Speaker’s Introduction:
Fei-Fei Li, co-director of Stanford Institute of Human-Centered AI (HAI), introduced Demis Hassabis, the founder and CEO of Alphabet DeepMind. Hassabis has a diverse background, starting in video game development before earning a PhD in cognitive neuroscience. He is a pioneer in the field of AI and has made significant contributions through DeepMind, such as advances in energy consumption, protein folding, and deep learning.

DeepMind’s Achievements:
DeepMind has achieved breakthroughs in various areas, including data center energy optimization, degenerative eye conditions, and protein folding through AlphaFold. They have made substantial progress in deep learning and reinforcement learning, even beating world champions in the ancient game of Go. DeepMind is currently exploring the application of AlphaFold techniques to nuclear fusion to combat the climate crisis.

Artificial General Intelligence:
Hassabis and his team are exploring artificial general intelligence (AGI), an epoch-defining technology with the potential to transform human lives. AGI aims to create intelligent systems capable of tackling a wide range of tasks, not limited to specific domains.

Stanford HAI’s Mission:
HAI was established in 2019 with the mission to serve as a forum for intellectual discourse on AI and its societal impact. It aims to foster dialogue between researchers, industry leaders, and policymakers to address the ethical and societal implications of AI.

Hassabis’s Talk Focus:
Hassabis’s talk will focus on using AI to accelerate scientific discovery and explore the current trend of generative models and large language models. He will also discuss the work DeepMind is doing in these areas, inviting questions and discussions afterward.

DeepMind’s Founding and Early Challenges:
DeepMind was founded in 2010 when AI was relatively unknown and gaining investment was difficult. The company’s mission was to understand intelligence and create artificial systems that benefit humanity.

Deep Reinforcement Learning and Atari Games:
DeepMind made significant progress in deep reinforcement learning, a method for training AI systems to play games by rewarding successful actions. The company’s Atari DQN system was the first to learn how to play Atari games solely by observing the pixels on the screen. This achievement demonstrated the potential of end-to-end learning systems for complex perceptual tasks.

Solving Complex Games and AlphaGo:
DeepMind’s AlphaGo system defeated the world’s top Go players in 2016 and 2017, demonstrating AI’s ability to master complex strategic games. AlphaZero later surpassed AlphaGo’s performance by learning to play multiple games, including Go, chess, and shogi, from scratch.

Recent Accomplishments and Future Directions:
DeepMind has achieved further breakthroughs in areas such as protein folding, language understanding, and robotics. The company continues to explore new frontiers in AI, including safe and ethical AI development. DeepMind’s work has the potential to revolutionize various fields and benefit society in numerous ways.

Initial Challenges of AlphaGo:
The game of Go has an astronomical number of possible positions, making brute-force solutions impossible. Top Go masters cannot explicitly explain the rules and heuristics they follow, unlike chess players. Expert system chess computers cannot be used for Go due to the lack of explicit rules.

AlphaGo’s Breakthrough:
AlphaGo famously won a million-dollar challenge match against the world’s top Go player in 2016. AlphaGo’s creative strategies, such as move 37 in game two, surprised Go experts. AlphaGo’s ability to develop novel strategies that humans would not consider demonstrates its potential for creativity.

Three Levels of Creativity in AI:
Interpolation: Averaging or combining existing ideas to create something new. Extrapolation: Generating new ideas based on existing knowledge or data. Invention: Creating something completely new and original.

Current Limitations of AI Creativity:
AI systems can achieve interpolation and extrapolation but struggle with invention. The challenge lies in expressing abstract conceptual instructions to AI systems.

AlphaZero’s Self-Play System:
AlphaZero started with a random play and gradually improved through self-play and reinforcement learning. The self-play system involved generating data, training a new version of the AI, and facing it against the previous version. After 17 epochs, AlphaZero went from random play to surpassing the world champion in Go.

Tree Search with a Model:
AlphaGo and AlphaZero used a tree search algorithm to find optimal moves. A model of the game allowed for narrowing down the search space to reasonable options. This enabled finding strong moves in a limited amount of thinking time.

Generalization to Other AI Problems:
The self-play system and tree search with a model can be applied to other AI problems. AI can be tasked with modeling a problem and then searching for solutions within the model. This approach can lead to creative and innovative solutions that humans may not consider.

DeepMind’s Breakthroughs in Games:
DeepMind has achieved significant milestones in various games, including Atari, AlphaGo, AlphaZero, and AlphaStar. AlphaStar’s victory in StarCraft II showcased AI’s ability to handle partially observable environments and long-term planning.

Games as a Means to an End:
While DeepMind’s work in games has been enjoyable, it has always been a means to an end, not an end in itself. The goal was to develop algorithms that could be applied to real-world problems.

Shifting Focus to Real-World Problems:
Recent years have seen a shift towards tackling challenging problems in industry and scientific discovery. Google products now incorporate DeepMind’s technology.

AlphaFold: Predicting Protein Structures:
AlphaFold is a program that predicts the 3D structure of a protein from its amino acid sequence. Proteins are essential for life and perform various functions in the body. The function of a protein depends on its 3D structure.

Introduction to the Protein Structure Prediction Problem:
Protein structure prediction involves determining the 3D structure of a protein based on its amino acid sequence. Understanding protein structure is crucial for drug discovery and disease research. Experimentally determining protein structures is time-consuming and challenging.

The Computational Approach to Protein Structure Prediction:
Christian Anfinsen proposed in 1972 that the amino acid sequence should determine a protein’s 3D structure. Protein structure prediction has been a grand challenge in biology for over 50 years. Demis Hassabis became aware of this problem as an undergrad and recognized its potential suitability for AI.

The Complexity of the Problem:
Leventhal’s paradox highlights the vast number of possible shapes a protein can take. In contrast, nature spontaneously folds proteins in milliseconds. Understanding how physics enables this rapid folding is a key challenge.

Background:
Protein folding is a complex process crucial for understanding protein structure and function. CASP competition is a prestigious event that challenges computational systems to predict protein structures.

AlphaFold’s Initial Participation:
In 2018, DeepMind entered the CASP competition with AlphaFold1, which won by a significant margin. AlphaFold1 demonstrated the potential of cutting-edge ML techniques in protein folding prediction.

Challenges and Innovations:
To achieve the desired accuracy, AlphaFold needed to improve from 60 GDT to 90 GDT (less than 1 angstrom error). AlphaFold2 was developed with a different approach and architecture, incorporating physics and chemical constraints.

Success in CASP and Recognition:
In the 2020 CASP competition, AlphaFold2 achieved remarkable results. The competition organizers acknowledged that the protein folding problem had been solved.

Impact and Applications:
AlphaFold’s methods were published, and the code was open-sourced to maximize impact. AlphaFold was used to generate protein structures for the entire human proteome and 20 model organisms. This significantly expanded the availability of protein structure data, particularly for understudied organisms like plants.

Conclusion:
AlphaFold’s groundbreaking achievements have revolutionized the field of protein folding prediction. The open-source availability of AlphaFold has accelerated research and enabled diverse applications across various scientific disciplines.

Database Collaboration:
AlphaFold’s predictions are stored in Uniprot and PDB databases, which are commonly used by biologists. EMBL-EBI assisted in creating a new database to host AlphaFold predictions and link it to other genetics databases.

Safety and Ethics:
DeepMind consulted over 30 experts in biology, pharma, biosecurity, and human rights to assess the potential risks and benefits of AlphaFold. The consensus was that the benefits of AlphaFold outweigh any potential risks.

Applications of AlphaFold:
AlphaFold has been used for various purposes, such as designing plastic-eating enzymes, studying antibiotic resistance, determining the structure of the nuclear pore complex, improving crop sustainability, developing malaria vaccines, and creating a molecular syringe for drug delivery.

Usage and Recognition:
Over a million researchers worldwide have used AlphaFold and its predictions. The AlphaFold methods paper has been cited 10,000 times in 18 months. AlphaFold has received accolades from scientific journals like Science and Nature.

Science and Digital Speed:
Demis Hassabis introduces the concept of “science at technology speed,” where scientific progress is accelerated through digital tools and methods. AlphaFold is presented as an example of science at digital speed, capable of folding proteins in milliseconds instead of years. This digital speed allows for rapid dissemination of scientific solutions to a global audience, making them accessible to researchers and practitioners worldwide.

Digital Biology:
Hassabis proposes the emergence of a new era of “digital biology,” where biology is viewed as an information processing system. He suggests that AI, with its ability to handle complexity and emergent phenomena, is well-suited to describe and understand biological systems. AlphaFold is seen as a proof of concept for this approach, heralding a new era of digital biology. Isomorphic Labs, a sister company to DeepMind, is established to further explore these technologies in biochemistry and drug discovery.

General Solution for Combinatorial Optimization:
Hassabis identifies a common thread in the work done by their systems: finding optimal solutions in vast combinatorial spaces. This approach involves learning a model of the environment and using it to guide a search process towards an objective function. This general solution can be applied to a wide range of problems, from games like AlphaGo to scientific challenges like protein folding and drug discovery.

Overview of AI Applications:
AI systems like AlphaGo have demonstrated success in complex domains like playing Go, which require searching through vast spaces of possibilities. Researchers are exploring the application of AI in drug discovery by searching for optimal compounds with desired properties. AI has been used to solve important conjectures in pure mathematics, improve quantum chemistry calculations, and control plasma in fusion reactors. AI systems have also been applied to improve weather forecasting, reduce energy consumption in data centers, and develop more realistic text-to-speech systems.

Large Language Models:
Researchers have made significant progress in developing large language models (LLMs) like Chinchilla, AlphaCode, Domingo, and GATO. GATO, a general token-to-token translator, can perform various tasks, including natural language processing, image description, and robot control. AI researchers are actively studying the safety and limitations of LLMs, exploring techniques like RLHF, attention retrieval, and memory to improve their performance.

Current Understanding of AGI:
The researcher believes that scale is important in the development of AGI, but it might not be sufficient on its own. Grounding, factuality, planning, memory, and reasoning are identified as key areas where further innovations are needed for AGI.

AI’s Incredible Potential and Dual-Use Nature:
Demis Hassabis believes that AI has the potential to address humanity’s greatest challenges and advance science. However, AI technologies can also be dual-use, posing potential risks.

DeepMind’s Ethics Charter and Responsible AI:
From its inception, DeepMind has prioritized ethics and responsibility in AI development. The company’s ethics charter, now part of Google’s AI principles, emphasizes the safe and beneficial deployment of AI.

Balancing Accuracy, Safety, and Engaging Interactions:
Sparrow, DeepMind’s best chatbot, aims to be helpful, accurate, and grounded in evidence. A trade-off exists between accuracy and safety on one hand, and engaging interactions on the other.

Dual-Use Nature of AI and the Need for Responsible Deployment:
Hassabis emphasizes the importance of deploying and using AI responsibly and safely. He stresses the need to ensure AI benefits everyone and addresses potential risks.

Critical Moment in Human History and the Move Fast and Break Things Approach:
Hassabis believes we are approaching a crucial moment in human history with AGI. He advises against the “move fast and break things” approach, advocating for a more cautious and responsible approach to AI development.

AI’s Unintended Consequences:
Social media’s unforeseen negative impacts serve as a lesson for AI development. AI’s potential impact surpasses that of social media, demanding more cautious and responsible approaches.

The Scientific Method for AI:
Advocating for the application of the scientific method in AI development. Thoughtful deliberation, rigorous testing, and data-driven analysis should be integral parts of the process.

Understanding Systems before Deployment:
Empirical testing is crucial, but haphazard implementation should be avoided. The scientific method helps comprehend systems, minimizing unforeseen issues.

AGI and the Precautionary Principle:
As AGI approaches, treating it with respect and exercising precautionary measures is essential. Boldness and responsibility must coexist in AI development.

AGI as a Universal Tool:
AGI’s potential as a versatile tool for understanding the universe, akin to the Hubble telescope.

Demis Hassabis’s Scientific Excitement:
Applications like AlphaFold and drug discovery captivate Hassabis’s scientific interests. Revolutionizing human health, disease management, and sustainability through AI.

Fascination with the Brain and Consciousness:
Demis Hassabis’s interest in the brain and consciousness led him to pursue a PhD in neuroscience. He was intrigued by the processes of thinking and creativity and wanted to understand why the brain makes certain mistakes. Reading philosophy and science fiction sparked his interest in big questions like consciousness, creativity, and dreaming.

AI as a Tool for Understanding the Human Mind:
Hassabis believes that building a general learning AI system can provide insights into the human mind and answer deep questions about consciousness, intelligence, and reality. By comparing the AI system to the human brain, he hopes to gain a better understanding of how we think, feel, and experience the world.

AI’s Potential to Aid Neuroscience:
Hassabis sees AI as a valuable tool for aiding neuroscience research. AI can be used for data analysis, modeling, and simulations to help scientists better understand the brain and its functions.

AI and the Search for Newtonian Laws of Intelligence and Physics:
Hassabis believes that AI has the potential to uncover new laws of intelligence and physics. He suggests that AI could help us understand quantum mechanics and quantum behavior, which are currently challenging concepts in physics.

The Role of Humans in Scientific Discovery:
Hassabis envisions a future where AI takes on the repetitive and routine tasks in scientific research, freeing up human scientists to focus on more creative and innovative aspects. He emphasizes the need for accurate and reliable AI systems to assist scientists effectively.

AI’s Role in Science:
AI systems cannot independently define the correct questions to ask, which is a critical aspect of scientific inquiry.

AlphaGo and Chess:
The invention of AlphaGo demonstrated that AI can achieve remarkable feats, but it still lacks the ability to create entirely new games like Go or chess.

Turing Machines and Classical Computers:
Demis Hassabis believes that Turing machines and classical computers have the potential to achieve more than previously thought. AlphaFold and AlphaGo exemplify this potential by tackling complex combinatorial spaces and modeling quantum systems efficiently.

Proteins as Quantum Systems:
Proteins are quantum systems at the atomic level. AlphaFold effectively mimics and models quantum systems, solving the protein folding problem in milliseconds.

Theoretical Implications:
Demis Hassabis is exploring the implications of AlphaFold’s success for theoretical computer science concepts like p equals mp.

Hypothesis Generation and Experiment Design:
Current AI systems lack the ability to generate hypotheses and design experiments independently. However, Demis Hassabis believes that future advances may enable AI to perform these tasks.

Norms for AI Safety:
Demis Hassabis emphasizes the importance of establishing norms for AI safety, especially in the context of powerful language models like LLMs. He has been advocating for these norms for some time and considers the current moment to be particularly crucial.

Future of AI and Societal Questions:
As AI capabilities advance rapidly, Demis Hassabis emphasizes the need to focus on safety and analysis, including better testing and collaboration for external partners to investigate AI systems independently. He highlights the importance of debating and discussing societal questions related to AI, such as the values, rules, and applications of AI systems, considering different cultural and geopolitical perspectives.

Collaboration and Partnerships:
Hassabis suggests exploring partnerships and collaborations to facilitate access to AI models for external partners, enabling independent investigation and testing before public release. He acknowledges the challenges of manual testing and emphasizes the need for automated testing methods for AI systems.

Robotics at DeepMind:
Hassabis believes robotics is a crucial application area for AI, with significant industrial and research implications. He views robotics as a grounding problem for fundamental AI research, helping to validate simulation work and push the boundaries of low data regimes and transfer learning. DeepMind engages in various robotics projects, including locomotion, robot arms, and soccer teams, to explore coordination and other challenges.

Technical Monoculture and DeepMind’s Broad Research Approach:
Demis Hassabis acknowledges the trend towards technical monoculture, where resources are increasingly concentrated on a limited range of directions. DeepMind has traditionally adopted a broad research approach, encompassing various fields from neuroscience to deep learning and reinforcement learning, and maintains a multidisciplinary team. Currently, DeepMind has shifted its emphasis towards large and pre-trained models due to their necessity and potential.

Demis Hassabis’s Approach to Asking the Right Questions in Research:
Hassabis has developed an intuitive ability to identify the right research questions over many years. He immerses himself in various domains, seeking to understand problems from experts’ perspectives and mapping them onto problems he knows. He utilizes an S-curve approach, selecting problems that are challenging but not insurmountable, and progressively tackles harder problems as capabilities improve. Brainstorming with multidisciplinary teams helps assess the suitability and potential of a research problem.

Potential for a Base Foundational Model for Scientific Research:
DeepMind’s work suggests the possibility of developing a general model that can assist in scientific research across various domains. Such a model could perform literature searches, suggest experiments, and provide other valuable insights.

Assessing When to Double Down or Give Up:
Hassabis uses various heuristics to determine when to continue investing in a research direction or when to abandon it. He considers the flow and quality of ideas generated by a diverse group of experts. In the case of AlphaFold, the abundance of new ideas and the potential for improvement prompted DeepMind to double down on the project.

Demis Hassabis’ Reflections on AlphaFold2 Development:
AlphaFold2’s success required changing teams and employing orthogonal approaches to overcome challenges. Hassabis experienced numerous “dark valleys” during the development process, highlighting the importance of perseverance and seeking alternative signs of progress.

Inspiration from Foldit Game:
Foldit, a citizen science game, demonstrated the ability of non-expert gamers to solve protein folding puzzles using their intuition. This inspired Hassabis to believe that AlphaFold2, which mimicked the intuition of top pattern matchers, could also solve the protein folding problem.

Additional Reasons for Pursuing AlphaFold2:
The existence of folded proteins in living organisms suggested that the problem was tractable, despite its apparent difficulty. Hassabis’ heuristics helped provide confidence in the project’s potential success.

Question from a Graduate Student:
A question was raised about how technologists and scientists can incorporate the perspectives and ideas of people from marginalized communities, especially those without a scientific background, into their research.

Multidisciplinary Approach:
DeepMind employs a multidisciplinary team of scientists, ethicists, social scientists, and philosophers to ensure a comprehensive and ethical approach to AI research and development. The company actively seeks diverse perspectives and inputs to inform the design, deployment, and impact assessment of AI systems.

Thoughtful Deliberation and Understanding:
DeepMind emphasizes the importance of thoughtful deliberation and understanding the consequences of AI systems, including both intended and unintended effects. The company recognizes the need to carefully consider the potential impact of AI on different segments of society and strives to design systems that benefit everyone.

Democratizing AI Research Resources:
Stanford University and the Human-Centered AI Institute (HAI) are leading efforts to democratize AI research resources in the United States. The National AI Research Resource bill aims to establish a federally funded platform for AI, compute, and data, with a focus on reaching traditionally underrepresented communities.

Long-Term Perspective and Founding Philosophies:
DeepMind’s long-term perspective and commitment to continuous innovation at the frontier of AI research are driven by core founding philosophies. These philosophies include a focus on solving important problems, taking risks, encouraging creativity, and fostering a collaborative and open culture.

Learning Systems and Deep Learning:
DeepMind focused on developing learning systems rather than expert systems, recognizing the potential of reinforcement learning and deep learning techniques.

Neuroscience Inspiration:
The company drew inspiration from neuroscience, incorporating concepts like memory replay and episodic memory into their AI systems.

Importance of Compute and GPUs:
DeepMind recognized the significance of compute power and leveraged GPUs, commonly used in gaming, for their AI research.

Simulations and Games as Test Platforms:
Instead of relying on robotics platforms, DeepMind chose simulations and games as test environments, allowing them to concentrate on software development.

Strategic Advantages of Being Based in London:
DeepMind’s location in London provided access to untapped talent in Europe and the UK, contributing to their success.

Global Representation in AI:
DeepMind’s UK-based headquarters and European outlook enabled them to represent Europe in the global AI landscape, bringing a unique perspective to the field.

Resisting the Move to Silicon Valley:
Despite pressure to relocate to Silicon Valley, DeepMind maintained its UK base, allowing them to foster a distinct viewpoint and contribute to the global AI conversation.