Google’s Shocking New AI Won a Medal in the Math Olympics - AlphaProof

Last week, something big happened at Google DeepMind's headquarters in London. They've got this ceremonial gong that they only ring for major breakthroughs. And this time it rang out for something truly unexpected. Back in 2016, they hit the gong for AlphaGo, an AI system that dominated the ancient game of Go, even defeating the best human players in the world. Then in 2017, they rang it again when AlphaZero announced another AI conquered the chess world, taking down human world champions. Fast forward to just last week and they brought out the gong once more to celebrate their latest achievement. Their AI just competed in the International Mathematical Olympiad, a competition usually reserved for the brightest young math whizzes from around the world. And get this, the AI performed so well it could have walked away with a silver medal if it were human. This isn't just a win for AI, it's it's a sign that computers are getting seriously good at tackling problems that have always been a human stronghold. So here's the deal. Last week, DeepMind's AI took part in the International Mathematical Olympiad, IMO, which is basically the Olympics for the world's brightest young mathematicians. This year, the event was held from July 11 to July 22 at the University of Bath, about 100 miles west of London. It's a huge deal in the math world, drawing 609 high school students from 108 countries, all competing to solve some of the most challenging math problems out there. The competition is fierce, with students vying for gold, silver and bronze medals. But here's where things get really interesting. For the first time ever, an AI system not only competed alongside these human math prodigies, but also performed well enough to earn a medal. Yeah, you heard that right. An AI system at the level of a silver medalist. It managed to solve four out of the six problems presented, earning a total of 28 points. While it didn't reach the top spot, it's still a groundbreaking achievement that has everyone talking. Dr. Pushmeet Kohli, one of the leads at Google DeepMind, described this accomplishment as a massive breakthrough in AI's ability to engage in mathematical reasoning. He went so far as to call it a phase transition, which is a fancy way of saying that this marks a transformative moment in how AI can be used in mathematics. It's not just about this one competition, it's about the broader implications for the future of AI and math. To make sure the AI's performance was judged fairly, DeepMind brought in two independent experts to assess its work. These weren't just any experts, they were Timothy Gowers, a mathematician at the University of Cambridge who's won the Fields Medal, which is like the Nobel Prize for mathematics, and Joseph Myers, a software developer who also happens to be a past IMO gold medalist. Myers has also served as the chair of the problem selection committee for the imo. So he knows his stuff. Both of these guys took a close look at the AI's solutions and they were impressed. Gowers even said that while his expectations were high going into this, the AI exceeded them in some areas. So let's talk about how the competition went down. The human competitors, these super smart high school students had to sit for two exams over two days. Each exam had three problems covering topics like algebra, geometry, combinatorics and number theory. They had just 4.5 hours per exam to solve these incredibly tough problems. Meanwhile, the AI was working away. Back at the DeepMind lab in London. With no time constraints, the researchers were watching closely. And every time the AI managed to solve a problem, they banged the gong in celebration. Now, to give you an idea of how tough these problems are, consider only one student, Hao Zishi from China, managed to get a perfect score of 42 points by solving all six problems. The US team won the overall competition with a total of 192 points, followed closely by China with 190 points. And remember, the AI earned 28 points by fully solving four two in algebra, one in geometry and one in number theory. However, it struggled with the two combinatorics problems, but still this performance was strong enough to earn a silver medal if it had been a human competitor. DeepMind's researchers are particularly excited about this achievement because it represents a significant step forward in AI's ability to tackle complex mathematical problems. For them, it's not just about how fast the AI can solve these problems, but about the fact that it can solve them at all. Dr. David Silver, another research scientist at DeepMind, pointed out that this marks a step change in the history of mathematics. It's the point where we move from computers only being able to solve very simple problems to computers being able to tackle problems that are on par with those solved by human experts. And in the future they might even go beyond that. DeepMind's work on applying AI to mathematics has been in the works for several years. They've been collaborating with world class research mathematicians to push the boundaries of what AI can do. Dr. Alex Davies, who leads Deep Minds Mathematics Initiative, is explained that mathematics requires a unique combination of abstract thinking, precise calculations and creative reasoning. This makes it a perfect test for AI Systems, especially for those aiming to achieve what's known as artificial general Intelligence. AGI is the ultimate goal in AI research, where a system can perform a wide range of tasks at or above human level. Math Olympiad problems have become a kind of Benchmark for testing AI's capabilities. Back in January, DeepMind introduced a system called Alpha Geometry, which was able to solve Olympiad level geometry problems at nearly the level of a human gold medalist. Fast forward a few months and Alpha Geometry 2 has now surpassed even the gold medalists in solving these types of problems, according to Thang Luong, the principal investigator riding on this success, DeepMind decided to take things up a notch for this year's IMO by bringing in a multidisciplinary disciplinary team to tackle a broader range of mathematical subjects. For this Olympiad, DeepMind actually had two teams working in parallel. One team was led by Thomas Hubert, a research engineer in London, while the other was led by Thang Luong and Kwak Le in Mountain View, California. These teams were stacked with talent, including a dozen IMO medalists. Dr. Luong even joked that this was by far the highest concentration of IMO medalists at Google. The AI that competed this year was a combination of Alpha Geometry and a new system called Alpha Proof, designed to handle a wide range of mathematical problems. Alpha Proof is particularly interesting because it incorporates a variety of AI technologies. One approach they used is called an informal reasoning system, which is based on natural language. This system leverages Google's large language model Gemini to understand and solve problems. It's good at recognizing patterns and suggesting the next steps. And while large language models are known for sometimes making stuff up, in this case the AI was able to stay focused and avoid too much creative wandering. Another key approach they used is a formal reasoning system, which is all about strict logic and code. This system uses a tool called Lean, which is a type of theorem prover and proof assistant. It ensures that every step the AI takes in solving a problem is logically sound and can be verified as correct. This is crucial in mathematics, where precision is everything. And then there's also a reinforcement learning algorithm, which is a type of AI that that learns by itself. This algorithm is based on the same technology that powered AlphaGo and AlphaZero. It's designed to keep learning and improving over time without needing a human teacher to guide it. Dr. Silver, who's in charge of reinforcement learning at DeepMind, explained that this kind of algorithm can scale indefinitely, meaning it can continue learning and solving increasingly complex problems. The idea is that eventually this AI AI could solve problems that are too difficult for even the best human mathematicians. And who knows, maybe one day it'll be able to tackle problems that humans haven't even thought of yet. But it's not all about AI taking over. The hope is that these AI systems will become valuable tools for mathematicians, helping them solve problems faster and more efficiently. Dr. Timothy Gowers, the field's medalist, isn't too worried about AI replacing human mathematicians anytime soon. He pointed out that there's still a long way to go before AI can handle the kind of high level research that human mathematicians are doing. But he also thinks that if AI can solve tough problems like those at the imo, then it won't be long before we start seeing AI tools that could be really useful in math research. If this technology keeps advancing, it could make math more accessible to everyone, speed up discoveries, and even help mathematicians think. Think outside the box. So yeah, that's why they banged the gong at Google DeepMind's headquarters. They're not just celebrating a victory, they're celebrating a new era in mathematics where AI isn't just a tool, but a real collaborator. And maybe, just maybe, a game changer for the entire field. Alright, I hope you enjoyed the video. If you did, please make sure to like the video and subscribe to the channel for more AI and tech news. Thanks for watching and I'll see you in the next one.