How Do Mathematicians Tackle Hard Problems? Thinking Methodologies from Lex Fridman's Conversations

The Underlying Logic of Problem-Solving: From "World Models" to "Students in the Driver's Seat"

In this episode of Lex Fridman's conversation, several mathematicians/theoretical computer scientists rarely dissected their own thinking processes. There's no mystical "moment of inspiration"—rather, it's mostly structured cognitive strategies. One interviewee mentioned that he relies heavily on world models for thinking: first forming a mental model of results in a certain subfield, then using that model to "color" surrounding fields, thereby building bridges between algorithmic thinking and other problems.

This kind of cross-domain transfer isn't a gift—it's the result of deliberate design. He admitted that when he first became an advisor, he made a mistake—giving students problems that were 100% within his own field. Students would come back with ideas, and he could shoot them down within seconds, whether right or wrong, and neither party enjoyed the process. Later, he "deflected the problems by 10 to 20 degrees": giving students papers that he wanted to understand but didn't have time to read, letting the students teach him in return. The key is reverse-engineering "why on earth does this thing even work" rather than printing through a paper line by line. Once you grasp the one or few key ideas that shift your perspective, your own organizational structure will grow, and you'll discover new solvable problems.

This experience points to a counterintuitive conclusion: put students in the driver's seat as early as possible, even if the problem isn't clearly defined yet, letting them be responsible for distilling, digesting, abstracting the essence, and using analogies that "don't pursue literal truth" to infer what might happen in adjacent fields. Technology's role here is to expand the set of accessible problems—previously you needed to hold tightly onto an expert who could explain the relevant field in your language, but now models can do something similar, albeit losing some of the human warmth.

World models, in cognitive science and artificial intelligence, typically refer to an individual's internal representation of how the world works. In the context of problem-solving, what mathematicians call a world model is a mental simulation of the core concepts, methods, and typical problem structures of a subfield. It differs from a checklist of factual knowledge—it's an operational "mental sandbox": you can run simulations in your head, predict whether an idea might work, and judge which paths are promising. Having a high-quality world model means you can quickly identify the essence of a problem and draw analogies to other fields. This model is gradually constructed through deep reading, repeated practice, and discussions with others, and is the foundation of cross-domain transfer.

The Mismatch of Timescales: Multi-Year Vision vs. Three-Month Anxiety

A sharp conflict is brought directly to the table in the conversation. Advisors are used to thinking about projects on a multi-year timescale: there's a meta-vision (e.g., how to apply learning theory techniques to quantum many-body systems), with many stepping stones along the way, and the most exciting problems often lie years in the future. But students are anxious about the "AI tidal wave," the future of work, universal basic income—they think in cycles of three to six months. When the two world models don't match, it's hard to find problems that both parties care about.

When asked whether he should encourage students to think long-term, the answer was honest: "I do, but they don't always listen." This is immediately followed by a harder judgment: only work on "future-proof" problems—the ones you'll still care about years from now. If you're working on a problem only because you're worried the next generation of models will solve it, you might as well go on vacation and not do that problem. This view is worth pausing to think about for anyone choosing a direction in the AI wave: are you solving problems, or are you racing against the iteration speed of models?

"Future-proof" problems refer to research directions that won't lose their significance due to technological iteration or short-term trend changes. The criterion is: even years from now, you'll still care about the answer to the problem itself, not just whether it brings publications or competitive victories. In the context of rapid AI development, many problems may be automatically solved by the next generation of models, but truly fundamental problems—such as understanding the nature of learning, the mathematical structure of complex systems—won't be devalued because of this. Choosing future-proof problems is essentially anchoring research in enduring human curiosity and the core contradictions of the discipline, rather than racing against the iteration speed of models.

Focus and Peripheral Vision: The Balancing Act of "Peripheral Vision" in Engineering Problems

Another interviewee (with a background in fluid dynamics) described the constant tug-of-war between two opposing states. One is extremely narrow engineering focus: putting on blinders, staring at one problem, not letting go, not relaxing—several people in the conversation have had problems they've held onto for years or even decades. The other is giving yourself psychological freedom to look around, be curious, and stumble upon solutions while not looking directly, like turning over stones. In his words, looking at a problem with peripheral vision sometimes makes it easier to solve.

This balance isn't chicken soup—it has concrete origins. For him, the "messiness" of biology and materials science—multi-scale, nonlinear, fuzzy goals—shares the same structural affinity with fluid mechanics. The fundamental mathematics developed to handle complexity does indeed transfer. In other words, the criterion for choosing adjacent fields isn't "hotness" but "sharing the same kind of messy structure."

This also explains why the term "adjacent fields" keeps appearing in the conversation: it's not a vague interdisciplinary slogan, but an operational problem-solving strategy—first use world models to draw a map of the field, then deliberately stand 10 to 20 degrees outside the boundary, using analogy and reverse-engineering to approach the core ideas, while keeping a problem you won't let go of for years as an anchor.

Looking at a problem with peripheral vision is a cognitive strategy that means not staring directly and intently at the problem, but letting attention be in a diffuse state, allowing the subconscious to freely associate. In psychology, this is related to the "incubation effect": when we temporarily set a problem aside, the brain continues to process information in the background, and偶然 external cues may trigger a solution. In mathematics and engineering, this manifests as suddenly getting inspiration while walking, reading unrelated fields, or chatting with others. Deliberately maintaining psychological freedom and allowing yourself to be curious in a "turning over stones" manner often bypasses the mental set of direct focus and reveals non-obvious solutions.

Implications for Researchers in the AI Era

Putting these pieces together, we can extract a fairly pragmatic methodology: use world models to organize cognitive maps; turn the learning process into an output process of "teaching others"; bet only on problems you'll still care about years from now; switch back and forth between extremely narrow focus and broad curiosity; and when choosing adjacent fields, look at structural similarity rather than surface popularity. These strategies don't depend on compute power, nor do they become obsolete with model versions—in a sense, they are themselves "future-proof."

One-line Summary

Mathematicians share problem-solving methodologies: world models, field deflection, future-proof problems, and balancing focus with peripheral vision.

Paragraph Summary

Based on Lex Fridman's conversations, this article distills the thinking strategies of mathematicians and theoretical computer scientists. Core points include: using world models to organize cognitive maps, building bridges in adjacent fields through 10-20 degree field deflection; putting students in the driver's seat early and letting them reverse-engineer key ideas; working only on future-proof problems you'll still care about years from now, avoiding racing against model iteration; maintaining balance between extremely narrow engineering focus and broad peripheral curiosity; and when choosing adjacent fields, looking at structural similarity rather than popularity. These strategies don't depend on compute power—they are themselves future-proof.