Using AI for Scientific Discovery: The Google DeepMind FunSearch Breakthrough

Google DeepMind's FunSearch is pushing the boundaries of generative AI, demonstrating how large language models can be used for scientific discovery by solving problems that have stumped researchers for decades.

May 1, 2026 10 min read

An abstract image of a neural network representing the process of using AI for scientific discovery with Google's FunSearch.

For years, the conversation around generative AI has been dominated by its creative and communicative abilities—writing poetry, generating stunning images, and holding surprisingly human-like conversations. But a new frontier is rapidly emerging, one that shifts the focus from artistic creation to empirical discovery. We are entering an era of using AI for scientific discovery, where these powerful models become partners in solving some of the most complex challenges in science and mathematics. This isn't science fiction; it's a revolution happening now, and Google DeepMind's FunSearch system is at the vanguard.

FunSearch represents a paradigm shift. Instead of just predicting the next word in a sentence, it leverages the underlying intelligence of large language models (LLMs) to generate novel, verifiable, and often surprising solutions to hard problems. By systematically exploring vast search spaces of possible answers, this approach has achieved what was previously thought to be years, if not decades, away: an AI making a genuine, published contribution to human knowledge. This article dives deep into how FunSearch works, the breakthrough it already has under its belt, and what it means for the future of research.

What is FunSearch and How Does It Work?

At its core, FunSearch (a name derived from "searching in function space") is a clever system that pairs a creative, code-generating LLM with a rigid, automated evaluator. It’s like having an incredibly imaginative brainstormer who throws out thousands of ideas, and a meticulous, tireless fact-checker who instantly tests each one for validity.

The process is elegant. The LLM is tasked not with writing an essay, but with writing small computer programs, or "functions." Each function represents a potential strategy for solving a given problem. The LLM, in this case Google

Key Takeaways

▸FunSearch is a new Google DeepMind system that pairs a large language model (LLM) with an evaluator to make verifiable scientific discoveries.
▸It works by having the LLM generate computer code as potential solutions, which are then automatically checked and scored.
▸FunSearch made a genuine breakthrough by solving the "cap set problem," a long-standing challenge in mathematics, with findings published in the journal Nature.
▸This approach of using AI for scientific discovery has huge potential in fields like drug discovery, material science, and algorithm optimization.
▸The future of research is a collaborative model where AI tools like FunSearch augment human scientists, accelerating the pace of discovery.

Frequently Asked Questions

What is Google DeepMind's FunSearch?+

FunSearch (short for "searching in function space") is a system by Google DeepMind that uses a large language model to write code that solves complex problems. By pairing the creative code generation of an LLM with a rigorous evaluator, it can discover new, verifiable knowledge in mathematics and science, moving beyond the typical text-based tasks of AI.

Did FunSearch actually make a new scientific discovery?+

Yes. FunSearch was used to attack the "cap set problem," a difficult challenge in combinatorics. It successfully discovered the largest constructions of cap sets ever found, representing a genuine mathematical breakthrough that was subsequently published in the prestigious scientific journal Nature.

Does this mean AI will replace scientists?+

No. The goal is not replacement but augmentation. Systems like FunSearch are powerful tools that can handle massive computational loads and explore vast solution spaces that are intractable for humans. The ideal future is one where scientists direct these AI tools to accelerate research and focus on higher-level strategic thinking and interpretation.

How does FunSearch avoid AI 'hallucinations'?+

FunSearch avoids hallucinations through its core design. The LLM's creative but potentially flawed code output is immediately passed to a separate, deterministic evaluator. This evaluator runs the code and scores its performance based on objective, pre-defined rules. Only the highest-scoring, correct programs are kept, effectively filtering out any incorrect or non-functional 'hallucinations'.