- Author
- Kristen Maevyn
- Published
- 14 October 2026
- Kind
- Personal essay, with a toy simulation and cited studies
- Original
- maevyn.codes/decoherence
- Subject
- Review, taste and expertise
- Themes
- How things break, Skill under automation
What it is
Kristen Maevyn, a personal-site essayist, defines a kind of tech debt called decoherence. The essay sets out to show that heavy AI use produces it unsustainably fast. It also argues against a fast, isolated intelligence explosion.
What it says
- Maevyn defines decoherence as locally valid but incompatible models piling up in one system.
- Decoherence grows with change, not with time, and its growth compounds.
- A toy codebase simulation shows a sweet spot, with a steep drop-off below it.
- The essay cites a study of 718 firms: more code, no more finished work.
- Maevyn argues AI cannot confirm better software coherence without leaving the data center.
- Maevyn fears Decoherence Collapse, where humans cannot maintain systems and agents may not repair them.
- The essay says decoherence can be reversed, since it only changes when systems change.
What evidence it gives
The essay offers a plumbing analogy and a graph simulation. The author admits the weights are arbitrary. It cites three studies on coding agents. It gives historical examples such as futures and Boolean algebra. It reports the author's own experience at past employers.
Interests and limits
This is one author's essay, and decoherence is a term the author coined. The simulation shows a curve can arise, not that it does. The cited studies are summarized, not examined. The author also states a skeptical view of AI agents. The essay has no test of its own on real codebases.
Connections
These links are this wiki's own reading, not claims the sources make.
- Linearizability! Refinement! Prophecy!. Maevyn describes locally valid models that no longer fit together. Hochstein shows a formal remedy on one queue: map one model's states onto another's. The map sometimes needs an invented prophecy variable. How large a system could afford such maps?
- Harness engineering: leveraging Codex in an agent-first world. Lopopolo's team runs background tasks that scan for drift and open cleanups. Maevyn says decoherence grows with change and compounds. A million generated lines make a good test of both claims. Can drift be cleaned faster than it accumulates?