Demis Hassabis arrived at the Google X campus carrying a banana. Out of a sense of decorum, I decided against making that photograph. Six weeks later, Google DeepMind released an image model called Nano Banana. I have regretted my restraint ever since.
It was July 2025. We settled into a dark lounge filled with chessboards and artifacts, a perfect place for the portraits. Demis spends much of his time in London, and I was glad to have this time with him. We talked about AI safety and the pressure to move faster.
He had helped establish how much these systems could accomplish. After studying computer science at Cambridge and cognitive neuroscience at University College London, he cofounded DeepMind in 2010. He wanted to understand how intelligence learns and build machines capable of doing more than the specific tasks their programmers had anticipated.
Games offered problems with clear rules and measurable results. DeepMind’s AlphaGo learned from human games and improved through play against itself, combining neural networks with search. In 2016, it defeated Go champion Lee Sedol by four games to one. Some of its moves surprised experts who had spent their lives studying the board. There were possibilities in the game that human players had scarcely considered.
The larger ambition was to apply learning systems to scientific questions. Proteins presented a formidable problem: knowing a chain’s amino acid sequence did not tell researchers how it would fold. Establishing that structure experimentally could take years. Across biology, sequences were accumulating far faster than structures could be determined.
With John Jumper and their colleagues, Demis helped develop AlphaFold2. In a major assessment in 2020, it predicted many protein structures with accuracy approaching experimental results. DeepMind and the European Bioinformatics Institute subsequently made predictions freely available through a database that, by 2022, contained models for more than 200 million proteins.
A researcher could now examine a proposed structure before deciding which experiment to undertake. Some regions of a model were more reliable than others, and predictions still needed testing. But laboratories everywhere had access to information that could change the questions they asked and how they pursued them. Demis and John shared half of the 2024 Nobel Prize in Chemistry for protein structure prediction; David Baker received the other half for protein design.
After our photographs in the lounge, we went upstairs to one of the humanoid robot laboratories. Robots were walking around, and Demis posed for a game of chess. Our arrival caused a stir. What had begun as a private photo session became something like a hundred Googlers with their Pixel phones out, filming Demis and me as we worked. Their admiration was unmistakable. I had come to photograph him and somehow become part of the spectacle.
Here was someone whose work had accelerated research across biology, concerned about the speed of his own field. Downstairs, Demis had told me that companies and nations would all be better off moving more slowly, allowing time to understand the risks. He also understood how much pressure there was to do the opposite.
A couple of weeks later, a package wrapped by hand arrived in the mail from Demis. Inside was a heavy metal cube, its faces etched with equations, and a well worn copy of David Deutsch’s The Fabric of Reality.
I often look at that cube. I wonder if there is a hidden code, or somewhere I might press, that would unlock its mystery. So far, it has kept its secrets.
