DeepMind's AlphaDev uses self-play to write faster sorting code, achieving speedups up to 70% on short sequences and 1.7% on large workloads. The AI treats coding like a game and generates assembly directly, with some algorithms already deployed in LLVM's C library.
Dear fellow scholars, this is two minute papers with Dr. Zsolnai. Wow, this is a historic moment today. Let me tell you why. DeepMind just introduced their new system called Alpha Dev and it is based on something they call AlphaZero. AlphaZero was an AI system that could learn to play chess, Go and Shogi by playing against itself. So how good is it? Well, it is incredible and I really mean that. This is an AI that can beat world champions in these games. Now, I hope you allow me to quote myself from 400 videos. The point of AlphaZero is not to solve chess or any of these games. Its main point is to show that a general AI can be created that can perform on a superhuman level on not one, but several different tasks at the same time. So it can learn as long as we formulate the problem. As a video game, for instance, it is easy to think about chess like we think about a game, because that's exactly what it is. If you win, you get some points, if you lose, you lose some points. Simple. However, what their new AI AlphaDev does is computer coding. Excuse me, what? How is coding a video game? Yes, it really imagines computer coding as a video game. You see, in the case of sorting numbers, the computer game level will be a series of numbers. And the actions that we perform with a controller in a game will be writing code. If this code is written well, it will sort these numbers correctly. In which case. Oh yes, that's right, we get a nice score. But let's add a twist to the game. Latency correctness is necessary, but the faster the code, the higher the score. I love it. Coding as a video game, that sounds nice. However, I doubt it will be able to say anything new as sorting is a very well studied area in computer science. This is one of the first things every computer scientist starts learning about. And there are hundreds and hundreds of algorithms published in this area. I don't expect that this will be able to say anything truly new here, but if it can at least match what humans do, that would be nice. And. What? Look at that. Am I dreaming or is this real? On shorter sequences, it is up to 70% faster and even on big workloads, it is about 1.7% faster than the currently used techniques. My goodness. And if that is at all possible, it gets even better. Get this. They have deployed this into the LLVM LIB C sorting library. What does that mean? Well, hold on to your papers, fellow scholars, because it likely means that it is now being used more than a trillion times a day. Wow. If someone finds out something new and useful in the area of physical manufacturing, it might very well take decades for a paper to see any use in the real world. However, here we are going from paper to product. Not in decades and not even in years. This is tech transfer in a matter of days. And this is an AI that can say, I got this, don't worry about it, just tell me what you wish to see and I'll write this software on my own. And don't forget this is a difficult game. It is like constructing a skyscraper from scratch. One wrong move and you are done. And it is not just as good as the collective intelligence of humanity, but it even exceeds it. My goodness, what a time to be alive. And interestingly, it also doesn't just write C, but it writes assembly code directly. This roughly means that it is really low to the ground. This is a bit closer to the metal, the hardware itself, proper coding skills from an AI that even rivals humans. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers. Because with the Lambda GPU cloud you can now get on demand H100 instances for just one 99 per hour. Yes, one 99 and they are one of the first cloud providers to offer publicly available on demand H100 access. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today. Thanks for watching and for your generous support and I'll see you next time.
