Author: Techub News Compilation

Introduction

At the beginning of 2025, Demis Hassabis, co-founder of the artificial intelligence research laboratory DeepMind and winner of the 2024 Nobel Prize in Chemistry, returned to his alma mater, the University of Cambridge's Computer Laboratory, to give a lecture titled "Accelerating Scientific Discovery with AI." As a core driver behind AlphaGo and the revolutionary protein structure prediction tool AlphaFold, Hassabis not only reviewed his personal journey from a gaming enthusiast to an AI pioneer in this talk, but also systematically elaborated on how DeepMind has successfully transferred the AI paradigm of tackling complex games to addressing significant scientific issues like protein folding, and envisioned the future of AI as a foundational tool in "digital biology" and the pathway to artificial general intelligence (AGI). This lecture provided a deep reflection on the last fifteen years of AI empowering the cutting edge of science and revealed the next core direction for DeepMind and the entire industry.

Summary

• Games as the "training ground" for AI: From Atari to Go, AI algorithms refined through self-play and reinforcement learning, at their core, possess the ability to efficiently search and solve vast combinatorial problems, laying the groundwork for tackling complex real-world issues.
• AlphaFold solved a "billion-year problem": Applying the gaming AI paradigm to the protein folding issue, AlphaFold 2 achieved atomic-level prediction accuracy, completing 200 million protein structure predictions equivalent to "a billion years of PhD work" within a year, and after being made freely available, it has fundamentally transformed structural biology research.
• AI is ushering in a new era of "digital biology": Hassabis believes that just as mathematics is the perfect descriptive language for physics, AI will become the perfect tool for understanding the complexities and emergent behaviors of biology, with his new company Isomorphic Labs aiming to reconstruct drug discovery from scratch based on this concept.
• The path to AGI: world models and general assistants: DeepMind's advancements in video generation (Veo), game world generation (Genie), and multimodal models (Gemini) aim to build a "world model" that understands the physical world and can plan actions, with the ultimate goal of developing a general AI assistant that can assist humans in the real world.
• Advancing transformative technology responsibly: Hassabis emphasized that in the face of exponentially advancing AI, the "move fast and break things" Silicon Valley approach must be abandoned in favor of the humility and caution of the scientific method, along with proactive safety research and extensive discussions with various sectors of society.

From the chessboard to the code of life: Gaming AI as the engine of scientific discovery

Demis Hassabis's journey in AI began with games, particularly chess. In his youth, he was fascinated not only by the game itself but also by the process of "thinking." Early home computers and chess programs made him realize for the first time that it was possible to teach "a pile of lifeless plastic" to engage in complex strategic games. This fascination drove him to study computer science at Cambridge, deeply influenced by the pioneering theories of Turing, Shannon, and others.

In 2010, when Hassabis co-founded DeepMind, he set forth an ambitious two-step mission: first, to solve the problem of intelligence; second, to use that intelligence to solve everything else. At that time, this sounded almost audacious. To achieve the first step, they chose games as the ideal "testing ground." Game environments have clear rules, explicit goals (scoring or winning), and can generate vast amounts of training data. In 2013, DeepMind's DQN system became the first to learn to play various Atari games directly from raw pixel input without any prior knowledge of game rules.

The real milestone came with the Go AI, AlphaGo. Hassabis pointed out that the possible board states in Go number up to 10 to the power of 170, far exceeding the total number of atoms in the universe, making it impossible to solve through brute-force search. The key to AlphaGo's success was its combination of deep neural networks and Monte Carlo tree search, and its ability to rapidly evolve through "self-play." The system began with random play, generating data through a massive number of self-matches to train a neural network that could evaluate board states and predict winning probabilities, ultimately iterating over this stronger version to replace the old one. Within 24 hours, the system could evolve from a novice to a level surpassing that of world champions. In 2016, AlphaGo defeated Lee Sedol, making a move (move 37) that was regarded as "unimaginable," showcasing AI's ability to discover new knowledge beyond millennia of human experience.

The subsequent AlphaZero further generalized this paradigm, mastering games like Go, chess, and shogi from scratch without human game records, relying solely on the rules. It not only defeated the strongest traditional chess engine, Stockfish, but also exhibited a beautiful style that emphasized maneuverability, dubbed "sacrificing pieces for positional advantage," influencing world-class players including Magnus Carlsen. Hassabis emphasized that these breakthroughs in games were not the ultimate goal but were meant to validate and refine a universal algorithmic framework applicable to broader fields.

AlphaFold: Applying gaming strategies to fold the "building blocks of life"

With the maturation of gaming AI technology, Hassabis immediately turned his attention to a scientific challenge he had long contemplated—the protein folding problem. Proteins are the building blocks of life, and their functions are largely determined by their three-dimensional structures. The "protein folding problem" involves predicting a protein's three-dimensional structure from its one-dimensional amino acid sequence. Due to the extraordinarily vast number of possible conformations (the "Levinthal's paradox"), progress on this problem had been sluggish for decades.

Hassabis believed that the protein folding problem perfectly aligned with the three criteria distilled from gaming AI: 1) an enormous combinatorial search space (too many possibilities to solve by brute-force); 2) a clear objective function (minimizing the deviation between predicted and actual structures); 3) a data-rich environment or one capable of generating data (existing protein structure databases).

In 2016, shortly after AlphaGo defeated Lee Sedol, DeepMind immediately launched the AlphaFold project. In 2018, AlphaFold 1 won its first crown in the Critical Assessment of Protein Structure Prediction (CASP) competition, being the first to introduce machine learning as a core method to the field, although it did not yet achieve atomic-level accuracy. The team learned from their experiences and redesigned the architecture, releasing AlphaFold 2 in 2020, which achieved atomic-level prediction accuracy comparable to experimental methods, with competition organizers declaring "the problem has been essentially solved."

AlphaFold 2 was not only accurate but also fast. Realizing this, DeepMind collaborated with the European Molecular Biology Laboratory to predict the structures of approximately 200 million known proteins within a year and made the database freely available to the world. Hassabis likened this achievement to completing "a billion years of PhD work" in just one year. This initiative significantly accelerated global biological research, from combating antibiotic resistance, studying neglected tropical diseases, designing enzymes to degrade plastics, to exploring reproductive mechanisms, making AlphaFold the standard tool in biomedical research, used by over 2 million researchers.

Since then, DeepMind has continued to iterate, releasing AlphaFold 3, capable of simulating protein interactions with other molecules (like DNA, RNA, and ligands), and the AlphaProtein system for "inverse folding," which designs entirely new proteins. Hassabis concluded that AlphaFold's success demonstrated that AI can transform previously intractable huge combinatorial search problems into solvable ones, with the key being enabling neural networks to learn the "topology" of the problem, efficiently guiding searches to find the optimal solution.

The era of digital biology and the all-science landscape empowered by AI

Based on AlphaFold's success, Hassabis proposed a grander vision: we are entering a new era of "digital biology." He believes that at its most fundamental level, biology is a system of information processing that resists the increase of surrounding entropy, and AI is the perfect language to describe and understand this extremely complex dynamic system with emergent behaviors, just as mathematics is to physics.

To put this idea into practice, DeepMind incubated Isomorphic Labs. The ambition of this company is to utilize AlphaFold and subsequent AI technologies to completely rethink the drug discovery process from scratch. Currently, the average development cycle for a new drug takes up to ten years and costs billions of dollars. Hassabis dreams of utilizing AI to shorten the drug discovery process from ten years to months or even weeks, achieving "digital speed science," much like how the discovery of protein structures was reduced from years to seconds. One of his long-term dreams is to create a "virtual cell" (for example, a yeast cell) that can simulate experiments on a computer, using predicted results to guide real-world wet lab verifications, thereby significantly reducing the costly and time-consuming trial-and-error search.

Hassabis pointed out that AI's empowerment in science extends far beyond biology. DeepMind's research has covered multiple domains: from detecting eye diseases through retinal scans, discovering new materials, aiding in nuclear fusion plasma control, and AI self-discovery of faster matrix multiplication algorithms, to improving weather forecasting and assisting quantum computing error correction. He encourages universities to actively promote interdisciplinary research connecting AI with various professional fields, believing that numerous breakthroughs will emerge from this in the next 5 to 10 years.

The path to AGI: world models, general assistants, and responsible development

Reflecting on DeepMind's original mission—to build AGI (artificial general intelligence), Hassabis shared the latest progress and thoughts. He believes that constructing a "world model" that understands the world is a key step. DeepMind's latest video generation model Veo can generate videos that conform to physical laws (like slicing tomatoes or bubbles forming when blueberries hit water) from text or a single image, indicating that AI systems can learn implicit understandings of real-world physics from vast amounts of video data.

Furthermore, the Genie 2 project can generate an interactive and controllable short game world based on textual instructions. Although currently the consistency lasts only a few seconds, this research aims to ultimately construct a model that can maintain consistency for extended periods, genuinely understanding the principles of how the world operates. These efforts converge with DeepMind's work on multimodal large models like Gemini.

Hassabis specifically mentioned his vision for the next generation of AI assistants—the "general assistant" (Project Astra), which may manifest as an intelligent agent on a phone or glasses, accompanying users to provide assistance in the real world, enhancing productivity and life experience. The next step in achieving this is integrating the searching and planning capabilities of agents like AlphaGo with general models like Gemini that understand the world, thus creating AI systems capable of planning and executing tasks in complex real-world environments, which is also the key to achieving significant advancements in robotics.

In the face of exponentially advancing AI technology, Hassabis solemnly emphasized the importance of responsible development. He recounted how DeepMind has been planning "scenarios after success" since its inception in 2010, considering how to ensure the safe and responsible deployment of transformative technology. He introduced technologies like SynthID for invisibly watermarking AI-generated content to tackle challenges posed by deepfakes. He believes that the "move fast and break things" creed of Silicon Valley does not apply to transformative technologies like AI, and should instead be replaced with humility, caution, and foresight as practiced in scientific methods. He praised initiatives like Britain's Bletchley Park's inaugural global AI safety summit, believing that it is crucial to involve governments, academia, industries, and civil society in discussions on how to maximize AI opportunities while minimizing its risks.

The ultimate conjecture: How far can classical computation go?

At the end of the talk, Hassabis posed a profound conjecture, tracing back to the foundations of computational theory he studied at Cambridge. The Turing machine defines the theoretical boundaries of classical computation. Today, quantum computing is highly anticipated, believed to be capable of solving many problems that classical computation finds difficult to handle.

However, Hassabis proposed an extraordinary conjecture: the capabilities of classical Turing machines (and the AI systems built upon them) may far exceed our past understandings. He cited AlphaFold as an example: proteins are essentially quantum systems, operating at atomic scales, leading one to believe that quantum simulation would be necessary to predict their structures. Yet AlphaFold achieved highly accurate approximations using classical neural networks. He conjectured that any patterns that can be generated or discovered in nature (i.e., patterns with real physical structures) could potentially be effectively discovered and modeled by classical learning algorithms like AlphaFold.

If this conjecture holds, it would have profound implications for quantum mechanics and even fundamental physics. Ultimately, this brings Hassabis back to the starting point of his AI journey: he has always believed that AGI constructed in this way will 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。