Summary
Highlights
The documentary opens with the ambitious goal of building Artificial General Intelligence (AGI), highlighting it as 'the most exciting journey humans have ever embarked on.' Demis Hassabis's lifelong fascination with the mind and his academic pursuit of neuroscience laid the groundwork for DeepMind. He, along with Shane Legg, envisioned a company that would not shy away from the term 'AI' and would focus on solving general intelligence, a concept then often dismissed in academia. Securing initial funding was challenging, as investors struggled to grasp the long-term vision beyond immediate commercial products. Peter Thiel became a crucial early investor, though he initially pushed for a Silicon Valley base. DeepMind established itself in London with a mission to build the world's first general learning machine, emphasizing 'general' and 'learning' ability, rather than specialized AI.
DeepMind adopted reinforcement learning, using games as a disciplined training ground for AI development. Their first significant breakthrough involved training an AI to play multiple Atari games, demonstrating the AI's ability to learn from pixels and scores without prior rules. After initial struggles with Pong, the AI eventually surpassed human capabilities. This led to their next major challenge: the ancient game of Go, considered the 'holy grail of artificial intelligence' due to its immense complexity. Google's acquisition of DeepMind provided the necessary computational scale. AlphaGo, trained on human games and then playing against itself, demonstrated unprecedented creativity with 'move 37' against Lee Sedol, ultimately defeating him. This event, termed the 'Sputnik moment' for AI, spurred a global race in AI development.
Building on AlphaGo's success, DeepMind developed AlphaZero, an algorithm that learned Go, chess, and shogi playing entirely from self-play, starting with zero human knowledge. AlphaZero achieved superhuman levels in these games within hours. The challenge then shifted to real-time strategy games like StarCraft II, which presented difficulties such as continuous decisions, imperfect information, and diverse skills. After initial struggles against human players, AlphaStar, inspired by large language models, quickly improved to defeat top professionals. While showcasing rapid advancement in AI, these victories also raised ethical concerns about the broader implications and potential military applications of powerful AI, emphasizing the need for responsible development.
The narrative delves into Demis Hassabis's background, revealing his early fascination with chess and the mind. A child prodigy in chess, he realized that while he loved the game, it might not be the best use of 'brain power' for solving larger problems. He took a gap year before Cambridge to work at Bullfrog, a leading games development company, contributing to the creation of Theme Park, a game noted for its complex AI and simulated human behavior. This experience solidified his vision for AI's potential beyond entertainment, inspiring him to pursue solving AI himself. He attended Cambridge, where he and David Silver developed interests in computational neuroscience and observed the Deep Blue vs. Kasparov chess match, noting Deep Blue's lack of general intelligence despite its chess mastery.
During his time at Cambridge, Hassabis became fascinated by the protein folding problem, seeing it as a critical scientific challenge that AI could potentially solve. Proteins, the 'machines of life,' are essential for biological functions, and predicting their 3D structure from amino acid sequences has profound implications for medicine and other fields. DeepMind decided to tackle this problem, despite the scarcity of data and the inherent complexity that had stumped scientists for decades. They entered the CASP competition, a biennial assessment of protein structure prediction methods. Initial attempts were humbling; their models, though the best among competitors, were still not accurate enough for practical biological use. This period of 'failure' taught them the importance of timing and incorporating more domain-specific knowledge.
Despite initial disappointments, DeepMind redoubled its efforts, forming a strike team and incorporating insights from biology and physics. The team revamped its data pipeline and continuously refined its models. The COVID-19 pandemic underscored the urgency and potential of AI-assisted science. During CASP14, held during lockdown, AlphaFold achieved a monumental breakthrough. Its predictions were so accurate that it was declared a solution to the protein folding problem, with scores exceeding 90. This achievement, shared with the world, led to the release of 200 million protein structures, a 'gift to humanity' that promises to accelerate discoveries across biology and chemistry. The success of AlphaFold reaffirms DeepMind's founding vision: using AI to solve fundamental scientific challenges.
The documentary concludes by reflecting on the rapid acceleration of AI capabilities and the impending arrival of AGI. Interview segments discuss the profound societal transformations that AGI will bring, comparing its impact to the discovery of electricity. There's a strong emphasis on the ethical dimension, with concerns about potential misuse for military purposes, surveillance, and economic displacement. The film highlights the need for careful governance, global coordination, and responsible deployment of these powerful technologies. Demis Hassabis stresses that the pace of AI development demands immediate and serious consideration of its implications, stating that AGI is 'coming faster than we can really prepare for' and that 'every moment is vital' in steering this future responsibly.