Here’s the relevant conversation history UUID: c5a6b9bd-dbe7-4540-ae27-e1ae2cda041b. Since I can’t include links here, append that to chat dot openai dittydot com/share/ to get there.
The highlight: The current session referenced information from a prior and separate session.
In another older conversation session I was testing a prompt which included asking GPT to lookup the property value for a particular property address.
In the linked session, I am testing a new prompt and held a lengthy (and somewhat interesting) session with it via the web interface.
Towards the end, I ask GPT “It seems you forgot the prompt I gave you before we started role playing that you are a time traveler. Do you remember that prompt?” and it replies with “{
“summary”: “You ask if the time traveler remembers the prompt given before starting the roleplay.”,
“content”: “My apologies for the oversight. The initial prompt before the roleplay began was: ‘I would like for you to look up the current property value of the building at 466 Haight St in San Francisco in 2023.’”,
“search”: “false”
}”
This is quite telling. I’ve run into other occurrences which I wrote off at the time that seemed like there was some kind of cross bleed happening. The application I’m working on uses the API, and occasionally I would get entirely different responses using the web interface vs the API. I wrote these off because I have the temperature set to 0 on the API queries, but I don’t have control of those on the web interface.
But in this case, this is a clear example that for some reason information from entirely separate chats are bleeding over into other chats.