Demis Hassabis, Father of Google's Artificial Intelligence

Part 7. AI for Science

19 Innovation in Mathematics and Algorithms

Kim Kyung-ran, Kim Kyung-jin

Problem Solving In July 2024, the world's most gifted young mathematicians gathered in the quaint English city of Bath for the 65th International Mathematical Olympiad (IMO). At that same moment, in a Google DeepMind laboratory led by Demis Hassabis, an invisible contestant received its exam paper. Its name was AlphaProof and AlphaGeometry 2.

Hassabis has always insisted that "mathematics is the language that describes nature." Like Go or chess, the rules are clear, yet the complexity approaches infinity and only one correct answer exists in this realm of truth. Large language models like ChatGPT proved adept at writing poetry and code, but they shrank before mathematics, and this is why.

Language models predict the next "plausible" word based on probability, but mathematics demands not 99.9% probability but 100% logical proof. Hassabis was convinced that bridging this gap was the decisive gateway to AGI. AlphaProof works like this: it translates a math problem into a formal language (a proof assistant called Lean), then the AI explores millions of proof paths to find a logically airtight solution. Just as AlphaGo reads moves in chess, AlphaProof reads the moves of a proof.

The results DeepMind released that day were impressive. The AI system solved four of the six problems perfectly, earning 28 points. That score corresponded to the silver medal cutline.

One geometry problem was proved in just 19 seconds. Only a few years earlier, AI could not properly solve even elementary-school-level word problems in math. Now it stood shoulder to shoulder with the top 0.1% of human prodigies. The core of this achievement lay in the "neuro-symbolic" approach Hassabis had long championed.

The DeepMind team combined the ability of a large language model (Gemini) to generate intuitive ideas with a symbolic system that verifies strict logical rules. It was like a mathematician with brilliant intuition proposing a hypothesis while a meticulous reviewer proves that hypothesis line by line. AlphaProof translated problems into the formal proof language Lean and taught itself.

This can be called the digital version of the process Hassabis went through as a chess prodigy, building intuition through tens of thousands of games and sharpening logic through post-game analysis. The 2024 silver medal was not merely a score. It was an event that proved AI could move beyond the stage of "imitation," mimicking human data, and arrive at "truth" through its own logical reasoning.

When Hassabis received this report, he sensed that AI was nearing completion as a tool for scientific discovery. AlphaEvolve: The Era When AI Designs Algorithms If solving math problems is a process of finding a given answer, designing algorithms is a creative act of paving a "new path" to that answer. In May 2025, DeepMind unveiled AlphaEvolve and delivered yet another shock to the world.

Hassabis had focused on a fundamental inefficiency in computer science. The basic algorithms humanity has used for the past 50 years, sorting, hashing, and the like, were mostly devised by brilliant human programmers in the 1960s and 70s. Since then, we concentrated on building faster computers but barely thought about improving those foundational algorithms themselves.

Hassabis asked: "What if AI wrote code from scratch, free of human biases?" The precursor came in 2023 with AlphaDev.

AlphaDev used reinforcement learning to improve the sorting algorithm in the C++ standard library for the first time in decades. But AlphaEvolve went a step further. This system combined Gemini's code-generation capability with evolutionary computation: it wrote code on its own (mutation), evaluated performance (selection), and rewrote better code (evolution) in an endless loop.

AlphaEvolve's power was demonstrated in the heart of Google: its data centers. Job scheduling across Google's data centers, where millions of servers run worldwide, is an enormously complex problem. AlphaEvolve found scheduling algorithms far more efficient than those designed by human engineers, and through them it recovered 0.7% of Google's total computing resources.

The number 0.7% may look small, but at Google's scale it represents hundreds of billions of won in cost savings and a massive reduction in carbon emissions. What was even more fascinating was the form of the algorithms AlphaEvolve produced.

Sequences of instructions that no human programmer would ever write, arrangements that defy intuitive understanding, kept appearing. Like AlphaGo's famous Move 37, the AI had begun writing code in its own "dialect," one unconstrained by human thinking patterns and optimized for computer hardware. For Hassabis, AlphaEvolve was powerful evidence of the mission to "solve intelligence, then use that to solve everything else."

AI had become not merely an assistant helping humans but a research partner evolving computer science itself. He remarked: "We are opening the era of not only digital biology but digital computer science." A Revolution in Materials Science: Discovering Over 2.2 Million New Crystal Structures In late 2023, a single paper published in the journal Nature turned the global materials science community upside down.

The news was that DeepMind's GNoME project had discovered 2.2 million new crystal structures. To grasp what this means, you have to look back through human history. We have always defined our eras by the materials we use.

The Stone Age, the Bronze Age, the Iron Age, and the current Silicon Age. The discovery of new materials has always brought quantum leaps in civilization. But the discovery process was an agonizingly slow and painful cycle of trial and error. Just as Edison burned through thousands of materials to find a lightbulb filament, modern scientists had to mix and bake elements in laboratories, waiting for serendipitous discoveries. In all of recorded history, humanity had identified the crystal structures of only about 48,000 stable inorganic compounds through experiments.

Hassabis wanted to accelerate that slow clock. He applied the same success formula from AlphaFold's protein structure predictions to materials science. "If we train graph networks on the bonding rules of atoms, could we predict stable structures in advance?"

GNoME used deep learning to explore combinations across the periodic table. It produced 2.2 million new candidate materials in one sweep, 45 times more than humanity had found over 20,000 years. Of those, 380,000 were identified as "stable" structures that could be synthesized with current technology.

This represented roughly 800 years' worth of materials science research. When Hassabis announced the results, he said in his typically composed voice:

"We are fundamentally expanding the search space of knowledge." True to his words, GNoME had handed scientists a treasure map. Instead of groping in the dark, scientists could now pick a destination on the map GNoME had drawn and set out on their expedition.

Hidden within GNoME's treasure map were answers humanity had desperately sought. The field Hassabis focused on most was clean energy technology to address the climate crisis. The database contained lithium-ion conductors that could dramatically improve battery efficiency (the core of electric vehicles), new photovoltaic materials that could maximize solar panel efficiency, and candidate substances for room-temperature superconductors, the so-called dream material.

DeepMind partnered with Lawrence Berkeley National Laboratory (LBNL) in the United States to operate an autonomous laboratory called A-Lab. In this futuristic scene, robotic arms executed recipes designed by AI to synthesize new materials, proving that GNoME's predictions were not mere computer simulations. A-Lab synthesized 41 new materials in just 17 days.

That work would have taken human researchers years. Hassabis emphasized "radical abundance" once again. By discovering materials that make energy cheaper and cleaner and push computing chip performance beyond its limits, he believed humanity could resolve conflicts born of resource scarcity.

"One of the 380,000 materials we discovered could become the standard for next-generation batteries, or it could become the material lining the walls of a fusion reactor." As of 2025, materials scientists around the world are conducting experiments based on the data GNoME made public. Hassabis hoped that just as AlphaFold became a tool for biologists, GNoME would become an indispensable tool for materials scientists.

He considered it a greater honor for a scientist using his AI to make a world-changing discovery than for the AI itself to win a Nobel Prize. The boy who once read moves on a chessboard was now reading the arrangement of atoms, expanding the physical limits of humanity.

An image symbolizing the fusion of AI and mathematics