Cognitive Atrophy

Recently I was out west spending time in the mountains with family, in the same part of the country where an old friend of mine lives. He’s part of a group of friends that texts constantly — every day, a running stream of jokes, links, and life updates.

You have the same group of friends, don’t you? The same messages. The same jokes.

Driving through town one morning I decided to stop by his house unannounced. No text ahead. No “you around?” Just knocked on the door the way people used to. I might have double-checked on the Find My Friends app to be sure he was there, but other than that, it was a cold drop-in.

We ended up on his porch, talking. No screens, no thread to scroll back through, nothing to do but sit there and have a conversation. He lives in close proximity to many of the other friends in my circle and in that text group. I asked in passing how often everybody gets together, assuming it was quite often, and to my surprise, he responded that it doesn’t happen much at all. In fact, it hasn’t happened for over a year. I was surprised and not surprised.

I’ve been chewing on that porch conversation ever since, because I don’t think it’s just about friendship. We are moving away from the analog. We are rolling downhill without brakes, and our minds can’t keep up.

I think it’s the same thing I’m watching happen at work, in interviews, in my own head. The convenient version of a thing quietly replaces the real version, and the muscle behind the real version starts to waste away.

I can’t help but believe that AI is a primary contributor to this cognitive atrophy. Not just in software — everywhere. AI is wicked smart, wickedly fast, and always available. That makes it a hell of a crutch. The side effect is quiet and serious: we stop burning knowledge into our own heads because the temptation to offload everything is too convenient.

Earlier in my career there was no easy button. You bought the O’Reilly book and actually read it. You read documentation pages top to bottom. You dug through Stack Overflow. You struggled, sometimes for days, on problems an LLM would now dispatch in seconds.

That struggle felt like the tax. It was actually the product. Out of it came confidence, domain knowledge, and technical ability stored where it matters most — in your own cortex. When you needed it, it was there. Nobody could rate-limit it, and it never lost the thread of the system you’d built, because you were the context.

Life outside of work had the same built-in throttle. Commercials interrupted television, so you took breaks whether you wanted to or not. Nothing was bingeable or streamable. You waited a week for the next episode. You were forced to wait because that was the nature of the environment, and in the waiting there was time to think, to process, to store.

That throttle is gone. The world is bingeable. AI is on tap. Friction is optional now, which means the struggle that builds a mind is optional too. Nothing external is going to insert the blocks and weights for you — the time to think, the time to process, the time to store memories. That’s on you now.

Not long ago I sat in an interview with a software engineer. The session was a mix of agentic coding and hand coding — a format I like, because it shows both sides of the current job.

On the first half he was excellent. He explained architecture clearly. His prompt engineering was thoughtful. He clearly knew how to direct the tools.

Then came a rather basic coding exercise, by hand. And the imbalance was huge. Not a stumble — a different engineer entirely.

I don’t think he was faking the first half. I think he simply hadn’t written code by hand in a long time. Two years ago, I suspect he would have done much better on that exercise. The skill didn’t disappear because he got worse. It atrophied because he stopped using it.

Here’s the honest part: I’m not even sure hand coding in an interview is still the right bar. Maybe it isn’t. But I am sure of this — a software developer should still be able to produce high-quality code without an LLM holding their hand. When that goes, the knowledge retention goes with it. And when the retention goes, you lose the one thing that makes you dangerous with these tools in the first place: your own context.

It takes work and energy to burn knowledge into your cerebral cortex. There’s no shortcut for it. You have to go through pain. You have to go through struggle. Out of that comes domain expertise — the deep, durable kind you can draw on years later without looking anything up.

LLMs are great, but they don’t have context the way a human mind has context. Their Achilles’ heel is working memory — how much they can hold and use at once before losing the thread. Our capacity to remember, retain, and retrieve is in a different league entirely. We have the advantage in spades.

The problem is we’re not putting in the work to store anything anymore. We’re skipping that step because it’s hard, and because the easy button is sitting right there. I’ve felt this myself. Skip the full document. Skip writing it by hand. Ask the model, take the answer, move on. It works — until the day you need the knowledge and it isn’t there, because it was never stored.

People don’t read full documents anymore. People don’t write code by hand. The muscle you refuse to use is the muscle you lose.

I hear it too often at work now. “Claude did this.” “AI did that.”

That phrase is a negative sign, and it concerns me every time.

Claude doesn’t do anything we don’t ask it to do. If you’re surprised at the outcome, you’ve handed over too much. You’ve offloaded your biggest asset — your context, your memory, your domain expertise — and put yourself in the passenger seat of your own work.

The same tell shows up in life. Texting instead of calling. Scrolling instead of showing up. Streaming instead of waiting. Each one is a small offload, and each one skips the step where the connection or the memory actually gets built. The porch conversation never happened over text, no matter how many messages we traded. It happened when I showed up.

So here’s what I’m changing.

Set aside time regularly in the void of an LLM — and in the analog world — to struggle the old-fashioned way. Not haphazardly. On purpose. Pick a topic that matters. Read the whole document. Write the code by hand. Practice the art of learning the hard way and let the struggle burn the information into your mind.

It’s going to really suck at first. That’s how I know it’s working. Over time it gets faster, because the skill of storing and retaining information is itself a skill — and it responds to training like any other.

The same discipline applies off the clock:

• Go out with friends instead of doom scrolling.
• Call someone to talk instead of texting.
• Stop by someone’s house unannounced just to have a conversation.
• Insert time to wait, reflect, write, and ponder.

Do this consistently and you don’t fall behind the people leaning hard on AI — you lap them. You 10x the 10x we’re already getting from LLMs, because you keep enough context in your own head to know exactly how to direct the tools instead of being directed by them.

The world isn’t going to throttle us anymore. We have to insert the throttle ourselves.