The questions we ask

Sat, 12 Sep 2026 17:54:29 +0100

Earlier this week, OpenAI announced that an internal model had produced a solution to the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems and a question that has remained unresolved for roughly 90 years. The result itself is remarkable, but I found the description of how they got there almost as interesting.

The system involved approximately 10,000 concurrent agents, which exchanged 2.7 million messages and generated approximately 130 billion output tokens. But the researchers did not simply give the system the problem and leave it to run; they tried different formulations, asked the agents to work on easier, related problems and moved resources between different lines of investigation. When the system produced a promising result for the Euler equations, the researchers recognised the significance and shifted effort towards Navier-Stokes, feeding the result back into it.

The system produced the proof while it was humans that played the important role in deciding what was worth exploring next. This process is an example of the role human judgement can play alongside increasingly capable models. My broader view is that this kind of judgement will become more valuable as producing answers becomes cheaper.

I remember my English teacher at secondary school telling my class that we tend to celebrate the people who answer difficult questions, but not necessairly those who ask the good ones. I’ve thought about the idea on and off ever since. As models become better at producing answers, the value of knowing what to ask becomes easier to see.

Knowing what to ask

For most of history, the ability to ask a difficult question and the ability to answer it have been closely connected. If you wanted to solve a hard problem in, say, mathematics or engineering, you generally needed enough expertise to both formulate the problem and do much of the work required to solve it.

AI is beginning to separate those two things (one might argue it already has). Models can already write code, analyse data, search large bodies of information and explore possible solutions far faster than an individual can. As those capabilities improve, the cost of producing candidate answers will continue to fall.

Lowering the cost of producing answers does not necessarily make expertise less valuable. It may change where the value of expertise sits. If answers become much cheaper and easier to produce, more of the difficult work moves towards deciding which questions are worth answering in the first place.

I expect this to matter most when choosing the wrong direction is expensive. If a model can answer a question in a few minutes, asking the wrong one does not cost much. But in research, engineering and other open-ended work, a poor choice can send someone down the wrong path for weeks or months. In these cases, knowing which question is worth pursuing becomes much more valuable.

This idea is not new. A paper from 1979 by Naomi Miyake and Donald Norman captures this well: To ask a question, one must know enough to know what is not known. It looked at how prior knowledge affected question asking while people were learning. When the material became difficult, people with more background knowledge asked more questions, while those with less background knowledge asked fewer. The researchers argued that without enough knowledge of a subject, it can be difficult even to identify what you do not understand.

The finding makes intuitive sense. Deep knowledge gives you a mental model of a domain against which you can notice gaps, inconsistencies and assumptions that might be wrong. Someone new to a subject can ask questions about what they do not understand, but it is much harder for them to identify what everybody else might overlook.

Research on expertise suggests something similar. In a well-known study of physicists, novices tended to group problems according to their surface features, while experts grouped the same problems according to the underlying physical principles needed to solve them. Expertise changes how a problem is represented before any attempt is made to solve it.

I think this distinction matters when applied to AI. Two people can have access to the same model, but the person with 20 years’ experience in a field may be able to direct it towards questions that would never occur to somebody encountering the subject for the first time. They may know why previous attempts failed, which assumptions tend to break down, when a result looks misleading and which apparently minor observation is in fact worth pursuing.

In that sense, a good question can contain a surprising amount of accumulated knowledge.

More than prompting

It’s worth stating that this is very different from being good at instructing a model. As models get better at interpreting vague instructions, the way we phrase things is likely to matter less. The skill more likely to stand the test of time is knowing which problems to give them and how to judge the results.

The researchers working on Navier-Stokes were making decisions of this kind. Is there an easier related problem that might reveal something useful? Which of several approaches look most promising? Does this intermediate result change where we should spend our finite compute? What information should one group of agents receive from another?

Questions like these depend on good judgement, and much of this judgement comes from understanding the domain well enough to recognise what matters.

There is, however, an obvious objection here. Models will get better at asking questions too. They can already generate hypotheses, suggest follow-up experiments and identify possible gaps within an argument. I also expect they will get better at judging which questions matter.

AI will also make domain knowledge easier to acquire. Someone new to a field can use a model to explain unfamiliar concepts, summarise previous work and surface questions they might otherwise have missed. I expect this to narrow the gap between newcomers and experts, but not eliminate the advantage that comes from having spent years working in the field.

We naturally celebrate the answer because it is the visible result. The Navier-Stokes proof will receive far more attention than the sequence of decisions which guided the system towards it. However, as models continue to become much better at producing answers, I think the ability to identify worthwhile questions will be increasingly important.

The people who benefit the most from these systems may not simply be those who know how to get an answer from them. They may be the ones who have spent enough time understanding a domain to know what is worth asking.

Important note: these are my current views, looking roughly 12 months ahead. Beyond this timeframe, things get hazy. Given the pace of progress, my opiniion may age quickly and I may very well change my mind. 🙂

Research on ./wiigg

The questions we ask

Knowing what to ask

More than prompting