientists whom I really respect but whose opinions on this particular topic I would find less weighty. For example, Terry Tao is a brilliant mathematician and has consulted on using AI to do research level math. But he was just a few months ago figuring out how to set up ChatGPT with VS Code, so I wouldn't expect him to be the most up to date on how AI is impacting science.
On the other hand, a Nobel laureate who has shown an aptitude for making important scientific discoveries is exactly the sort of person I believe would be able to talk knowledgeably about how to carve up science problems and about how to apply the current generation of AI to solve them. Especially if they've seen the internal world of a company that has a top tier foundational model and a track record for making scientific discoveries. Because in that case they see how science is being done if you have unlimited funds and access to some of the smartest scientists on the planet.
In contrast, I would be much more skeptical of a similar claim made by a company like OpenAI or Microsoft that doesn't have the same track record of producing new science.
For those reasons I don't think it's true that the claim would be more credible from someone with more distance (and hence less expertise). And I think similar claims made in other contexts would strike most people as bizarre. For example, if I said that claims about medicine are more credible if they come from people who aren't medical doctors.
His title is in the video thumbnail.
> But I suppose that in the current age where everyone just wants to get rich these courtesies no longer matter.
Surely you agree this is a veiled ad hominem
Bring any tool you wish, but the problem is very well defined and very real.
If anything I think they're more important! Whether or not it works out for Navier-Stokes, this kind of thing is an extremely plausible avenue of approach and could yield interesting singularities for other major equations. I am however extremely concerned about public understanding. I know you are well aware that this is worlds away from the speculative technologies like 'mathematical superintelligence' but, if it works out, it'll be like a nuclear bomb of misinformation about AI and math.
I have no idea what solar research experimentation looks like in detail, is it theoretically possible to build similar setups for that use case? Where exactly is the bottleneck?
So it depends on if the same machinery can be used for more general material science research and testing work.
Not sure if I should finish with /s or /fear or /uncertainty
Can you say more about this? Nothing about this approach seems very amazing to me. Construct an approximate solution by some numerical method (in this case neural networks), prove that a solution which is close enough to satisfying the equation can be perturbed to an exact solution. Does the second half use some nonstandard method?
1. Proving existence to a differential equation using a numerical approximation is quite a bit more difficult than it seems at first. For Euler or NS, it seems almost absurd. Not only do you need a rather substantial amount of control over the linearisation, you need a way to rigorously control a posteriori error. This is easy for polynomials, but doing it for other models requires serious techniques that have only been created recently (rigorous quadrature via interval arithmetic).
2. Further to that, neural networks are far from an obvious choice. They are not exactly integrable and it is certainly not clear a priori that their biases would help to search for a blowup solution. In fact, I would have initially said it was a poor choice for the task. You also need to get it to a certain degree of precision and that's not easy.
3. The parameterisation of the self-similar solution turns out to be completely appropriate for a neural network. To be fair, solutions of this type have been considered before, so I'm willing to chalk this down to luck.
It's difficult to explain how challenging it is to fully derive a plan for a computationally-assisted proof of this magnitude unless you've tried it yourself on a new problem. At the end it seems completely obvious, but only after exhausting countless dead ends first.
The current LLM approach makes huge assumptions, including that training only on text prediction is enough to simulate true intelligence. That may or may not be a valid assumption, but it could be enough for the LLM to make one seemingly small change that ends up running away from us faster than we would realize.
More importantly, I think we agree a sudden run away is possible, we don't have safeguards in place, and the risks are extremely high should that happen.
You could call it artificial ant colony optimization.
People come up with such ideas all the time. Sorry, but nothing groundbreaking here.
Perhaps some much needed context. Mathematicians are not stupid; we are very much aware of all the existing forms of genetic optimisation algorithms, cross entropy method, etc. Nothing works on these problems, or at least not well at scale. As I said, state of the art for many of these was SAT-related. The problem is that the heuristic used for exploring new solutions always required very careful consideration as the naive ones rarely worked well.
Here, the transformer is proving effective at searching for good heuristics, far more so than any other existing technique. In this sense, it is achieving far, far. far better performance in optimisation than previous approaches. That is a breakthrough, at least for us mathematicians. If this doesn't constitute improvements in optimisation, I don't know what does.
Saying it's "just meta heuristic relying on local search" is akin to saying these tasks are "just optimisation". If it's so procedural, why weren't we making ground on these things before?
Also, by the way, a :facepalm: is not exactly the pinnacle of academic rebuttal, no matter how wrong I could have been.
It’s just that the paper cited is no different than any other paper in the meta-heuristic community.
Some idea for guiding the local search. Some limited sample results. No promises on bounds or generalizability of the method.
If this is ground breaking, then every legitimate meta heuristic paper in the past 50 years was also ground breaking.
I will change my mind if I see a wide set of benchmark results where it consistently beats or is even head-to-head with the SoTA. Then we would know that we have a game changer.
This isn't true, it was a major accomplishment but by far the easiest part of Perelman's papers and not actually even part of the proof of the Poincaré conjecture.
Happy to be corrected on this; I wasn't an active mathematician at the time, so everything I know comes from other accounts.
Moreover I'd strongly dispute that there was any particular point where it was clear that the problem was "near a solution." The W functional is an example where experts could very quickly verify that Perelman had made at least one major new discovery about the Ricci flow. But the proof of the Poincaré conjecture was on another order of complexity and required different (more complicated) new results about Ricci flow.
The key difference is that this one approach (and the improvements that follow up on it) seems to be generating counterexamples and sharper estimates at such a rapid speed that they simply can't publish them fast enough. It almost seems pointless to do so given how easy they now are to find. Mathematicians are not stupid, we know about meta-heuristics. It takes forever to get a single result with existing techniques, so long that each result is often its own paper. Getting everything in the PatternBoost paper took a few months, as I understand. Local search is problem-specific, but the point is that the NN is doing a lot of heavy-lifting so the choice of local search is not as critical.
Here is a frivolous example mentioned in the paper itself: largest subset of a 2D lattice with no embedded isosceles triangle. For larger lattices (I don't remember the exact size under consideration), SoTA are SAT solvers which work up to n=32. Previous meta-heuristics cannot even do this, of course, you can try if you don't believe me. PatternBoost has just recently gotten to ~96% of the expected size for n=64 and is still improving by the day. Once it reaches 100%, there are techniques to show local optimality. Does this work as a benchmark? There are plenty more in there and unpublished.
For more serious cases that are difficult to explain here, the group in Sydney have counterexamples to hundreds of graph-theory and rep. theory conjectures that have stood for many decades. I also disagree on the "no different" aspect; Geordie Williamson is a very strong mathematician, and does not tend to jump on trivial things. He is very receptive to discussion on these matters, so you can ask him yourself how this is actually a game-changer downstream.
Yes, it is a meta-heuristic. But almost all meta-heuristics have been useless so far for these problems. This one is not, and for the people downstream, that's really all that matters.