Interview
DeepMind's Pushmeet Kohli on AI's Scientific Revolution
- AI is expected to transition from passive question-answering to actively formulating queries, evolving into a scientific method that redefines the scope of solvable problems rather than merely accelerating existing discovery.
- Systems like Alpha Evolve will discover entirely new algorithms and human-interpretable code that outperform expert-designed solutions, operating across programming languages such as C++, Python, and Verilog with fewer function calls and without requiring researcher-provided templates.
- Future advancements will leverage coupled large language models and evaluators to uncover hidden mathematical truths, unsolved decades-old problems, and unrecognized symmetries, with performance scaling driven by inference-time and test-time compute alongside model architecture.
- AI will drastically reduce protein target understanding time in drug discovery from years to weeks or months, while simultaneously accelerating progress in healthcare, material science, energy (including fusion and room-temperature superconductors), and coding.
- Multi-agent systems, including configurations like Co-Scientist, will extract deeper insights from the "tail of the distribution" over extended processing periods of up to days, eventually reaching days of processing as computation increases.
- The primary bottlenecks for future adoption involve validating digital results in the physical world and ensuring technology accessibility, alongside the critical development of calibrated uncertainty to signal model confidence and potential errors.
- Human roles will shift toward defining sophisticated optimization goals such as chip efficiency, fault tolerance, and cooling requirements, while robotics is projected to see significant medium-to-long-term impact, with humanoid robots leveraging infrastructure designed for the human form.
- The impact of tools like Alpha Evolve is anticipated to mirror AlphaFold by advancing fields, accelerating work, and democratizing access, leading to a future where Nobel Prizes are increasingly awarded to teams where AI is an indispensable collaborator.
- Gemini models will improve the effectiveness of sampling and searching for solutions to hard mathematical and computational problems, while work on generator-evaluator and multi-agent architectures will continue to scale and evolve.