Yup. That's a fair point to bring. We use the OpenAI APIs with `gpt-3`. No particular fine-tuning done, or anything special than prompt-building and some invocation parameters that encourage more diverse answers.