Comprehension Debt

Most conversations about AI cost are really about routing: cheap models for the easy work, expensive ones for the hard work. The assumption is that the routing keeps the bill under control. That only holds as long as today's price gap between cheap and capable models survives, and I don't think it will — capable models keep getting cheaper, and the open ones are closing the distance fast.

So when the model is nearly free and routing barely moves the number, where does cost discipline actually go? It stops being about which model you buy. It becomes about how much work you let a model do unchecked, and how much of the resulting codebase your team can still genuinely follow.

Here's the shape of it in practice: point a coding agent at a goal, let it work for a few hours, and it comes back with a wall of changes too large to read line by line, let alone verify by hand. The gap between what got shipped and what a person actually understands is what I call comprehension debt: work that exists, runs, even passes its tests, without anyone on the team able to account for why it's built the way it's built.

The obvious fix is to have AI check AI. That only goes so far. A checker built on the same kind of model shares the writer's blind spots; it can hallucinate agreement as easily as it hallucinates code. Eventually the chain bottoms out at a person deciding whether to trust what came back, and that person can only vouch for what they still understand. Models keep getting cheaper. That judgment doesn't.

The scarce resource was never really the compute. It's the number of people in the room who can look at what got built and say, honestly, “I understand this well enough to stand behind it.” That number doesn't grow just because the tokens got cheap.

I haven't fully worked out what the fix looks like. Different teams will land on different combinations of review gates and slower merges, but naming the debt is the first move, because a team that can't see it will keep mistaking velocity for understanding. So it's worth asking plainly: for the AI-written code shipping through your team today, where does the trust actually bottom out — a senior reviewer, a test suite, or honestly nowhere yet?