DeepMind's Pushmeet Kohli: No "move fast, break things" approach to AI
At the Hindustan Times Leadership Summit 2025, Pushmeet Kohli, Vice President of Science and Strategic Initiatives at Google DeepMind, underlined a simple stance: build AI responsibly, for real impact, without the "move fast, break things" mindset.
Fifteen years in, DeepMind's bet is clear. Treat intelligence as a scientific problem, push the boundaries of human knowledge, and pick problems where AI changes how work gets done - not in small steps, but in leaps.
Science-first, impact-focused
"The organisation has science embedded in its DNA," Kohli said. That shows up in projects like AlphaFold, which reframed protein structure prediction and opened up new paths in drug discovery, enzyme design, and agriculture.
Before AlphaFold, it could take years to determine a single protein's structure. Now, researchers can explore high-confidence predictions at scale. If you work with proteins, the AlphaFold Protein Structure Database is a practical starting point.
General models, specialized systems, and efficiency
On model strategy, Kohli drew a line between broad LLMs and targeted systems. The goal isn't to apply one model to everything; it's to build the most competent systems for the hardest problems, and measure intelligence by how quickly and reliably a model accomplishes a task.
Less data and less supervision are key priorities. DeepMind continues to advance specialized systems (like AlphaFold) while also improving general models. Kohli pointed to the Gemini family as evidence of progress across a range of tasks.
Trust, uncertainty, and failure modes
Can scientists trust AI outputs? "AI is still making some mistakes," Kohli noted. The critical skill is knowing when the model might be wrong - and surfacing that early.
AlphaFold, for instance, communicates uncertainty so researchers know where to be cautious. For modern LLMs, DeepMind is building tools to flag hallucinations and warn users. This mindset will matter as teams push into energy modeling, material discovery, and other high-stakes domains.
What this means for your lab
- Prefer systems that expose confidence estimates. Treat model outputs like any instrument reading: validate with controls and orthogonal methods.
- Track data efficiency. Models that learn with less data and supervision reduce labeling overhead and speed iteration.
- Evaluate on time-to-solution and calibration, not just benchmarks. Stress-test reasoning, generalization, and failure detection.
- Explore agentic workflows for experiment planning, literature triage, and protocol execution - with audit trails and safety checks.
- Expect momentum in structural biology and healthcare. Kohli highlighted strong adoption - including 180,000 users of AlphaFold in India - as a signal that access matters.
Looking ahead
Kohli expects acceleration in structural biology, healthcare, drug discovery, and the rise of agentic systems that can take on more tasks end-to-end. A major theme is democratisation - getting high-quality tools into more hands, across more regions.
The ethos is steady: build competence, prove reliability, and scale responsibly. That's how AI becomes a force multiplier for science - and how labs avoid costly dead ends.
Resource: Explore structures and confidence metrics in the AlphaFold Protein Structure Database.
Upskilling: For researchers building AI fluency across roles, see our curated options by job at Complete AI Training.
Your membership also unlocks:
