We just discovered faster sorting algorithms!

DeepMind built AlphaDev, which discovered faster sorting algorithms using Deep Reinforcement Learning.

📚 My 3 favorite Machine Learning books:

Disclaimer: Some of the links included in this description are affiliate links where I'll earn a small commission if you purchase something. There's no cost to you.

[ricosrealm]

I wouldn't say it found a new algorithm, it just optimized the existing ones. The AI made the sorting library slightly more efficient (1.7%) by adjusting assembly code instructions. Essentially AlphaDev is like a reinforcement learning compiler AI that found the best way to optimize the sorting library.

[smvk]

sounds cool, but would be nice to have the link of the paper and how much % improvement are they talking about ?

[cimmik]

Was this video generated by VideoGPT? Because it spends 1:32 on giving one information: an AI generated a sorting algorithm. Nothing about the algorithm nor the AI

[enantiodromia]

I would have enjoyed a thorough explanation of the details of the algorithms.

[WilliamDye-willdye]

IMO there's a deep reason why we have never found a simple optimal sorting algorithm: even in pure mathematics, nature does not have a strongly preferred order. We choose to focus on certain arrangements and the operators which work well with them, but we're cherry picking a few tiny subsets which happened to be useful and/or easier for our brains to comprehend. In other words, abstract set theory does not have a baked-in preferred order, we pick the orders we find useful out of abstracts set theory.

[GEMSofGOD_com]

My immediate thought: space (which is about O(n) sorting, ideally) and operation time have different weights, one can even project hardware development into the future, but yeah, hardware instructions are to be taken in focus for some desired fitness

[mael]

Your intense hand movements are making the video harder to watch, visually.

[timgabby]

Beware "breakthroughs", unless Ping..

[titusfx]

Let's see the paper, but with comparison, the lower possible algorithm is nlogn, and it is impossible to get lower that.