The point of hard problems

When OpenAI announced that one of its internal models had proposed a solution to the Navier–Stokes Millennium Prize Problem, it sounded like a big deal. Out of curiosity, I read reactions from some of the smartest people I follow on X.

But a few days later, twenty-five Fields Medalists wrote a public letter about what they called a severe misalignment between AI companies and mathematicians. Their view was that the point of hard problems is not only the solution, but how we arrive at it—and that a solution also matters because of the system that surrounds it. Problems help train students, organize inquiry, generate new concepts, and connect one generation of mathematicians to another.

I would not have thought about mathematics this way. But I can relate to it in my own line of work.

The point of planning

As an engineering manager, I witness, participate in, and lead many product-planning cycles. There is planning of various kinds, at different intervals, but it tends to produce similar artifacts: product requirements, designs, technical plans, estimates, roadmaps.

Having been at the same company for about a decade now, I have seen the planning process change a few times. Through all that change, I have come to believe that the primary benefit of planning is not the outputs at the end but the thinking and collaboration required to create them.

One way to see this is to imagine deleting an artifact after all the work that went into it. It can often be recreated relatively easily because the good things that came before it—not just individually, but organizationally—are still there: the conversations it spawned, the relationships it formed, and the clarity it brought. That thinking and effort make the result easier to trust, explain, and stand behind. They also create ownership and accountability because people know not just what the decision was, but why it was made and that they made it.

Time required to plan still matters. But two common views about this seem wrong to me. The first treats reducing the time to create artifacts as the goal. Faster generation of artifacts does not necessarily produce faster understanding. The second claims that we should get rid of artifacts altogether so we can start building sooner. That also misses why the artifacts existed in the first place. Without intentionality, both approaches can remove thinking and collaboration from the process in favor of building fast.

With AI, we may have separated the artifact from the work that made it trustworthy. In the process, we may also have made another separation clearer: what the work represents to the business and what doing the work means to the individual.

Growing up in India, I often heard my parents say that I would value money only after I had worked hard to earn it. Their point was that effort changes our relationship with the result.

A mathematical proof, a product plan, and a piece of software are different things. But they can be mistaken in the same way. When we reduce an activity to its most measurable output, we can optimize that output while losing some of the activity’s value.

I see the same pattern in how we think about food. Once we reduce food to calories, cooking, cleaning, and sharing a meal can look like overhead. An instant drink or packaged meal may deliver the measurable output more efficiently while replacing the broader system that gives eating some of its social and personal value.

Means and outcomes

Part of the difficulty in processing this is that an activity does not have one fixed meaning. Even the same person can value different kinds of work for different reasons.

I love spending time alone reading. That is when my creative juices flow. I then write to shape those ideas. I run and bike because I love them, not particularly because of the health benefits. The experience itself—the effort, the solitude, the attention, and the perceived growth—is part of what I want.

Paperwork is different. I hate it, and I would happily use almost any reliable tool that gave me the result without demanding my time. Some household chores sit somewhere in between. I find them boring, but I like the physical effort they require and the way they keep my attention in the present.

So as I think about all this, activities, it seems, can be viewed from three different angles:

• What we get. The result: a clean room, a healthy body and mind, a working piece of software, a proof.
• What we experience. The pleasure, absorption, companionship, frustration, dignity, mastery, purpose, or belonging contained in doing the activity.
• What we become. The skill, judgment, character, confidence, relationships, and collective capacity that develop through repeated practice—and the identity, pride, and self-belief we form around them.

The value of AI is easy to see when it removes drudgery, such as the administrative work many of us would happily give away. It becomes difficult when removing the labor also removes part of the experience or meaning. In other words, it's easy to see and agree on the value of AI when all we care about is what we get. But not so much when it's entangled with what we experience and what we become.

Despite being married to a biking enthusiast, my wife sometimes wonders why anyone would bike up a mountain when one could drive to the top and enjoy the same view. From her perspective, the summit is the destination. For me, climbing is why I do it. I may not even stop to enjoy the view at the top.

I make the same mistake. She can spend hours talking on the phone, and I sometimes wonder what “better” thing she might have done with that time. But that is because I misunderstand the point of small talk and gossip. The value is in the act of conversation. Sometimes, what it is about almost does not matter.

In both cases, the person observing from outside mistakes the purpose of the activity.

When differences are allowed and when they are not

Our different preferences and worldviews do not seem to be problems in and of themselves. We do not need to fully understand what motivates each other, or why. We only need to recognize that those motivations matter and make room for each other’s choices.

But differences of opinion around something as influential as AI play out differently.

AI is not merely a tool that individuals use however they choose. It is also a tool that organizations allocate to individuals and govern through policies. They can use it to redefine work for everyone. Work that people relate to as a craft, a source of judgment, a collaboration, or part of their identity can be reclassified from above as output-producing factory. In most cases, the people deciding have different incentives from the people who must live with the consequences of those decisions day in and day out.

A company benefits from producing good software faster. Leaders are responsible for keeping the business competitive. But the outputs easiest for an organization to count—documents completed, features shipped, cycle time reduced—are not necessarily the only things that contribute to the success of the organization over the long term.

When designers are encouraged to generate code, product managers to generate designs, and engineers to review and ship all of it faster, the boundaries between disciplines appear to dissolve. That democratization is liberating. But not always in the most intuitive way. Rushing to collapse those boundaries with only a superficial understanding of another function can move unfinished thinking and unintended consequences downstream. Agent-generated code may look complete to its author and look like technical debt to the engineer accountable for operating it.

Each discipline sees nuances that are difficult to perceive from outside because those nuances were acquired through practice. The so-called taste and judgment of an expert are often the accumulated experience of years of thinking, effort, failure, and attention.

So the choices are not between giving specialists the exclusive authority to decide who contributes to their domain and breaking open the floodgates. The choice is how to strike the right balance: how to widen access without giving up quality, craft, judgment, or sanity. And doing that requires intentionality.

Both sides are true

The hard part about this debate is that both sides are right because they are optimizing for different things. The mathematicians’ word for that is apt in that sense: misalignment.

AI has undoubtedly lowered barriers to writing, design, programming, and pretty much all other kinds of knowledge or creative work. For someone who could not previously do those things, that feels like freedom.

The experts can also be right that a shiny output may still be mostly fluff. At the same time, they may also be protecting identity, authority, or status. Human motives are rarely separable.

Work offers people mastery, community, dignity, and a story about who they are. At the same time, personal and emotional reasons cannot by themselves justify preserving an inefficient system or denying useful capabilities to others.

That is why the debate is less about what AI can or cannot do and more about what we should and should not do with it, and how we adapt to it, because it is a force that is going to change how we work across every possible domain of knowledge or creative work.

What must remain when output becomes cheap?

We do not seem to be well equipped for this transition. It is happening faster than any individual or system can make sense of it. That is why Dario Amodei’s phrase “pace the frontier” feels like a good way to put it. Pacing gives people and institutions time to think, decide what they want the technology to mean, and adapt with some deliberation.

I do not think anybody knows what matters when outputs become cheap. It's all conjectures at this point, and it will take time to answer that in the broader sense. But it is something each of us can ask ourselves now.

Context and sources

1. “In Defense of Strategy”, Packy McCormick’s essay on the relationship between strategy and execution.
2. “A Severe Misalignment of AI in Mathematics”, the declaration published by Terence Tao and signed initially by twenty-five Fields Medalists.
3. OpenAI’s account of its proposed Navier–Stokes solution. This draft uses “proposed solution” because the announcement was recent and independent evaluation was still developing at the time of writing.
4. “The Mathematicians Rebel Against AI”, Marginal Revolution’s commentary on the mathematicians’ declaration and the debate around adapting to AI.
5. “We Must Pace the Frontier”, Dario Amodei’s proposal for balancing continued frontier-AI development with the time and safeguards needed to manage its risks.
6. Dario Amodei’s original post introducing “pace the frontier”.