Using AI for real is becoming a privilege

Picture two people in front of the same screen.

The first asks ChatGPT to fix an email. It works, they close the tab and get on with their day.

The second opens an AI agent and explains what they want to build. While they get a coffee, the agent writes the code, runs the tests and leaves them a list of things to review.

Both would say they "use AI". Both think they know what it can do. And they live in different worlds.

01How many are on each side

Andrej Karpathy, one of the most listened to voices in artificial intelligence, recently shared a funnel that explains it well. The figures are his estimates to illustrate the idea, not a study:

1. 01
   ~75% barely use AI at all. They know it from the news.
2. 02
   ~20% use it casually: emails, translations, summaries.
3. 03
   ~0.2% hand real work to agents: code, tests, whole products.
4. 04
   ~0.00006% work at the frontier, with systems that are not public yet.

Look at the third number again. Two in every thousand people.

02Why so few?

It would be comfortable to think it comes down to motivation or intelligence. It does not.

To go from fixing emails to delegating work you need, first, money. The good tools are paid for, almost always in dollars. If you earn in pesos, that subscription does not cost what it costs in the United States.

You need time. Learning to work with agents means trying, getting it wrong and trying again. If you come home from a ten hour shift, that time does not exist.

You need English, because almost everything written about these tools comes out in English first.

And you need a job that lets you delegate. In some you can give an agent a task and review the result. In others you are paid by the hour, or you cannot take a single piece of data out of the company.

None of that measures talent. It measures what currency you are paid in, how much free time you have and what job you landed. However you look at it, that is social class.

03Why we never agree

This explains something I hear all the time. Someone says AI "makes things up and is useless". Someone else says it saves them days of work.

They are both right. They are just not using the same thing.

04What I see from software

I build software with agents every day, so I am writing from the lucky side of the gap. And I want to be clear about one thing: delegating is not the same as stepping away.

An agent can generate hundreds of files in minutes. That does not guarantee the system is secure, easy to maintain or that it solves the business problem. If you delegate without understanding what was generated, you pile up technical debt faster than ever. I went into more detail in this note on vibe coding.

Among developers, the difference now is also who knows how to give context, set limits and review with judgment.

05What I keep thinking about

I do not have a solution. But I think we need to stop talking about AI as if we all started from the same place.

If you are already on the side that delegates, the question is not how much code you can write, but how much you can build without losing control of what you are building.

And if you are on the other side, what is holding you back today: the price, the time, or not knowing where to start?

Inspired by a reflection from Andrej Karpathy (@karpathy) on the growing gap in how people understand and adopt language models. This note is also available in Spanish.