How Organizations make adaptations to ChatGPT? Fine-tune or extension?
To best harness ChatGPT power, we may need to adapt to an organization's domain problem effectively. And we often hear the buzzword "fine-tuning," which refers to making an additionally labeled dataset to let the model learn from it. It has been a hot topic in the OpenAI community and at the time I wrote this, I am not aware there is a publicized generally available feature on this.
In Pingping Consultation's view, "Fine-tuning" is inherently a problematic approach, for small organizations. "Fine-tuning" is like giving a kid a new skill, like riding a bike. However, in ChatGPT's training, the AI skills are acquired through reinforcement learning, something like having that kid riding a bike on various landscapes, sometimes he/she may fall, but next time will learn not. Because of the repeated fall-and-learn, ChatGPT native skills are "battle-hardened." Do you hope it will pick up new things quickly? Well, I'm not sure, but I guarantee that your limited fine-tuning dataset will always be an overestimation of its performance.
Another approach is doing "extension." Say, there are ChatGPT answers that resemble a template, you can capture that, analyze it, and assign slots to the parameters. Basically, it is an "intent category," using the words of Conversational AI design; however, it is a structured "intent category" that consists of a compositional structure, rather than just a binary category. Then traditional Chatbot design notion follows, such as dispatching pre-defined dialogues.
The benefit of the "extension" approach is, you can attach business logic that's truly controlled by you as an individual organization. You can have your own copywriting for your in-house dialogue.
But then, the problem shifted to,
(1) how do you capture a template, analyze it, and assign slots to the parameters?
(2) it sounds simple when we think about the similarity to old Conversational AI dialogue; however, the transformation from the existing ChatGPT answer to org-customized dialogues is rule-based and it is not straightforward to do that.
Takeaway
Discussed some potential ways to do org integration.