The Answer Factory

After the Iron Man paradox: what happens when we solve problems faster than we grow the minds that solve them?

Imagine that tomorrow morning an alien lands at a mathematics department, hands over a correct solution to a celebrated open problem, and leaves.

The proof checks out. The alien does not take questions. Apparently, even advanced civilizations have calendar issues.

What, exactly, has the department gained?

A theorem, certainly. Perhaps a magnificent one. But suppose the alien has taken away every failed attempt, every revealing obstruction, every unexpected connection that led to it. The mathematicians have the destination. Much of the intellectual territory remains unexplored.

This is the uncomfortable possibility in Terence Tao ’s recent reflection on AI and the Navier–Stokes regularity problem: an autonomous system could resolve a famous question while keeping the process largely private, leaving the wider field with far less understanding than a more open, participatory effort might have generated. [Tao’s post].

Last month, in the Iron Man paradox, I explored how AI could amplify an experienced person’s capabilities while reducing the experiences that replenish them. Tao’s argument takes that concern to another scale.

What happens when an entire field becomes more productive while fewer people participate in the experiences that develop its next generation of judgment?

A hard research problem can produce several things simultaneously: a result, new methods, better questions and people capable of asking them. The result is the easiest to announce. The others may determine what becomes possible next.

Consider a failed approach. To a system rewarded for finishing, it can look like wasted computation. To a researcher, its failure can reveal a constraint that applies to an entire family of approaches. That constraint might become the useful result. It can travel to another problem, another discipline, another person.

Once the famous question is settled, however, exploring those abandoned routes becomes harder to justify. Attention moves. Funding moves. “We would like to reconstruct several instructive mistakes” is a difficult opening sentence for a grant application.

This is what makes Tao’s concern more consequential than nostalgia for the blackboard. A private discovery process can withhold benefits that would otherwise spread beyond the team producing the answer. The company gets a result; the research community may lose an occasion to learn.

That does not make an opaque solution worthless. A correct construction can itself provoke new ideas, and AI may help explain or simplify it. The question is how much understanding remains recoverable, by whom, and at what cost. A learning opportunity can be diminished without disappearing completely.

There is a revealing precedent in research on invention. Tian Heong Chan , Jürgen Mihm and Manuel Sosa examined more than 1.5 million US design and utility patents. Teams had advantages for inventions that could be divided into components; lone inventors could excel at more integrated inventions. Their collaboration histories mattered: prior work with others helped explain later success alone. [Research paper]; [INSEAD’s account].

The celebrated solo inventor may therefore be carrying an invisible team. Former collaborators survive inside the work as standards, questions, habits and ways of seeing.

That creates a possibility worth investigating in the AI economy. Today’s extraordinary solo performance may draw on intellectual resources accumulated under working conditions that tomorrow’s professionals encounter less often. We could mistake the successful use of an inheritance for proof that we have found a sustainable way to replenish it.

The patent research does not establish that AI causes such a decline. It shows why the history behind apparently independent performance deserves attention.

Nor is my earlier intuition about seniority a universal law. Erik Brynjolfsson , Li and Raymond’s customer-support research found larger AI benefits for novices and lower-skilled workers. In some settings, AI helps distribute expertise. In open-ended research and design, my hypothesis is that experience can improve several linked choices—framing, representation, evaluation—so that benefits compound across the work. “Exponential” remains a metaphor, not an established curve. [Generative AI at Work].

The sustainability question survives either way: what does using the system teach its user?

I see the tension most clearly when I run work in parallel. With one experiment, I can follow the important decisions. With several, I increasingly inspect summaries, assess risks and approve trade-offs. These activities require judgment. They also give me a different education from working through the mechanism myself.

Parallelism increases the number of decisions faster than it increases my capacity to digest them. Ten agents can produce ten streams of work. My nervous system remains stubbornly on the individual plan.

The danger is easy to miss because delivery keeps improving. Another prototype works. Another report appears. The next task becomes more tempting than understanding the last one. Time saved becomes time committed.

This can narrow both depth and breadth. I learn less about how my own solution works, while reading mainly what the current project requires. My questions become increasingly efficient at finding more of what I already know to ask for.

Yet analogy needs material from elsewhere. To connect a biological feedback loop to a control problem, I must encounter both and understand enough of their behavior to distinguish a useful correspondence from a decorative metaphor. Breadth supplies possibilities. Depth helps decide whether they hold.

Knowledge also needs sedimentation: occasions to revisit an idea, explain it, apply it and discover its limits. A summary can expose me to a concept. Using that concept under resistance can make it part of how I think.

This is where Anima Anandkumar ’s physics-first approach resonates with me. In the post that prompted this reflection, she describes using AI guided by physics to locate an approximate solution, then mathematical analysis to establish the necessary bounds. It offers a candidate behavior whose mechanisms and constraints can be investigated.

The scientific status matters. Tao’s September 7 discussion describes the Euler work by Adarsh Ganeshram , Valentin Duruisseaux and Anima Anandkumar as locating a numerically stable candidate, with a major rigorous stability argument still required. The candidate and a completed proof are distinct achievements. [Tao’s assessment].

Physics does not automatically make a neural network interpretable, and symbolic mathematics can illuminate mechanisms beautifully. What interests me is a research process that gives humans something meaningful to perturb, question and understand. The choice of AI method matters partly through the learning opportunities the surrounding process preserves.

There is, however, an objection that deserves a serious answer. If machines discover, verify and maintain the result, why must humans understand it?

In many situations, the result is enough for its immediate purpose. We already depend on technologies we could not reproduce individually. Requiring everyone to master every mechanism would end specialization shortly before lunch.

But we still need to ask where understanding resides and whether it remains accessible. If a provider disappears or a system encounters unfamiliar conditions, independent expertise and recoverable methods give us options. Dependence becomes more consequential when those alternatives erode.

Verification also has a scope. A formal proof establishes a stated proposition within a formal system. An engineering test establishes behavior under specified conditions. Neither automatically establishes that the proposition captures our intention, the tests cover every important situation, or the outcome is one we should want.

Even a world of perfectly reliable machines leaves us the question of purpose.

And purpose includes taste: elegance, beauty, eccentricity, the unusual solution someone cares enough to defend before it scores well. An optimizer can pursue all of these. Rationality does not require beige furniture. But if many organizations adopt similar models, objectives and evaluators, convergence can become the easiest outcome to reward.

There is an empirical warning here. In a short-story experiment, Anil Doshi and Oliver Hauser found that AI-generated ideas improved evaluations of individual stories while making the stories more similar to one another. This is evidence from a particular task, not a forecast for all culture. It nevertheless shows that improved individual output and reduced collective variety can coexist. [Science Advances].

The concern extends beyond whether machines can generate surprising things. Can we still develop and articulate preferences that resist the prevailing score? Will we exercise that capacity when approval is faster, safer and rewarded more consistently?

AI can help here too. It can introduce unfamiliar research, expose weak analogies and support demanding learning. But those benefits depend on how we organize the work. A workflow devoted to completion will not automatically preserve them.

For research, that means making consequential alternatives, failed experiments, assumptions and verification evidence available alongside the result. A polished retrospective is useful only insofar as it faithfully accounts for what happened. For individuals, it means choosing important decisions to understand deeply and spending some of the time saved on people and ideas outside the current assignment.

For teams and institutions, it means recognizing that a project develops people as well as artifacts. After a substantial effort, we should be able to point to something the participants can now explain, recognize or attempt that they could not before. Learning deserves evidence of its own.

The Iron Man paradox began with the person inside the suit. The next question concerns the community that teaches people to build, question and redesign the suits.

I want machines that help us discover things we could never discover alone. I want us to emerge from those discoveries with more ways to think, more freedom to choose and more capacity to disagree.

Otherwise, we risk becoming answer bees: exquisitely productive, endlessly busy, and increasingly dependent on someone else to decide which flowers matter.