[KongchangAI]
· 1 min read· 778 words

Why Building Math Muscles Still Matters in the AI Era: An MIT Professor's Deep Dive

Why Building Math Muscles Still Matters in the AI Era: An MIT Professor's Deep Dive

MIT professor Urschel argues that math fundamentals are not obsolete in the AI era—they are more critical than ever.

Based on a Q&A with MIT Assistant Professor John Urschel, this article explores the value of math fundamentals in the AI era. Urschel introduces the concept of "math muscles"—cognitive resilience built through repeated training in breaking down structures, building models, and verifying assumptions—rather than calculation speed or formula memorization. The article notes that AI lowers the threshold for getting results but raises the threshold for judging them. The long-term returns of math training lie in the cross-domain transfer of logical reasoning, quantitative analysis, and abstract modeling, which are shifting from bonus points to essentials. For learners, the reasonable stance is to use AI to aid understanding and check reasoning, while completing core derivations and proofs independently.

Q&A: Why building math muscles remains important in the AI era

AI tools are reshaping every industry. Calculators have long been able to handle complex computations, and symbolic computation software can derive formulas. Against this backdrop, is it still meaningful to spend years practicing hand calculations, proofs, and derivations? MIT Assistant Professor John Urschel gives an affirmative answer in a Q&A: mathematical fundamentals are not obsolete—they have actually become more critical in the AI era.

What Exactly Are "Math Muscles"?

The "math muscles" Urschel refers to are not simply calculation speed or formula memorization, but a kind of cognitive resilience forged through repeated training—the ability to break down the structure of unfamiliar problems, build models, and verify hypotheses.

Just as athletes need strength training to maintain physical condition, math training shapes the brain's "underlying fitness" for handling abstract problems. This fitness cannot be obtained by clicking an AI's "generate" button, because it relies on the cognitive traces left by long-term active thinking. Urschel himself is both a mathematics researcher and a former professional football player, so his use of the "muscle" metaphor clearly carries personal experience.

Behind the "math muscles" metaphor lies the schema construction theory in cognitive science. A schema refers to the structured knowledge network formed in the brain after repeated exposure to similar problems. The core difference between experts and novices is not working memory capacity, but that experts possess richer, more tightly organized schemas that allow them to quickly recognize problem patterns and invoke the appropriate strategies. Math training is a typical scenario for schema construction: each time a learner completes a proof or derivation, they reinforce the connections among "conditions—transformations—goals." In addition, the retrieval practice effect in psychology shows that actively retrieving information from memory consolidates long-term memory more effectively than passive reading. Hand calculations and derivations force the brain to perform high-frequency retrieval, and this is precisely the cognitive mechanism that makes "muscles" hard for AI to replace.

Can AI Tools Replace Math Training?

AI can quickly produce answers, but judging their correctness, analyzing their applicable conditions, and extracting insights from them still require humans. A person without math training, when faced with an AI-generated derivation, can hardly distinguish which step is a reasonable approximation and which step is a hidden assumption.

In other words, AI lowers the threshold for "getting results" but raises the threshold for "judging results." As tools become more powerful, the capability gap between users and tools actually amplifies risk—the less you understand, the more easily you are persuaded by seemingly professional errors.

When large language models generate mathematical derivations, they are essentially predicting the next most likely token sequence rather than truly performing logical deduction. This means their outputs may contain seemingly plausible but actually incorrect reasoning steps, the so-called "hallucinated reasoning." Models often produce the correct final answer while introducing imprecise approximations or unfounded assumptions in intermediate steps—a phenomenon especially prominent in proof problems requiring multi-step deduction. Judging whether such outputs are reliable requires users to identify logical leaps, check boundary conditions, and trace the sources of assumptions. In other words, the quality of AI output depends on the user's verification ability, and that ability comes precisely from the accumulation of training in completing derivations independently. Without such training, AI's "fluent output" can instead create a dangerous illusion of authority.

The Long-Term Returns of Learning Math

Urschel emphasizes that the value of math training lies not in immediate application, but in the way of thinking it shapes, which transfers to other domains. Logical reasoning, quantitative analysis, and abstract modeling continue to play a role in research, engineering, finance, and even everyday decision-making.

In the AI era, the information environment is more complex, and data-driven decision-making is increasingly common. Being able to understand the assumptions behind models, identify statistical traps, and assess the reliability of conclusions—these abilities are shifting from "bonus points" to "essentials." And the starting point for these abilities is a solid foundation in mathematics.

Practical Advice for Learners

When facing AI tools, math learners need not reject them, nor should they depend on them. The reasonable stance is: use AI to aid understanding of concepts, check lines of reasoning, and broaden perspectives, but the core derivation and proof processes still need to be completed by oneself.

Urschel's view points to a simple conclusion: the more powerful the tools, the more users need the ability to judge them. Training math muscles is essentially laying the foundation for this judgment. In this sense, the AI era is not the end of mathematics but a demand for higher mathematical literacy.

(This article is based on a Q&A with MIT Assistant Professor John Urschel.)

#math muscles#AI era math education#John Urschel MIT#math fundamentals#AI tools and learning#schema construction#verification ability
Share:

Related articles