Cognitive Debt
Originally a 2–3 min video — also on LinkedIn / TikTok / YouTube · @allemaar
We use cookies to understand how you use this site and improve your experience.
Originally a 2–3 min video — also on LinkedIn / TikTok / YouTube · @allemaar
You're paying for a genius and making them count inventory.
Last video I talked about what happens when JSON breaks. One bad bracket, everything dies. But here's the thing. Even when it works perfectly, you're still paying a cost. You just can't see it.
AI models have a context window. Think of it as a fixed amount of working memory. Everything the model reads, everything it generates, all has to fit inside that window.
It's like doing mental math. You can hold maybe four or five numbers in your head at once. That's your budget. If you're spending some of that budget remembering where you put your keys, you have less left for the actual math.
When a model generates JSON, part of that budget goes to formatting. Tracking brackets. Remembering nesting depth. Am I inside an object or an array? How deep am I? Does the comma go here?
None of that is thinking about your problem. That's bookkeeping. And it comes out of the same budget as reasoning.
Think about a pianist performing a concerto. Now put a metronome in their ear that clicks at the wrong tempo. They can still play. They might even play well. But some part of their attention is always fighting that click. Always correcting. Always spending effort on something that has nothing to do with the music.
That's what JSON does to an AI model. The output comes back. It's valid. It looks fine. But there's a headache you only notice when it stops. The reasoning could have been deeper. The model had less room to think because part of its budget was spent keeping the structure from collapsing.
That's cognitive debt. Not an error. Not a crash. Just... less intelligence than you paid for.
And this compounds. When you have AI agents talking to other AI agents, each one generating JSON, parsing it, generating more, that debt stacks at every step.
Every hop in the pipeline. Every nested response. Every bracket the model has to track. The quality loss is invisible at any single step. But across ten steps? Twenty? You're running the entire pipeline on a fraction of the intelligence you think you're using.
Remove the debt. Watch what comes back.
The block problem was about what happens when JSON breaks. Cognitive debt is about what it costs even when it doesn't. The thinking that never happened. The reasoning that had no room.
Loading comments…