Timothy Gowers and Peter Sarnak credit large language models with substantial ability but see limits when genuinely new ideas are required. The systems are strong at combining known methods.
Timothy Gowers argues that current models are very good at connecting known techniques and trying many solution paths in parallel. What they lack is intuition for the few viable routes within a very large search space.
Peter Sarnak frames the same limit differently: a system can derive results from existing theory but fails to develop, from an elementary question, the abstractions on which major proofs rest. DeepMind researcher Tom Zahavy calls this bottleneck the invention of new premises for which no linguistic template exists.
For operational use this is a useful dividing line. Tasks that consist of combining, reformulating and checking known patterns run reliably. Tasks that must first define a new problem do not – regardless of how the model scores on leaderboards.
What this means for decision-makers
- Sort planned AI initiatives by whether they apply known patterns or must define new questions; only the first group is plannable today.
- Expect language models to accelerate execution, not strategy formation.
- Keep expert review in place for outputs in areas without a clear correctness test.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made