DeepMind uses its AlphaTensor AI to solve a matrix multiplication problem by building a 3D TensorGame board game, beating a record that has stood for 50+ years
MIT Technology ReviewWill Douglas Heaven
Context & Ripple Effects
AlphaTensor is DeepMind running a proven internal playbook on a new target. The lab first showed that self-play reinforcement learning could surpass human-taught systems when its Go-playing agent beat its own predecessor without human input (trained solely by reinforcement learning), then generalized the recipe across chess, shogi, and Go with AlphaZero.
With AlphaTensor, the same technique is pointed at algorithm discovery itself: DeepMind recast matrix multiplication as a custom 3D TensorGame and let the system search for moves no human had found, toppling a record that stood for over 50 years. The follow-on coverage makes clear this was not a one-off — the approach later produced AlphaGeometry2 solving 84% of recent IMO geometry problems, confirming game-formulated search as a repeatable research method.
First-order effects
Matrix multiplication underpins most large-scale AI computation, so an algorithmically cheaper method directly lowers the arithmetic cost of training and inference workloads that depend on it.
DeepMind gains a demonstrated template for attacking open mathematical problems: formulate them as games, then let self-play search find solutions beyond the published human record.
Second-order effects
Other AI labs face pressure to add automated algorithm-discovery teams, since rivals who let models search the algorithm space can compound software-level efficiency gains on top of their hardware budgets.
Hardware vendors' optimization roadmaps gain a new variable: if discovered algorithms change which operations dominate workloads, chip designs tuned to today's standard routines may need revisiting.
Third-order effects
If the pattern holds — TensorGame for algebraic algorithms, AlphaGeometry and AlphaGeometry2 for Olympiad mathematics — reinforcement learning shifts from a tool that plays games to a systematic engine for producing new mathematical and computational knowledge.
Algorithm discovery becoming machine-searchable erodes the assumption that foundational efficiency limits are fixed by human mathematicians, redistributing advantage toward labs with the compute and game-design expertise to run such searches.
The trend: DeepMind is converting game-based self-play reinforcement learning from board-game mastery into a general engine for discovering algorithms and mathematical results, with each successive system widening the domain it can attack.
Since 1969 Strassen's algorithm has famously stood as the fastest way to multiply 2 matrices - but with #AlphaTensor we've found a new algorithm that's faster, with potential to improve efficiency by 10-20% across trillions of calculations per day! https://dpmd.ai/...
Apparently AI has found an algorithm that does 4x4 matrix multiply in 47 muls. Best traditional algorithms so far: Winograd = 48 muls. Strassen = 49 muls. We save one multiply :) https://twitter.com/...
Today in @Nature: #AlphaTensor, an AI system for discovering novel, efficient, and exact algorithms for matrix multiplication - a building block of modern computations. AlphaTensor finds faster algorithms for many matrix sizes: https://dpmd.ai/... & https://dpmd.ai/... 1/ https:/…
Alright well I guess I gotta rewrite my draft for my big “AI content generation: What's next?” post. Thanks guys. 😒 Can we just pause the advances until I get this stuff out? https://twitter.com/... https://twitter.com/...
When does computer search count as creativity? Deep Mind found a way to solve one of the most basic math problems better than any living mathematician https://www.technologyreview.com/ ... @erikbryn @amcafee
In other Google AI news from the last 12 hours, I guess an AI that was built to play board games figured out how to beat a 50-year-old record in complex math as a side project? Was today the last day before performance reviews or something? https://www.technologyreview.com/ ...
Mind-blowing result from @DeepMind but also plausible since @JeffDean & friends first published their fast data structures optimization paper — Indexes are Models. I share some lessons about that there, including https://deliprao.com/... https://twitter.com/... https://twitter.co…
Speeding up this calculation could have a big impact on thousands of everyday computer tasks, cutting costs and saving energy! https://www.technologyreview.com/ ...
A better algorithm found a better algorithm that makes the original algorithm itself better. That is, AlphaTensor improved matrix multiplication which itself improves AlphaTensor. Concrete argument for why superhuman intelligence may be able to bootstrap itself. https://twitter.c…
Aaaaaaaand just like that a @DeepMind AI system called AlphaTensor finds a way to make GPUs and TPUs 10-20% faster at matrix multiplication (the ultimate workhorse calculation) https://www.technologyreview.com/ ...
AI is casually overthrowing fifty year old best practices for the most common operation performed by your GPU This is *astoundingly* based https://twitter.com/...
Deepmind was able to find matrix multiplication algorithms that meet or beat Strassen's and other records by up to 10%: https://www.deepmind.com/... https://github.com/... https://www.nature.com/...
Matrix multiplication may not sound as sexy as “protein design”, or “self-driving cars”, but make no mistake, it is at the core of pretty much anything computational. Deepmind's AlphaTensor can discover and outperform state-of-the art algorithms 👇 https://www.nature.com/...
It's no longer possible to keep track of everything happening in the AI space. Every day there are now several interesting papers and there's just no way for a human to read them all. Perhaps an AI could. https://www.nature.com/...
Woah! “AlphaTensor discovered a way to multiply together two four-by-four matrices that is faster than a method devised in 1969 by the German mathematician Volker Strassen, which nobody had been able to improve on since” https://www.technologyreview.com/ ... https://twitter.com/.…
You know #AlphaFold for predicting protein structure. Now there's #AlphaTensor with automated discovery of novel and efficient algorithms #AI @Nature https://www.nature.com/... https://www.deepmind.com/... https://twitter.com/...
Starting from scratch, #AlphaTensor gradually improves from finding the “class-room” algorithm to discovering historical algorithms such as Strassen's - eventually surpassing them. This work is an example of how AI can advance the field of algorithmic discovery. 2/