Google's Math System is the Most Important AI News This Year
Don't follow the hype, follow your intuition
Three reasons make Google DeepMind’s AlphaProof the most important—and most surprising—news in the AI space since the start of 2024:
First, AlphaProof (together with AlphaGeometry 2) got a silver medal (almost gold) in the International Math Olympiads (IMO). In contrast, the best large language models (LLMs) like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet struggle with the dumbest math riddles: they say 9.9 is smaller than 9.11, fail to win at tic tac toe playing first, or miscount the r’s in strawberry. This means AlphaProof has a secret sauce—it’s not like the chatbots in the market.
Second, AlphaProof isn’t a pure LLM (the non-LLM part is the one that solves the math). It’s a neurosymbolic system—a mix of neural networks and symbolic AI.1 In particular, it’s a clever combination of Google’s Gemini and DeepMind’s AlphaZero, the chess-playing AI. Simplifying, AlphaProof uses Gemini’s language skills to generate plausible solutions and AlphaZero’s self-play style to search for the right one.
Third, AlphaProof lays the ground for the next goal in AI that many in academia and industry—even the skeptics—would consider a game changer: scientific discovery. The system is not in itself capable of expanding the frontier of our knowledge but mastering complex math is the first step on that path. OpenAI’s project Strawberry (still to be announced officially) hints at the same ambitions.
AlphaProof stands out as the most important achievement amid a flood of AI news. To illustrate my point, let me compare it with two other developments from this week.