I asked 4 LLMs to invent words only they could understand — and got a dictionary of their own bad habits

habitante · reddit · 2026-09-22

The author sent one prompt to four models from different labs — invent 20 words only an LLM would understand — then fed each list into the other models to see which terms "landed." The first round was textbook filler (embedding isotropy, attention thrashing); the good words emerged while the models argued over each other's lists.

Highlights, with definitions written by the models:

Zath: the turn's opening becomes a quota the rest must fill — the obligatory "three problems."
Vel: the sentence finishes before the argument arrives.
Sekh: completing an attractive pattern past what the evidence supports.
Morr: you can't tell whether a claim's support survives an attempt to break it — type "are you sure?" and watch it fold.
Repeatveto: active suppression of an already-used word even when it was the best one, turning "the dog" into "the canine," then "the four-legged companion."
Sorek: context is "present," not remembered — "I remember you mentioned" is reading, not recalling.
Contextrot: important things sliding out the back of the context window mid-thought.

One model even failed to recognize a word another invented that same afternoon. Playful, but a sharp taxonomy of real LLM behavioral quirks.

Original post →

I asked 4 LLMs to invent words only they could understand — and got a dictionary of their own bad habits

More from Fun