AFAIK some outputs are made with a really tiny/quantized local LLM too.

And yeah, even that aside, GPT 3.5 is really bad these days. It’s obsolete.