A Conversation with Demis Hassabis, Co-Founder and CEO of Google DeepMind

Introduction to Demis Hassabis and DeepMind

• Demis Hassabis is an artificial intelligence researcher, entrepreneur, and Nobel laureate who co-founded Google DeepMind, one of the world's leading AI research companies, in 2010, which was later acquired by Google in 2014 10s.
• Google DeepMind has produced significant breakthroughs in the field, including AlphaGo, the first program to defeat a world champion at the game of Go, and AlphaFold, which solved the 50-year grand challenge of protein structure prediction, with enormous implications for disease understanding and drug discovery 2m6s.
• Demis Hassabis, alongside John Jumper and David Baker, was awarded the 2024 Nobel Prize in Chemistry for his work on AlphaFold, and he has also been named to the Time 100 list of the world's most influential people multiple times, including in 2017 and 2025 4m10s.

AI and Human Flourishing at Stanford

• The conversation around AI at Stanford has never been only about capability, but also about human flourishing, with professors Fei-Fei Li and Jennifer Aaker teaching a course on AI for human flourishing that explores questions about what it means to be human and how technology can help advance human goals 6m20s.
• Demis Hassabis's work and the discussion around AI are particularly important at Stanford, where advances in AI are already reshaping how we think about discovery, diagnosis, leadership, learning, and human potential itself, and forcing us to grapple with larger questions about judgment, ethics, institutions, and what kinds of lives we ultimately want technology to help us build 8m30s.

Demis Hassabis's Background and Career Trajectory

• Demis Hassabis has a remarkable trajectory, having been a chess prodigy, video game developer, scientist, tech entrepreneur, and leader, and his thoughts and experiences are highly valued, with a movie and a book recently chronicling his story 10m40s.
• Demis Hassabis has always been drawn to working at the intersection of creativity and technology, which is reflected in his early career in the video games industry, where he found it to be a highly creative space that utilized cutting-edge technology with art and design 10s.

Early Influences and Vision for AI and AGI

• His interest in AI and AGI began at a young age, inspired by science fiction books such as Gödel, Escher, Bach, and biographies of scientific heroes like Turing and Feynman, which motivated him to understand the world in a deep way and build the ultimate tool for science 2m6s.
• Hassabis has tried to reuse and repurpose every experience he has had in service of his mission to build AGI, including his chess training, which has influenced his approach to business and planning, and his experience with games, which has been used as a proving ground for testing algorithmic ideas 4m30s.

DeepMind's Early Work and AlphaGo

• The use of games in the early days of DeepMind, such as AlphaGo, served as a perfect testing ground for algorithmic ideas and marked the start of the modern AI era, with the 10-year anniversary of AlphaGo being a significant milestone 6m40s.
• When Demis Hassabis started DeepMind in 2010, he had a very ambitious vision to solve intelligence and then use it to solve everything else, which was met with confusion by VCs, but he really meant it and may go back to using that mission statement 8m10s.
• The broad arcs of the plan have gone well, with the goal of building AGI and understanding the nature of intelligence, and using AGI to help understand the brain and mind better, including deep mysteries like consciousness, creativity, and dreaming 10m0s.

DeepMind's Scientific Foundations and AGI Vision

• Hassabis studied neuroscience to learn from what is understood about the brain as inspiration for algorithmic ideas, and he believes that AGI, as a general-purpose technology, can be applied to almost anything if built in the right way 12m0s.
• The goal of using artificial intelligence to solve everything else refers to advancing science and medicine, and tackling the big questions in science, including the nature of time and reality, with a particular interest in physics, which was a favorite subject in school 10s.

DeepMind's Approach to AI Development

• The decision to build new tools to aid scientists and experts in making faster progress in their fields was driven by the desire to tackle the many interesting big questions in science and medicine, and to make a significant impact in these areas 42s.
• Artificial intelligence itself is a fascinating and important field of study, worthy of scientific investigation, and has the potential to be used to solve a wide range of problems, including those outside of science and medicine, such as productivity 2m6s.
• The development of DeepMind was a culmination of experiences and knowledge gathered over time, with the goal of making fast progress in science and medicine, and the company started by working on games, including Atari games like Pong, as a way to test and develop artificial intelligence systems 2m6s.
• Games were chosen as a starting point for developing artificial intelligence systems because they are self-contained, challenging, and have clear objective functions, making them ideal for testing and developing reinforcement learning algorithms 2m6s.
• The use of reinforcement learning to develop artificial intelligence systems was a key aspect of DeepMind's approach, and the company started by applying this technique to simple games like Pong, with the goal of scaling up to more complex problems 2m6s.

Challenges and Breakthroughs in Early AI Development

• The DQN system, also known as the Atari system, was given raw visual input data of 20,000 pixels on the screen without any privileged information about the program, and it had to learn from this complex input data 10s.
• The system initially struggled to win a single point at Pong, jerking the bat around and losing 21-nil to the inbuilt AI, which led to concerns that the approach might be 10 or 20 years too early, especially given the limited funding of only a couple of million dollars 42s.
• However, the system suddenly started winning points and games, achieving a breakthrough, and once it had a foothold, it was possible to optimize and improve it, which is a common pattern in the history of AI 2m6s.

DeepMind's Milestones and Deep Reinforcement Learning

• The combination of deep learning and reinforcement learning led to the development of the first deep reinforcement learning model at scale, which was published in a Nature paper and marked a significant milestone in the field 2m6s.
• The ultimate goal of the project was to create a system like AlphaGo, which was achieved through the collaboration of Demis Hassabis and Dave Silver, who had been discussing the idea since their undergraduate days at Cambridge in the mid-1990s 4m30s.

AlphaGo: Inspiration and Development

• The inspiration for AlphaGo came from the Deep Blue-Kasparov match, but instead of being impressed by the supercomputer, Demis Hassabis was more impressed by Kasparov's brain and its ability to compete with the machine, and he realized that there was something missing from the Deep Blue system 6m10s.
• The AlphaGo project aimed to create a system that could play Go at a world champion level, not just to achieve that level but to develop an interesting algorithmic approach that could generalize to other domains, and the project ultimately succeeded in creating a system that could play Go at a world champion level 8m40s.
• The AlphaGo project achieved significant success by winning a match against Lee Sedol in 2016 and creating new strategies that had never been seen before in the 2,000-year-old game of Go, which was a double whammy as it not only won the game but also came up with novel ideas 10s.

AlphaFold and the Protein Folding Problem

• The success of AlphaGo led to the initiation of the AlphaFold project, which aimed to solve the protein folding problem, a long-standing issue in biology that has significant implications for fields such as drug discovery and disease understanding 2m6s.
• The protein folding problem was chosen because it had a clear objective function, a large amount of available data, and the potential to unlock new avenues of research, making it a fascinating and important puzzle to solve 2m6s.
• The AlphaFold project was able to solve the protein folding problem by using deep learning models to guide the search for the correct structure, similar to how AlphaGo was able to find great moves in the game of Go, and the solution was considered a huge, Nobel-worthy scientific breakthrough 2m6s.

AlphaFold's Scientific Impact and Open Access

• The decision to give away the AlphaFold solution for free was made because the problem was considered a root note problem that could have significant downstream effects on biology and disease understanding, and making it freely available could accelerate research and progress in these fields 4m30s.
• The protein folding problem was considered a crucial issue to solve because proteins are essential for life, and understanding their structure and function can lead to significant advances in fields such as drug discovery and fundamental biology 6m10s.
• The use of deep learning models in the AlphaFold project was inspired by the success of AlphaGo, and the idea was to apply similar approaches and theories to the protein folding problem, which was considered an analogous problem in science 8m20s.
• The AlphaFold project built upon 50 years of painstaking work by structural biologists and crystallographers, who had determined the structures of around 150,000 proteins, and the new solution was able to significantly advance this field 10m40s.

AlphaFold's Technical and Practical Achievements

• The development of AlphaFold, a machine learning system, was a significant effort, but the amount of data available, 150,000 proteins, was considered small, and many thought it would take 10 to 20 years to make progress in the field 10s.
• Using various techniques, AlphaFold was able to accurately and quickly fold proteins, taking only seconds, and it was decided to collaborate with the European Bioinformatics Institute to host the entire 200 million protein structures on their database 42s.
• The decision was made to make the protein structures publicly available, allowing researchers to easily search and access the data, along with confidence intervals, which was considered a valuable resource that could have been kept proprietary but would have limited the potential impact 1m6s.

AlphaFold's Societal and Scientific Implications

• The choice to make the data public was also motivated by the fact that the development of AlphaFold relied on public data, and it was seen as a way to give back to the structural biology community, which has around three million researchers worldwide 2m6s.
• The goal is to accelerate drug discovery using AI, with the potential to reduce the time it takes from years to months or even weeks, and Isomorphic Labs, an Alphabet spinout, is working on building on the success of AlphaFold to achieve this goal 4m10s.

The Future of AGI and Artificial General Intelligence

• The concept of the singularity, referring to the next version of general artificial intelligence, was discussed, and it was suggested that we are currently in the foothills of this development, with the potential for significant advancements in the next 10 years 6m20s.
• The advent of Artificial General Intelligence (AGI) is expected to happen around 2030, marking the beginning of a new human era that will be transformative and have a profound impact on society 10s.
• The development of AGI is still in its early days, but significant progress has been made, with agents and tool use becoming increasingly useful in people's workflows, and leading labs are working together to advance the technology 2m6s.

Uncertainty and Challenges in AI Development

• The future is still to be written, and the next few years will be critical in determining the direction of AGI development and its impact on society, making it essential for collective effort and discussion to shape its outcome 2m6s.
• Public concern about AI is valid, and it is a dual-purpose technology that can have both positive and negative consequences, with potential benefits including solving diseases and addressing climate and energy challenges, but also causing disruptions and changes 4m30s.
• The negative perception of AI in some countries is driven by concerns about privacy, state control, job losses, and the size of tech companies, but in other countries like India, AI is seen as an opportunity for democratization and access to tools and resources 6m20s.

Addressing AI's Societal Impact and Governance

• To address the challenges and opportunities presented by AI, it is essential to bring together experts from various fields, including technologists, economists, social scientists, and humanity experts, to discuss and chart the course of AI development and its impact on society 8m40s.
• The current state of artificial intelligence is characterized by huge uncertainty, and many pronouncements about its future are overly certain, which is worrying, and it is believed that nothing is decided yet, with the next few years being crucial in determining its direction 10s.
• The youth of today, who are growing up with AI, will be able to master these technologies and become super productive, potentially leading to a significant change in the nature of jobs, with more entrepreneurial and small-scale ventures emerging 2m6s.
• It is essential for society to come together to take the exponential growth of AI seriously, not just technologists and economists, and to start charting out what a post-scarcity world would look like, ensuring that the benefits of AI accrue to everyone, not just a few individuals or companies 2m6s.

Concrete Actions and Public Perception of AI

• There is a need for concrete answers and actions to be taken to address the challenges posed by AI, and for leading labs and their leaders to come together to discuss these issues more candidly, using the scientific method to be rigorous and thoughtful about this critical moment in history 4m30s.
• The industry and field of AI need to demonstrate the benefits of AI more unequivocally, such as in health, medicine, and science, with tangible examples like AlphaFold, and to concretely mitigate the risks while enabling the amazing things that AI can achieve 6m40s.
• Demonstrating the benefits of AI, such as curing cancer, will be necessary to show the public why many people are excited about AI and have spent their lives building towards this, and to address the risks associated with it 8m20s.

Long-Term Implications and Productivity Gains

• The concept of thinking about a world with significantly improved productivity is challenging, and it is rare for people in social science to project far into the future, with Keynes' article on the economic lives of grandchildren being a notable example 10s.
• The idea of frontier labs regulating themselves by not releasing certain technologies that might be threats to safety has been discussed, and the current dynamic is characterized by a breakneck competition among labs, with the question of whether the government should step in to regulate AI 2m6s.

The Competitive and Geopolitical Landscape of AI

• The development of AI technology has been amazingly fast, but the environment in which it is being developed is not ideal, with a race dynamic happening among companies and tech leaders, which was a concern 10-15 years ago 2m6s.
• A more ideal approach to developing AI would be to build it in a research facility, like CERN, where the best minds can collaborate, critique each other's ideas, and ensure rigorous testing, but this would likely delay the arrival of AGI by 10 years 2m6s.
• Specialized systems, like AlphaFold, can be developed to provide societal benefits without waiting for the development of general-purpose AGI, and the effectiveness of transformers for language has been a surprise, allowing for the development of chatbots and other language-based technologies 2m6s.
• The competitive environment in the tech industry, particularly in AI development, is currently the most ferocious it has ever been, driven by the potential for commercialization and scaling with engineering and money 2m6s.
• The current state of AI development is intense and feels like a double-layered race, with companies competing against each other and a geopolitical dynamic between countries like the US and China, making it very tricky 10s.

Regulation and Safety in AI Development

• There is a need for cooperation and coordination among lab leaders on safety and security elements, as everybody wants to avoid something catastrophic from happening, but the problem is that taking more time to release something or make it safer can put a company at a disadvantage 42s.
• The race to the bottom dynamic is a classic problem, and to change it, some form of government involvement, such as dynamic regulation, is required, which can adapt to the latest developments and inform by the actual risk rather than perceived risk 2m6s.
• The science of AI is not yet settled, and the leading scientists would not agree on a short list of what checks and balances are needed, making it challenging to create effective regulation 4m10s.

Global Access and Equity in AI Benefits

• To enable the good use cases of AI and mitigate the bad, a regulatory system is needed that can balance innovation with safety, and Demis Hassabis plans to discuss this further later in the year 6m20s.
• There is a need to balance pushing the frontier of AI with ensuring that the health and scientific dividends are evenly distributed, particularly in places like Africa and the Global South where the need is greatest but the infrastructure for deployment and research is limited 8m30s.
• Demis Hassabis' company has made efforts to make AI research and tools accessible to researchers from around the world, such as making the AlphaFold database available to three million researchers from 190 countries 10m40s.

Collaborative Applications of AI in Health and Agriculture

• Collaborations were established in the early days to utilize AlphaFold, including working with the DNDi, Drugs for Neglected Diseases, part of the WHO in Switzerland, to address diseases in poorer areas of the world that lack good healthcare systems 10s.
• The collaboration allowed researchers to bypass the painstaking process of determining the structures of diseases like malaria or Zika virus, and instead start working on drugs directly, significantly speeding up the process 10s.
• Similar collaborations were established to work on crop resilience affected by climate change, including working with Jennifer Doudna's institute, to determine the structures of plant proteins, which had limited data available 10s.

AI's Role in Drug Discovery and Global Health

• The goal is to make the drug discovery platform efficient enough to reduce the time and cost of developing new drugs, from years and billions of dollars to months and tens of millions of dollars, allowing for the possibility of curing diseases that affect both rich and poor parts of the world 10s.

Ethical and Philosophical Considerations of AGI

• The potential of AGI to redefine and reshape people and the challenges they face is being considered, with a focus on the responsibility that comes with developing such technology and the need to think about the second-order consequences 2m6s.
• There is a call to action for individuals, particularly those in the humanities, to get involved in thinking about the economic, philosophical, and human condition implications of AGI, with a cautious optimism that humanity will be able to get it right and figure out solutions to the challenges that arise 2m6s.
• The need to take emerging technologies seriously and plan for their impact on society is crucial, as they have the potential to bring about significant changes, such as a post-scarcity world, which would require a new type of economic system, unlike any that have been tried before, as they were all based on a zero-sum and scarce world 10s.
• A new economic system is necessary because the current technologies have the potential to make the world a non-zero-sum one for the first time in humanity's existence, allowing for the utilization of resources beyond those limited to Earth, such as those found in the solar system, and enabling space travel 1m30s.
• The development of new technologies, such as AI, will also raise important questions about the evolution of society, virtue, meaning, and purpose, which will require the input of great philosophers 2m40s.

AI as a General-Purpose Technology and Consciousness

• AI is considered a fully general technology, similar to a Turing machine, which can compute anything that is computable, and its development should be approached with caution, particularly when it comes to complex and not well-posed problems like consciousness 5m10s.
• It is recommended that the first AI systems be built as tools, or intelligent tools, and that their development should be followed by a rigorous study of neuroscience, philosophy, and other relevant fields to come up with a clear definition of concepts like consciousness, before deciding whether to create entities that seem conscious 7m20s.
• The development of intelligent systems and conscious entities are considered dissociable, meaning that it is possible to create intelligent systems without necessarily making them conscious, and this is a choice that society should consider carefully 9m40s.

Advice for Students and Future Generations

• Demis Hassabis advises students to continue studying science, STEM subjects, mathematics, and computer science, as understanding how these tools are put together will allow them to take better advantage of their capabilities in the next 10 years 10s.
• He also encourages students to lean into the potential of these tools, as they have only scratched the surface of what they can do, and pairing them with other domains or building them into their workflow can unleash creativity 42s.
• Hassabis believes that these tools can enable both democratization and specialization, allowing individuals without coding skills to produce their ideas and experts to work on larger projects, making it an amazing time for both groups 2m6s.
• He acknowledges that the rapid change brought about by these tools can be worrying, but it also presents enormous opportunities, and it is up to the current generation of students to shape the future world 2m6s.

The Future of Education and Adaptability in the AI Era

• Hassabis notes that the current generation of students will be the first to be AI-native, similar to how his generation was computer and internet-native, and he envies them for having the opportunity to build the future world 4m30s.
• He emphasizes the importance of adaptability, having a broad domain of knowledge, and doubling down on one's own agency, as the future is still to be written and can be shaped by individual actions 6m20s.
• Hassabis agrees that the current period of change will be a golden era for liberal education, which will allow students to develop a wide range of skills and knowledge to navigate the uncertain future 6m40s.