Constraining an agent to use simplified technical language collapses the space of possible responses, limiting the probability of novel output and capability overhang tapping.

Depending on how much you want to anthropomorphise language models, there’s a coherent argument to be made that linguistic relativity does in fact apply to LLMs, and that enforcing linguistic constraints reduces the fitness of a given response.

This is fine for the last mile of problems, when a more deterministic response is needed, but further upstream it may actually neuter the upside of a technology like LLMs.