Introduction of the Event:
The chapter introduces a lecture hosted by Oxford University’s Institute for Ethics in AI as part of the Obert C. Tanner Lectures on AI and Human Values.

AI Institute at Oxford:
The Institute brings together leading philosophers and experts from various fields with AI researchers, developers, and users. Professor John Tersoulis serves as the director of the Institute.

Ubiquity of AI in Modern Times:
AI has become prevalent in various aspects of our daily lives, including smartphones, games, cars, drug discovery, search engines, and translation tools.

Recent Advancements in AI:
Modern machine learning algorithms, vast amounts of data, and powerful computing hardware have contributed to significant achievements in AI.

Demis Hassabis:
Demis Hassabis, CEO and co-founder of DeepMind, is the speaker for the evening’s lecture.

Intellectual Journey of Demis Hassabis:
Hassabis had a remarkable career journey, starting as a chess prodigy and becoming a successful computer game developer. He holds a double first in computer science from Cambridge. His fascination with the human brain led him to pursue a PhD in Cognitive Neuroscience and a Henry Wellcome Postdoctoral Research Fellowship. His research focused on imagination, memory, and amnesia, published in leading journals like Nature and Science.

Formation of DeepMind:
Hassabis combined his interests in computing and neuroscience to establish DeepMind in 2010. DeepMind’s ambitious goal is to solve intelligence and use it to address various global challenges.

Games as a Training Ground for AI Algorithms:
DeepMind initially used games as a context to develop and test new AI ideas using machine learning methods inspired by neuroscience. Games, like arcade games and notably Go, provided an ideal environment for training and evaluating AI algorithms.

Ultimate Aim of DeepMind:
Despite the success in games, the ultimate objective of DeepMind is to create general learning systems capable of tackling real-world problems.

DeepMind’s Vision:
To build AGI (Artificial General Intelligence) and use it to advance science and benefit humanity. Step one: solve intelligence; step two: use it to solve everything else.

Challenges of Traditional AI:
Logic systems and expert systems are limited in generalization and unexpected situations.

Learning Systems:
Can learn solutions from first principles and generalize to new tasks. Inspired by systems neuroscience and validated through comparisons to the brain. DeepMind’s approach is based on learning systems.

Deep Reinforcement Learning:
Fuses deep learning (model building) with reinforcement learning (goal-seeking and reward-maximizing). Enables discovery of new knowledge through trial and error using models. Agent system observes the environment, builds an internal model, selects an action, and iterates.

Reinforcement Learning:
Reinforcement learning is a type of active learning where an AI system interacts with its environment, makes decisions, and learns from the consequences of those decisions. The AI system’s decisions determine the experiences and data it receives, further shaping its learning. Reinforcement learning is inspired by the learning mechanisms found in mammalian brains, including humans.

Theoretical and Practical Complexities:
Despite the simplicity of its diagram, reinforcement learning poses significant theoretical and practical challenges that need to be solved.

Dopamine Neurons and General Artificial Intelligence:
Reinforcement learning was found to be implemented by dopamine neurons in the brain in the late 1990s. This suggests that reinforcement learning could be a path towards general artificial intelligence.

AlphaGo as a Landmark Achievement:
AlphaGo was a program that successfully beat the world champion at the game of Go, a significant milestone in AI research.

Reinterpreting AlphaGo:
The speaker reflects on AlphaGo in hindsight and offers a simpler and more general explanation for its success.

Introduction to Go:
Go is a complex and esoteric board game played in China, Japan, Korea, and other Asian countries. Go has resisted traditional logic system and expert system approaches due to the vast search space and difficulty in creating an evaluation function.

Intuition in Go:
Go players rely on intuition rather than explicit calculation to make decisions. Intuition is not something mysterious, but rather information learned through experience and stored in the association cortices. The conscious part of the brain cannot access these representations, making it difficult to explicitly code intuition in logic code.

AlphaGo and Intuition:
AlphaGo was a series of programs that attempted to approximate intuition in learning systems. AlphaGo Zero learned to play Go without human data and AlphaZero could play any two-player game.

Training Process:
Neural network is trained through self-play and starts with zero knowledge. Initial version (V1) makes random moves. V1 plays 100,000 games against itself, creating a data set.

Data Set Generation:
The data set is used to train a new version (V2) of the network. V2 is trained to predict the winner and likely moves from a given position.

Mini-Tournament Evaluation:
V1 and V2 compete in a mini-tournament (100 games). If V2 achieves a 55% win rate, it replaces V1.

Network Evolution and Improvement:
If V2 does not reach the 55% threshold, it continues generating data. The process repeats with V3, V4, and so on until the network consistently outperforms V1.

Neural Network and Classical Search Algorithm:
AlphaGo combines the neural network system with a Monte Carlo Tree Search algorithm. The neural network constrains the search space to likely and effective moves. It evaluates each node’s position with its evaluation function.

Evolutionary Speed and Observability:
The training process is incredibly fast, with noticeable improvements within a day. The network evolves from random to better than world champion level in 17-18 generations.

AlphaGo and AlphaZero:
AlphaGo and AlphaZero are computer programs developed by DeepMind that excel in playing board games. AlphaGo is known for its victory over Lee Sedol, a top-ranked Go player, in a million-dollar challenge match in 2016. AlphaZero was developed to generalize the techniques of AlphaGo to all two-player games, including Go, chess, and shogi.

AlphaGo’s Influence on Go:
AlphaGo’s unconventional strategy of playing a move on the fifth line from the edge of the board (move 37 in game two) revolutionized the way Go players think about the game. This move, which was previously considered unthinkable and discouraged by Go masters, proved to be a strategic decision that influenced the outcome of the game. AlphaGo’s innovative strategies have prompted Go players to adopt new approaches and techniques, leading to a broader understanding of the game’s dynamics.

AlphaZero’s Dominance in Two-Player Games:
AlphaZero’s success in two-player games such as chess, shogi, and Go demonstrates its ability to learn and master complex strategies with minimal training. AlphaZero surpassed the performance of the best handcrafted systems, including Stockfish in chess, within just four hours of training. This achievement highlights the potential of AlphaZero and similar AI systems to excel in a wide range of games and domains.

Generalization of AlphaGo’s Techniques:
The development of AlphaZero marked a significant step in generalizing the techniques used by AlphaGo to a broader range of games. This generalization allowed AlphaZero to achieve superhuman performance in multiple games, showcasing the versatility and adaptability of its approach. The success of AlphaZero opens up new avenues for research in game theory, artificial intelligence, and the development of more sophisticated AI systems.

New Heights in AI Gaming:
AlphaZero, an AI system developed by DeepMind, surpassed AlphaGo and AlphaZero in the games of Go and Japanese Chess Shogi.

AlphaZero’s Impact on Chess:
AlphaZero’s victory over the top chess program, Stockfish, sparked a discussion about the potential for further improvement in chess.

The Aesthetic Appeal of AlphaZero’s Style:
AlphaZero’s approach to chess is distinct from conventional chess programs, prioritizing mobility over materiality. This results in visually appealing and aesthetically pleasing games, with AlphaZero often sacrificing pieces to gain mobility.

Zugswang and the Immortal Zugswang Game:
AlphaZero demonstrated its prowess in a game against Stockfish, showcasing a unique position known as “Zugswang.” In Zugswang, any move made by one player worsens their position, highlighting AlphaZero’s ability to outmaneuver its opponents.

Implications for Chess and AI Research:
AlphaZero’s achievements underscore the potential for AI systems to revolutionize games like chess. It opens up new avenues for research in AI and machine learning, particularly in the pursuit of more creative and strategic approaches to complex games.

AlphaZero’s Innovations:
AlphaZero is a groundbreaking AI system that plays Chess, Go, and Shogi at a world-champion level. It doesn’t rely on hard-coded rules, allowing it to evaluate positions contextually and sacrifice pieces strategically. It learns to balance various factors automatically, overcoming the limitations of handcrafted rule-based systems. AlphaZero is more efficient than traditional search engines, evaluating tens of thousands of moves per decision, compared to millions for Stockfish.

AlphaZero’s Influence:
It inspired world champions like Magnus Carlsen to incorporate its ideas into their play. Garry Kasparov praised AlphaZero’s style as reflecting the truth. AlphaZero’s success in games laid the groundwork for general AI systems to solve real-world problems.

Criteria for AI in Scientific Discovery:
Massive combinatorial search spaces, where traditional methods fail. Clear objective functions for optimization and hill climbing. Availability of data or an accurate simulator for data generation.

Protein Folding as an Ideal Challenge:
Protein folding involves predicting the 3D structure of proteins from their amino acid sequence. Protein structure is crucial for understanding protein function. Experimental methods for protein structure determination are time-consuming and challenging. Christian Anfinsen’s conjecture in 1972 suggested that the amino acid sequence determines the 3D structure.

Background on the Protein Folding Problem:
Protein folding is a complex process that determines the three-dimensional structure of a protein. The specific part of protein folding that scientists are interested in is protein structure prediction. Exhaustively sampling all possible conformations of a protein is computationally intractable.

Challenges in Solving the Protein Folding Problem:
Leventhal’s calculation estimated approximately 10^300 possible conformations of an average-sized protein. The large number of possible conformations makes it challenging to find the correct structure using computational methods.

Personal Interest in Solving the Protein Folding Problem:
The speaker first encountered the protein folding problem as an undergraduate in Cambridge. The speaker was fascinated by the problem and its potential applications in biology.

The Game Foldit and Citizen Science:
Foldit is a citizen science game that turns protein folding into a puzzle game. Gamers were able to fold proteins and some even discovered new protein structures that were published in scientific journals.

Intuition and Counterintuitive Folds:
The speaker noticed that some gamers were able to make counterintuitive folds of the protein backbone, which led to the correct structure. This observation suggested that it might be possible to mimic the intuition of these gamers using AI.

AlphaFold Project and CASP Competition:
The AlphaFold project was initiated to work on the protein folding problem. CASP is a biennial protein folding competition that provides a blind prediction assessment. Experimentalists hold back their protein structures and provide them to the competition organizers, who then release them to the participants. Participants have a week to submit their predictions, which are then scored against the ground truth.

AlphaFold2’s Revolutionary Advances:
AlphaFold2 marked a breakthrough in protein structure prediction, revolutionizing the field by dramatically improving accuracy. It brought cutting-edge machine learning techniques, similar to those used in AlphaGo, as the core of the system.

Achieving Atomic Accuracy:
AlphaFold2 achieved atomic accuracy in CAS14, with a root mean squared error of 0.96 angstroms, surpassing experimental techniques. This remarkable precision made AlphaFold2 a reliable tool for structure prediction, competitive with and even surpassing experimental methods.

Key Technical Advancements:
AlphaFold2’s success was attributed to several critical innovations: It employed an end-to-end system, directly predicting the 3D structure from the amino acid sequence. It utilized an attention-based neural network to infer the implicit graph structure of the amino acid residues, removing biases from computer vision techniques. It required a large number of component algorithms, each essential for achieving the desired accuracy.

Significance of AlphaFold2:
AlphaFold2’s accomplishments led to the declaration that the structure prediction problem had been solved. Its predictions demonstrated exquisite accuracy, revealing the intricate beauty of proteins as nanomachines. The system’s success opened up new possibilities for research and drug discovery, enabling scientists to study and manipulate proteins with unprecedented precision.

AlphaFold2: A Breakthrough in Protein Structure Prediction:
AlphaFold2, a remarkable system developed by DeepMind, revolutionized protein structure prediction, outperforming its predecessor, AlphaFold1, in speed and accuracy.

Multidisciplinary Collaboration:
The development of AlphaFold2 involved a diverse team of experts, including biologists, physicists, chemists, and machine learners, highlighting the importance of cross-disciplinary work in AI for sciences.

Generalization vs. Specificity:
While AlphaFold1 and AlphaFold2 were initially designed for generality, the approach shifted towards specificity for solving real-world biological problems, prioritizing output over generality.

Efficient Predictions and Accessibility:
AlphaFold2’s training required only two weeks on a modest setup, and subsequent predictions could be made in minutes or seconds, enabling rapid and accessible protein structure predictions.

Folding the Human Proteome:
DeepMind undertook the ambitious task of predicting the structures of all proteins in the human body (human proteome), significantly expanding the coverage of protein structures beyond experimental methods.

Accuracy Levels and Unstructured Proteins:
AlphaFold2 achieved high accuracy in predicting protein structures, reaching 36% of the human proteome at very high accuracy (less than one angstrom error). Unstructured proteins, which lack a defined shape, posed a challenge for AlphaFold2, leading to the development of disorder predictors to identify these proteins.

Confidence Estimation:
AlphaFold2 introduced a confidence estimation mechanism, allowing biologists to evaluate the quality of predictions and identify regions that require experimental verification.

Expansion to Model Organisms:
AlphaFold2 was extended to cover 20 model organisms, including plants and disease-causing organisms, providing valuable structural information for researchers in various fields.

Data Accessibility and Collaboration:
DeepMind partnered with EMBL-EBI to host and distribute AlphaFold’s data, ensuring free and unrestricted access to researchers worldwide.

Safety and Ethics Considerations:
Before releasing AlphaFold, DeepMind consulted with experts to address safety and ethical concerns, emphasizing the importance of responsible AI practices.

Neglected Tropical Diseases Prioritization:
DeepMind prioritized protein structure predictions for neglected tropical diseases, aiming to aid drug discovery efforts for diseases affecting the developing world.

Early Applications and Success Stories:
Researchers have already utilized AlphaFold for various applications, including studying the nuclear pore complex and developing disorder predictors. AlphaFold has also been adopted in industry for drug discovery, demonstrating its practical utility.

Pioneering Digital Biology:
AlphaFold has revolutionized protein structure prediction, enabling the analysis of millions of proteins and advancing our understanding of biology. Digital biology utilizes AI and computational methods to process biological information, offering new insights and potential breakthroughs.

AI for Drug Discovery:
Isomorphic Labs, a spin-out company from DeepMind, focuses on accelerating drug discovery using AI and computational methods. Aiming to reduce the time from target identification to candidate drug development by an order of magnitude.

Extensive Applications of AI:
DeepMind collaborates with Google product teams to integrate AI technology into various products. Applications include data center energy optimization, text-to-speech systems, video compression, and recommendation systems. Cutting-edge large language models, such as AlphaCode, Chinchilla, Flamingo, and Gato, demonstrate diverse capabilities in programming, language understanding, vision, and robotics.

Responsible and Ethical AI Development:
DeepMind emphasizes responsible and safe AI development, considering the potential benefits and risks of AI technology. The company’s values of pioneering and responsibility guide its approach to AI research and deployment. Wide-ranging debates and engagement with institutes like the Institute of Ethics are crucial for shaping the responsible use of AI.

Diversity and Inclusion:
DeepMind prioritizes diversity and inclusion in its workforce and sponsors initiatives to increase representation in AI research and development. Efforts include scholarships, diversity chairs, academic institution partnerships, and funding conferences like the Deep Learning Indaba.

Ethics and Safety as Core Principles:
DeepMind’s mission emphasizes ethics and safety, recognizing the potential dual-use nature of AI technology. The company helped draft Google’s AI principles to guide the responsible development and deployment of AI systems. Thought leadership in strategy, risk, ethics, and safety topics is a key focus for DeepMind.

Rejecting “Move Fast and Break Things” Approach:
DeepMind advocates against the “move fast and break things” approach, particularly for powerful dual-use technologies like AI. Live A-B testing in the real world can have significant societal consequences, especially at large scales. The scientific method, with its emphasis on rigorous experimentation and hypothesis testing, is a more responsible and effective approach to AI development.

Scientific Method for AI Development:
Use the scientific method for developing AI, including careful thought, foresight, hypothesis generation, controlled testing, and empirical data analysis.

Importance of Control Testing:
Controlled testing is crucial for understanding the system and its behavior. Engineers may not initially grasp the significance of control tests, but scientists and researchers understand their importance.

Rigorous Testing Before Deployment:
Conduct controlled testing in controlled environments before deploying AI at scale. This approach allows for a better understanding of the system and potential issues.

Treating AI with Respect and Caution:
AI is a powerful technology that demands respect, precaution, and humbleness. We must treat AI with the seriousness it deserves, given its potential impact on humanity.

AI as a General-Purpose Tool for Scientific Understanding:
AI has the potential to be the ultimate general-purpose tool for scientific understanding. It can help scientists better comprehend the universe and our place in it.

Tension Between Tool and Colleague:
The concern arises that if AI outperforms humans in various cognitive tasks, can it still be considered a tool or would it become a colleague. The concept of respecting AI may require a different form of respect that challenges the original objective of treating it as a tool.

Determining Benefits to Humanity:
Determining what benefits humanity is a complex task. Different perspectives exist, such as whether military applications of AI benefit humanity.

Division of Labor in Decision-Making:
The responsibility of making decisions about AI’s benefits and applications should not solely rest on developers, researchers, and corporations. Governments should play a role in resolving these issues, considering the broader implications for society.

Comparing AI’s Abilities to Human Performance:
Measuring the capabilities of AI becomes challenging due to the vast range of potential tasks it might encounter. Comparing AI’s performance with humans across various tasks can provide an approximation of its capabilities. However, there remains the possibility of missing out on specific cognitive abilities, such as creativity.

AI’s Potential Contribution to Neuroscience:
AI can be applied to neuroscience to gain a better understanding of the human brain and its functions. This application of AI allows neuroscientists to explore and uncover the mysteries of the human mind. It provides a unique opportunity to study complex mental processes such as dreaming, creativity, and emotions.

Building AI to Understand Human Mysteries:
Creating AI and intelligent artifacts can help scientists identify what is missing in these systems compared to humans. This process aids in the scientific exploration of big questions such as free will and consciousness. Building AI systems with consciousness and free will raises ethical and philosophical questions.

Approaching AI Development with Caution:
Initially, AI systems should be developed as tools without consciousness or free will. Introducing these complex elements requires a cautious and gradual approach. Ethical considerations and multidisciplinary collaboration involving philosophers, ethicists, theologians, and humanities scholars are crucial.

The Ongoing Journey of Understanding Human Minds:
The development of AI presents an exciting journey of scientific discovery and exploration. Understanding the capabilities and limitations of AI can shed light on the complexities of the human mind. This endeavor requires a comprehensive approach that encompasses both scientific research and philosophical inquiry.

Dual Use Considerations:
DeepMind has an Institutional Review Committee (IRC) to assess research projects and consider potential dual-use risks. The IRC involves diverse experts, including ethicists, philosophers, legal professionals, and rotating senior researchers. The IRC works with research teams to evaluate projects, identify concerns, and suggest safeguards or modifications. Open sourcing alone is not seen as a solution, as dangerous systems could be exploited by bad actors. DeepMind acknowledges the complexity of ethical questions surrounding dual use and the need for ongoing learning and adaptation.

Lessons from Industry:
DeepMind’s organizational setup was designed to combine the best of startups (energy, creativity, pace, nimbleness) and academia (blue sky thinking, ambitious thinking). The acquisition by Google added scale and resources to the mix, enabling projects of significant impact. The key lesson is to have big ambitions, recognize the potential for scaling, and understand the consequences and responsibilities that come with it. DeepMind’s unique organizational culture and processes could serve as a blueprint for other grand projects.

Metrics for a DeepMind Social Network:
A DeepMind social network would likely focus on metrics related to user well-being, engagement, and positive impact. It would aim to avoid metrics that could lead to addictive or harmful behaviors or create echo chambers. The network would prioritize metrics that promote diverse and respectful interactions, accurate information, and constructive dialogue. DeepMind’s approach to metrics would be guided by ethical considerations and a commitment to minimizing negative consequences.

Social Networks and Moral Obligations:
The speaker expresses skepticism about the value of weak ties in social networks compared to deeper, more meaningful relationships. He emphasizes the need for careful consideration of metrics and outcomes before optimizing social networks. The scientific method should be employed to thoroughly analyze the consequences and potential risks of building social networks.

AI’s Role in Moral and Political Domains:
The speaker acknowledges the challenges of applying AI to moral and political issues due to the complexity of human motivations and behaviors. He suggests the possibility of using AI to create experimental test beds or simulations for political science and economics. This could facilitate low-stakes exploration of different political systems and market dynamics.

AI as an Emulation of the Scientific Method:
The speaker proposes the development of AI systems that emulate the scientific method. Such systems would be capable of generating hypotheses, conducting experiments, and refining their understanding over time. This could lead to advancements in science and other fields by automating and accelerating the research process.

Scientific Method as a Foundation for AI:
The speaker highlights the importance of the scientific method as a guiding principle for AI development. He emphasizes the need for rigorous experimentation, data collection, and hypothesis testing to ensure the accuracy and reliability of AI systems. By adhering to the scientific method, AI can be used to solve complex problems and contribute to human knowledge.

Scientific Method in Neural Networks:
Neural networks don’t follow the scientific method in the traditional sense, but they use its approach. Current systems are often black boxes, but efforts are being made to make them more transparent.

Neuroscience Techniques for Understanding AI:
The goal is to understand artificial minds as well as we understand the real brain. Access to every artificial neuron and complete control over experimental conditions provide unique opportunities for study. Techniques like fMRI and single-cell recording can be adapted for AI analysis.

Challenges in Studying AI Systems:
AI systems change over time, making long-term studies difficult. Researchers are now reaching a point where systems are interesting enough to warrant dedicated study.

Future of AI Understanding:
Over the next decade, there will be a significant increase in our understanding of AI systems. The goal is to make AI systems as transparent and understandable as the real brain.