Questions

This is the question that started Noviq.

The premise is simple: intelligence — the ability to reason through a problem, to have a concept explained until it makes sense, to get a second opinion — has always been distributed unequally. Not because some people have more of it, but because the infrastructure for developing it (good teachers, good schools, patient tutors, access to books) has been concentrated in certain geographies and certain income levels.

AI changes the cost structure of that infrastructure dramatically. A model that can answer a question, explain a concept ten different ways, and never get impatient with a confused student costs roughly the same to run for a student in Addis as for a student in Amsterdam.

I do not think this automatically produces equality. Infrastructure alone does not. But it removes one of the most stubborn constraints — and I find it genuinely hard to overestimate what becomes possible when that constraint is gone.

Every product decision is an implicit answer to this question.

The payment methods you support, the languages your UI speaks, the connectivity conditions you design for, the price point you choose — each of these is a statement about whose reality you have modeled. There is no neutral default.

I have come to think that most software built in global tech centers is built, honestly, for the people who built it. The user it imagines is a version of the builder. This is not malicious — it is the natural result of building what you understand.

The interesting design challenge is to build for someone whose context is genuinely different from yours. That requires a kind of epistemic humility that is harder to cultivate than any technical skill.

The philosophical answer is genuinely unclear to me. The practical answer, for the work I do, has to be: it does not matter.

If a model can explain photosynthesis to a 14-year-old in Amharic, in a way that causes her to understand photosynthesis — what exactly is missing? The Chinese Room argument says: the system is just manipulating symbols. But so, at some level of description, are neurons.

I suspect "understanding" is not a binary property. It is a family of capacities — generalization, analogy, error correction, knowing what you do not know — and current models have some of these more than others.

What I am more confident about: the question of whether the model understands is less important than the question of whether the person using it comes away understanding more. That is the metric I care about.