Written by: Techub News Compilation

Introduction

In December 2024, during Nobel Prize award week held in Stockholm, Sweden, Demis Hassabis, co-founder and CEO of Google DeepMind, delivered his Nobel lecture as one of the recipients of the 2024 Nobel Prize in Chemistry. This lecture not only summarized his team's groundbreaking achievements in the field of protein structure prediction (AlphaFold) but also provided an in-depth retrospective and outlook on his decades-long journey in artificial intelligence research. Hassabis began by discussing his childhood obsession with chess and early computers, weaving together DeepMind's unique path of using games as a testing ground to ultimately tackle major scientific challenges, while boldly envisioning the future landscape where AI as a core tool reshapes biology, materials science, and even the entire scientific discovery process.

Summary

• Games as the "perfect testing ground" for AI: DeepMind initially demonstrated the powerful capabilities of “self-learning” AI through systems like AlphaGo in complex games such as Go, with the core idea being to allow AI to learn strategies through massive self-play instead of relying on hard-coded rules.
• Conquering the century-old challenge of "protein folding": Hassabis's team applied the "learning model guided search" methodology validated in game AI to biology, developing AlphaFold 2, which achieved atomic-level precision in computationally predicting protein 3D structures for the first time, deemed by the academic community as the “solution” to this problem.
• From static structures to dynamic interactions: The recently released AlphaFold 3 marks the transition from predicting static structures of individual proteins to simulating interactions between proteins, DNA, RNA, and ligands, opening doors for applications like drug discovery.
• Towards a "digital biology" and an AI-driven new scientific era: Hassabis proposed that AI could become the "perfect language" for describing extremely complex biological systems and predicted that we are entering an era of “digital biology,” where AI will accelerate the entire scientific process from basic research to drug discovery.
• Responsibilities of AI and the grand vision for AGI: He emphasized the need to develop and use AI, a dual-use technology, responsibly and safely, and believes that if safely guided, artificial general intelligence (AGI) could become the ultimate universal tool for humanity to understand the universe and its own position within it.

From the chessboard to the code of life: The unique journey of an AI pioneer

Demis Hassabis's journey in AI began with a chessboard. He started playing chess at age four and quickly entered competitive arenas, having represented the England youth team. This early experience had a profound impact on him. He recalls that at a serious age for chess, he became fascinated with “thinking itself” — how does the human brain generate plans, strategies, and creativity? This curiosity drove him to explore the nature of thought.

A pivotal turning point came at a chess training camp during his teenage years. There, he encountered early chess computers, such as the “Fidelity Chess Challenger,” which required manually pressing LED lights on the board to move pieces. Hassabis admitted that at that time, he was more fascinated not by using machines to improve his game but by the fact that “someone could make a lifeless piece of plastic play chess with me so well.” This curiosity led him to read related books, and at the age of 11, he wrote his first AI program using his Commodore Amiga 500 computer — a game that played “Othello.” When this program defeated his then 5-year-old brother, he felt a great shock: one could actually create something capable of performing a task independently of oneself. From that moment on, he formed an inseparable bond with computers and AI.

More than twenty years later, games remain at the core of Hassabis's AI methodology. In 2010, he co-founded DeepMind with the mission of building artificial general intelligence (AGI). They view games as the “perfect testing ground” for AI because vast amounts of data can easily be generated on computers (the systems can play against themselves), and objectives are clear (to win a game or achieve a high score), facilitating the measurement of algorithmic progress. The history of game AI is long, from pioneers like Alan Turing and Claude Shannon to later systems like IBM's Deep Blue, with Go being regarded as the pinnacle challenge in this field.

The complexity of Go is unimaginable, with possible positions reaching as high as 10 to the power of 170, exceeding the total number of atoms in the universe, making any brute-force exhaustive method unfeasible. In 2016, DeepMind's AlphaGo system made headlines by defeating world champion Lee Sedol of South Korea with a score of 4:1, marking a milestone in the history of AI. Hassabis emphasized that AlphaGo's victory was significant not just for the outcome but for showcasing unprecedented creativity. For instance, its famous “37th move” was an unprecedented move in top Go matches, yet it became a critical turning point in the game, as if AlphaGo had already foreseen the board position a hundred moves ahead.

How was this breakthrough achieved? The key lies in “self-learning.” AlphaGo and its more general successor, AlphaZero, were not directly programmed with solutions, but instead learned through millions of self-play games, gradually refining strategies from mistakes. In this process, the system constructed an effective Go neural network model. This model could efficiently guide the search process, allowing it to avoid traversing all possible branches in the game or any complex space. Hassabis explained using a tree diagram that in any Go position (node), the model evaluates which side is more likely to win and focuses on the most likely moves (represented by blue lines), ultimately outputting the best move path it has found (purple line) within a limited time. This methodology of "learning models to guide combinatorial search" became a technical cornerstone for later solving real-world scientific problems.

Cracking the "Holy Grail" of biology: How AlphaFold achieves atomic-level precision predictions

After conquering the pinnacle of game AI, Hassabis's team turned their focus to significant challenges in the real world, particularly in the scientific domain. They sought problems suitable for their AI methodology based on three main criteria: first, whether the problem can be framed as finding paths in a vast combinatorial search space; second, whether a clear optimization objective function can be defined; and third, whether there is sufficient data to train the neural network model or efficient simulators to generate synthetic data.

For Hassabis personally, the "protein folding problem," which had lingered in his mind since his undergraduate days at Cambridge University, was always at the top of the list. Proteins are the building blocks of life, and nearly all biological processes depend on them. The function of a protein is determined by its complex three-dimensional structure, which is encoded solely by its one-dimensional sequence of amino acids. The 1972 Nobel Prize winner Christian Anfinsen hypothesized that proteins theoretically adopt conformations that minimize free energy, hence the "protein folding problem" initiated a fifty-year quest for computational solutions.

Why is this problem so difficult? Experimentally determining protein structures can often take months or years. The Levinthal paradox indicates that a typical-sized protein could have up to 10 to the power of 300 possible conformations, a number so vast that even exhausting all time since the beginning of the universe would be insufficient. However, in nature, proteins can spontaneously fold in milliseconds. This gives researchers hope: there must be some efficient guiding mechanism.

DeepMind decided to tackle this challenge, partly due to the availability of data for learning and clear assessment benchmarks. About 170,000 protein structures accumulated from decades of experimental work are stored in the Protein Data Bank (PDB). Additionally, the CASP (Critical Assessment of Protein Structure Prediction) competition, held every two years since 1994 by Professor John Moult and others, provided a blind test golden standard. It is widely accepted in the academic community that for computational predictions to be useful to experimentalists, the average error must be less than 1 angstrom (atomic width), which equals atomic-level precision.

DeepMind's journey began in the 2018 CASP13 competition. Their debut AlphaFold 1 system significantly improved prediction accuracy and became the first to introduce machine learning as a core component in this field, yet it did not reach atomic-level precision. Subsequently, the team learned from their experiences and designed a whole new AlphaFold 2 system from scratch, which broke through the atomic precision barrier in the 2020 CASP14 competition, leading the competition organizers to announce that the problem "has basically been solved."

Hassabis noted that AlphaFold 2's success was not due to a “silver bullet.” It integrated more than a dozen innovations, forming a complex mixed system. One key was directly embedding evolutionary information and physical constraints into the architecture, combined with the learning component. He emphasized that the success of the AlphaFold team was largely due to its multidisciplinary composition — including professional biologists, chemists, and machine learning and engineering experts. It is an end-to-end system that outputs 3D structure predictions directly from amino acid sequences and iteratively optimizes through a "recycling" stage, progressively enhancing prediction accuracy.

After achieving this technical breakthrough, the team focused on maximizing its impact. AlphaFold is not only accurate but also exceedingly fast, capable of predicting the structures of all known sequences of proteins (about 214 million). DeepMind partnered with the European Molecular Biology Laboratory (EMBL-EBI) to open-source AlphaFold and established the AlphaFold database, freely providing all predicted structures to researchers worldwide. Before release, the team consulted over 30 biosafety and ethics experts to ensure that the benefits outweigh the risks. To date, more than 2 million researchers have utilized AlphaFold, with related papers cited over 30,000 times, becoming one of the standard tools for biologists, accelerating scientific discovery in fields such as addressing plastic pollution, researching neglected diseases, and analyzing the nuclear pore complex structure.

AI for Science: Towards a new paradigm of digital biology and scientific discovery

In the latter part of his lecture, Hassabis expanded the perspective from the specific achievement of protein folding to explore the broader significance of such AI work and the future of AI in science.

He believes that from AlphaGo to AlphaFold, the essence of their systems has been to find optimal solutions in vast combinatorial search spaces. The method involves learning environmental models from data or simulations and then using that model to guide the search process according to optimization goals. This is a highly generalizable solution framework that can apply to many problems. For example, one could replace the nodes in the Go move search tree with chemical compound designs, transforming the search tree into finding the best molecules or drug compounds in chemical space. The same techniques can be entirely applicable to such problems.

This perhaps means we are entering an era of “digital biology” as Hassabis describes it. He has long believed that biology can fundamentally be viewed as an information processing system, albeit one that is extraordinarily complex and emergent. Mathematics serves as the perfect language for describing physical phenomena, while AI may similarly become the perfect language for describing biology. AlphaFold may be a starting point; looking back in 10 or 20 years, it could be seen as the hallmark that opened this era of digital biology.

To promote this vision, DeepMind incubated a new company called Isomorphic Labs a few years ago, aimed at reimagining the drug discovery process from first principles based on AlphaFold, with the goal of shortening this typically multi-year, costly process to mere months or even weeks. Hassabis likened this acceleration to "doing science at digital speed." He has a long-term dream: perhaps one day to simulate an entire virtual cell, rather than just the interaction of a single protein or a handful of proteins.

At the conclusion of his Nobel lecture, Hassabis proposed a more forward-looking and even slightly provocative idea. Since developing AlphaGo, he has been contemplating the limits of classical computing systems. In the current debate over the superiority of quantum versus classical computing, he posits that the capabilities of classical Turing machines (the foundation of modern computing) may far exceed what we previously imagined. They can build good models of problems or environments through extensive “pre-computation," enabling efficient exploration of solution spaces within polynomial time. He thus puts forward a preliminary hypothesis: any patterns or structures that can be generated or discovered in nature could potentially be efficiently discovered and modeled by classical learning algorithms. This doesn't mean that classical algorithms can solve all problems (for instance, abstract problems like integer factorization), but natural systems like proteins, or perhaps even materials, likely possess intrinsic structures that can be learned through such processes. If classical systems can, therefore, simulate certain types of quantum systems, this could have significant implications for fields like complexity theory and information theory in fundamental physics.

In closing, Hassabis showcased DeepMind's extensive application of AI technologies across various scientific domains, including medical image analysis, interpreting genetic variants, controlling nuclear fusion plasmas, discovering faster matrix multiplication algorithms, developing cutting-edge weather prediction systems, and identifying thousands of new materials. He believes that with these powerful new AI tools, we may be entering a new golden age of scientific discovery.

In concluding his lecture, Hassabis reiterated his lifelong belief in AI’s potential to help tackle humanity's most significant challenges. However, he stressed that AI is a dual-use technology that must be developed and used responsibly and safely for the benefit of all. This requires close collaboration among scientists, technology experts, and a wide array of stakeholders including government, academia, and civil society. He believes that transformative technologies like AGI are as significant as the inventions of fire or electricity, necessitating extreme caution and foresight to guide their development. If this technology can be safely harnessed, AGI may ultimately become the ultimate universal tool for us to understand the universe and our place within it.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。