The Thinking Game | Full documentary | Tribeca Film Festival official selection

The Genesis of DeepMind and the Pursuit of AGI [00:30]

• DeepMind was founded with the explicit mission to build a "general learning machine" or Artificial General Intelligence (AGI).
• The founders believed that achieving AGI was the most exciting and important journey humanity had ever undertaken, comparable to the advent of electricity or fire.
• Their vision extended beyond AI itself, aiming to use AI as an ultimate tool to solve the world's most complex scientific problems.
• Early academic circles were often dismissive of AI research, leading the founders to establish a company to pursue their goals.
• Securing initial funding was challenging, with investors often viewing the ambitious goal as akin to buying a lottery ticket.
• Peter Thiel was an early significant investor, insisting on a Silicon Valley presence for talent and infrastructure.

"If you're really going to take that seriously, there isn't a lot of time. Life's very short."

DeepMind's Early Breakthroughs: Mastering Games with Deep Learning [09:01]

• DeepMind employed a reinforcement learning setup where an AI agent interacts with an environment to learn.
• Games were chosen as the ideal training ground for AI development due to their structured environments and clear objectives.
• The goal was to create a single algorithm capable of learning to play a multitude of games, mirroring human general intelligence.
• An early success was developing an agent that learned to play various Atari games by observing pixels and maximizing scores, combining reinforcement learning with deep learning.
• The agent's ability to learn end-to-end, without explicit programming of rules or controls, was a significant breakthrough.
• The agent's progression from struggling with simple games like Pong to mastering them and eventually exceeding human performance was a key validation of their approach.

"No one had ever combined those two things together at scale to do anything impressive. And we needed to prove out this thesis."

The AlphaGo Revolution: Conquering the Game of Go [15:00]

• The game of Go was identified as the "holy grail of artificial intelligence" due to its complexity, with more possible board configurations than atoms in the universe.
• AlphaGo was developed using reinforcement learning and deep learning, trained on 100,000 human games and then playing millions of games against itself.
• A pivotal moment occurred when AlphaGo played Lee Sedol, one of the world's top Go players, in a highly anticipated match.
• AlphaGo made an "original move" (move 37) that surprised commentators and was deemed to have a 1 in 10,000 probability of being played by a human.
• AlphaGo's victory over Lee Sedol was described as a "Sputnik moment" for AI, triggering a global race in AI development.

"The game of Go has been studied for thousands of years and AlphaGo discovered something completely new."

The Birth of AlphaZero and the Quest for True Generality [21:05]

• AlphaZero was developed to strip away all human knowledge, learning solely from self-play to achieve a more elegant and general approach to intelligence.
• Unlike AlphaGo, which was initially trained on human data, AlphaZero started with "zero human knowledge."
• The system could learn to play any two-player perfect information game, including chess, with remarkable speed and efficiency.
• AlphaZero demonstrated superhuman performance in chess within hours of starting its training, discovering its own attacking styles.
• This demonstrated the potential for AI to achieve mastery in complex domains without prior human expertise.

"Zero meaning having zero human knowledge in the loop. Instead of learning from human data, it learned from its own games."

The Early Years and Vision of Demis Hassabis [38:48]

• Demis Hassabis's early fascination with computers and games led him to win a competition to work at Bullfrog Productions, a leading game developer.
• His work on "Theme Park" involved simulating the complex behaviors of virtual people, a pioneering use of AI in game development.
• This success in creating nuanced simulations sparked an intuition about AI's potential beyond entertainment, specifically its utility in solving real-world problems.
• Hassabis harbored a lifelong ambition to be the person who solves AI, a goal he pursued through his studies and entrepreneurial ventures.
• He famously turned down a million-pound offer to defer university, choosing to pursue his academic and research goals.

"The human player set out the layout of the theme park and designed the roller coaster and set the prices in the chip shop. What I was working on was the behaviors of the people. They were autonomous and that was the AI in this case."

AlphaFold: Cracking the Protein Folding Problem and Ushering in AI-Assisted Science [47:50]

• The protein folding problem, predicting a protein's 3D structure from its amino acid sequence, had been a grand challenge in biology for over 50 years.
• Solving this problem had immense potential for drug discovery, understanding diseases like Alzheimer's, and developing new therapies.
• DeepMind applied AI to this challenge, entering the CASP (Critical Assessment of protein Structure Prediction) competition.
• AlphaFold achieved state-of-the-art accuracy, significantly outperforming other methods and demonstrating the power of AI in scientific discovery.
• The subsequent release of 200 million protein structures made freely available to the scientific community marked a paradigm shift, enabling accelerated biological research globally.

"After half a century, we finally have a solution to the protein folding problem."

The Future of AGI and Societal Impact [01:00:01]

• The advent of AGI is viewed as a transformative event that will divide human history, offering tools to reinvent civilization.
• There is a race towards AGI, with the potential for it to arrive faster than society can prepare, raising concerns about responsible development and governance.
• AI's increasing sophistication, including human-like interactions and generation of realistic content, blurs the lines between human and artificial intelligence.
• Ethical considerations are paramount, including how to imbue AI with human values, the potential for misuse in warfare and surveillance, and the societal displacement caused by automation.
• The development of AI is a complex interplay of technical, ethical, and societal factors, demanding cautious and thoughtful deployment.

"It really feels like we're in a race to AGI. The prototypes and the models that we are developing now are actually transforming the space of what we know about intelligence."