Hassabis: "AI is the Ultimate Tool for Science" – Advancing Towards AGI

Demis Hassabis, the founder of DeepMind and recipient of the Nobel Prize in Chemistry, shared his insights on the role of AI in science and his vision for the future in a recently published interview. Hassabis views AI as “the ultimate tool to push the frontiers of knowledge,” and elaborated on its potential applications in addressing fundamental problems in physics and life sciences.

The Roots of a Scientist

Hassabis identifies himself as “first and foremost a scientist.” Early in his career, he worked as an engineer, developing game AI and studying computer science at Cambridge University. However, it was during his doctoral studies in cognitive neuroscience, where he researched memory and imagination, that he realized the scientific method of hypothesizing and testing through experiments suited him best.

Despite his scientific mindset, Hassabis maintains a practical, hands-on approach to constructing tools and experiments for hypothesis testing, which he attributes to his “engineer’s brain.” He cites AlphaGo and AlphaFold as prime examples of this way of thinking.

AI as the Ultimate Tool

Hassabis highlights humanity’s unresolved fundamental scientific questions, such as the nature of time, the essence of gravity, and its unification with the quantum world. He notes that little progress has been made on these issues over the past 50 years, largely due to their extreme complexity.

He argues that AI could serve as the ultimate tool to tackle these challenges by aiding in the processing of immense datasets and proposing new hypotheses. Hassabis also posits a hypothesis that information might be the most fundamental unit in physics, even more so than matter or energy.

He points out that many systems in nature possess stable structures that neural networks can learn. AlphaFold, he explains, has demonstrated the capability of classical computers to simulate complex natural systems beyond traditional understanding. He even speculates that future classical systems may model the universe itself, potentially uncovering simpler foundational descriptions than existing quantum theories.

Compressing Knowledge and Narrowing the

Search Space

When discussing the limits of computation, Hassabis references concepts like Kolmogorov complexity, the P=NP problem, and the halting problem. He leans toward the belief that P≠NP but emphasizes that AI is opening new pathways for problem-solving.

Central to this is the compression of knowledge. By compressing large-scale pre-computed knowledge into neural networks, the search space during the testing phase can be narrowed, allowing near-optimal solutions to be achieved within polynomial time.

He cites protein folding and the game of Go as concrete examples. Proteins have around 10^300 possible folding configurations, while Go has about 10^170 possible moves—both numbers far beyond the scope of exhaustive search. AI has succeeded in solving these problems by learning patterns and focusing only on a small subset of possibilities, revolutionizing not just technology but also our understanding of reality itself.

How AlphaFold Transformed Life Sciences

AlphaFold solved the 50-year-old biological challenge of protein folding. By analyzing data on the structures of 150,000 known proteins accumulated over 30 to 40 years, it learned the rules of protein folding and achieved atomic-level precision in predicting structures.

AlphaFold can now output the structure of an unknown protein within seconds. The team has completed predictions for 200 million known proteins and made the results freely available to researchers worldwide. The latest version, AlphaFold 3, has advanced to predict interactions among nearly all major molecular entities.

Future goals include gradually expanding simulations to biological pathways, simple cells, organelles, and ultimately, entire living organisms. Once a virtual cell is perfected, computational experiments will far outpace traditional wet labs, fundamentally transforming biological research and drug development.

Criteria for Selecting Scientific Problems

Hassabis’ team selects scientific problems based on three criteria: the problem must involve a vast combinatorial space, have sufficient data available or the ability to generate data through simulation, and possess a clear optimization goal.

Many scientific problems fit this structure, and the team prioritizes solving “root node” problems that could open up new fields of study. Protein folding is a prime example of such a problem.

Even in fields like mathematics, which may appear to lack clear structures, Hassabis sees a richness of structure derived from real-world descriptions. He believes the intuition of mathematicians, built through long-term tacit knowledge, could ultimately be modeled by AI.

Two Current Weaknesses of AI

Hassabis candidly acknowledges two major weaknesses in current AI systems.

First is the absence of true deep creativity. Unlike human scientists, AI cannot yet pose genuinely valuable and original scientific questions, such as the Riemann hypothesis. Formulating the right questions remains one of the most challenging aspects of science, and in this respect, AI lags significantly behind humans.

Second is the lack of consistency in output. While AI can solve complex problems like International Math Olympiad questions, it might falter on basic elementary-level math problems if the question is framed differently. Such inconsistencies do not occur with human scientists.

Defining AGI and the Road to Realization

Hassabis defines Artificial General Intelligence (AGI) as a system possessing all the cognitive abilities of the human mind. Whether AGI can be achieved by merely scaling current models remains an open question. He suggests that a few more significant breakthroughs, akin to Transformer models or deep reinforcement learning, may be necessary.

The foundational models of today will likely form the core framework of future AGI, but Hassabis believes its realization will require both symbolic creativity and rigorous scientific validation. AI must not only propose valuable new theories in physics or mathematics but also invent elegant new games like Go, while undergoing relentless testing to ensure all its abilities are free of flaws.

Principles for AI Safety and Future Vision

Hassabis envisions AI as a powerful tool for scientists over the next decade, capable of increasing the efficiency of hypothesis generation by tenfold or even a hundredfold. In the long term, AI will gradually evolve into a true partner in scientific research.

He emphasizes the importance of responsible AI development, which requires strict adherence to the scientific method, rigorous testing and validation, and proactive consideration of secondary effects before deployment. Hassabis argues that the scientific method is better suited to managing the profound impact of AI than the “build first, ask questions later” culture often associated with hackers.

Hassabis highlights two core benefits he hopes AI will bring to humanity. The first is improving human health, particularly by overcoming diseases and extending benefits to underserved regions. The second is addressing climate and energy challenges through the development of new materials and technologies, optimizing infrastructure and energy systems, and modeling and analyzing ecological changes.

At the current pace of development, Hassabis hopes humanity can safely achieve AGI by 2050, benefiting all of humanity. This would usher in a new golden age of scientific discovery and a post-AGI civilization of unparalleled material and intellectual abundance.

Throughout the interview, Hassabis made it clear that he regards AI not merely as a technical tool, but as an entity fundamentally extending humanity’s intellectual exploration. The success of AlphaFold is a vivid testament to this vision, with AI solving the biological challenge of protein structure prediction and providing tangible evidence for his aspirations.