Catalyzing breakthroughs in science
By proving it can navigate the immense search space of the Go board, AlphaGo has demonstrated the potential of artificial intelligence to lend a hand us better understand the enormous complexity of the physical world. We started by trying to solve the protein folding problem, which has been a major challenge for the last 50 years: predicting the 3D structure of proteins – information that is crucial to understanding diseases and developing modern drugs.
In 2020, we finally solved this long-standing scientific problem with our AlphaFold 2 system. From there, we assembled the structures of all 200 million proteins known to science and made them available to scientists in an open-source database. Currently, the tool is used by over 3 million researchers around the world AlphaFold Database accelerate their significant work on everything from malaria vaccines to plastic-eating enzymes. In 2024, it was the honor of a lifetime for John Jumper and me to receive the Nobel Prize in Chemistry for leading this project on behalf of the entire AlphaFold team.
Since AlphaGo’s victory, we have applied its groundbreaking approach to many other areas of science and mathematics, including:
Mathematical reasoning: The most direct descendant of the AlphaGo architecture, AlphaProof learned to prove formal mathematical theorems using a combination of language models and AlphaZero’s reinforcement learning and search algorithms. With AlphaGeometry 2, it was the first system to achieve the medal standard (silver) at the International Mathematical Olympiad (IMO), proving that AlphaGo methods can unlock advanced mathematical reasoning and lay the foundation for our most powerful general models.
Gemini, our largest and most productive model, has recently gone even further. An advanced version of Deep Think mode achieved gold medal status at IMO 2025 using an AlphaGo-inspired approach. Deep Think has been around ever since applied to even more convoluted, open challenges in science and engineering.
Algorithm discovery: Just like AlphaGo searched for the best move in the game, our AlphaEvolve coding agent explores the computer code space to discover more productive algorithms. He had his own moment where he found a novel way to do matrix multiplication, the basic mathematical operation that powers almost all newfangled neural networks. AlphaEvolve is currently being tested on a variety of problems, from data center optimization to quantum computing.
Scientific cooperation: We integrate the search and reasoning principles introduced in AlphaGo in AI Collaborator. By having agents “debate” scientific ideas and hypotheses, the system serves as a collaborator capable of the exacting thinking needed to identify patterns in data and solve convoluted problems. In validation studies at Imperial College Londonanalyzed decades of literature and independently arrived at the same antimicrobial resistance hypothesis that researchers had spent years developing and testing experimentally.
We’ve also used artificial intelligence to better understand the genome, accelerate fusion energy research, improve weather prediction, and more.
While our scientific models are impressive, they are highly specialized. To achieve fundamental breakthroughs like generating unlimited amounts of neat energy or solving diseases we don’t understand today, we need general AI systems that can find the underlying structure and connections between different topic areas and lend a hand us come up with modern hypotheses just like the best scientists do.
The future of intelligence
For AI to be truly general, it must understand the physical world. From the beginning, we built Gemini to be multimodal, so that it could understand not only language, but also audio, video, images and code to build a model of the world.
To think and reason in these modalities, the latest Gemini models employ some of the techniques we pioneered with AlphaGo and AlphaZero.
The next generation of artificial intelligence systems will also need to be able to employ specialized tools. For example, if the model needed to know the structure of a protein, it could employ AlphaFold to do so.
We believe that the combination of Gemini world models, AlphaGo search and planning techniques, and the employ of specialized AI tools will prove crucial to AGI.
True creativity is a key skill that such an AGI system would need to possess. Move 37 was a glimpse of AI’s out-of-the-box thinking potential, but a true original invention will require more than that. It would have to not only invent a novel Go strategy, as AlphaGo did so impressively, but also create a game as deep and elegant and worthy of study as Go.
Ten years after AlphaGo’s legendary victory, our ultimate goal is already on the horizon. The original spark that first emerged in the 37 Movement became a catalyst for groundbreaking discoveries that are now converging, paving the way for AGI and ushering in a modern golden age of scientific discovery.