Kristen Maevyn published "The Decoherence Trap." Decoherence, in Maevyn's usage, is what a system accumulates when it holds several models that are each valid on their own and incompatible with each other.

The cost of any change stops being the cost of the change and becomes the cost of translating between the models and guarding against side effects.

Four claims follow.

The cost grows only with change, so a system left alone neither improves nor degrades.

The cost compounds, because each hurried change makes the next one harder and therefore more hurried.

The cost's most common source is the unfinished migration, where the old way and the new way both stay in the walls and the overlap is a third system more complex than either.

And because costs scale with velocity rather than time, anything that raises the rate of change raises the rate of decoherence with it, which is where the AI argument enters the conversation.

Lehman's laws of software evolution said much the same thing fifty years ago. The name is new, and the name is probably useful, because it lets you point at the overlap itself as the debt rather than at either system in it.

The back half of the essay argues that this same property bounds how far AI can improve software on its own, since coherence is the thing that only reality can test.

It's interesting, and it feels like two essays that got mashed together, but it'd probably worth the read.