A key question in artificial intelligence is how often models go beyond just regurgitating and remixing what they have learned and produce truly novel ideas or insights.
A new project from Google DeepMind shows that with a few clever tweaks these models can at least surpass human expertise designing certain types of algorithms—including ones that are useful for advancing AI itself.
The company’s latest AI project, called AlphaEvolve, combines the coding skills of its Gemini AI model with a method for testing the effectiveness of new algorithms and an evolutionary method for producing new designs.
AlphaEvolve came up with more efficient algorithms for several kinds of computation, including a method for calculations involving matrices that betters an approach called the Strassen algorithm that has been relied upon for 56 years. The new approach improves the computational efficiency by reducing the number of calculations required to produce a result.
DeepMind also used AlphaEvolve to come up with better algorithms for several real-world problems including scheduling tasks inside datacenters, sketching out the design of computer chips, and optimizing the design of the algorithms used to build large language models like Gemini itself.
“These are three critical elements of the modern AI ecosystem,” says Pushmeet Kohli, head of AI for science at DeepMind. “This superhuman coding agent is able to take on certain tasks and go much beyond what is known in terms of solutions for them.”
Matej Balog, one of the research leads on AlphaEvolve, says that it is often difficult to know if a large language model has come up with a truly novel piece of writing or code, but it is possible to show that no person has come up with a better solution to certain problems. “We have shown very precisely that you can discover something that’s provably new and provably correct,” Balog says. “You can be really certain that what you have found couldn’t have been in the training data.”
Sanjeev Arora, a scientist at Princeton University specializing in algorithm design, says that the advancements made by AlphaEvolve are relatively small and only apply to algorithms that involve searching through a space of potential answers. But he adds: “search is a pretty general idea applicable to many settings.”
AI-powered coding is starting to change the way developers and companies write software. The latest AI models make it trivial for novices to build simple apps and websites, and some experienced developers are using AI to automate more of their work.
AlphaEvolve demonstrates the potential for AI to come up with completely novel ideas through continual experimentation and evaluation. DeepMind and other AI companies hope that AI agents will gradually learn to exhibit more general ingenuity in many areas, perhaps eventually generating ingenious solutions to a business problem or novel insights when given a particular problem.
Josh Alman, an assistant professor at Columbia University who works on algorithm design, says that AlphaEvolve does appear to be generating novel ideas rather than remixing stuff it’s learned during training. “It has to be doing something new and not just regurgitating,” he says.
The DeepMind researchers found that they could sometimes give an idea for an algorithm as a prompt and produce interesting new results. Alman says this raises the prospect that human scientists could collaborate with a system like AlphaZero. “That seems really exciting to me,” he says.
AlphaEvolve is not the only DeepMind program to demonstrate real ingenuity. The company’s famous boardgame-playing program AlphaZero was able to devise original moves and strategies through its own form of experimentation. Balog says that the evolutionary approach used by his group could be coupled with the reinforcement learning method employed in AlphaZero—a process that lets a program learn through positive and negative feedback—to create something that explores new ideas in other areas.
Two previous DeepMind projects also used AI to push the boundaries of computer science. AlphaTensor, from 2022, used the reinforcement learning method to produce novel algorithms. Fun Search, from 2024, used an evolutionary method to generate more efficient code for a given problem.
Neil Thompson, a scientist at MIT who studies the way algorithms affect technological progress, says that a key question is not just whether AI algorithms can exhibit original ideas, but how generally this may apply to scientific research and innovation.
“If these capabilities can be used to tackle bigger, less tightly-scoped problems, it has the potential to accelerate innovation—and thus prosperity,” Thompson says.
What do you make of AlphaEvolve? What novel problems would you like to see AI take on? Let me know by emailing [email protected] or adding to the comments section below.
