AI reshapes scientific discovery with Pushmeet Kohli

DeepMind's head of AI for Science, Pushmeet Kohli, has sparked a revolution in how we understand scientific discovery. In a recent interview, Kohli outlines how artificial intelligence is fundamentally changing scientific research across disciplines—from protein folding to material science. This transformation isn't just accelerating existing research methods; it's creating entirely new approaches to solving humanity's most complex scientific challenges.

Key insights from Kohli's perspective:

• AI as scientific collaborator: AI systems like AlphaFold aren't just tools but active participants in scientific discovery, capable of generating novel hypotheses and approaches humans might miss.

• Breaking disciplinary boundaries: The most exciting breakthroughs happen at the intersection of AI and multiple scientific domains, creating opportunities for cross-pollination that traditional research structures often miss.

• Democratizing scientific progress: AI models that condense specialized knowledge make scientific exploration more accessible to researchers from diverse backgrounds, potentially accelerating discovery rates.

• Human-AI scientific symbiosis: Rather than replacing scientists, AI augments human creativity and intuition, handling computational complexity while humans guide investigation with contextual understanding.

The emergence of a new scientific method

Perhaps Kohli's most profound observation is that AI represents an evolution in the scientific method itself. For centuries, the approach of hypothesis formation, testing, and refinement has been relatively stable. Now, AI introduces what might be considered a fourth paradigm of scientific discovery—where models can analyze vast datasets, identify patterns beyond human perception, and generate novel hypotheses at unprecedented scale.

This matters immensely in our current context. Many of humanity's most pressing challenges—climate change, disease, sustainable energy—require processing information at scales and complexities that overwhelm traditional approaches. As Kohli notes, protein folding was considered a computational problem that might take centuries to solve, yet DeepMind's AlphaFold effectively solved it in just a few years. This suggests similar breakthroughs might be possible across other seemingly intractable scientific domains.

Beyond the interview: real-world implications

While Kohli presents compelling examples from DeepMind's portfolio, other groundbreaking applications deserve mention. Consider Climate TRACE, which combines AI with satellite imagery to track global greenhouse gas