DeepMind’s AlphaEvolve AI: History In The Making!

Today you might be witnessing history in the  making. I am so excited to tell you about this,   I cannot wait any longer. This new  paper called AlphaEvolve and it is   absolute insanity. I had the honor of  having a look at it earlier before its   release. And it’s been days now  and my head is still spinning. You see, about 900 of our videos ago or 9  years ago, this was the state of the art in AI.    Yes. Now you might think I flipped out  because a neural network learned to rotate   this car into a frontal position, or that  it sorted these lines. But no. Not at all.    I flipped out because it learned not just to  solve a task at hand, but because it learned   create an algorithm to do that. I thought  this concept might make it big one day. And now, 9 years later, finally, this is the day.   Here is AlphaEvolve. An evolutionary coding agent   that grows algorithms and computer code  out of nothing. It is absolutely crazy. It was given 50 open problems in mathematics,  geometry, combinatorics and more. And in 75% of   the cases, it was able to rediscover the best  solution the smartest humans were able to come   up with so far. That is already insanity, but  that’s nothing, because in 20% of the cases,   it even improved upon previous solutions. Yes,  an AI that is able to push humanity forward   in a way that no human can. I think this is a  historic moment, but get this — there is more. It also discovered a faster matrix multiplication  algorithm than one of the landmark techniques,   which is called the Strassen algorithm.   That is ridiculous. This area is so old,   so saturated with so many works, it is almost  impossible to come up with something that is   just a tiny bit, maybe 1% better. They say  that no one was able to do that in 56 years,   and after the 56 years, not a human  did it. But an AI technique did. Wow. This means that we are in for a crazy AI ride  where these methods will start to improve   things that we thought are impossible  to improve. For instance, don’t forget,   AI techniques themselves are also running on  matrix multiplication, so what does it mean? Well, now hold on to your papers  Fellow Scholars and check this out:   it learned to improve circuit designs for  chips meant to run these AI techniques,   and it also improved its own training algorithm.   Yes, you heard it correctly. It is able to create   a better piece of hardware that it runs  on, and it also improves its own code. That sounds insane. So, how does it do it? Well, first, you can mark a piece of code that you  want to evolve, some logic to decide whether a new   algorithm is better or worse, and any comments  that you might have. And then, off it goes in   an evolutionary loop. Then, a new piece of code  emerges, that gets better and better over time. So — is that new? Nope. No-no-no. Evolutionary algorithms have existed for a  while now. For instance, you see my genetic   algorithm building up the Mona Lisa from  a bunch of triangles. You start out with   a random configuration, and it over time,  evolves into the correct solution. It very   loosely mimics how evolution works in nature. The  link is in the description. I wrote this by hand,   and it shares some very rough, basic concepts with  AlphaEvolve. But AlphaEvolve is a supercharged,   AI-infused variant of that. It is space  technology compared to this program. By the way, AlphaEvolve also solved a variant of  the kissing problem too. What is that about? Well,   this is not about a robot doing…whatever  the heck it is doing here. Nope. This is   one of the crazy names mathematicians like to  come up with. And if you think these spheres   kissing each other is so weird, well then, look  at this. Oh my goodness. This is the hairy ball   theorem. It doesn’t have anything to do with  this new AI, but I couldn’t resist mentioning   it. Mathematicians and their weird naming  schemes. You’ll get used to it, don’t worry. So, what happens now? Well, I think I  have an idea, and it’s pretty insane.    Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. I think they are going to Alpha all  the things. But I hear you asking,   Károly, what do you mean by that? Well, earlier DeepMind came up with AlphaGo. An

AI-based technique that was able  to play Go on a superhuman level. Then, AlphaStar, to play the strategy game  Starcraft at the very least on a level of a   human champion. Perhaps even better.   But both of these are the wrong way   to think about their Alpha project.   These are all experiments to create   generally intelligent AI techniques  that can learn and do so much more   than just play Go and StarCraft. And  AlphaEvolve is finally one of these. This one learned something so much more general  — computer coding. And it can code up so much   more than just Chess or Go. You see, computer  code underpins almost everything we do. So,   yes, with that, it will be able to design  better hardware for itself to run on, and   it gets better — perhaps even the next version of  AlphaEvolve will be written by AlphaEvolve itself. Get it now? Sir Demis Hassabis, DeepMind’s CEO  keeps saying that step number one is solving   intelligence. And step number two is using  this intelligence to solve everything else. He says the following: “I think one  day, maybe we can cure all disease   with the help of AI. Maybe within the  next decade. I don't see why not. ” Once again: cure all disease perhaps  in the next decade. Sounds crazy,   right? But don’t forget, this is not coming  from some guy, this is coming from a person   who just won a Nobel Prize in chemistry with  his AI technique. You know what? I could not   imagine how this could be true, but with  AlphaEvolve, you know, now I am a believer. So, let’s Alpha all the things, and with  the help of human ingenuity plus AI,   we might be able to cure all disease in  the next decade. What a time to be alive!