The end of the why loop: curiosity, AI, and what stays valuable
Dixon, R.J., 2026. The end of the why loop: curiosity, AI, and what stays valuable. [online] robdixon.ai, 15 September. Media. In brief. A ten-year-old taught at school that there are four states of matter asked an AI the same question. He came back with fifteen... eleven confirmed and four still debated. Nobody set him the task. For the first time in human history a child's "why" can run all the way to the end, and that is the clearest signal of what has actually changed: execution is becoming close to free, so the value moves to knowing why a thing should exist and judging whether it landed. That work is harder, not easier. This is the record of my guest appearance on episode 14 of the Better With Humans Podcast, hosted by Shaun Phillips, published on 15 September 2026, and the argument I made on it.
Shaun Phillips hosts the Better With Humans Podcast, where he sits down with people navigating what emerging technology is doing to creativity, to thinking, and to everyday life. He had me on as his guest for episode 14. We talked for forty-four minutes, and it turned into the clearest run at my own argument I have managed out loud.
So this is the record of that conversation. The episode is below, and it is the better version, because Shaun asks sharper questions than I would have thought to ask myself. What follows is the argument written out, for anyone who would rather read than listen.
We got to it through a wall that every parent hits. A child asks why, you answer, they ask why again, you answer again, and somewhere around the fourth or fifth round you run out of patience, or you run out of answers, and you say go and ask your mum. The loop closes. Not because the child stopped being curious, but because you stopped being able to feed it.
That wall has just moved, and I don't think we have begun to reckon with what it means.
Fifteen states of matter
I was driving my son Rehman to school. He's ten. No phones in the car, so we talk. I asked him what he was doing at the moment and he said the states of matter. I said I know about that, there are three, solid, liquid and gas. He said no Dad, there are four, you've forgotten plasma.
I wasn't sure. I didn't want to tell him something I couldn't stand behind, so at the traffic lights I quietly started typing it into Google. He spotted me. "Are you typing that into Google?" And then, with the particular relish of a ten-year-old: "I've already put it into ChatGPT. There are actually fifteen states of matter. Four of them are in the national curriculum. Eleven are confirmed by scientists and four are still debated."
Nobody set him that task. There was no homework. He had learned something at school, gone home, gone on a journey with it, found that the answer was bigger than the lesson, and then gone one step further and worked out why the curriculum stops at four. He understood the topic better than I did, and he understood something about the shape of knowledge as well.
The second story matters just as much, and it points the other way. Rehman has been learning Python. He got stuck. I said, why don't you screenshot it into Claude. He said, I want to work it out. Two days he spent on that problem. He cracked it himself, explained it to me, which I only half followed, and only then checked it against the AI, because he wanted to know he was right.
So the tool did not make him lazy. It gave him something to push against.
What we actually lost when execution got cheap
Put those two stories next to the argument I keep making in my papers and they stop being charming anecdotes.
Any piece of work you bring into the world has three parts. There is the purpose, the reason the thing should exist at all. There is the execution, the actual making of it. And there is the judgement, the moment you look at the finished thing and decide whether it meets the purpose, or whether it needs another pass. That is the PEJ framework, and I've set it out properly elsewhere (Dixon, 2026b; see also the paper behind the framework).
Run a time and motion study on someone writing a book before 30 November 2022, the day ChatGPT was released to the public (OpenAI, 2022), and execution takes something like 99 per cent of the effort. You could not produce a hundred versions of a book, so purpose and judgement had to be built into the writing itself, inseparable from it.
Here's the part people miss. That 99 per cent was doing a second job. Nobody wrote a book by accident. Nobody stood up and did an hour of comedy by accident. The sheer effort was itself the quality filter: if someone had committed that much of their life to making a thing, there was probably a reason for it, and they had probably formed some judgement along the way.
Now: write me a book on cucumbers. Done. Write me an hour of stand-up. Done. The execution is close to free, and it is getting freer. David Sacks, the White House AI adviser, reckons capability is compounding at three to four times a year on each of three fronts, the models, the chips, and the deployed compute, which multiplies out to roughly 27 to 64 times a year and something near a million times over four years (Sacks, 2025). Take the exact figure as a claim rather than a measurement. The direction is not really in dispute, and it compounds again with the individual's own growing literacy, because a more capable person pointed at a more capable model gets a multiplier on a multiplier.
But the filter is gone. Did the world need a twenty-thousand-word book about cucumbers? Was it any good?
In a world of abundant quantity, quality carries a premium. And quality is exactly the two things the machine did not do: a reason to exist, and the judgement to know which of the hundred versions is the one. That is the shift from a knowledge-constrained economy to a purpose-constrained one, which is the argument of the longer paper (Dixon, 2026d; the short version is From knowledge to purpose).
The uncomfortable bit: this is harder, not easier
The popular story about AI is that it will make us lazy and stupid. The popular story in the other direction is that it will make everything effortless. I think both are wrong, and in the same way.
AI is a multiplier of whatever you already are. If what you want is the easiest possible life with the least possible effort, it will hand you exactly that, and you'll get away with more than you used to, because it raises the bottom bar. If you're curious, or ambitious, or you want to build something, it will take you an extraordinarily long way. It multiplies the purpose it is pointed at. It does not supply one.
And for anyone who does point it somewhere, the work gets heavier, not lighter. Execution is a learned, repetitive skill. Purpose and judgement are not. Asking why am I doing this, and is this actually any good, burns far more mental fuel than doing the thing ever did. Anyone who has gone properly AI-enabled knows the feeling: every idea you have can now become real in an afternoon, and you are suddenly drowning in outputs you have to read, assess and send back. Your production goes up. Your thinking load goes up faster.
Which brings me back to Rehman and the two days on the Python problem. He chose the harder path on purpose, because he wanted the understanding, not the answer. That instinct is the whole ballgame.
What holds its value
In the paper I set out five capacities that keep their worth when machines can do the rest (Dixon, 2026d).
Purpose. Deciding what should exist. This is just another word for innovation.
Judgement. Deciding what is fit to go out with your name on it.
Empathy. Empathy only has value when there is another mortal human on the other end of it. An AI can emulate it, perhaps better, faster and more often than any person. It still will not be your partner noticing how your day went. In a restaurant, in an interview, in a negotiation, the human on the other side is the point, not the interface.
Authenticity. This is why live music is booming in an age of infinite streaming. The more abundant the copy, the more the original is worth.
Continuous learning. A category of its own, and the one most people skip. If the ground is moving this fast, understanding it is not something you finish at twenty-one. Perhaps we end up with four days of work and a fifth spent keeping up. Something like that is coming.
Notice what this does to the wage table. Today the high pay sits with deep-expertise execution: the lawyers, the surgeons, the coders, the analysts, the people who spent a decade learning to do a hard thing reliably. Those are not the high-paid jobs of the future. The value moves to the people with ideas, with taste, and with the judgement to point expensive compute at the right problem.
Everyone gets promoted
The most useful analogy for working with AI is not a computer. It's a new starter.
You don't send a new colleague's work straight out to clients. You give them a bit of purpose, you let them execute, and you apply judgement to what comes back. Over time you learn what they can be trusted with, and you loosen the leash. Then they get promoted and do the same thing for someone below them, and their name is now on work they didn't personally produce.
That is exactly the relationship you form with a model, and it means we all already have the skill set. So I don't think jobs disappear. I think every job changes and everyone gets promoted. We delegated arithmetic to the calculator and deterministic process to the spreadsheet. Now we delegate generative, non-deterministic knowledge work, and we move up a level. Then probably up again (Dixon, 2026e takes that nesting seriously, and asks who owns purpose when the systems themselves become agents).
What to do on Monday
Roughly six billion people, about 74 per cent of humanity, were online by the end of 2025 (ITU, 2025). A very large number of them are children, and a growing share of those children now have a curiosity loop with no wall at the end of it. I genuinely don't know what that generation builds. I'd like to be around to find out.
For the rest of us, the instruction is smaller and more boring than people want it to be. Reading about AI, talking about AI, going to conferences on AI and making lists about AI changes nothing. It is an enabler of building, and you only get good at building by building.
So: find a use case. Build one thing. Then build another. You will end up somewhere you could not have planned, which is rather the point, and it is not a bad description of how innovation has always worked (see Our Viking moment).
No individual in history has had more opportunity for agency, mobility and reinvention than you have right now, and it hardly costs anything. The mainstream story is gloomier than the facts warrant. This is the best time to be alive.
The full conversation with Shaun Phillips runs to forty-four minutes and is on the Better With Humans Podcast, or on YouTube.
If you're building AI capability in an organisation rather than on your own, that's what we do at Dixon AI.
References
Better With Humans Podcast, 2026. AI and Curiosity: Rob Dixon on What Humans Keep. Episode 14, host S. Phillips. [podcast] Better With Humans, 15 September. Available at: podcast.betterwithhumans.com [Accessed 19 September 2026].
Also published as AI and Curiosity: What Happens When Kids Never Stop Asking. [online video] Better With Humans Podcast, YouTube, 15 September. Available at: youtube.com [Accessed 19 September 2026].
Dixon, R.J., 2024. Our Viking moment: why the winning strategy changed on 30 November 2022. [online] robdixon.ai, 4 September. Available at: robdixon.ai/post/our-viking-moment-strategy-change-ai [Accessed 19 September 2026].
Dixon, R.J., 2026b. Purpose, Execution, Judgement: A Framework for Human-AI Collaboration and Organisational AI Capability in the Post-ChatGPT Era. SSRN Working Paper, 10 March. Available at: dx.doi.org/10.2139/ssrn.6282200 [Accessed 19 September 2026].
Dixon, R.J., 2026d. From Knowledge to Purpose: The Shifting Bottleneck of Value Creation and Five Domains of Human Value as We Approach AGI. SSRN Working Paper, 23 March. Available at: dx.doi.org/10.2139/ssrn.6396778 [Accessed 19 September 2026].
Dixon, R.J., 2026e. AI Innovation Architecture: Nested Purpose, Closed-Loop Dynamics, and the Chief Agency Officer. SSRN Working Paper, 23 April. Available at: dx.doi.org/10.2139/ssrn.6546178 [Accessed 19 September 2026].
International Telecommunication Union, 2025. Facts and Figures 2025: Internet use. [online] ITU, 15 October. Available at: itu.int [Accessed 19 September 2026].
OpenAI, 2022. Introducing ChatGPT. [online] OpenAI, 30 November. Available at: openai.com/index/chatgpt [Accessed 19 September 2026].
Sacks, D., 2025. The 1,000,000x boost. [online] X, May. Available at: x.com/DavidSacks [Accessed 19 September 2026].
Rob Dixon is Founder and CEO of Dixon AI Ltd and lectures at Manchester Metropolitan University. He writes and speaks on AI capability, strategy, and human-AI collaboration. ORCID [0009-0008-0548-1321](https://orcid.org/0009-0008-0548-1321). On [LinkedIn](https://www.linkedin.com/in/robert--dixon), and training and transformation work at [dixonai.com](https://www.dixonai.com/).
Comments