DeepMind AI hits silver-medal standard in International Mathematical Olympiad

Google’s DeepMind says its AlphaProof and AlphaGeometry 2 AI models are breaking new ground in mathematical reasoning — an Achilles’ heel of AI chatbots.

Google DeepMind says its artificial intelligence models are now at a “silver-level standard” in complex mathematics, after solving four out of six problems of this year’s International Mathematical Olympiad — one of the toughest mathematics competitions in the world. 

On July 25, Google DeepMind stated it had achieved breakthroughs in solving advanced reasoning problems in mathematics for AlphaProof and AlphaGeometry 2.

Mathematics has been deemed a crucial benchmark for AI development. Complex maths such as geometry requires intuitive and creative problem-solving and sophisticated reasoning skills. 

AlphaProof is a new reinforcement learning-based system for formal math reasoning, and AlphaGeometry 2 is an improved version of a geometry-solving system.

The two systems solved four out of six problems from this year’s International Mathematical Olympiad (IMO), “achieving the same level as a silver medalist in the competition for the first time,” the firm stated.

The IMO is the oldest, largest, and most prestigious competition for young mathematicians. It has been held annually since 1959, and this year’s UK-based event saw more than a hundred countries participate.

Google explained the Olympiad has become widely recognized as a “grand challenge in machine learning and an aspirational benchmark for measuring an AI system’s advanced mathematical reasoning capabilities.”

IMO gold medalist Professor Sir Timothy Gowers commented:

“The fact that the program can come up with a non-obvious construction like this is very impressive and well beyond what I thought was state of the art.”

Google previously dabbled in mathematical AI models with a system called AlphaGo that learned to play the board game Go and defeated the legendary Go player Lee Sedol in 2016.

Researchers have now combined the successor to AlphaGo, known as AlphaZero, with pre-trained language models to make AlphaProof. It trains by solving millions of problems translated into the formal programming language “Lean.”

Meanwhile, AlphaGeometry 2 is a neuro-symbolic hybrid system based on Google’s Gemini AI model, with improved capabilities in geometry problem-solving.

The company stated that its AI teams are “continuing to explore multiple AI approaches for advancing mathematical reasoning and plan to release more technical details on AlphaProof soon.”

According to a Q2 earnings filing from Google parent Alphabet, there has been a lot of spending on AI-related research at DeepMind. “Alphabet-level activities” loss, which caused by primarily by money spent on AI research and development, reached $2.3 billion for Q2, almost double that of the same period last year.

Related: 4 things Google Gemini users will be able to do soon

Meanwhile, AI competition has continued to heat up.

According to a July 15 Reuters report, ChatGPT maker OpenAI is working on a novel approach to its AI models in a project code-named “Strawberry.” The model is reportedly designed to dramatically improve AI reasoning capabilities and enable autonomous deep internet research.

Additionally, on July 25, OpenAI announced the launch of an AI-powered search engine prototype called SearchGPT.

On July 23, Meta CEO Mark Zuckerberg said the firm was taking the next steps toward open-source AI becoming the industry standard while announcing the release of its latest model, Llama 3.1.