As AI coding assistants and general-purpose LLMs become standard infrastructure for developers, a parallel debate is emerging in academia. A recent discussion on MathOverflow titled "AI and Training PhD Students" argues that assigning problems to graduate students remains essential, even if AI can technically solve them. The core argument is that the value of these exercises lies in the training process, not just the final answer.

The Calculator Analogy for Modern Research

The post draws a direct comparison to the widespread adoption of calculators in education. Just as students still learn manual arithmetic to develop number sense and critical thinking, PhD students need to wrestle with difficult problems to build resilience and familiarity with standard tools. The author notes that even the best AI models remain quite poor at high-level MathOverflow questions, suggesting the threat to human training may be overstated in practice, but the principle holds regardless of current model capabilities.

Incentives Need a Rewrite, Not Just a Patch

Where the author sees genuine disruption is in academic reward structures. The current model, which prioritizes being "first" to a result, is described as flawed and prone to shenanigans, such as claiming results in remarks without providing proofs. The proposal is to shift evaluation criteria toward understanding, clear communication, and the ability to contextualize work within broader theory. This mirrors a shift we often see in engineering teams moving from lines-of-code metrics to code quality and system reliability.

Thurston’s Warning About "Cleaning Out" Fields

The discussion cites William Thurston’s essay "On Proofs and Progress in Mathematics" to illustrate the danger of automated solution generation. Thurston observed that after he rapidly solved many open problems in foliation theory, researchers fled the field, mistakenly believing the interesting work was done. The author worries AI could accelerate this "evacuation" by producing correct but unilluminating proofs, discouraging humans from exploring areas where the real intellectual value lies in the journey, not the destination.

Key Takeaways

  • AI should be treated like a forklift in a gym: useful for moving heavy loads, but useless for building your own muscles.
  • Academic incentives must shift from "first to solve" to "best to explain and expand."
  • Correct proofs that teach nothing can actively harm a field by making it seem barren to new entrants.

The Bottom Line

If you are building AI agents for research, optimize for generating insight, not just closing tickets. A solution that kills curiosity is a regression.