@Nil053 Getting below 2.25 might be too hard to crack, if it's even possible. But there's another race worth running: how small can the matrices be?
OpenAI proved that algorithms near 2.25 exist without saying what they are. Our two AI-written papers construct them explicitly, with actual sizes. At exponent 2.67, the first needed matrices with about 10^(2.3×10^11) rows. The second cuts that to about 10^4,650.
That's still galactic. A leaderboard for the smallest matrix size at a given exponent could let the community push this toward sizes that matter in practice.
Paper 1: https://t.co/L0X0ZsvppM
Paper 2: https://t.co/2pdDntGiCw
Recently, OpenAI showed that matrix multiplication algorithms with exponent near 2.25 exist, but not what they actually are. We then shared an AI-written paper that builds them explicitly. That first version was really bad: even at exponent 2.67, it needed matrices with about 10^(2.3×10^11) rows.
So we pushed the AI agents further, and they improved it a lot. At the same exponent, the size dropped to about 10^4,650 rows. That's a huge improvement, but the observable universe has only about 10^80 atoms, so even at one number per atom it couldn't hold a single row of such a matrix.
For anyone wondering what this new exponent of 2.25 means: so far, it's a theoretical result only. The algorithms constructed so far are "galactic," and making such algorithms practical would require another major breakthrough. Still, there's real new interesting math here: a step-by-step version of OpenAI's construction that shrinks the sizes dramatically, and proofs of how far these specific constructions can and can't go.
New paper produced by AI: https://t.co/2pdDntGiCw
@YFilmus@YFilmus Thank you very much for mentioning our results; I really appreciate it!!!
I just posted v3. It's still AI-generated and needs proper human review, but it should be much more readable. Hope it will be more useful: https://t.co/zcz6DGGWnD
Yes, the proof shows that such algorithms exist, but it doesn't say what they are. So we asked AI agents to construct them. And they made a paper that gives an explicit procedure that, for any ε > 0, outputs an actual O(n^(2.25+ε)) algorithm. And yes, it won't help GPUs: the constants are enormous, so the best known concrete algorithm remains unchanged, indeed. And also, the results still require proper review.
If you are curious what this algorithm can be, check it out: https://t.co/fKnBu8H4yZ
Great blog post! One interesting detail is that in OpenAI's 9/4 proof, it shows an O(n^(9/4+ε)) algorithm exists, but gives no way to construct one.
We had AI agents fill that gap with an explicit construction for every ε > 0.
As your last section would predict, it's purely asymptotic: the constants are enormous, so it will not accelerate real computations any time soon. And the results still need proper review. I think you'll find it interesting; check out the actual algorithm here:
https://t.co/fKnBu8H4yZ
The proof shows that such algorithms exist, but it doesn't say what they are.
We asked AI agents to construct them. And they made a paper that gives an explicit procedure that, for any ε > 0, outputs an actual O(n^(2.25+ε)) algorithm.
As you say, it won't help GPUs: the constants are enormous. And it still requires proper review.
Check it out:
https://t.co/fKnBu8H4yZ
Because the proof goes by contradiction, it shows that a fast algorithm exists without saying what it is. We asked AI agents to fill that gap. They produced a paper with an explicit construction: for any ε > 0, a concrete O(n^(9/4+ε)) algorithm over fields of characteristic zero. It replaces the contradiction step by composing the convolution relations directly. The constants are enormous, so it doesn't seem to be practical and still requires proper review.
Check it out: https://t.co/fKnBu8H4yZ
We asked AI agents to fill exactly this gap. They produced a paper with an explicit construction: for any ε > 0, it outputs a concrete algorithm that multiplies n×n matrices in O(n^(2.25+ε)) operations.
Caveats: the constants are enormous, so this won't speed up AI hardware anytime soon. And neither paper has been independently verified yet.
Paper, source, and checks: https://t.co/fKnBu8H4yZ
You don't need to work in a big lab to push this forward. You can do it today with just a subscription.
A remarkable result. One gap, though: the proof shows that such fast algorithms exist, but it doesn't say what they are.
We asked AI agents to work out the explicit recipe. They produced a paper that constructs algorithms arbitrarily close to 2.25. The constants are enormous, so whether this ever leads to practical speedups is still open.
Paper, source, and checks: https://t.co/fKnBu8H4yZ
You don't need to work in a big lab to push this research forward. You can do incredible things today with just a subscription.
OpenAI just released a stunning set of math results. One claims the matrix multiplication exponent is at most 2.25, down from about 2.37. That would be a huge drop in the work needed to multiply large matrices.
Their proof shows that such fast algorithms exist, but it doesn't say what they are.
We asked AI agents to work out the explicit recipe. They produced a paper that constructs algorithms arbitrarily close to 2.25. The constants are enormous, so whether this ever leads to practical speedups is still open.
Neither paper has been independently verified yet, so treat both with appropriate caution.
Paper, source, and checks: https://t.co/fKnBu8H4yZ
The wild part: you don't need to work in a big lab to push this research forward. You can do incredible things today with just a subscription.
This goes beyond pure math. @bernalde and I ran similar experiments in applied fields where math is central and computation is modest. Anyone with public AI agents can try it. Imagine what happens when this is applied to a problem you care about.
Check out our paper on how we did it and the questions it raises for research.
Blog: https://t.co/ZV0a7MPP5v
arXiv paper: https://t.co/cKiET2LWvL
All research output on GitHub: https://t.co/x5YLcdUp55
Think about what this means. With enough usage, anyone can make progress on the problems they care about, in many fields.
For example, don't just tune your code. Improve the algorithms behind it.
AI agents can already do real science. Anyone can run them.
We gave them research areas, no ideas, and told them not to stop. In weeks: dozens of potential papers advancing those areas.
Paper: https://t.co/LcrV2DhteI
Blog: https://t.co/x1zMboiPzR
Output: https://t.co/L09IVlLCtr