The Man Behind The Most Useful Thing AI Has Ever Done

Demis Hassabis encountered the protein folding problem as an undergraduate at Cambridge in the late 1990s. A biologist friend described it as «the equivalent of Fermat's last theorem but for biology» — a grand challenge to predict the 3D structure of proteins from their amino acid sequences. Proteins make all of biology possible, and their 3D shapes determine their functions. But figuring out those structures had required shooting X-rays at crystals for years at a cost of hundreds of thousands of dollars per protein. Hassabis recognized it as the kind of problem AI might one day crack, and one that would unlock massive downstream benefits in drug discovery and disease research.

By 2021, AlphaFold had solved it. In a meeting captured on camera, Hassabis realized mid-discussion that the system could fold proteins in seconds — fast enough to process all 200 million proteins known to science in about a year. His team skipped the traditional approach of setting up a server where scientists request individual structures and instead folded everything and released it for free. That decision made the crucial biological data that had taken decades to accumulate suddenly available to every researcher on Earth. The impact was immediate: scientists working on wheat, malaria, and obscure organisms suddenly had access to structural data they could never have afforded to generate themselves.

Today, over 3 million scientists use AlphaFold, and a pharma executive told Hassabis that «almost every drug developed from now on will have probably used AlphaFold in its process.» The team keeps the database updated as new organisms are sequenced. Hassabis points to the nuclear pore complex — one of the body's largest and most important proteins — as his favorite example: within a year of AlphaFold's release, teams used it to finally determine the structure of this massive gateway protein that controls what enters and exits cell nuclei. It was fundamental biology, but the kind that opens doors to future treatments.

On March 10, 2016, AlphaGo faced world champion Lee Sedol in a match watched by 200 million people. In game two, it played Move 37 — a stone on the fifth line, early in the game, violating a fundamental rule Go masters teach beginners. The move was so unexpected that Sedol sat with his head in his hands. A hundred moves later, it became clear: that stone was in exactly the right place to win the game. It was not just surprising; it was the critical move, placed with what seemed like preternatural foresight. The move has since changed how professional Go players approach the game.

For Hassabis, Move 37 was the moment he had been waiting for. Unlike Deep Blue, which executed rules programmed by chess grandmasters, AlphaGo learned by playing against itself. It started with human games from the internet, then used Monte Carlo tree search to explore beyond human knowledge. The intelligence wasn't hardcoded by programmers — it emerged from the learning process. That distinction mattered enormously: it meant the system could generalize, discover, create. It was the signal that AI was ready to tackle scientific problems, not just games.

AlphaGo evolved into AlphaZero, which removed all human knowledge from the training process. Starting completely from scratch with only the rules of the game, AlphaZero played 100,000 games against itself, trained a new version on that data, then repeated the cycle. After 16–17 generations — spanning a single day — it went from random play to better than world champion level. Hassabis watched it happen live with chess: random in the morning, better than grandmasters by teatime, better than the world champion by dinner. AlphaZero discovered novel chess strategies that even the best brute-force engines like Stockfish had never found.

Demis Hassabis founded DeepMind with a plan: build AGI carefully, scientifically, like CERN. Take decades if necessary. In the meantime, deploy narrow AI systems like AlphaFold to deliver tangible benefits to humanity — cures for cancer, new energy sources, solutions to climate change. He sold to Google specifically because they promised to let DeepMind focus on science, not products. For years, that's exactly what happened. DeepMind pursued moonshots while Google handled consumer AI.

Then ChatGPT launched. OpenAI scaled a transformer-based language model, released it as a research experiment, and watched it go unexpectedly viral. Hassabis admits even he didn't realize people would find it so useful despite its flaws — the hallucinations, the limitations. But they did. Google declared «code red.» Hassabis became head of all Google AI, including the consumer products he hadn't been focused on. The careful, contemplative approach he'd envisioned evaporated. The industry entered a «ferocious commercial pressure race,» compounded by U.S.–China geopolitical competition. Progress accelerated to lightning speed.

Hassabis sees trade-offs in both directions. The benefits: faster progress overall, millions of people getting access to cutting-edge AI within months of its creation, and society gradually normalizing to the technology rather than facing a sudden shock when AGI arrives. The stress-testing from millions of users also builds more robust systems. The costs: the philosophical contemplation is gone, the scientific method rushed, and the world is now locked into competitive dynamics that prioritize speed over safety. He's pragmatic about it — «we have to deal with the world as we find it» — but it's clear this isn't the timeline he would have chosen.

Demis Hassabis grew up reading Ian Banks' Culture series, which depicts a post-AGI civilization a thousand years in the future. He thinks some of it could happen in fifty years. The key is solving what he calls «root node problems» — challenges like AlphaFold that, once cracked, unlock entire branches of the knowledge tree. Fusion energy is one. Room-temperature superconductors at atmospheric pressure combined with optimal batteries is another. If AI solves the energy problem — creating essentially free, clean, renewable power — it removes the primary cost barrier to space exploration. Rocket fuel could be synthesized from seawater using fusion-powered desalination plants.

With energy solved, asteroid mining becomes economically viable. Mercury, conveniently made of the right materials and in the right place, could support Dyson spheres capturing solar energy at scale. Humanity would cure all major diseases, extending lifespans dramatically. We would live healthier, longer, and expand into the cosmos. Hassabis describes it as «bringing consciousness to the rest of the galaxy» — maximum human flourishing enabled by AI breakthroughs in the coming decades. He believes it's plausible, not fantasy.

But he's equally candid about the path dependencies. This optimistic future requires navigating the AGI transition safely, which means solving the guardrail problem for autonomous agents, preventing bad actors from weaponizing dual-use technologies, and maintaining international cooperation on safety research. The timeline is tight: the systems are already becoming more capable and autonomous. Whether we get the sci-fi utopia or something darker depends on decisions being made right now, in labs like his, by people balancing scientific ambition against commercial pressure and geopolitical rivalry.