Summary
Highlights
Demis Hassabis voices his lifelong goal: to solve artificial general intelligence (AGI) and use AI as a tool for solving the world's most complex scientific problems. He likens this endeavor to the advent of electricity or fire in its significance. Shane Legg and Demis, both obsessed with AGI, decided to form a company to pursue this academically challenging and financially risky goal, facing skepticism from investors.
DeepMind was founded with the mission to build the world's first general learning machine. The early days involved operating in 'stealth mode' due to the unconventional nature of their research. Peter Thiel became their first major investor, though Demis insisted on keeping the company in London rather than Silicon Valley to foster a long-term research culture. They decided that games would be the perfect training ground for AI development, aiming for a single algorithm that could play multiple Atari games without prior knowledge of the rules.
DeepMind developed a system that combined reinforcement learning with deep learning, a novel approach. Initially struggling with Pong, their AI eventually surpassed human performance. This led to breakthroughs in games like Breakout, where the AI discovered optimal, never-before-seen strategies. The development of AlphaGo, which beat the world's best Go player, Lee Sedol, was a significant milestone, seen as a 'Sputnik moment' that sparked an AI race, particularly in China. Subsequent research led to AlphaZero, an algorithm that learned to play games like Go, chess, and shogi from scratch, achieving superhuman levels in hours.
Demis Hassabis's fascination with the mind began in childhood, leading him to excel in chess at a young age. His bohemian parents supported his talents, though the intense pressure of professional chess led him to question its ultimate purpose. A pivotal moment came during an international chess tournament where his opponent tricked him into resigning a drawn game, making him realize the immense intellectual power being 'wasted' on chess, igniting a desire to solve more profound problems.
DeepMind sought to develop AI capable of generality and learning in novel situations. They used simulated environments to train AI agents, observing how they learned to navigate and perform tasks, similar to how children learn. The popular game Starcraft presented a new challenge, requiring diverse skills, complex image processing, and strategic decision-making in real-time with imperfect information. DeepMind's AlphaStar eventually beat professional Starcraft players, showcasing the AI's advanced capabilities.
The success of AI in games like Starcraft brought into focus the ethical implications of such powerful technology. Concerns about military applications, surveillance, and autonomous weaponry were raised. Demis Hassabis emphasizes the need for society to control these technologies and not to 'move fast and break things' with AI, highlighting the irreversible consequences. The comparison to the Manhattan Project and Oppenheimer's dilemma is drawn, underscoring the responsibility of scientists in developing powerful technologies.
Demis's early experience at Bullfrog, where he helped develop the successful game Theme Park, illustrated AI's potential to create complex, interactive simulations. He then decided to pursue a university education at Cambridge, despite a lucrative job offer, to study neuroscience and further explore intelligence. This ambition later led DeepMind to tackle the protein folding problem, initially inspired by conversations at Cambridge, recognizing its immense potential to impact medicine and science.
DeepMind embarked on solving the protein folding problem, despite its difficulty and the limited available data. They entered the CASP competition, a biennial assessment of protein structure prediction. Their initial attempts were not as successful as hoped, leading to a humbling realization that even being the 'best in the world' at a difficult problem doesn't mean it's solved practically. This led to a period of doubt and re-evaluation, pushing the team to innovate further.
DeepMind redoubled their efforts, forming a 'strike team' to tackle protein folding with renewed vigor and incorporating deeper biological domain knowledge. Despite initial setbacks and the challenges posed by the COVID-19 pandemic, AlphaFold eventually achieved a major breakthrough in CASP 14, delivering predictions with unprecedented accuracy. This achievement effectively solved the 50-year-old protein folding problem, allowing scientists to predict protein structures rapidly. DeepMind then made these predictions publicly available, releasing structural data for 200 million proteins, which is seen as a 'gift to humanity' with the potential to revolutionize biological and medical research.
The documentary concludes by emphasizing the accelerating pace of AI innovation and the imminent arrival of AGI. It highlights the profound implications of AGI for humanity, likening its potential impact to the discovery of electricity. The importance of responsible development, ethical considerations, and global coordination in governing AI is stressed, with Demis Hassabis reiterating his lifelong dedication to this transformative field. The film ends with a demonstration of a more advanced Alpha, showcasing its ability to understand and interact with complex concepts.