This is a spin off of this post. The idea is that when generating responses, Kagi Assistant could take advantage of all of its LLM offerings by generating two responses instead of one. There's a number of ways this could work:
- Either two models could generate a response and the 'better' one could be shown to the user.
- The models could 'work-together' for a while, thinking back and forth between each other and eventually coming to a response that could be shown to the user.
- You could do a generator/adversarial setup. Where one model generates a response and another critics it. Then the first generates a better one incorporating the critic's feedback. They could go back and forth for a bit before being satisfied with the answer.
- Etc...
A feature like this could improve Kagi assistants responses and be a feature that isn't available on other LLM app offerings.
The user would interact with the assistant as normal. But the responses would potentially be better. Or you could display the back-and-forth to the user in someway, maybe making their assistant chats look like group chats.
