AI coding assistants are the Exponential Field

2026-01-25

Arthur C. Clarke's story "Superiority" describes the defeat of a technologically superior spacefaring civilization in a war because they rushed out new technologies to the detriment of their old but reliable technologies. The last of the Wunderwaffen described in the story is the "Exponential Field," a device which distorted space around a ship for stealth and defense. As with the others, this failed, for what is a very interesting reason to me now:

It was never possible to restore the initial state exactly. Switching the Field on and off was equivalent to an elongation and contraction of the ship carrying the generator, but there was a hysteretic effect, as it were, and the initial condition was never quite reproducible, owing to all the thousands of electrical changes and movements of mass aboard the ship while the Field was on. These asymmetries and distortions were cumulative, and though they seldom amounted to more than a fraction of one per cent, that was quite enough. It meant that the precision ranging equipment and the tuned circuits in the communication apparatus were thrown completely out of adjustment. Any single ship could never detect the change—only when it compared its equipment with that of another vessel, or tried to communicate with it, could it tell what had happened.

It is impossible to describe the resultant chaos. Not a single component of one ship could be expected with certainty to work aboard another. The very nuts and bolts were no longer interchangeable, and the supply position became quite impossible.

I have been reluctant to use things like Claude Code and Codex for this exact reason. Today, I gave one a try. After a single round of breaking up a god-module -- housekeeping! not even logic -- by Codex, the first line of output from a model would fail to stream to the frontend. There was no reason this bug should have been introduced, as the only thing I had asked it to do was move code around.

I chose to roll it back at that point. Chatbots are one thing, but letting LLMs run wild in your codebase, even in managed mode, is quite another.