Demis Hassabis explained how Google DeepMind's AI is already changing science, climate, and technology
He acknowledges limits: some problems—such as the factorization of large prime numbers—could require quantum computers. Meanwhile, in physics, biology, and chemistry, current networks are already achieving a lot.
The CEO of DeepMind envisions a future where AI will enable the creation of unique and adaptive video games. Each player will experience a different story, with characters that evolve in real time.
Generative AI will make possible changing environments, improvised missions, and believable dialogues. "The experience will be co-created with the user," Hassabis said, who was already seeking this in the 90s, when he programmed games.
He also considers that video games are social and emotional laboratories, where one can fail, learn, and explore without real consequences.
For Hassabis, the great challenge of science with AI is to formulate relevant questions. Today, systems still can't do this well, even if they solve problems once posed.
"Choosing the right question is the most difficult thing," he stated. He also suggested that achieving true AGI (artificial general intelligence) will require these technologies to match human abilities in thousands of tasks.
He spoke of "lighthouse moments," such as inventing a new game like Go, or making an original scientific conjecture. That would mark that we've crossed the AGI frontier.
Demis Hassabis, CEO of Google DeepMind and Nobel laureate in Chemistry, anticipates a scientific revolution driven by AI. His developments are already impacting biology, climate, video games, and more.
In a conversation with the Lex Fridman podcast, the researcher detailed how artificial intelligence is beginning to solve challenges that once seemed impossible.
Hassabis highlighted that AlphaFold solved a key mystery in biology: the 3D structure of proteins. Its new version, AlphaFold3, already models interactions with RNA and DNA.
Another key development is Veo, capable of generating hyper-realistic videos that simulate liquids, textures, and physical laws with scientific precision. This tool enables machines to learn directly from videos.
In addition, the WeatherNext system outperforms traditional weather prediction models. Thanks to deep neural networks, it anticipates hurricanes and extreme phenomena with unprecedented speed and accuracy.
Hassabis proposes a bold idea: that any natural pattern, if it has structure, can be discovered and modeled by AI. This is based on the fact that evolution generates systems with order, not randomness.
"If natural systems have structure, then a neural network can learn it," he stated. According to him, information is the essential unit of the universe, even above matter or energy.
