But the news that really stood out to me was one that didn’t get as much attention as it should have. It has the potential to introduce more powerful artificial intelligence and scientific breakthroughs than were previously possible.
Last Thursday, Google DeepMind announced that it had built AI systems that can solve complex mathematical problems. The systems – called AlphaProof and AlphaGeometry 2 – worked together to successfully solve four of six problems from this year’s International Mathematical Olympiad, a prestigious competition for high school students. Their performance was equivalent to winning a silver medal. It is the first time that any AI system has achieved such a high success rate on this type of problem. My colleague Rhiannon Williams has the news here.
Mathematics! I can already imagine your eyes shining. But bear with me. This announcement is not just about math. In fact, it marks an exciting new development in the kind of artificial intelligence we can now create. AI search engines that you can chat with may add the illusion of intelligence, but systems like Google DeepMind could enhance the true intelligence of AI. For this reason, building systems that are better at math has been a goal for many artificial intelligence labs, such as OpenAI.
This is because mathematics is a benchmark for reasoning. To complete these exercises aimed at high school students, the AI system needed to do very complex things, such as planning to understand and solve abstract problems. The systems were also able to generalize, allowing them to solve a whole range of different problems in various branches of mathematics.
“What we saw here is that you can combine [reinforcement learning] that was so successful at things like AlphaGo with large language models and produced something that is extremely capable in the text space,” said David Silver, principal researcher at Google DeepMind and arguably the pioneer of deep reinforcement learning, in a press briefing. In this case, this feature was used to build programs in the Lean computer language that represent mathematical proofs. He says the International Mathematical Olympiad represents a test of what is possible and paves the way for further discoveries.
That same recipe could be applied to any situation with really clear, verified reward signals for reinforcement learning algorithms and a clear way to measure correctness like you can in math, Silver said. A possible application would be coding, for example.
Now for an obligatory reality check: AlphaProof and AlphaGeometry 2 can still only solve hard high school problems. This is a far cry from the extremely difficult problems that top human mathematicians can solve. Google DeepMind emphasized that its tool, at this point, has not added anything to the body of mathematical knowledge that humans have produced. But that wasn’t the point.
“We aim to provide a system that can prove anything,” Silver said. Consider an AI system as reliable as a calculator, for example, that can provide proofs for many difficult problems or verify tests for computer software or scientific experiments. Or maybe build better AI tutors that can provide feedback on exam results or data-check news articles.