Question on why use ChatGPT rather than a model that can be fine tuned

[userrand]

Language models are starting to integrate into a lot of software such as Notion, Microsoft software, and Google software from their recent announcements.

I am not sure how these integrations are done or how they avoid the situation where the AI produces code that does not work.

One possibility might be that they take whatever the AI gives as a first draft and then automatically send errors and iterate until the answer is correct. This could be quite expensive to do and might not be very reliable.

Another possibility might be that they fine-tune large language models for that purpose. As I understand, this can not be done with ChatGPT but can be done with other models. The question is then whether using ChatGPT for a long-term project rather than a fine-tuned model is the better option. Perhaps one could also connect the conversational skills of ChatGPT to a fine-tuned model specialized in unity code along the ideas of langchain https://github.com/hwchase17/langchain.

There is clearly an expense in fine-tuning but I suppose a discussion on which model to choose could be interesting for those interested in contributing to the project.

[keijiro]

The simple answer is that because I'm just an AI newbie. I have no idea about how to train language models. I don't know any other practical language models.