An idea for multi-response battles

[endolith]

Almost all of the battles are single-response "chats":

The responses of the two models are often so different that there is no meaningful way to continue the conversation with both with a single response. Would this be possible?

1. User sends a message
2. Both AIs respond, and responses are shown side-by-side in random order
3. User chooses which response they prefer
4. The selected AI wins a battle
5. The better response is kept as the conversation context, and the bad response deleted, and the good response is given back to both AIs as the context for the following message. (Meaning the AI thinks it gave the good response to the last message even if it didn't.)
6. User responds again
7. Both AIs respond again, orientation is randomized.
8. User chooses better response
9. Chosen AI wins another battle
10. etc.

Then we can know better how models handle long conversations. (Though the worse AI is "cheating" by seeing the better AIs responses, so maybe it needs a handicap / weighted win or something to make it more fair.)

I think this "editing history" is possible with all the local LLMs, and with the OpenAI ChatCompletion API, but maybe not with all APIs. Maybe it could still be used for a subset of models, though?