Over the past few months I have used AI to build this website, the question bank on it, and the series of articles you are reading now.
The method was not to write code line by line myself. It was to explain clearly what I wanted, understand what the AI handed back, judge whether it was right, and ask it to change things. That, I think, is exactly the skill students most need to learn now.
Like most people, I started out using AI only to look things up and tidy sentences, treating it as a slightly better search engine. That shift is not unusual. Andrej Karpathy, one of OpenAI's founding members and formerly Tesla's head of AI, wrote on X on 11 October that less than a year ago he was still typing code by hand, character by character.
The point of what he wrote, though, was a funnel.
The four-tier funnel
Karpathy divides the world into four tiers by how much contact people have had with AI. He is clear that the numbers are rough estimates:
- Tier one: around three-quarters of the world's population, who have barely come into contact with AI at all.
- Tier two: one to two billion people who occasionally use a free version of ChatGPT or something like it, treating it as a better Google or a writing aid.
- Tier three: around 20 million people, roughly one in every five hundred. Mostly in mathematics and programming, they have seen first-hand that large projects which used to take weeks or months can now be done by AI from a single set of instructions.
- Tier four: around 5,000 people inside the leading AI companies, with access to systems that are not yet public. This tier is about frontier labs, and is not what this article is about.
The gap in understanding between these tiers, he says, is widening. The same tool that looks clumsy and not very useful from tier two looks, from tier three, like it is improving fast enough to make you dizzy.
Where are Hong Kong's children?
Tier one is not Hong Kong's problem. Almost none of the students I have taught have never used ChatGPT or something like it.
But having used it is not the same as having seen what it can do. Most of them use AI to look something up, to explain a single question, or to polish an English essay. That is tier two.
Some people use a smartphone only for calls or WhatsApp. Others use the same phone to run an online shop, or even to shoot a film that ends up screening at an international festival. The phone is the same. The difference is whether anyone gave it something bigger to do.
The gap that matters is between tier two and tier three.
What moves people up
Karpathy gives three reasons.
First, the size of the problem. In his words, "a question with a paragraph answer barely stresses the system." A simple question only gets a simple answer, and that tells you nothing about what AI can do. You need to be able to come up with large, difficult problems that you genuinely care about before AI shows what it is really capable of.
Second, AI's ability is uneven. It is strongest in fields that are digital, have clear right and wrong answers, and are commercially valuable. Mathematics and programming are exactly those fields.
Third, access. Free, paid and internal versions show you different things.
Of the three, the third is the easiest for a parent to fix: pay for an upgrade. But I think it is also the least important. If a child only asks AI questions that a paragraph can answer, the free version and the paid one will look much the same.
What decides which tier a child is in is the first.
What I think
What schools give students is mostly questions with a set answer, finished in a paragraph. That is not the school's fault. It is what exams are.
But it does mean that a child who only uses AI for homework is likely to stay in tier two. Worse, a child who uses AI to do the homework skips the thinking the homework was there to train.
When I built this website, nobody gave me a question. What to build, how far to take it, what to change: I had to decide all of it. AI did most of the work, but two things could only come from me: deciding what to build, and noticing when something was wrong.
Once, the article list that AI had written for this site looked perfectly normal in a browser, and it turned out Google could not read what was in it.
Another time, I had a technical problem to solve. The AI walked me through it confidently, step by step, but after nearly two hours the problem was not solved and I felt I was getting further away. At that point I stopped, thought it through again, and asked the AI to try a different direction. One simple step solved it.
Those two experiences taught me that using AI well comes down to how you think, and how you instruct it.
If you want to help your child move up, I have three suggestions.
Start from something to make, not from a tool. It is easy for a parent to ask first which AI tool their child should learn. I think that is the wrong question. Tools keep changing, but the things that are genuinely worth doing don't. Find something your child wants to make, and the tools will follow.
Teach them to check. Whatever AI hands back, a child needs to be able to judge whether it has been done right. This is where the foundations matter. If you don't understand the mathematics, you can't see where the AI went wrong. If you don't understand what a question is asking, you can't tell whether the AI has answered something else. The better AI gets, the more this is worth.
Give it time. A real project is not finished in an evening. The first version is usually a mess and needs several rounds of changes. That process is itself the way up.
What a child usually lacks is not the tool, and it is not that they aren't clever enough. It is something worth spending a few weeks on.
One thing to try this week
Together with your child, find something they genuinely care about that can't be answered in a sentence or two. A small game they want to make, or a set of information they want to organise.
Then encourage them to use AI to build it.