DeepMind Documentary: General AI Is Greater Than Thermal Power, Human Civilization Will Be Redefined

DeepMind documentary “The Thinking Game” is now available for free, chronicling Demis Hassabis’s lifelong pursuit of Artificial General Intelligence (AGI). Hassabis believes that AGI is more important than the advent of electricity or fire. He warns that AGI is about to be born, marking a watershed moment in human history: “Our next generation will live in a brand-new world, and every moment counts.”

Demis Hassabis’s Lifelong Mission for AGI

(Source: YouTube)

As a Cambridge prodigy who won a local chess championship at the age of six, Demis Hassabis decided early on to dedicate his life to AGI research because he wanted to solve a problem that had plagued the biology community for 50 years: protein folding. It may be hard to imagine now, but at the time, many in the venture capital and academic worlds were highly skeptical of AGI technology. The former thought Demis’s ideas were pie in the sky, while the latter felt combining neuroscience and machine learning was not “pure” science.

Founded in 2010, DeepMind’s early fundraising journey was fraught with difficulties until they met renowned angel investor Peter Thiel. Although Thiel became DeepMind’s main backer, he insisted the team move to Silicon Valley. Demis adamantly chose to stay in London, believing it had a unique talent pool and that the fast-fail, quick-pivot culture of Silicon Valley was ill-suited for the long-term research required to achieve AGI.

This decision highlights Hassabis’s deep understanding of AGI research. AGI isn’t a consumer product that can be rapidly iterated, but a long-term project requiring foundational scientific breakthroughs. While Silicon Valley’s startup culture emphasizes rapid market validation and business models, the value of AGI research may not become evident for decades. By insisting on staying in London, Hassabis protected DeepMind’s research purity.

Hassabis likens AGI to humanity’s discovery of fire, a metaphor with profound meaning. The discovery of fire enabled humans to cook food, stay warm, create light, and smelt metals, fundamentally changing the trajectory of human civilization. Hassabis believes AGI will have an equal or even greater impact, as it is not just a tool, but an intelligence capable of self-learning and creativity.

From Games to Go: Breakthroughs with DQN and AlphaGo

(Source: DeepMind)

After DeepMind was established in London, it gathered a group of dreamers. To train AI, they decided to use “games” as their testing ground, since games are perfect controlled environments. They combined deep learning and reinforcement learning to create the DQN model, letting AI play Atari’s Pong without teaching it the rules—just asking it to watch pixels and pursue a high score.

Initially, the AI couldn’t return a single ball, causing the team to doubt whether AGI was just a fantasy. But suddenly, the AI started scoring. They then had the AI play Breakout. After hundreds of games, the AI figured out the tunnel strategy—digging through the wall at the side so the ball would bounce above the bricks. Most importantly, this was a solution the machine independently discovered, without any human preset.

This proved that DeepMind had successfully created a general learning system that could adapt to different environments, a huge breakthrough in AGI development. It wasn’t just about teaching machines to play games—it proved that machines could independently discover strategies and solutions without human guidance. Such autonomous learning is a core feature of AGI.

Despite breakthroughs in machine learning, computing power became a bottleneck. To accelerate AGI, DeepMind eventually agreed to be acquired by Google for about £400 million, hoping to retain their research independence. With Google’s computing resources, DeepMind turned its focus to Go, a game originating from China long considered the holy grail for AI.

AlphaGo was born and faced off against the world’s top Go player Lee Sedol. AlphaGo played the now-legendary Move 37—a move considered almost impossible for humans, demonstrating not just computational skill but creativity. Lee Sedol’s defeat shocked the world, especially in China, where it was likened to the Sputnik moment, awakening global awareness of AI and sparking an AI version of the “space race.”

Four Milestones in DeepMind’s Technological Evolution

DQN Model: Blending deep learning and reinforcement learning, AI autonomously discovers game strategies

AlphaGo: Defeats human Go champions, displaying creativity and intuition

AlphaZero: Completely abandons human knowledge, learns purely by self-play

AlphaFold: Solves the protein folding problem, wins Nobel Prize in Chemistry

While AlphaGo was powerful, it mainly relied on human game data for learning. DeepMind then developed AlphaZero—a more elegant algorithm that completely abandoned human knowledge and learned to play solely through self-play. Starting from scratch, AlphaZero mastered chess, shogi, and go within a day, even demonstrating playing styles never seen before by humans—proving that machines can surpass thousands of years of human wisdom through pure experience.

AlphaFold Solves Protein Folding and Wins Nobel Prize

Games were just the testing ground; Demis’s real ambition was to use AI to solve scientific problems—most notably “protein folding.” Protein folding has long been one of humanity’s greatest biological mysteries. If humans could predict protein structures, drug development and disease treatment would accelerate. To test AI’s potential in biology, DeepMind formed the AlphaFold team and entered CASP (Critical Assessment of protein Structure Prediction).

In CASP13 in 2018, AlphaFold won but wasn’t accurate enough for real-world biology research, leaving the team humbled and aware that scientific problems are far more complex than games. Determined not to give up, Demis doubled down on AlphaFold research during the COVID-19 pandemic, assembling a strike team combining physics expertise and machine learning. They worked tirelessly under tough lockdown conditions.

Finally, in CASP14 in 2020, AlphaFold achieved stunning results. The scientific community agreed that the protein folding problem had been substantially solved. DeepMind then made a bold decision—not to commercialize this achievement, but to release the predicted structures for over 200 million proteins, nearly all known protein sequences on Earth, to the public for free. Demis and fellow researcher John Jumper were awarded the 2024 Nobel Prize in Chemistry for this feat.

Countdown to AGI: Responsible Stewardship Is Urgent

Since the advent of large language model (LLM) chatbots like ChatGPT, generative AI has, in just three years, completely rewritten the division of labor in programming and creative work. Now, LLM-based AI products like ChatGPT, Gemini, Grok, and others are allowing ordinary people to experience the impact of AI firsthand. Next, the era of AGI is imminent—a watershed for humanity.

Demis says technology is neutral, but how humans use it determines good or evil. He once demanded that Google promise never to use DeepMind’s technology for military surveillance and emphasized that a “move fast and break things” attitude isn’t acceptable. He believes AGI is too powerful—if it ever gets out of control, the consequences would be unimaginable.

Hassabis warns: “AGI is about to be born, and our next generation will live in a brand-new world. With AI, everything will be different. If you want to manage AI responsibly, every moment counts. I’ve lived my whole life for this moment.” This sense of urgency reflects DeepMind’s deep awareness of AGI’s potential risks. Just as fire can be used for cooking or destruction, AGI can solve humanity’s greatest challenges—or bring unprecedented risks.

As we stand at the critical juncture of generative AI’s explosion and look toward the rise of AI agents, now may be the best time to review the journey of AGI development and reflect on the future.