Accelerate scientific breakthroughs
By proving that it can navigate the vast search space of a Go board, AlphaGo has demonstrated the potential of AI to help us better understand the vast complexity of the physical world. We started by solving the protein folding problem, a grand 50-year challenge of predicting the 3D structure of proteins, information critical to understanding disease and developing new drugs.
In 2020, we finally solved this long-standing scientific problem with the AlphaFold 2 system. From there, they collapsed the structures of all 200 million proteins known to science and made them freely available to scientists in an open-source database. Today, more than 3 million researchers around the world are using the AlphaFold database to accelerate important research on everything from malaria vaccines to plastic-eating enzymes. And in 2024, on behalf of the entire AlphaFold team, it was the honor of a lifetime for John Jumper, who led this project, and I to win the Nobel Prize in Chemistry.
Since AlphaGo’s victory, we have applied its groundbreaking approach to many other areas of science and mathematics, including:
Mathematical Reasoning: AlphaProof, the most direct descendant of the AlphaGo architecture, has learned how to prove formal mathematical statements using a combination of language models and AlphaZero’s reinforcement learning and search algorithms. Along with AlphaGeometry 2, it became the first system to achieve medal standard (silver) at the International Mathematics Olympiad (IMO), proving that AlphaGo’s methods can unlock advanced mathematical reasoning and lay the foundation for the most capable general models.
Gemini, our largest and most capable model, has recently been further improved. An advanced version of Deep Think mode used an approach inspired by AlphaGo to achieve gold medal level performance in 2025 IMO. Since then, Deep Think has been applied to more complex and open-ended challenges across science and engineering.
Algorithm Discovery: Just as AlphaGo searches for the best move in the game, our coding agency AlphaEvolve explores the space of computer code to discover more efficient algorithms. The product had its own Move 37 moment, discovering a new way to multiply matrices, a fundamental mathematical operation that powers nearly all modern neural networks. AlphaEvolve is currently being tested on a variety of problems ranging from data center optimization to quantum computing.
Scientific collaboration: We are integrating the exploration and inference principles pioneered at AlphaGo into AI Collaborative Scientists. By having agents “discuss” scientific ideas and hypotheses, the system acts as a collaborator that can identify patterns in data and perform the rigorous thinking needed to solve sophisticated problems. A validation study at Imperial College London analyzed decades of literature and independently arrived at the same hypothesis on antimicrobial resistance that researchers have spent years developing and testing experimentally.
We have also used AI to better understand the genome, advance fusion energy research, and improve weather forecasting.
Our scientific models are impressive, but highly specialized. Achieving fundamental breakthroughs, such as creating unlimited clean energy or solving currently ununderstood diseases, requires general-purpose AI systems that can help you discover underlying structures and connections between different subject areas and come up with new hypotheses like a good scientist.
The future of intelligence
For AI to be truly versatile, it must understand the physical world. We built Gemini to be multimodal from the beginning, allowing it to understand not only language but also audio, video, images, and code to build models of the world.
To think and reason across these modalities, the latest Gemini models use several techniques pioneered in AlphaGo and AlphaZero.
Next-generation AI systems will also need to be able to invoke specialized tools. For example, if your model needs to know the structure of a protein, you can use AlphaFold for that.
We believe that the combination of Gemini’s world model, AlphaGo’s search and planning technology, and the use of specialized AI tools will prove important for AGI.
True creativity is a key capability that such AGI systems must exhibit. Move 37 was a glimpse into the potential of AI to think outside the box, but truly original inventions will require something more. We will not only need to invent novel Go strategies, as AlphaGo has so successfully done, but we will also need to invent games that are actually as deep and elegant as Go, and worthy of study.
Ten years after AlphaGo’s legendary victory, our ultimate goal is within reach. The creative sparks first seen in Move 37 led to breakthroughs that are now coming together to pave the way for AGI and usher in a new golden age of scientific discovery.
