A key problem in AI is that the frequency of models is more than just the frequency of what they learn and remix and generate truly novel ideas or insights.

A new project from Google DeepMind shows that with some clever tweaks, these models can at least go beyond human expertise to design certain types of algorithms, including those that advance AI itself.

The company's latest AI project, called Alphaevolve, combines the coding skills of its Gemini AI model with an evolutionary approach to testing the effectiveness of new algorithms and producing new designs.

Alphaevolve has proposed more efficient algorithms for several calculations, including a calculation method involving matrices that can improve a method called Strassen algorithm, which has been dependent on for 56 years. The new method improves computational efficiency by reducing the number of computations required to produce results.

DeepMind also uses Alphaevolve to propose better algorithms for several real-world problems, including scheduling tasks within the data center, outlining the design of computer chips, and optimizing the design of algorithms used to build large language models such as Gemini itself.

“These are three key elements of the modern AI ecosystem,” said Pushmeet Kohli, head of AI science at DeepMind. “This Superman code agent is able to take on certain tasks and is beyond the visibility of the solution.”

Matej Balog is one of Alphaevolve's research, and he said it's often difficult to know whether large language models come up with really novel writing or code, but it can prove that no one can provide a better solution to some problems. "We've proved very precisely that you can find something that proves new and proven," Balog said. "You can really be sure what you find is impossible in training data."

Princeton University scientist Sanjeev Arora specializes in algorithm design, says Alphaevolve's advancement is relatively small and only applies to algorithms involving searches through potential answers. But he added: "Search is a fairly common idea that applies to many settings."

AI-driven coding is beginning to change the way developers and companies write software. The latest AI models make it trivial for beginners to build simple applications and websites, and some experienced developers are using AI to automate more work.

Alphaevolve demonstrates the potential of AI to come up with completely novel ideas through continuous experimentation and evaluation. DeepMind and other AI companies hope that AI agents will gradually learn to show more universal creativity in many areas, ultimately producing clever solutions to business problems or novel insights when given specific problems.

Josh Alman, an assistant professor at Columbia University who works in algorithm design, said Alphaevolve does seem to be generating novel ideas rather than what he learned during training. “It has to do something new, not just refute,” he said.