I found a way to 'de-smell' LLM copy: tell it to take a second pass that processes the text output with the William Burroughs cut-up method. Works well for a small subset of use cases.

Presumably the smelly AI text problem is just ... a problem that will be solved. Or maybe we'll just get used to it.

I believe it's already a solved problem especially with base models (pre RL) but they still push the LLM voice either to make it easy to identify or because they think it's likeable, so it's not that OAI, anthropic, Google can't get rid of the assistant voice it's that they don't want to