What is Knowledge in the Age of AI: AGI and the Future of Academia and Society as Envisioned by the CEO of Google DeepMind

Throughout history, humanity has acquired knowledge and understood the world through various methods such as philosophy, science, and art. However, the rapid progress of artificial intelligence (AI) in recent years is poised to fundamentally change the very nature of knowledge. In this article, based on the lecture "The Future of Knowledge" delivered at the Institute for Advanced Study (IAS) in Princeton by Sir Demis Hassabis, CEO of Google DeepMind and 2024 Nobel Laureate in Chemistry, we will explain the possibilities and challenges brought about by AI in an accessible manner. By touching upon technical content while incorporating specific examples and quotes, we hope this will serve as material for reflecting on "what knowledge is in the age of AI."

1. The Historical Context of IAS and AI

1-1. Perspectives on Computing at the IAS

The Institute for Advanced Study (IAS) is a research institution that has been exploring the "frontiers of knowledge" since the 1940s. As moderator David Nermberg mentioned at the beginning of the lecture, J. Robert Oppenheimer, John von Neumann, and others were active at the IAS and established the modern computer structure (the von Neumann architecture).

Alongside the atomic bomb, they referred to the influence of computers as a "monster large enough to change history." Von Neumann's wife, Klari, said, "Humanity must catch up with what it has created," and an anecdote is told of him being so excited that he could not sleep, symbolizing that anxiety.

Oppenheimer stated that "national security depends not only on technology but also on ethics, social organization, and emotion," and aimed to "maintain humanity" by making the IAS an interdisciplinary space. It is precisely because of this background that the IAS became a historical holy ground not only for art and physics but also for computer science and AI research.

1-2. The Intersection of AI and Human Knowledge

In the lecture, while noting that "AI is still a form of 'human discovery'," it was emphasized that, as its name suggests, it "holds the potential to destroy human categories."

Hassabis posed the question, "How deep is the power of computers, and how will they change human knowledge and humanity?" and presented a perspective for considering the history of the IAS and current AI research in connection with one another.

2. The Life of Demis Hassabis and the Path to AI

2-1. From Chess Prodigy to Gamer

Demis Hassabis began playing chess at the age of four and was skilled enough to represent the UK as a junior player. In the lecture, Hassabis said, "Zero-sum games (games where there is a clear winner and loser) are too simple. At 13, I felt I wanted to master intelligence more."

He had been interested in the "process of thinking itself" since childhood, and the key was early chess computers. Seeing a physical chess-board-shaped computer from the 80s up close, he was shocked that "this lump of plastic is playing chess just like a human," and from then on, his interest in programming sprouted.

2-2. Home Computers and Game Development

Around the age of 13, he used the prize money from a chess tournament to buy home computers like the ZX Spectrum and Commodore Amiga, and began programming himself. During his high school and university years, he was involved in the development of simulation games (such as Theme Park) and created game characters using AI.

Although the games of that time had "extremely low CPU performance and very primitive AI," he succeeded in creating different experiences for each player by implementing a mechanism where "character reactions change according to the player's actions." He says that it was here he became convinced that "if it can be scaled, AI will become an incredibly powerful tool."

2-3. Neuroscience and the Founding of DeepMind

In university and graduate school, he majored in neuroscience and computer science, and in his doctoral program, he researched the "neural basis of memory and imagination." He presented the hypothesis that memory is reconstructed "constructively" rather than as images, and that imagination uses the same mechanism, and he had been interested in the "comparison between artificial intelligence and human intelligence" since then.

In 2010, while at University College London (UCL), Hassabis founded DeepMind with his co-researchers. Their representative mission is the ambitious one of "solving the problem of intelligence and using that intelligence to solve all other difficulties." With almost no research funding, they initially launched as a company out of a desire to "pursue a sense of speed that cannot be learned in academia."

3. DeepMind's Major Achievements and Applications to Science

3-1. AlphaGo: The Impact of Go AI

• AlphaGo is a Go program that first defeated world champion Lee Sedol in 2016. This victory delivered the shock that "computers can reach the pinnacle of human strategic thinking."

• In his lecture, Hassabis explained, "Go has a history of thousands of years and is one of the most complex games ever developed by humans. Because it has excellent metrics (win/loss conditions) and can generate massive amounts of self-play data, it was a gateway for AI development." In fact, AlphaGo discovered new strategies that surpassed existing Go theory through a combination of convolutional neural networks, reinforcement learning, and Monte Carlo tree search.

3-2. AlphaFold: The Protein Structure Prediction Revolution

• AlphaFold is an AI system that solved a major challenge of over 50 years in protein 3D structure prediction. While it previously took one graduate student several years to analyze a single structure using experimental methods (such as X-ray crystallography or NMR), AlphaFold "learned from 150,000 existing structures and then predicted and filtered approximately 1 million additional structures to add to its training set."

• The lecture mentioned that "AlphaFold eventually reached a level where it could predict all protein structures in nature in just a few years. This is equivalent to over 100 million person-years of human effort," emphasizing the importance of utilizing vast experimental data funded by public research grants (such as the NSF and NIH). AlphaFold's success embodied DeepMind's initial goal of 'AI for science.'

3-3. Isomorphic Labs: Expansion into Drug Discovery

• Applying AlphaFold's technology further, Isomorphic Labs was established. It is an attempt to discover and optimize new drug candidates in a significantly shorter time and at a lower cost than traditional research by integrally performing the drug discovery process—protein structure analysis, drug candidate design, and toxicity prediction—using AI.

• Hassabis says, "Since AlphaFold created the foundational model, we just need to move to the next stage of 'how to design molecules and prevent off-target binding in the body.'" This is expected to lead to major breakthroughs in disease treatment and biotechnology in general.

4. The Philosophy of Problem Formulation and Application to Science

4-1. Selection Criteria for Problems Tackled by AI

The problems DeepMind focuses on share the following three commonalities:

1. The existence of large-scale, high-quality data

   • For AlphaGo, synthetic data from self-play was used, and for AlphaFold, the training set consisted of 150,000 structures obtained through experiments, plus curated data of 300,000 predicted structures.

   • Having a structured 'true dataset' allows for the alignment of simulation and real-world data distributions, making it possible to improve model accuracy.

2. The ability to define clear evaluation metrics

   • Problems where it is clear 'what should be optimized,' such as wins and losses in Go or atomic coordinate error in protein structure prediction, satisfy the conditions that make it easy for AI to learn toward a goal.

   • It is said that for many problems in nature (such as the minimization of free energy), if the 'function to be optimized' can be formulated, a path to solving them with AI will open up.

3. The combinatorial space is extremely vast and cannot be solved by brute force

   • It is said that there are on average 10^300 possible three-dimensional structures for proteins, and conventional computational methods cannot 'find the needle in the haystack.' AI uses 'models to approximate the space and intelligent search strategies (reinforcement learning and Monte Carlo tree search)' to efficiently find optimal solutions.

4-2. Fields of Application: Biology, Mathematics, and Physics

• Biology: Following the aforementioned AlphaFold, drug discovery pipelines based on structural prediction and simulation are being built for areas such as drug development, genome editing, and biomaterial design. Hassabis states, 'For example, AI can also be applied to physics fields such as searching for candidates for room-temperature superconductors, climate and weather forecasting, and nuclear fusion plasma control.'

• Mathematics: Through projects like 'AlphaProof,' DeepMind is attempting to formalize mathematical propositions and automatically construct proofs. Because mathematics possesses 'verifiable answers' and 'ease of synthetic data generation,' it is a field well-suited for AI learning and evaluation.

• Physics: In climate modeling and materials design (such as searching for novel catalysts), there is an expectation that AI will be used to rapidly solve 'free energy minimization problems derived from fundamental equations.' Hassabis emphasized the universality of this approach, stating, 'Interesting phenomena in the natural world all have huge combinatorial spaces and can be formulated as optimization problems.'

5. AGI (Artificial General Intelligence) and Multimodal Models

5-1. Modal vs. Multimodal vs. Generality

• In its early days, DeepMind focused on modal (domain-specific) problems such as 'Go' and 'protein structure.' However, the ultimate goal is to 'build a general-purpose intelligent agent (AGI) close to human intelligence.'

• The key is to create a world model that can understand and simulate multimodal information, including language, video, audio, and mathematical formulas. Hassabis aims to give machines a function equivalent to the 'mental simulations' that humans perform daily, and the 'neural basis of memory and imagination' that he studied during his PhD is being applied to current model design.

5-2. Development Status of Project Astra and Gemini

• As of 2024, DeepMind/Google is developing Gemini, a state-of-the-art general-purpose foundation model, and these models are designed to handle not only text but also video and audio. In a demonstration, when AI was asked to generate a '10-second video of a tomato being cut on a cutting board,' it reached a stage where it could realistically reproduce even the cut surface and the scattering of water droplets.

• Additionally, Project Astra is underway as a research project, with a view toward understanding and predicting 'everyday contexts and intuitive physical phenomena' for applications in robot control and general-purpose digital assistants. This makes the prospect of 'AI simulating the physical and social world and working in cooperation with humans to complete tasks' a reality.

6. AI Risks and Social Frameworks

6-1. Concerns Regarding Malicious Use and Agentification

As a 'dual-use' nature of AI, both 'benevolent scientific research and medical applications' and 'misuse by malicious states and individuals' always exist. Hassabis emphasized, 'We want to use it for public interest purposes such as medicine and climate forecasting, but at the same time, there are risks such as cyber weapons and fake videos.'

Furthermore, he mentioned that 'the next phase of AI is the rise of autonomous agents.' When AI that makes decisions and acts autonomously in the real world—rather than agents operating in Go or programmed environments—emerges, risks that humans cannot fully control may become apparent.

6-2. Ethics, Governance, and New Institutional Design

As Oppenheimer and others demonstrated through nuclear research, Hassabis stated that 'unless ethical and social systems are developed alongside technological innovation, humanity could face catastrophic consequences.'

• Specifically, he proposes the following organizations and systems:

  1. An international joint research institute specialized in AI (CERN-style)

     • A platform where researchers gather to openly research and verify AI as a fundamental science.

  2. Establishment of an independent AI agency (IAEA-style)

     • An agency independent of governments and corporations that monitors and regulates dangerous research and projects.

  3. Building a global governance council of experts (a technical UN)

     • A multi-stakeholder organization that formulates international rules and ethical standards to ensure the peaceful use of AI technology and prevent risks.

7. Hassabis's Approach to Problem Selection and Future Outlook

7-1. Hypothesis on the Similarity Between the Natural World and AI Learning

• In his Nobel lecture, Hassabis proposed the hypothesis that "patterns generated in the natural world can be efficiently discovered and modeled by appropriate learning algorithms." This suggests that "AI may be effective not only for life, but also for a wide range of natural phenomena, such as geological phenomena and the structure of the universe."

• The background to his reaching this hypothesis lies in the logic that "much of the natural world forms stable patterns through evolution and physical processes -> there is structure, not randomness -> if there is structure, it can be modeled through learning." For example, fields that can be approached with modern AI technology are expanding, such as searching for candidates for room-temperature superconductors and estimating environmental conditions for extraterrestrial life.

7-2. The Future Relationship Between the Human Brain and AI

In his lecture, he stated that "it is highly likely that the human brain also operates with a structure equivalent to a classical computer (Turing machine)." Regarding the "quantum consciousness theory" advocated by Roger Penrose and others, he maintains the position that there is currently no solid evidence for it, and that "the human brain can be digitally simulated."

Hassabis says, "The very fact that a monkey brain like ours can produce complex engineering achievements like a 747 passenger jet is proof in itself, suggesting the possibility that AI could replicate and transcend the human brain." This is also a message that "human intelligence is surprisingly flexible and general-purpose, and questioning its theoretical limits is the ultimate theme of AGI research."

8. Recommendations for Society and Academia

8-1. Division of Roles Between Industry and Academia

• Hassabis points out that "currently, many resources for AI research (especially computing and engineering power) are concentrated in corporations." On the other hand, he emphasized that academia is required to play the role of "using large-scale models generated by corporations to perform analysis, behavioral constraints, and evaluations of social impact."

• He recommends that academia should invest resources in areas such as "creating benchmarks to interpret and verify the behavior of AI models," "countermeasures against inaccurate information generated by AI (such as fake news)," and "research on the psychological and social impact of human-AI collaboration," rather than just "chasing after what companies are making," but instead "advancing research ahead of new challenges."

8-2. Contributions from Philosophy and Social Sciences

• It is essential to build "interdisciplinary AI governance" that incorporates insights from philosophy, ethics, and economics, rather than just pursuing technical aspects. A place where researchers with diverse backgrounds gather, such as the IAS, is best suited for objectively evaluating the social transformation brought about by AI and exploring ways to operate technology while protecting "human purposes and values."

• Hassabis stated, 'For example, by using neuroscience and psychological methods (such as fMRI and neural network analysis) to elucidate an AI's "ability to lie" or "ability to intentionally generate misinformation" during poker play, we can proactively grasp the risk of AI engaging in social manipulation at or above the human level.' This emphasizes the importance of an academic approach: 'not just fearing AI as a threat, but managing risks through deep understanding.'

Sir Demis Hassabis's lecture, 'The Future of Knowledge,' provided many insights, such as 'what impact AI will have on the formation of knowledge,' 'what problems should be chosen and how they should be addressed,' and 'how to suppress risks while returning technical achievements to society.' The points of this article are summarized below.

AI is not merely a technical advancement; it poses fundamental questions regarding how humans define, acquire, and share knowledge. I hope this article helps readers think about 'what knowledge is in the age of AI.'

Recommended Articles