NeatRead

Once you understand compounding, you'll never use AI tools the same way.

Focus
Text size
Line spacing
Column width
Theme
Page turning

Swipe left or right, tap either side, or press your reader's page buttons. Auto turns pages in the E-ink theme.

Exit focus

Listen to this post · 6 min

Read aloud by a synthetic voice.

I found out I'd forgotten how to write an email on a Tuesday.

Nothing dramatic. A client had pushed back on a quote, and I needed to reply with something firm but friendly. Four sentences, maybe five. I opened a blank message and just sat there. My hands went to the other tab out of habit. The chatbot was already open, cursor blinking, ready to do it for me.

I closed the tab and wrote it myself. It took twenty minutes. Two years ago it would have taken five.

That bothered me for a week, and it led me back to an idea I thought I already understood.

The boring math that runs everything

You've probably seen the compounding chart. Get 1% better every day for a year and you end up about 37 times better. Get 1% worse every day and you end up with roughly 3% of what you started with.

Most of us file that under money or the gym. I did too. What I'd missed is that compounding doesn't care what it's applied to. It works on anything you repeat. And it works in both directions, without asking you which one you'd like.

AI tools are the most repeated thing in my working day now. I use them dozens of times before lunch. So the real question isn't whether AI is good or bad. It's which way each of those dozens of tiny uses is compounding.

What the research actually shows

I went looking for evidence, half expecting a moral panic. What I found was more interesting than that.

In 2023, researchers from Harvard and BCG gave 758 consultants access to GPT-4 and watched what happened. On tasks the AI was good at, the consultants finished 12.2% more work, 25.1% faster, and their output was rated more than 40% higher in quality. That's a huge gain.

But the same study had a catch. On a task that sat just outside what the AI could do well, the consultants using it got the answer right less often than the ones working alone. The researchers called it a "jagged frontier." The tool makes you better on one side of a line you can't see, and worse on the other side.

Then there's the MIT Media Lab study from 2025. Fifty-four people wrote essays over several months, some with ChatGPT, some with a search engine, some with nothing. The ChatGPT group felt the least ownership of their essays. In the first session, 83% of them had trouble quoting a single line from the essay they'd just written. None of them got a quote right.

I read that and thought about my email.

A Microsoft and Carnegie Mellon survey of 319 knowledge workers found the pattern I was most afraid of. The more people trusted the AI, the less critical thinking they put into checking its work. The people who trusted their own skills were the ones who still pushed back.

To be fair, none of these studies proves AI "makes you dumb." The MIT team said so themselves, and the sample was small. But put them together and the shape is hard to ignore. Using AI well compounds your ability. Handing your thinking to it compounds your dependence. The tool is the same either way. What differs is the habit.

Three things that changed once I saw it this way

I stopped measuring AI by the minutes it saves today. A shortcut that saves ten minutes now but costs me a skill over a year is a bad trade. I just can't see the bill yet. So now I ask a different question: will I be better or worse at this in twelve months if I keep doing it this way?

I started treating my first draft as the part that matters. The value was never in having a draft. It was in the struggle of making one. When I write the messy first version myself and then ask the AI to tear it apart, I get faster and sharper each month. When I ask it for the first draft, I get a decent draft and nothing else.

I began building things that pile up. A prompt I type once disappears. A prompt I save, refine and reuse gets better every week. The same goes for a notes file I feed into each project, or a checklist the AI helps me improve after every mistake. These are small deposits, and they earn interest.

What I actually do now

None of this is clever. It's just applied consistently.

  • Think first, then ask. I write my own answer, even a bad one, before I open the chat. Two minutes is enough. That way the AI is checking my thinking instead of replacing it.
  • Ask it to quiz me, not tell me. When I'm learning something, I ask for five questions instead of a summary. It's slower, and it actually sticks.
  • Keep one running file. Prompts that worked, mistakes the AI made, the rules I want it to follow. I add something almost every day. Three months in, it's the most valuable document I own.
  • Do one thing a week by hand. Usually something I used to be good at: a short post, a spreadsheet formula, a tricky reply. It's the reps, not the result.
  • Double-check anything I can't judge. If I couldn't spot a wrong answer myself, I don't trust the right-looking one. That's exactly the side of the jagged frontier where people get burned.

It felt awkward for the first week. By the third week I'd stopped noticing. Most of the benefit showed up quietly, the way compounding always does. I wasn't suddenly brilliant. I just noticed one day that I wasn't stuck anymore.

The question I keep on a sticky note

I don't think the people who do best with AI over the next ten years will be the ones who use it the most. They'll be the ones whose daily use leaves them a little sharper each time instead of a little softer.

You won't notice either direction this week, or this month. That's the trap and the gift of compounding. By the time it's obvious, it's already done.

So here's the note on my monitor, right under the webcam:

Is this making me better at the work, or just better at skipping it?

Ask it once a day. That's the whole system.

Sources

Get NeatRead's best posts in one weekly email

One email a week, no password. Unsubscribe anytime.

or continue with email

Back to top

Comments

No comments yet. Start the conversation.

Sign in to comment. It takes an email address, no password.