Demis Hassabis on the Path to AGI, Scientific Discovery, and the Future of AI

Google DeepMind CEO and Nobel Laureate Demis Hassabis discusses the breakthroughs, challenges, and future vision for artificial general intelligence (AGI) and its transformative impact on science and creativity.

The Nobel Moment and DeepMind's Role at Google

Receiving the Nobel Prize was a "surreal" moment for Demis Hassabis, a culmination of a lifelong ambition to use AI for scientific acceleration. The recognition for AlphaFold's impact on protein folding was a landmark event, not just for his team but for the entire field of AI.

Today, Google DeepMind acts as the "engine room" for Alphabet. The division, comprising around 5,000 predominantly engineers and PhD researchers, is responsible for core AI models like Gemini. These models are now integrated across Google's products, from Search to Workspace, impacting billions of users.

Genie: Building Interactive Worlds from Text Prompts

A stunning demonstration of progress is Genie, a generative model that creates interactive environments from a single text prompt. Unlike traditional video games built on pre-programmed physics engines like Unity or Unreal, Genie's worlds are generated on the fly from its understanding of intuitive physics, learned by watching millions of videos.

This capability is a critical step toward AGI. For an AI to be truly general and useful in applications like robotics or smart glasses, it must understand the dynamics and physics of the world around us. Genie represents a foundational "world model" that brings this understanding closer to reality.

The Future of Robotics and an "Android-like" OS

The path to practical robotics is being paved by multimodal models like Gemini. Hassabis describes systems where you can simply instruct a robot with language, and it will interpret that command into physical motor movements. This is powered by the model's real-world understanding, not just robotic-specific programming.

Google DeepMind is exploring a dual strategy: a vertically integrated approach with specific robot designs and a broader "Android-like" play to create a universal operating system layer for robotics. Hassabis believes the humanoid form factor will be crucial for navigating a world designed for humans, though specialized robots will dominate industrial settings.

He predicts a "wow moment" for robotics is coming in the next few years, with millions of robots eventually helping to increase societal productivity once the algorithms and hardware mature in tandem.

AI as the Ultimate Tool for Scientific Discovery

The primary driver for Hassabis's work remains using AI to accelerate scientific discovery. From AlphaFold and material design to controlling fusion reactors, AI is already tackling complex scientific problems. However, a key ingredient is still missing: true creativity.

Today's AI can prove given theorems but cannot yet formulate new conjectures or revolutionary theories like Einstein's special relativity. Hassabis defines this creative, intuitive leap—the ability to spot patterns and draw analogies across disciplines—as a fundamental test for AGI.

He pushes back on the idea that current models are "PhD-level," noting that while they possess some PhD-level capabilities, they still lack the general, consistent reasoning across domains that defines true intelligence.

Democratizing and Supercharging Creativity

Tools like Imagen 3 and Veo are revolutionizing content creation. They democratize complex tasks like image and video editing, allowing anyone to create through simple instructions rather than years of learning complex software. Simultaneously, they supercharge top professionals, enabling directors and artists to iterate 10x or 100x faster.

Hassabis envisions a future of entertainment with elements of co-creation, where audiences can dive into and shape worlds conceived by visionary creators. This represents not just better tools, but potentially a new art form.

Isomorphic Labs: Revolutionizing Drug Discovery

Building on the AlphaFold breakthrough, Isomorphic Labs is developing a suite of AI models to tackle the entire drug discovery process. The goal is to reduce the timeline for discovering new drug candidates from a decade down to mere days or weeks by designing compounds with high efficacy and minimal side effects.

The approach uses hybrid models that combine probabilistic learning from data with deterministic rules of chemistry and physics. This ensures the AI's predictions are not just statistically likely but also physically plausible. Partnerships with pharmaceutical giants like Eli Lilly and Novartis are underway, with programs in cancer and immunology expected to enter pre-clinical phases soon.

Addressing the AI Energy Challenge

While the energy demand of large-scale AI training is significant, Hassabis highlights a parallel trend of drastic efficiency gains. Techniques like distillation, where a smaller model is trained to mimic a larger one, have improved serving efficiency by 10x to 100x in just two years.

He argues that the net effect of AI on energy will be positive. The energy used by AI systems will be far outweighed by the gains in efficiency they unlock—from optimizing electrical grids to designing new energy sources and materials.

The Next Decade: A New Golden Age of Science

Looking ahead 10 years, Hassabis believes we will have achieved AGI. This will usher in a "new renaissance" or "golden era for science," where AI acts as humanity's ultimate partner in solving its greatest challenges, from disease and energy to understanding the universe itself.