Unraveling the Enigma: Google's AI Reshapes Code Beyond Human Comprehension

0:00Today on the podcast, we're going to be talking about some major advancements and some very interesting ways Google has actually gotten AI to improve itself. And in addition, we're going to talk about where we see AI going and what these advancements actually mean for progress in AI in general and how we see it growing. So the first thing I want to bring up is I read a really interesting tweet recently by Adam Dan, and he said that the reason the AI boom is being underestimated is the GPU slash TPU shortage. The shortage is causing all kinds of limits on the product rollout and models training that are not visible.

0:38Instead, all we see is NVIDIA spiking in prices. Things will accelerate once supply meets demand. Elon Musk actually replied to that and said, true. Also, many other neural network accelerator chips are also under development. NVIDIA will not have a monopoly on large-scale training and inference forever. And actually, I've had a debate with people recently about this. Some people say that NVIDIA is going to have some serious competition from Intel and a lot of other companies. So it'll be interesting to see if they can stay on top. But I think the overarching, you know, concept here is that we may not be seen like the biggest the AI boom gets, this thing might have a long way to go in what it can do.

1:17And right now, we kind of have this big bottleneck where there really is a massive GPU shortage, these things cost an astronomical amount of money. And when the competition catches up, I think that, and there's more resources available, I think we're going to see a lot of really incredible innovation in AI in general. So today I want to talk about one thing that Google is doing to actually kind of overcome this major hardware shortage we're seeing right now, which I think is really interesting. And regardless of, you know, the fact that eventually there may be more GPUs and we might ramp up production of those kind of things, this new system is actually really important.

1:54So what happened is the fact that DeepMind's AI branch essentially just presented something called AlphaDeb. So essentially, it's a reinforcement learning agent that can discover faster sorting algorithms on its own. And it goes beyond just sorting algorithms. We'll get to that in a second. Essentially, AlphaDeb's advanced computer science algorithm outperforms, I would say, decades of human benchmarks for engineers and scientists. So the new algorithms are already part of two Stanford C++ coding libraries, or standard, not Stanford, and they're being used trillions of times per day by programmers worldwide.

2:29So in addition to that, it also has demonstrated potential for enhancing other computer science algorithms like hashing. So let me explain what that means and why this is important. The first thing I want to say is in DeepMind's kind of blog post about this, they said, digital society is driving increasing demand for computation and energy use. For the last five decades, we relied on improvements in hardware to keep pace. But as microchips approach their physical limit, it's critical to improve the code that runs on them to make computing more powerful and sustainable. This is especially important for the algorithms that make up the code running trillions of times a day.

3:06So I think this is really interesting. They bring up a point that we really push microchips to what many are saying their absolute physical limits. And so now we just need to make them more effective and bigger, and yada yada. But in reality, you know, DeepMind is saying that we actually can make improvements, we can overcome some of these, you know, limitations set by the physical speed or whatever of a microchip by improving the code that is running on them. So this is a really interesting step. So essentially, what they did is they wrote an algorithm that essentially allowed AI to create its own algorithms and programs to improve its own code base.

3:43So the thing that they specifically focused it on at the beginning was a sorting algorithm. So sorting data is something that happens behind the scenes of pretty much all of our digital interactions. So when you're searching for something online, or if you're using an app on your phone, your phone or your computer is using sorting to organize the data that gives you the best results. So like your, you know, your news feed on social media, or anywhere else, these are all sorting algorithms. So in the past, sorting was something that had been done, like, and when I say the past, I mean, like, way back in, you know, ancient times is something that was done by hand, right?

4:18You know, a library 1000 years ago was hand organized alphabetically. And that sounds crazy and archaic. So nowadays, obviously, computers are doing this for us. And what's interesting is we kind of had copied similar, similar methods for sorting things in code as you know, the archaic ways that it's been done in the past. And so what we're using AI to do now, or what Google used AI to do now, is to train the AI to come up with new ways to sort that we hadn't considered in the past. So AlphaDev didn't try to improve kind of the old way that we were using for these sorting algorithms. Instead, it started from scratch.

4:57So it looked at the basic instructions that computers use to do their tasks. And these instructions, which are usually overlooked by computer programmers were used to find a new way to sort data that is a lot faster and a lot and uses a lot less energy so to train itself to do this this whole kind of process alpha dev treated the task of sorting like a game so the goal was to build the fastest most efficient method of sorting so every time essentially they built the ai so and told it was playing a game and told it that each time it added a new step to the method it checked to see if it was correct and faster and if it was it got a point.

5:36And then the game eventually was won when it discovered the absolute fastest program and the fastest algorithm for sorting things, the fastest sorting algorithm. So the results were super amazing. Alphadeb's new sorting algorithm that they came up with was up to 75 times faster for smaller tasks and was about 1.7 times faster for really big tasks. So these improvements are being translated into a popular computer language called C++. And this makes them available to computer programmers all over the world. This is open source. Google and DeepMind allowed anyone to use this. And what's really impressive with this is it didn't just find a way to, you know, it didn't just improve the current way that these sorting algorithms work.

6:21It found smarter ways to do it. So the AI introduced a shortcut and a way to skip a step each time each time they did the sorting algorithm. And this is a massive time saver. It saves money, it saves computation time, saves energy, electricity. And this is, I think, really important because these sorting algorithms are used billions of times a day when you do almost anything on your computer or anything on your phone. There's billions of these steps are being used. And so this is actually saving quite a lot of computational power. So AlphaDev, in addition to that, also improved another process called hashing, which is essentially another way that computers organize and find data.

7:03So they made this about 30 % faster with using the same kind of method where they told it it was a game and they got points by if it could improve the algorithm. And so this new method is now available for millions of computer programmers around the world and it's been used and it's used countless times a day. And so I think this is just really impressive that they were overall able to improve for smaller tasks, this sorting algorithm by 70 % and also hashing by 30%. And I think that kind of looking forward at where this goes, Alphadep's, you know, these discoveries that the AI made are really the first step in using AI to make computers and devices work faster and more effectively based off of code alone and not just, you know, getting better hardware, right, faster GPUs or faster chips in there.

7:52But we're learning how to write better algorithms and make the code better. So it's actually less computationally heavy on the current hardware that we have. So I think it's really impressive. The team is now exploring how AlphaDev can optimize even more processes in a lot of different computer languages like C++. And I think that through these breakthroughs, AlphaDev not only makes the current methods better, but they're also able to find new ways of doing things that, you know, humans or I mean, people have not really thought of before. And I think that's one of the powers of AI is that we can ask it to improve the processes we have.

8:29And if we can give it more leniency or more leeway to kind of have more creativity and freedom with how it accomplishes that task, I will be surprised by the way it actually approaches some of these things and the way that it's actually able to accomplish this so