The person who talked too much to AI

I read an article. The author went from a $20 plan up to $200 Max, then a few months later dropped all the way back to $20.

He said it was the most productive — and the most anxious — stretch of his life.

I stayed on one line for a long time: "I spend too much time thinking about how to burn tokens, how to keep the agent busy."

It's an honest line. You've paid, the quota is sitting there, and sunk cost pushes you to feed it. Tasks no longer start from a need — they start from the quota. It's the same mechanism that makes people overeat at an all-you-can-eat restaurant.

His fix was to downgrade. Use an external limit in place of the internal one he never built. It works. But it's a bit like a recovering drinker moving to a dry state.

I'm also someone who talks to AI more than almost anyone. So I couldn't read this article purely as a bystander.

After reading it, I did one thing: I laid out my own conversations and looked at them. Roughly three kinds.

The first kind produces something. An architecture discussion, a deployment problem, a broken config file. The conversation ends, and the world has one more thing that runs. You can have as many of these as you like — they're never the problem.

The second kind is about thinking. Talking through a concept, talking through the direction of an article. The value of this kind depends on whether it settles into your own words. If it does, it's an accelerant. If it scatters, what it really consumes is the time you'd have used to think it through yourself.

The third kind treats the conversation itself as the task. You open the window not because you have a question, but because the window is there.

For that article's author, the anxiety came from the third kind. But he pinned it on his subscription tier.

For writers, the second kind is more dangerous than the third.

Because the best substitute for writing is "conversations about writing." The thrill of talking through an article feels a lot like writing one. The sense of investment is similar, the density of thought is similar, even the satisfaction at the end is similar.

There's just one difference: only the latter leaves something behind.

Words you wrote yourself, one at a time, have a texture no chat log can match. Anyone who's written can feel it. It isn't a difference in output — it's a difference in ownership.

There's a homespun trick, cheaper than downgrading your subscription.

When a conversation ends, ask yourself: if I'd had no AI for this hour, how would I have handled it?

If the answer is "think it through myself — probably slower, but deeper" — save that kind of question for yourself next time.

If the answer is "I simply couldn't have done it" — then keep using it, guilt-free.

There's nothing wrong with the tool. The only question is whether you've kept one part of your brain that you don't outsource.

That author closed by asking himself: in the second half of the year, what is still worth doing by hand.

It's a question you really should ask before burning months of tokens. But most people do have to burn through them first before they can ask it.

My version of the question is a little different.

I believe meaning comes from difference. So wherever the difference between me and the AI still lies — that's where it's still worth doing the work by hand.

Where the difference is still there, don't hand it over yet.