Curiosity, not coding: what AI cannot replace in human thinking

Executive overview

AI tools improve at executing instructions, but cannot determine what to work on. The bottleneck shifts from doing to directing — whoever can describe the task, evaluate the output, and ask better questions becomes more valuable.

The job AI cannot replace is the ability to be curious and go after hard problems.

Staying relevant alongside AI

• Describing tasks clearly to AI is the core skill — bad programmers get replaced, great ones who can assess and fix AI output are in more demand
• Question everything; critical thinking matters more than memorising what AI already knows
• Follow genuine curiosity rather than chasing skills AI might make obsolete
• Use AI to accelerate learning, not as a substitute for forming your own questions

Building a mental map of unknowns

• Noticing gaps in your understanding — and holding them open — is how hard problems get solved later
• Growing up around factory machinery, Anandkumar tracked what she didn't know and connected it when new knowledge arrived
• That habit of mapping unknowns is more durable than any specific technical skill

Neural operators: AI that understands physics

• Traditional weather simulation requires supercomputers; neural operators replicate physical behaviour far faster
• A model trained on partial differential equations can run on a consumer GPU what previously needed a supercomputer
• The AI weather model outperformed traditional forecasts — months after experts predicted AI would need a decade to do so
• Key insight: ML is not constrained by what others consider difficult, only by whether data and method design are possible

Human agency in the AI era

• Human agency means deciding what tasks AI performs, evaluating its outputs, and steering it over time
• AI can amplify curiosity or kill it depending on how it is used
• Scientists tackle open problems by definition — harder problems remain, so the role doesn't disappear
• The real bottleneck in science is not new ideas but slow, expensive real-world testing; AI that encodes physical knowledge can replace those experiments