Introduction of Demis Hassabis:
Martin Vetterli, President of EPFL, introduced Demis Hassabis, the founder of DeepMind and Isomorphic Labs. He highlighted Hassabis’ impressive academic achievements, including his early success in chess, completing A-level exams at a young age, and graduating with a double first in computer science from Cambridge University.

Hassabis’ Career and Contributions to AI:
After graduating, Hassabis co-founded a video gaming company, Elixir Studios, before obtaining a PhD in Cognitive Neuroscience from UCL, where he studied the connection between imagination and memory recall. He then co-founded DeepMind, a leading AI company that has made significant advancements in the field.

AlphaGo and Protein Folding:
In 2016, DeepMind’s AlphaGo program became the first computer program to defeat a professional human Go player in a five-game match. Hassabis also led the development of AlphaFold, a protein folding prediction tool that has revolutionized the field of protein science.

Honorary Doctorate from EPFL:
EPFL awarded Hassabis an honorary doctorate in recognition of his visionary leadership in artificial intelligence and neuroscience. The award recognizes Hassabis’ groundbreaking innovations in AI-driven scientific research and his contributions to advancing the frontiers of science.

Hassabis’ Vision for AI in Scientific Discovery:
Hassabis emphasized the potential of AI to accelerate scientific discovery. He shared his perspective on how AI can contribute to addressing fundamental scientific problems and highlighted the importance of bringing together hardware advancements, algorithm developments, and insights from neuroscience.

DeepMind’s Early Success with Atari Games:
DeepMind’s first major breakthrough was with classic Atari games in 2013. The DQN (Deep Q-Network) system demonstrated the power of learning systems by learning to play various Atari games without being explicitly given the rules or criteria. DQN achieved superhuman performance in Atari games, surpassing the best human players.

AlphaGo’s Achievement in Go:
AlphaGo, developed by DeepMind in 2015-2016, made a significant impact in the field of AI. AlphaGo became the first computer program to defeat a professional human Go player in a five-game match, demonstrating the capabilities of AI in complex strategic games.

AlphaGo’s Victory and Its Significance:
AlphaGo, developed by DeepMind, made history by defeating the world-renowned Go champion Lee Sedol in 2016. The complexity of Go, with its vast number of possible board positions, makes brute-force search methods impractical. Unlike chess, Go’s evaluation function is challenging to define due to its esoteric nature and reliance on intuition.

AlphaGo’s Innovative Strategies and Move 37:
AlphaGo surprised the Go world with its unconventional strategies, including the infamous Move 37 in Game 2. Move 37, played near the center of the board early in the game, defied traditional Go strategies and astounded professional commentators. This move sparked discussions about creativity and intuition in AI, highlighting AlphaGo’s ability to explore new patterns and concepts.

Creativity and Intuition in AI:
Creativity in AI can be categorized into three levels: interpolation, extrapolation, and invention. Interpolation involves averaging or combining known patterns to produce new instances. Extrapolation extends these patterns beyond the training data to explore new possibilities, as seen in AlphaGo’s Move 37. Invention, the highest level of creativity, remains a challenge for AI, as it requires abstract thinking and out-of-the-box ideas.

Limitations of Current AI Systems:
Despite AlphaGo’s groundbreaking achievements, it could not invent Go itself, demonstrating the limitations of current AI systems. The high-level specifications required for game invention are challenging for AI to understand and fulfill.

Understanding AlphaGo’s Functioning:
AlphaGo combines various techniques, including deep neural networks, reinforcement learning, and Monte Carlo tree search. Deep neural networks learn to evaluate board positions and predict the outcome of moves. Reinforcement learning allows AlphaGo to improve its strategies through self-play and interaction with other AI systems. Monte Carlo tree search explores possible moves and outcomes to find the most promising actions.

Background:
AlphaGo and its successors, AlphaGo Zero and AlphaZero, are notable for their exceptional performance in various two-player board games, surpassing human world champion levels.

Process Overview:
The process involves training a neural network to play games and improve its performance over time.

Creating Version 1:
The initial neural network (Version 1) is initialized to play randomly and serves as a starting point for improvement.

Game Play and Data Collection:
Version 1 plays a large number of games (e.g., 100,000 games) against itself or other players. This gameplay generates a dataset of positions and outcomes from these games.

Training Version 2:
A second neural network (Version 2) is trained using the dataset collected from Version 1’s games. Version 2 learns to predict the probability of winning from a given position and the likely moves in that position.

Game Tournament and Version Replacement:
Version 2 is pitted against its predecessor (Version 1) in a 100-game tournament. If Version 2 wins a significant majority of games (e.g., 55% win rate), it replaces Version 1 as the current system.

Repeating the Cycle:
The process of game play, data collection, and training is repeated with the new version (e.g., Version 3). This cycle continues until a new version fails to perform significantly better than its predecessor.

Convergence and Mastery:
Eventually, after a sufficient number of cycles (e.g., 17 times for Go), the system reaches a point where it consistently outperforms all human players in the game.

Using Games to Advance AI Algorithms:
DeepMind utilized games to efficiently test and improve algorithmic ideas, leading to significant breakthroughs like DQN, AlphaGo, AlphaZero, and AlphaStar. Games provided a fast and effective way to make progress in developing sophisticated algorithms.

Applying AI to Real-World Problems:
DeepMind’s goal has always been to use AI as a tool to accelerate scientific discovery and solve challenging real-world problems. The company’s technology has already been integrated into various Google systems, including data center management, voice recognition, and text-to-speech.

Protein Folding and AlphaFold2:
DeepMind recognized the importance of solving the protein folding problem, a long-standing challenge in biology. Protein folding is crucial for understanding protein function and structure, which are essential for various biological processes. AlphaFold2 achieved atomic accuracy in protein structure prediction, a significant milestone in computational biology. This breakthrough was the result of years of work, starting from AlphaFold 1, which showed promising results in CASP 13.

CASP Competition and Protein Structure Prediction:
CASP is a prestigious competition that evaluates computational systems for protein structure prediction. AlphaFold2’s success in CASP 14 led to the organizers declaring that the protein folding problem had been “solved.” The competition has been running for nearly 30 years and serves as a benchmark for progress in protein structure prediction.

Significance of AlphaFold2’s Achievement:
AlphaFold2’s atomic accuracy in protein structure prediction is a major breakthrough that has revolutionized the field of computational biology. It opens up new possibilities for understanding protein function, designing drugs, and advancing various biological research areas.

AlphaFold’s Iterative Approach to Protein Structure Prediction:
AlphaFold takes iterative steps to predict protein structures with remarkable accuracy. The prediction process involves multiple stages, gradually refining the structure until a final prediction is reached. The complexity and beauty of the predicted structures are astounding and a testament to AlphaFold’s capabilities.

Sharing AlphaFold Predictions with the World:
AlphaFold’s predictions were made freely available through a database collaboration with EMBL and EBI. The database contains structures for nearly every protein known to science, enabling researchers to access structural information quickly and easily. Safety and ethical considerations were taken into account before releasing the database, with input from experts in various fields.

AlphaFold’s Wide-Ranging Applications:
AlphaFold has been used in diverse fields, including drug discovery, antibiotic resistance research, and neglected tropical disease studies. It has facilitated the design of plastic-eating enzymes to tackle plastic pollution. The database has been utilized in fundamental research, such as studying the nuclear pore complex and developing molecular protein syringes.

Science at Digital Speed:
AlphaFold exemplifies the concept of “science at digital speed,” providing solutions and disseminating information at remarkable speeds. The fast and scalable nature of AlphaFold allows for rapid access to protein structures, unlike traditional experimental methods that can take months or years. The digital dissemination of information enables immediate propagation of scientific breakthroughs to researchers worldwide. AlphaFold serves as a prime example of this phenomenon, showcasing the potential for digital technology to accelerate scientific progress.

Novel Era of Digital Biology:
Biology interpreted as an intricate information processing system. AI holds potential as a potent framework for comprehending and modeling complex biological phenomena. AlphaFold seen as a pioneering illustration of this potential.

Isomorphic Labs:
A new spin-out company founded to revolutionize drug discovery through AI. Aims to tackle fundamental issues in drug design using AI as the primary approach. Leveraging collaborations with DeepMind and Alphabet’s resources for long-term and ambitious research.

Multidisciplinary Approach:
Strong emphasis on interdisciplinary research teams at Isomorphic. Team composition includes experts in AI, chemistry, biology, physics, and engineering. An impressive scientific advisory board comprising Nobel laureates.

Expanding AlphaFold-like Systems:
Aiming to develop numerous additional systems like AlphaFold. These systems will target different parts of the drug discovery chain.

AI as a General Tool:
AI can be universally applied to navigate vast combinatorial spaces. Algorithm design to explore these spaces effectively is critical. AlphaFold as an example of successful algorithm design.

Generalization and Transfer Learning:
AI can generalize knowledge gained in one context to other similar situations. Transfer learning as a powerful technique for AI efficiency.

AI for Societal Benefit:
AI’s potential to address global challenges such as climate change and healthcare. Importance of responsible and ethical AI development.

Pioneering Responsibly with AI:
AI has the potential to solve humanity’s greatest challenges, but it needs to be built responsibly and safely. Ethical considerations should be central to AI development from the beginning. Google has an ethics charter and committee to guide AI development and deployment. Thoughtful deliberation, hypothesis generation, and rigorous testing are crucial before deploying AI systems.

AGI and the Scientific Method:
AGI (Artificial General Intelligence) is approaching rapidly, requiring careful consideration. The scientific method should be used to approach AGI development and deployment. Move at pace, but avoid breaking things by using foresight and empirical data. AGI requires exceptional care and a balance between boldness and responsibility.

From Games to Proteins:
Demis Hassabis’s early interest in chess and AI led him to computer games development. His games often featured AI as a core gameplay element, such as Theme Park. Seeing people enjoy AI-driven interactions inspired him to pursue AI research.

Inspiration for Complex Systems:
Demis Hassabis and David Silver discussed programming Go during their time at Cambridge. Their fascination with Go and reinforcement learning led to the development of AlphaGo.

AlphaFold and Drug Development:
AlphaFold, a computational approach, has the potential to revolutionize drug development by accelerating the process and reducing costs. With faster development times and lower costs, it may become feasible to address diseases that are currently underserved by the pharmaceutical industry.

Ethical Considerations:
Some investors in the pharmaceutical industry prioritize profits over ethical considerations. Isomorphic Labs aims to make medication more accessible and affordable for the entire human population, including those in developing countries. Open-sourcing medication for diseases that primarily affect the developing world is a potential goal.

Unforeseen Challenges and Surprises:
Challenges: Enumerating all the heuristics for complex tasks like Go proved to be impractical. Surprises: The learning approach to AI proved to be more effective than expected, even for complex tasks like Go.

Hard Moments and Rebuilding:
One of the biggest challenges in developing AlphaFold was reaching atomic accuracy. After six months of effort, the team realized the system couldn’t progress further and had to start over, re-architecting everything. The new system, AlphaFold 2, took a long time to surpass the performance of AlphaFold 1, requiring confidence and persistence.

Language Modeling’s Surprise:
Language modeling has been a surprising success, with transformers and RLHF making it relatively easy compared to other AI tasks.

Addressing Inequality:
DeepMind is actively addressing inequality in AI education and opportunities. They sponsor hundreds of master’s scholarships for underrepresented backgrounds and encourage governments to do the same. They have funded chairs in machine learning at universities and provided a $5 million sponsorship to train top African students in machine learning.

AI’s Potential for Humanity:
AI has the potential to solve many intractable problems, including energy sources, material design, biology, and drug design. AI can contribute to radical abundance and equality, as well as help solve societal problems like climate change.

AI in Infrastructure:
AI can be used to optimize existing infrastructure, such as data centers, transport, and power grids, leading to significant efficiency gains. AI can also help design new solutions for waste management, recycling, and pollution reduction.

Balancing Open Source and Safety in AI:
Open-sourcing AI models has been a key factor in the field’s progress. However, as AI systems become more powerful and general, there are concerns about safety and potential misuse by bad actors. Balancing open access with safety considerations will be a challenge in the future.

Balancing Open Science and Access Restrictions:
Balancing the benefits of open science (e.g., faster progress, external review) with the risks of access by bad actors is a challenge. Controlled access to AI systems through departments, universities, or known academics may be one solution.

Developing Interpretability Analysis:
Interpretability analysis of AI systems is important for understanding their properties and limitations. More research is needed in this area to develop robust evaluation systems.

Evaluation Tests for Safe AI Systems:
Evaluation tests can help ensure that AI systems are safe and have desired properties before open-source release. These tests can provide guardrails around AI systems to limit potential misuse.

Challenges in Evaluation Systems:
Developing robust evaluation systems for AI systems is challenging due to a lack of knowledge about what these systems should be.