Multi-Model Message Attribution/Context Within Threads
No idea on feasibility given reliance on external APIs, but it would be convenient if each model understood which messages it did or didn’t send within a thread.
Scenario
I select GPT-4o as the model, and I ask “Why is the sky blue?”
I wait for, and eventually vet GPT-4o’s response.
Skeptical of GPT-4o’s response, I switch the model to DeepSeek within the same thread, and ask “what do you think about GPT-4o’s answer?”
current behavior
From what I can tell, DeepSeek doesn’t understand that GPT-4o sent the previous response. Each model assumes ownership of every reply sent in response to user messages within a thread, irregardless of which model actually sent each reply.
This ostensibly has a negative effect on each model’s reasoning capabilities, as they spend time defending/justifying/correcting something they never said.
Log in to comment and vote
Comments2
Salmon Jaguar
Cool stuff, but if it is added I would like to be able to opt out of it since switching models transparently is very useful (example, getting a Gemini model to analyze a codebase and come up with a plan/doc, the transparently using Claude to execute said plan), I don't know how if it could affect the models negatively if they become aware of each message’s origin, so I might be just overly cautious here.
Aquamarine Butter
Guess I didn’t really put much thought into the opposite functionality. I agree, could just be a toggle within a thread.