#Google #DeepMind used a #largelanguagemodel (#LLM) to solve an unsolvable #math problem
First #AlphaTensor found a way to speed up a calculation at the heart of many different kinds of #code, beating a 50-year record. Then #AlphaDev found ways to make key algorithms used trillions of times a day run faster.
They had to throw away most of what it produced but there was gold among the garbage. https://www.technologyreview.com/2023/12/14/1085318/google-deepmind-large-language-model-solve-unsolvable-math-problem-cap-set/
#AlphaTensor
The application of game-playing GPU to ML matrix multiplication changed ML. ...
Google DeepMind - AlphaZero - AlphaTensor
Discovering faster matrix multiplication algorithms with reinforcement learning
https://www.nature.com/articles/s41586-022-05172-4
Discovering novel algorithms with AlphaTensor https://triangleagency.co.uk/discovering-novel-algorithms-with-alphatensor/?utm_source=dlvr.it&utm_medium=mastodon #TheTriangleAgencyNews #Algorithms #AlphaTensor #Discovering
@achim I agree that most probably #ChatGPT won't help us solve the climate crisis or world hunger. At least not in a direct way. But I also deem it a bit unfair to dismiss all AI technologies as baseless hype without any scientific merit. #AlphaFold changed protein folding research forever, and #AlphaTensor recently found a faster matrix multiplication algorithm (https://www.heise.de/news/AlphaTensor-AI-system-speeds-up-matrix-multiplication-with-new-algorithm-7288780.html). Thus, I would be more than happy to see the advent of a coding AI advancing #OpenSource developments. :)
#ITByte: #Google #DeepMind’s #AlphaTensor, pioneer in AI-driven mathematical innovation, developed a novel matrix multiplication #Algorithm.
Know more...
Tomorrow 5pm Alhussein Fawzi
will talk about #AlphaTensor discovering faster matrix multiplication algorithms, the #Maths breakthrough from DeepMind with reinforcement learning.
Register in person (London) or online:
https://eventbrite.com/e/imperial-college-icarl-seminars-alhussein-fawzi-deepmind-tickets-451788921827?aff=estw&utm-campaign=social&utm-content=attendeeshare&utm-medium=discovery&utm-source=tw&utm-term=checkoutwidge
#ArtificialIntelligence
#ImperialCollegeLondon
One week for Alhussein Fawzi
talk at the @icarl research spotlight session about #AlphaTensor @Nature the recent #Math breakthrough from
DeepMind through reinforcement learning.
You can attend in person (London) or online. Register here!
https://eventbrite.com/e/imperial-college-icarl-seminars-alhussein-fawzi-deepmind-tickets-451788921827?aff=estw&utm-campaign=social&utm-content=attendeeshare&utm-medium=discovery&utm-source=tw&utm-term=checkoutwidge
When ever I see a cool new result in #ComputationalMath, I like to see if I can replicate it. So, last month when that Nature article came out about #MatrixMultiplication formulas from #AlphaTensor I set out see if I could get their formulas and verify them symbolically.
I was able to do that and of course they were right. But I was excited to see Kauers and Moosbauer publish a response a couple days later. So, here's their results replicated in a Maple Jupyter notebook https://github.com/johnpmay/MapleSnippets/blob/main/KMtoFFM.ipynb
Finally got a chance to review the #AlphaTensor paper from @AlhusseinFawzi and team @DeepMind. Interesting piece and trajectory.
But the response from @ManuelKauers and @jakobmoosbauer in their paper "FBHHRBNRSSSHK-ALGORITHM", beating Deepmind's 5x5 matmul by 1.
The abstract reads somewhere between a tongue and cheek, "lets poke fun at this" comment and a literal shitpost, and I’m not over it.
The question of course is how much @googlecloud time does @DeepMind now get to take back the record?
@ben
Families of specialized hardware becomes normalized and increases flops/sec by >10x.
Matrix algebra becomes 50% more efficient with new algorithms #AlphaTensor.
Such specifications get abstracted into #Jax … TF and Torch become deprecated.
Early proof-of-concept advancements in RL lead by #selfsupervised learning on how-to videos.
Generative AI becomes commonplace in CGI, GameDev, WebDev, Marketing for economic reasons, displacing many creatives.
Thought you might enjoy this summary of #AlphaTensor by #DeepMind. They applied #AlphaGo (neural net guided Monte Carlo tree search) to the problem of simplifying #LinearAlgebra expressions and speeding up matrix multiplication:
https://fgiesen.wordpress.com/2022/10/06/on-alphatensors-new-matrix-multiplication-algorithms/
4x4 matrix multiplication #algorithms:
* Brute force: 64 multiplications
* Strassen/Winograd algo: 49 mult
* AlphaTensor discovery: 47 mult
Don't do this at home. Modern PC hardware likes the direct algo.
RT @DeepMind: 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://t.co/E18DezRPTL
Client Info
Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst