Demis Hassabis (DeepMind Co-founder) – Demis Hassabis, DeepMind CEO (UC Berkeley) (Oct 2023)


Chapters

00:00:03 Foundation and Breakthroughs in DeepMind's Artificial Intelligence
00:05:54 Levels of Creativity in Artificial Intelligence
00:09:46 Evolution of AlphaZero: From Random Play to Superhuman Performance
00:12:06 Search and Decision-Making Strategies in Reinforcement Learning for Complex Games and Beyond
00:16:03 Protein Structure Prediction: Challenges, Progress, and AlphaFold's Breakthrough
00:19:57 Journey to Atomic Accuracy: The Evolution of AlphaFold

Abstract



Revolutionizing AI and Biology: The DeepMind Odyssey from AlphaGo to AlphaFold

In a groundbreaking journey, DeepMind, led by visionaries Demis Hassabis and Shane Legg, has transformed artificial intelligence and biology. Beginning with the revolutionary AlphaGo, which challenged conventional wisdom in AI through creative strategies like the famous Move 37, DeepMind evolved into developing AlphaFold, a tool that revolutionized the understanding of protein structures. This article delves into DeepMind’s inception, the development of AlphaGo and AlphaFold, and their profound impact on AI and scientific research. It highlights the creative potential of AI, exemplified by AlphaGo, and the transformative power of AlphaFold in biology.



DeepMind’s Inception: Pioneering Artificial General Intelligence

In the early days of AI research, Demis Hassabis and Shane Legg faced skepticism and resistance from academia, with many believing that achieving human-level artificial intelligence was impossible. However, they dared to pursue artificial general intelligence (AGI) at UCL, combining deep learning, reinforcement learning, and computational power, influenced significantly by Hassabis’ background in neuroscience.



The Apollo Program for AI: Establishing DeepMind

Inspired by the ambitious Apollo space program, Hassabis and Legg co-founded DeepMind as a company to accelerate AI progress, believing that a dedicated and well-resourced environment would facilitate faster advancements. Their objective was clear: to accelerate AI development through an intensive, resource-backed effort, mirroring the spirit and scale of Apollo.



Early Triumphs: The Impact of Gaming on AI

DeepMind initially focused on games, achieving significant breakthroughs with DQN, a system learning directly from raw pixels without explicit instructions or rules. This phase was pivotal, not just for testing AI capabilities but also for shaping DeepMind’s mission towards responsible and beneficial AI.



AlphaGo: Surpassing Human Intelligence in Go

AlphaGo’s victory over Go legend Lee Sedol was a defining moment for AI. It wasn’t just the win but the inventive strategies employed, like the unforeseen Move 37, that showcased AI’s creative and adaptive capacities.



AlphaGo’s Creative Essence

Hassabis identifies three creativity levels: interpolation, extrapolation, and true invention. AlphaGo’s Move 37 exemplified extrapolation, stepping beyond known strategies to create something unprecedented, hinting at AI’s potential for true invention.



From AlphaGo to AlphaZero: The Evolution of Learning

AlphaZero advanced beyond its predecessors by mastering games without human data. It refined its strategies through self-play, demonstrating extraordinary learning capabilities, significantly in a complex domain like Go.

Modeling Go and Guiding Search:

– AlphaGo’s neural networks serve as reliable models for Go, predicting likely moves and chances of winning in various positions.

– These models guide the search process, narrowing down the vast search space in the game of Go, focusing on useful parts of the tree.

– AlphaZero systems search tens of thousands of moves per decision, significantly less than brute force systems like Stockfish, but still more than human grandmasters.

AlphaZero and its successor, AlphaZero, use deep neural networks to learn from data and make decisions. AlphaGo initially required human games to learn, but AlphaZero was able to learn and play any two-player game without human data. AlphaZero’s name reflects its ability to learn and play games purely from scratch, without relying on human games for bootstrapping.



AlphaStar: Navigating Partial Information

AlphaStar and Partial Information:

– AlphaStar achieved success in the real-time strategy game StarCraft, which features partial information and fog of war.

– This advancement tested the AI’s ability to handle uncertainty and decision-making with incomplete information, pushing the boundaries of AI in more realistic environments.

DeepMind’s AlphaStar tackled StarCraft, a game with incomplete information, pushing AI capabilities closer to real-world scenarios. This achievement underscored AI’s ability to handle uncertainty and complexity.



Transition to Real-World Challenges: Protein Folding

Post-AlphaGo, DeepMind shifted focus to real-world problems, notably protein folding. This move signaled a transition from theoretical AI challenges to practical, impactful applications in biology and medicine.

Moving Beyond Games:

– After AlphaGo’s victory, the focus shifted from games to real-world applications, particularly in scientific challenges.

– Protein folding, predicting the 3D structure of a protein from its amino acid sequence, emerged as a top priority due to its significance in biology and medicine.



AlphaFold: A Milestone in Protein Structure Prediction

AlphaFold emerged as a key player in protein structure prediction, particularly in the CASP competition. AlphaFold1, despite its limitations, marked a significant advancement, and AlphaFold2’s end-to-end architecture brought unprecedented accuracy in predicting protein structures. AlphaFold2’s iterative refinement process led to achieving atomic accuracy in protein structure prediction. This breakthrough, reaching the accuracy threshold set by CASP, was not only a victory for DeepMind but also a major stride for scientific research, with profound implications in drug discovery and biology.

Physics, Nature, and Protein Folding:

– Protein folding occurs spontaneously in nature, suggesting that the problem is not inherently intractable.

– Modeling sufficient aspects of physics and biology should allow AI to solve the protein folding problem.



DeepMind’s Enduring Legacy

DeepMind’s journey from AlphaGo to AlphaFold epitomizes the fusion of AI and human ingenuity. By transcending traditional boundaries, DeepMind has not only advanced AI’s frontiers but has also paved new pathways in scientific discovery and problem-solving, marking a new era where AI and human creativity coalesce to unravel some of the most intricate mysteries of both the digital and natural worlds.


Notes by: ChannelCapacity999