Exclusive DeepMind Founder Interview: From "Behind the Curve" Mockery to Nobel Prize Win – Demis Hassabis Reviews Google’s AI Strategy - 36Kr

Over the past few years, large language models have almost become synonymous with AI. From ChatGPT to Gemini developed by Google DeepMind, from Claude built by Anthropic to China's DeepSeek, most public discussions have centered on chatbots, reasoning capabilities, and generated content. But if you ask Demis Hassabis, CEO of Google DeepMind and 2024 Nobel laureate in Chemistry, what a truly world-changing AI should look like, his answer has never changed.

In a recent interview with tech journalist Cleo Abram, he emphasized what he believes is AI's most essential value: it is not about generating more content, but about helping humanity understand the world, accelerating scientific discoveries, and ultimately solving critical issues such as cancer, new materials, and new energy sources.

He entered the AI field over 30 years ago with this original aspiration. He has long been fascinated by scientific questions such as "what is consciousness" and "how does the world work", and in his view, AI is the ultimate tool to help scientists process massive amounts of information and discover underlying patterns. Therefore, rather than developing an increasingly intelligent chatbot, he hopes AI can step into laboratories to help humanity solve problems that shape the course of civilization.

However, in the race for large language models, Google has indeed fallen behind. That's why in the interview, he also mentioned the challenges he faces as Google's AI leader, as well as the tremendous value he sees in competitors like ChatGPT.

The "Mission" of the Nobel Prize-Winning Product:

Bringing AI into the Laboratory

AlphaGo first made the world realize that AI could beat Go champions, but it was AlphaFold that truly earned DeepMind the Nobel Prize: it solved a problem that had plagued the biology community for more than half a century — protein folding.

Hassabis first encountered this problem when he was an undergraduate at the University of Cambridge. Back then, many of his biology classmates were obsessed with a puzzle known as "Fermat's Last Theorem in biology": whether it was possible to accurately predict the final 3D folded structure of a protein solely from its 1D amino acid sequence.

This problem is so critical because proteins are involved in almost all life activities in the human body. A protein's actual function does not depend on its amino acid composition, but on its final 3D folded structure. Different structures lead to completely different functions. Once misfolding occurs, it can cause various diseases including cancer, Alzheimer's, and Parkinson's.

However, for decades, determining a protein's 3D structure could only rely on experimental methods such as X-ray crystallography and cryo-electron microscopy. Researchers often spent years and hundreds of thousands of dollars to resolve a single protein structure. Meanwhile, human gene sequencing technology advanced rapidly, and a huge number of new protein sequences were discovered every year — the speed of experimental analysis could never keep up with the growth of data.

Hassabis realized very early on that this is exactly the type of problem AI is best at solving. Protein folding is essentially a massive pattern recognition problem, and AI is the most effective tool for discovering complex patterns. But in the late 1990s, AI was far from having today's capabilities, so he had to put this idea aside temporarily.

After DeepMind was founded, the opportunity finally arrived. AlphaFold could not only predict protein structures, but did so at an astonishing speed.

But Hassabis had an ambitious vision. He overrode differing opinions not just to provide scientists with a problem-solving method, but to solve the problem entirely. As he mentioned in his autobiography, "Achieving great feats requires a certain degree of impracticality."

In his conversation with tech journalist Cleo Abram, Hassabis recalled an interesting detail: following industry conventions, the team originally planned to build an online server for scientists around the world to submit protein sequences, which the system would then compute one by one and return the results. This had been the standard model in the protein prediction field for over 40 years, as a single prediction often took several days.

But during an internal meeting, he suddenly picked up his phone and did the math: if AlphaFold could complete one protein prediction every ten seconds, how long would it take to calculate all the roughly 200 million known natural proteins on Earth?

The answer turned out to be only about a year.

He decided there was no need to wait for scientists to submit applications one by one. Instead, the team would directly compute all known proteins worldwide and make the results freely accessible.

This seemingly crazy decision quickly became the team's goal. DeepMind partnered with the European Bioinformatics Institute to build the AlphaFold Database, making prediction results freely available globally. The database has since expanded continuously, now covering almost all protein structures known to the scientific community, and is constantly updated with new gene sequencing results. For any researcher, simply entering a protein name instantly retrieves data that would have taken years of experiments to obtain in the past.

Hassabis brought protein research into the "search era". More than 3 million researchers worldwide now use AlphaFold. A scientist at a pharmaceutical company once told Hassabis that he believes almost every new drug in the future will use AlphaFold during its R&D process.

In fact, AlphaFold has changed more than just medicine. Hassabis pointed out an easily overlooked field: plant science.

Compared to animals, many plants have far more complex genomes, but due to limited research funding, botanists often lack sufficient resources to resolve protein structures. Now with AlphaFold, botanists can finally devote more time to solving problems such as improving crop disease resistance, enhancing drought tolerance, and helping agriculture adapt to climate change.

Another group of beneficiaries of AlphaFold are researchers studying long-neglected diseases. Diseases such as malaria, Chagas disease, and leishmaniasis primarily affect developing countries, and due to limited commercial returns, large pharmaceutical companies have invested relatively little in them. After AlphaFold made relevant protein structures freely available, some non-profit research institutions with limited resources could directly enter the drug screening stage, drastically reducing upfront research costs.

Among all the achievements, Hassabis's favorite case is the nuclear pore complex. It is one of the largest protein complexes in the human body and the central "gatekeeper" for information exchange between the nucleus and cytoplasm. Due to its enormous size and complex structure, scientists had struggled for decades to fully resolve its 3D structure. Less than a year after AlphaFold was released, a research team combined its experimental predictions to successfully reconstruct the complete structure of this giant protein complex, providing a critical foundation for understanding the mechanism of cellular material transport.

Today, Hassabis speaks confidently to the media, but he also acknowledges that AlphaFold is only the first step in drug development.

Knowing what a protein looks like does not mean a drug already exists. The real challenge is to find a chemical molecule that can precisely bind to the target protein without interacting with the other 20,000+ proteins in the human body, minimizing toxic side effects as much as possible.

That's why Google's parent company Alphabet later founded Isomorphic Labs, aiming to advance AlphaFold from "predicting proteins" to "designing drugs".

The new AI system can not only predict protein structures, but also design candidate compounds, simulate their binding effects with target proteins, and simultaneously detect whether they will accidentally bind to other proteins and cause side effects — which, in Hassabis's view, is the most worthwhile direction for AI investment.

Reflections as Google's AI Leader:

Has Google Fallen Behind?

AlphaFold represents the ideal form of AI in Hassabis's mind, but the emergence of ChatGPT completely shifted his work priorities.

It is impossible for Google not to feel envious of ChatGPT's success and global influence.

Google actually wanted to build chatbots very early on. DeepMind and Google Research had long been researching similar large language models. In 2017, Google published the paper "Attention Is All You Need", proposing the famous Transformer architecture. Transformer is the technical foundation for all today's large language models, and systems like ChatGPT, Gemini, Claude, and DeepSeek are all built on the Transformer framework.

ChatGPT and Claude often compete fiercely with each other, but Google's Gemini has drawn little public attention. In this battle, Google has consistently fallen short, failing to launch an application-level large language model with influence comparable to ChatGPT.

Hassabis admitted in the interview that at the time, many labs including Google underestimated user acceptance of AI. The reason was simple: people on the front lines of R&D were more focused on the model's flaws.

Researchers worried that AI would confidently output nonsense, that logical reasoning would be unstable, and that factual accuracy could not be guaranteed. But ordinary users did not use AI that way. They used AI to summarize documents, organize meeting minutes, write emails, and learn knowledge. When the model made occasional mistakes, users were willing to tolerate them.

Hassabis confessed that in his original vision, he would have preferred AI to stay in the laboratory longer, developing gradually in a more cautious and scientific manner before being publicly released.

However, the explosive success of ChatGPT pushed the entire AI industry into a new competitive phase. As both a scientist and a manager of a large tech company, Hassabis has to advance fundamental research projects like AlphaFold and Gemini, while also facing competitive pressure from Google's consumer-grade AI products.

In Hassabis's view, this competition is not entirely negative. The upside is that AI is spreading at an unprecedented speed. Once upon a time, cutting-edge technologies in laboratories might take years for ordinary people to access; today, the gap between public-facing products like ChatGPT and Gemini and the latest lab models could be as short as a few months.

Hassabis believes that people can personally understand what AI is good at and what it is not, and society as a whole can gradually adapt to the coming AGI era, rather than suddenly facing a completely unfamiliar new world one day. Another huge benefit is that the real-world feedback generated by hundreds of millions of users every day has become an irreplaceable "stress test" that no laboratory can replicate, allowing development teams to truly identify remaining vulnerabilities in the model and continuously improve the system.

According to Hassabis, while commercial competition disrupted his originally research-focused development rhythm, it also accelerated AI's maturation.

What he is truly worried about, more than competition itself, are the long-term risks posed by AI.

For example, public discussions have focused on issues like deepfakes and disinformation generated by AI. DeepMind has already launched SynthID, which adds digital watermarks to AI-generated images and videos to help platforms, governments, and the public identify AI content. Hassabis has even called on all generative AI companies to build similar mechanisms to make AI content traceable.

Hassabis argues that the most important research topic going forward is not making AI smarter, but ensuring it always operates in line with human-defined goals, does not bypass safety restrictions, and does not exhibit unintended behaviors caused by incomplete goal design. This requires broader collaboration among leading labs, governments, and academia.

When asked what young people should focus on today, Hassabis does not advise everyone to train the next generation of foundation models. On the contrary, he believes that large models themselves are rapidly becoming a type of infrastructure.

Models such as Gemini, AlphaFold, AlphaGenome, and Veo are continuously unlocking new capabilities, and the real opportunity lies in how to combine these capabilities and apply them to new industries and new problems. Hassabis encourages young entrepreneurs to familiarize themselves with these tools as much as possible, then build new platforms or products on top of them.

During the interview, tech journalist Cleo Abram mentioned that Dr. Jennifer Doudna, the pioneer of CRISPR gene editing, asked her to convey a question to Hassabis. DeepMind's newly released AlphaGenome can read very long gene sequences and predict whether a mutation will occur at a single-letter position in the sequence. Once a mutation triggers a chain reaction, CRISPR can enter cells to repair it. Many products can work in tandem with DeepMind's product line.

He suggests that more young people do not need to develop models or engines from scratch, but instead stand on the shoulders of giants. For example, using tools like Gemini (Google's multimodal AI assistant), Veo (Google's AI video generation model), and Nano Banana (Google's AI image model) to build companies worth billions of dollars. Products like OpenClaw can also provide plenty of inspiration.