DeepMind’s Latest AI System, AlphaGeometry, Aces High-School Math
(Bloomberg) -- Google DeepMind, Alphabet Inc.’s research division, said it has taken a “crucial step” towards making artificial intelligence as capable as humans. It involves solving high-school math problems.
Most Read from Bloomberg
Florida Governor DeSantis Drops Out of 2024 Race, Endorses Trump
Hong Kong Stocks at 36% Discount Show True Depth of China Gloom
Morgan Stanley, JPMorgan Say Buy the Dip After Treasury Rout
Stocks Rise as Wall Street Builds on Record Highs: Markets Wrap
In a paper published in Nature this week, DeepMind introduced a system called AlphaGeometry that proves complex geometry theorems better than prior computer programs. Researchers at the company described this as a major breakthrough of an AI applying reasoning and planning to tasks, something even today’s advanced models can’t do — but something scientists believe is necessary for a hypothetical future artificial general intelligence, or AGI.
“I considered this one of the grand challenges for AI,” Quoc Le, a distinguished scientist at Google and one of the paper’s co-authors, said in an interview. “A couple years ago, I would have said it was impossible.”
Last year, Google merged its research division with London-based DeepMind to better compete in the high stakes race to build AI products. Google invested heavily in the technology for years, but was widely seen as falling behind in the emerging field of generative AI, chiefly to OpenAI and Microsoft Corp. In December, Google released Gemini, an AI model that underpins its chatbot Bard and which the company says can outcompete OpenAI’s GPT-4 in certain benchmarks.
These generative AI tools are quite good at chatting naturally or producing images, but they struggle with planning or handling problems that require several steps, said Le. That’s why they’re bad at solving geometry and other advanced mathematics. In the Nature paper, AlphaGeometry tackles a set of problems from the Mathematical Olympiad, an international competition, performing nearly as well as the top high-school competitors.
AI systems have struggled with this largely because there’s not enough available data to train them to learn on their own — unlike chatbots, which can glean information from the abundant text online. To solve this, DeepMind fabricated its own dataset, creating 100 million unique examples of geometric constructs for the AI system to absorb. Making this so-called synthetic data is expensive: Google’s scientists deployed 100,000 central processing units for three days to create the dataset. The company declined to share the cost.
AlphaGeometry will not be immediately incorporated into Gemini, but Thang Luong, another co-author, said he imagines the system could eventually help serve as a mathematics tutor within services like Bard. Google has faced criticism for not moving some its research advances into commercial products quickly enough.
Google decided to open-source AlphaGeometry, giving the code and model freely to others — an approach that potentially increases the risk of the technology being misused by bad actors. Le said he was not concerned about this in light of the company’s risk assessment of the system and the current progress of the science. “If you look very closely at AI today and what it can do, you’d be surprised that it’s not very good in the field of mathematics,” he said.
Most Read from Bloomberg Businessweek
©2024 Bloomberg L.P.
