Keywords: DeepMind · AlphaEvolve · matrix multiplication exponent · combination loss analysis · exact rational certificate · Carnegie Mellon · arXiv 2608.16884
---
How to Read This
In 1969, Volker Strassen reduced the arithmetic complexity of multiplying two n×n matrices from n³ to O(n^log₂7) ≈ O(n^2.807), opening a 57-year chase to find the true value of the matrix multiplication exponent ω. The lower bound is 2 (you must at least read every element), and Strassen's 2.807 has never been refuted — only pushed down. On August 17, 2026, a DeepMind-led team — Emilien Dupont, Marvin Eisenberger, Borislav Kozlovskii, Abbas Mehrabian, Francisco J. R. Ruiz, Abigail See, Renfei Zhou (CMU), Josh Alman (Columbia), Virginia Vassilevska Williams (MIT), and Matej Balog — pushed the upper bound from 2.371339 (Alman et al., SODA 2025) down to ω < 2.371177. The difference between the two records is only 0.000162, which sounds like a rounding error — but on this track, every drop in the sixth decimal place corresponds to a genuinely new mathematical construction. This is not "AI independently proving mathematics," but a collaboration sample of "humans rewriting the mathematics, AI rewriting the optimizer's code" — worth unpacking in detail.
1. What the Optimization Problem Looks Like
Where do upper bounds on ω come from? All progress over the past half century has come from the same mother lode — the laser method, most recently in the form of combination loss analysis (Duan et al. 2022, Williams et al. 2024, Alman et al. 2025). This approach translates "can ω be pushed to some value" into "can one find a feasible point in a non-convex optimization problem." It sounds like a common industrial black-box optimization problem, but the essential difficulty is: how low a bound you can prove depends entirely on how large you can make the search space; and as the search space grows, the combinatorial variables explode exponentially — which is why previous work could only handle ℓ* = 3.
Picture the optimization problem as a non-convex black box with hundreds of knobs; each knob setting corresponds to a certificate, and the certificate's value is the ω upper bound implied by that setting. Finding better knob combinations is one of the most painful shapes in non-convex optimization. The previous record (SODA 2025), ω < 2.371339, was held together only by Alman pushing this method's asymmetry to the extreme.
2. Three Layers: From 2.371339 to 2.371177
The DeepMind short paper submitted to arXiv on 2026-08-17 (2608.16884) splits the work into three parts. Each layer alone is not "sexy," but stacked together they advance the bound by 0.000162:
**Layer 1: Algebraic restructuring to make a larger ℓ* solvable.** They reformulate the original optimization problem at an algebraic level so that a larger parameter setting becomes tractable. The paper calls this parameter ℓ*; the previous record used ℓ* = 3, this work pushes it to **ℓ* = 4. This is pure mathematics — AI had no hand in it.
Layer 2: Rewriting the solver with JAX + automatic differentiation + Sinkhorn-Knopp normalization. Previous solvers used sequential quadratic programming (SQP). DeepMind switched to a gradient-based solver written in JAX, using automatic differentiation for gradients and Sinkhorn-Knopp for normalization (repeatedly projecting intermediate matrices back onto the doubly stochastic matrices). This layer is numerical-methods work — AI still had no hand in it — but after swapping tools, the shape of the feasible region changed.
Layer 3: Handing the solver to AlphaEvolve. AlphaEvolve is DeepMind's own Gemini-driven coding agent, designed specifically to "rewrite the optimizer's code itself." In other words, it continually rewrites the solver code into new variants, runs them, checks what upper bound comes out, and rewrites the code again. The paper states explicitly: "AlphaEvolve does not prove anything about ω; it rewrites the optimizer code that searches for upper bounds on ω."
3. Why This Matters
This track has a distinctive iteration rhythm. Each new upper bound means either a new combinatorial structure has appeared, or an existing one has been mined more deeply. Every advance is a genuinely new construction, not a small tweak from parameter fiddling. This 0.000162 sounds tiny, but it represents "another step down the staircase of the past 57 years."
More crucially, this result comes with an exact rational arithmetic certificate. DeepMind rewrote the final numerical solution using rational numbers and re-verified it under exact arithmetic — to "eliminate the possibility of floating-point error." Any ω upper bound of 2.371177 computed with IEEE floating point could be suspected of being 2.3711769999... rounded up; writing it in rationals makes it independently verifiable with no rounding. This is exactly the posture such "AI moved one digit on a 70-year constant" news items should adopt: certificates must be re-runnable by third parties.
Up to this point, DeepMind did what mathematicians do — except they made the boundary of "AI contributing decisive ideas as a collaborator" land, for the first time, on a "machine-verifiable certificate." This is what distinguishes it from the Sum-of-Squares conjecture (July 19) and the Crouzeix conjecture (July 20) episodes: those were "humans wrote the proofs, AI helped find ideas," while this one is "mathematicians set the problem, AI rewrites the solver, the certificate is machine-checkable." Three modes now appear side by side in the late summer of 2026.
4. Three Honest Limitations
First, the result has not yet been independently reproduced. The ω < 2.371177 from AlphaEvolve is DeepMind's own computation. We must wait for someone on arXiv to replicate it, or for the authors to release the verification repository. The paper mentions that verification code and the found feasible solution are in a "code repository in preparation" — not yet released. Until that step happens, independent cross-validation has to wait.
Second, this upper bound will almost certainly not make real matrix multiplication faster. ω is an infimum, and the "crossover point" with existing algorithms — the matrix size beyond which these methods beat schoolbook multiplication — is far beyond any size used in real computation. These are galactic algorithms; however beautiful the asymptotic bound, no BLAS library will be updated. In other words, there is zero impact on the speed of training any model or running any real workload.
Third, AlphaEvolve's credit is scoped narrowly. What it rewrote is not the "proof path for ω" but "the optimizer's code." The problem formulation, the parameter space, and the feasible combinatorial methods were all written by humans; the machine searched an already-defined search space. On this point, headlines like "AI advances a 57-year-old math problem" are somewhat misleading; the accurate description is "DeepMind + mathematicians collaborating to scale the search to a new magnitude." But this is still a crucial advance, because enlarging the search scale has been a necessary condition for every step forward in those 57 years.
5. What It Means
This track has one notable feature: the intervals between records keep shrinking. From Pan's 2.781 (1978), to Coppersmith-Winograd's 2.3755 (1990, later jokingly called the "CW convex optimization bound"), to Stothers' 2.387 (2010), to Alman-Williams' 2.372869 (2020), to Alman's 2.371339 (2024), to AlphaEvolve's 2.371177 (2026-08). Each step is hard, but with AI in the loop, search scale has become a new tool: previously, human optimization algorithms were choked by their own compute; now AI can run thousands of iterations in parallel while rewriting code. The pattern of "AI doesn't prove mathematics, but accelerates the tools of mathematics" will likely recur on more records.
The arXiv abstract page lists no institution; the team consists mainly of nine DeepMind researchers plus one each from Carnegie Mellon, Columbia, and MIT. First author Emilien Dupont is a DeepMind research scientist long working on neural network interpretation and AlphaEvolve-style automated code discovery; corresponding author Matej Balog is also at DeepMind; Josh Alman and Virginia Vassilevska Williams are precisely the people who set the previous record of 2.371339 — and co-authors on this paper. This is why the seemingly "sudden small advance" is in fact the same math circle and the same AI circle extending a single chain of collaboration.**
---
References
- arXiv 2608.16884: Improving the matrix multiplication exponent with modern optimization and AlphaEvolve (DeepMind / CMU / Columbia / MIT, submitted 2026-08-17 17:59 UTC): https://arxiv.org/abs/2608.16884
- AI Brief 18 August 2026: a sharper matrix multiplication exponent, with AlphaEvolve in the loop (independent technical commentary, 2026-08-18): https://muhammad-ahmed.com/blog/ai-brief-matrix-multiplication-exponent
- DEV.to: AlphaEvolve helped tighten the matrix multiplication exponent, and the proof was checked in exact arithmetic (independent technical write-up, 2026-08): https://dev.to/breachprotocol/alphaevolve-helped-tighten-the-matrix-multiplication-exponent-and-the-proof-was-checked-in-exact-2fhi
- Alman, Duan, Williams, Xu et al.: More asymmetry yields faster matrix multiplication (SODA 2025, source of the previous record 2.371339)
- Duan, Wu, Zhou et al.: Fast Matrix Multiplication via Group Leaders (starting point of combination loss analysis, 2022)
- Williams, Xu, Xu, Zhou: New Bounds for Matrix Multiplication with applications to GRH and beyond (2024)