Analyzing AlphaFold’s Solution to Protein Folding | Google DeepMind AI History

Chess, Go, Starcraft, and now protein folding. I'm John Coogan and today we're talking about DeepMind's pioneering work in artificial intelligence and what it means for the future. The big news that I want to cover today is about DeepMind's AlphaFold 2. The short version is that back in November of 2020, researchers at DeepMind created an artificial intelligence program that solved the protein folding problem, which has plagued biologists for over 50 years. The long version is much more interesting though, so stick with me here while I cover some background on the company as well as the protein folding problem itself. The whole story here is super interesting because while this is clearly some of the most advanced technology humans have ever created, there is really very little consensus as to what the impact of this will be over the next decade. Let's dive into it. DeepMind was founded in 2010 specifically to try and build a general purpose AI that can be useful and effective for almost anything. They quickly brought in some of the most high profile investors, including Peter Thiel and Elon Musk. After just four years in business, Google bought DeepMind for $500 million. But the structure and goals of the company remain largely unchanged. Despite being a for profit enterprise, DeepMind has focused on developing AI systems that can win at common games. They started by playing Atari games and quickly surpassed human abilities there. And in 2014, they moved on to the strategy board game Go. Go has been played for more than two millennia, and many Go experts thought it would be impossible for a computer to ever beat a world champion Go player. This is because Go is an extremely complex game. Although the rules are simple, a standard Go board is a 19x19 grid which yields more legal positions than there are atoms in the universe. Simply put, it's impossible for any computer, no matter how powerful, to calculate every potential move and then determine the best strategy. The best Go players have amazing intuition for what strategies will work, and computers have historically not been able to develop real intuition. But DeepMind had a new approach to this problem called deep reinforcement learning. This can get a little bit confusing, but it's actually fairly simple if we break each piece down. Learning means that the designers of the system will not be manually defining the features that the program will optimize for. Essentially, the system learns what's important to pay attention to from the raw data directly. Reinforcement means that the system will learn by experience instead of by example. This means that you don't need to provide the system with examples of what a correct answer looks like. Instead you. You provide a goal called a Loss function that the system optimizes towards. And lastly, the deep part refers to the fact that the system has multiple layers of neurons that serve to predict outcomes. This all mimics the way humans learn in the real world. Our brains have neurons that fire electrical signals that help us perceive the world and intuitively understand cause and effect. This approach also lends itself well to games, because games have a clear goal that the algorithm can optimize towards winning. And games can be simulated extremely quickly in software. So when DeepMind trained its deep reinforcement learning system on go, it was able to play millions of games of Go in a very short amount of time. And this approach worked. In 2016, DeepMind's artificial intelligence, AlphaGo, beat the world champion Lee Sedal in a five game match. The Go community was shocked by the performance of AlphaGo. Over 60 million people watched the match and Lee Sedol, who had predicted that he would win in a landslide, said that he misjudged the capabilities of AlphaGo and felt powerless. It wasn't long before DeepMind tackled the even more complex game of StarCraft II, which is unique because the player doesn't have perfect information about the pieces on the board like in Go, and the game is played in real time. DeepMind's AlphaStar system defeated two professional StarCraft players in 2019 using the same fundamental approach of self trained artificial intelligence as their AlphaGo system. While it's certainly a great demonstration of computing power, winning a game doesn't have all that much practical value. And that's where the protein folding problem comes in. Proteins underpin every biological process in every living thing. Each protein has an intricate 3D shape that defines what it does and how it works. Biologists have cataloged over 20 million proteins, but pinning down the exact shape of these has been extremely difficult. Each protein is made up of a string of amino acids. Interactions between these amino acids make the protein fold, leading it to find a particular shape out of essentially limitless possibilities. For decades, scientists have been working on a way to deduce a protein shape just from its amino acids. This is one of the major challenges of our understanding of biology. Using a similar approach to their Go system, DeepMind trained a new network on the hundred thousand or so proteins that we do know the shape of. The results of DeepMind's latest system, called AlphaFold 2, are really remarkable. The efficacy of these protein folding systems is measured by what's called the global distance test. And AlphaFold 2 scores remarkably high. A score above 90 is considered roughly equivalent to the true shape and should be usable by researchers, and AlphaFold2 has essentially reached that threshold. This is important because determining the shape of proteins is extremely useful to biologists, and experimental techniques are expensive and time consuming. Most estimates put the cost of solving a single protein shape around $120,000, and it usually requires a full year to do it. Having a tool like AlphaFold could speed up work being done in biotech and help researchers understand more about diseases. It's a huge step forward. A big part of what enabled the jump in efficacy between Alphafold 1 and 2 was switching the core technology from convolutional neural networks to transformers. Both of these models still fall under the broader deep learning category, but transformers are unique in that they can effectively simulate attention, just like a human. OpenAI's GPT3 model produces remarkably realistic natural language results using transformers, and the approach has clearly been spreading quickly throughout the deep learning community. But at the end of the day, AlphaFold is just a tool. So I want to be clear about the implications here and make sure we aren't getting ahead of ourselves. Part of what makes all this possible is that potential protein structures can be simulated very quickly, not unlike simulating a GO match. When it comes to testing different drugs in human patients, these approaches are unlikely to be as effective. We don't have a way to simulate the full biological system that makes up the human body, so you can't just test millions of different drug candidates in seconds on a computer. That being said, this tool is going to speed up a lot of work, and it will almost certainly have a big impact on the lives of researchers. Now, that doesn't mean that this research can't produce entirely novel results, not by any means. There is a great moment during AlphaGo's match against Lee Sedol that I think can be really instructive here. In Game two, the AI made an extremely unexpected move that no human player would ever think to make. Lee Sedol was completely taken aback and actually left the room to take a break and think about what just happened. It seemed like an error, and the announcers were confused as to why AlphaGo would make that particular decision. But after playing out the entire game, AlphaGo won, and that unpredictable move proved to be effective. And this is what I think is so powerful about these new AI systems. They aren't just taking human strategies and performing them with more clinical precision. They are actually developing new strategies altogether. Since AlphaGo beat Lee Sedol, several top GO players have trained with the AI and improved their game significantly. And Garry Kasparov, the chess master who was defeated by Deep Blue in 1997, has said that a good human plus a machine is the best combination. My hope is that these artificial intelligence systems can work alongside humans to develop new strategies that humans wouldn't think to consider. That's the real promise of this technology in my mind. But what about the threat of AI? Last week I talked about how Elon Musk had been watching DeepMind's progress and was worried about AI becoming sentient and taking over the world world. It's a true doomsday scenario straight out of the Terminator, but fortunately we aren't quite there yet. These AI systems are still in need of lots of data and are purpose built for very specific tasks. The Economist recently covered artificial intelligence in their Technology Quarterly and highlighted a number of headwinds that AI must still overcome. These include the difficulty of collecting data in the real world, the high cost of training each model, and unpredictable human factors. While it's certainly possible that humanity will eventually face an existential threat from artificial general intelligence, I think that weaponized AI can become destructive far sooner than it can develop sentience. As I covered in my video about the Solar Winds hack, cybersecurity is increasingly becoming a primary battleground where world powers wage war. AI enabled hacks could wreak havoc on critical infrastructure, and there's always the potential for these viruses to spread beyond the creator's initial target. The Stuxnet worm was originally aimed at Iranian nuclear facilities, but it wound up infecting over 200,000 computers. Still, we are clearly in the early stages of AI development. The progress is exciting, but it's still narrowly applicable. Self driving cars are not yet widely available, and voice assistants still routinely fail to answer simple questions. So I think we should press on and continue to push AI as far as we can. So how does all of this pertain to startups? Well, there are a few things that I can think of. First, there's been a lot of discussion about breaking up big tech companies. Elizabeth Warren led the charge from the left, while Peter Thiel and Josh Hawley have led it from the right. Generally, startup founders fall into two camps on this issue. On the one hand, if you break up Google, they might be easier to compete with, and startups could potentially get bigger faster. But on the other hand, without massive tech companies, it would be hard to land billion dollar acquisition offers like we've seen with Waze, Instagram and WhatsApp. Returning to AlphaFold if DeepMind's solution to the protein folding problem actually has the practical applications that many people think it will, it kind of makes the case for not breaking Google up. This type of work is extremely resource intensive, and it doesn't look like academic institutions are anywhere close to matching the results of AlphaFold. Having a big tech company that is willing to pour billions of dollars into research initiatives that may not turn a profit seems like a good thing on balance, but we'll have to wait and see. The issue of antitrust in big tech is extremely complex and will have to be the subject of a future video, but I want to know what you think. Let me know in the comments what you think about DeepMind and AlphaFold, and if you enjoyed the video, please consider subscribing to the channel. I've also put together a playlist of other great videos about DeepMind that you can check out in the description. Thanks for watching.