#ChatGPT Unexpected Model Switching

7 messages · Page 1 of 1 (latest)

uneven owl
#

Bug Report:

I have noticed that conversations with GPT-4 on ChatGPT will unexpectedly switch to the GPT-3 model (with the attempted usage of Plugins) after:
A) switching to another conversation in ChatGPT and switching back to your original conversation
OR
B) switching to the "New chat" UI and then switching to your original conversation

Steps to reproduce:

  1. have GPT premium account and plugins access
  2. start a new conversation with GPT-4
  3. send a message to your conversation with GPT-4 (message contents are irrelevant)
  4. switch to another conversation on ChatGPT (any) OR click the "New chat" button
  5. switch BACK to your original conversation with GPT-4 (the one started in step 2)
  6. send a message to your original conversation with GPT-4

Expected result:
The continuation of my pre-existing conversation with the GPT-4 model. Default ChatGPT behavior.

Actual results:

  1. (after step 5, before step 6) at the top of our message history "Model: GPT-4" switches to "Model: Plugins Alpha"
  2. (after step 6) text reading "The previous model used in this conversation is unavailable. We've switched you to the latest default model" appears above the next GPT response (note, this message is inaccurate: it switches to GPT-3, not GPT-3.5, which is the current default model as of 4/27/2023)
    3.ChatGPT responds with its GPT-3 model (Green OpenAI icon, no longer refers to itself as "based on GPT-4", specifically refers to itself as GPT-3 and not 3.5)
  3. the user is unable to continue their original interaction with the GPT-4 model, conversation continues using GPT-3

I am not able to link my video demo & narration of this behavior due to Discord rules (video is on my google drive because I used Screencastify) but feel free to reach out if you are interested.

Additional information
OS: Windows
Browser: Chrome
ChatGPT Account: Premium
ChatGPT Plugins: Code Interpreter, Plugins (access to store enabled but none installed)
GPT Models: 4, 3, text-davinci-002

rose obsidian
#

I have also found GPT-4 to have developed a bug where it switches the model for no discernible reason. I believe it is associated with long posts/content windows.

#

There have been many posts about The previous model used in this conversation is unavailable. We've switched you to the latest default model in the last 24 hrs.

kindred arrow
#

I'm also getting the same issue. At first I thought it was because I hit my "25 in 3 hours cap" and then got worried that they disabled the pop-up feature that asks if you want to switch models before doing it and just started doing it automatically, which annoyed me as I'd rather wait for my timer to reset because I don't want to use 3.5.

But then I opened a new chat and it allowed me to use 4 again for a few prompts before it did the same issue and auto-switched me to 3.5 again. Has happened in both new and old chats, but doesn't seem to be because of my rate cap as sometimes I can still create new chats with 4.

Probably the most annoying thing is that once it auto switches to 3.5, there's no way to switch it back to 4, which means my old conversations I've been saving to interact with again at a later time now have to be restarted in a new convo since some of them were auto-switched to 3.5

broken swift
#

I just noticed the same problem when switching between instances of the chat.

If I start using GPT4 and don't start a new chat or switch to an old one then it seems to keep responding with GPT4, however, as OP says, if I start a new chat or go to a previous instance, when I return to my original GPT4 prompt I find the model has switched to GPT3.5 with no way to change back.

This is frustrating as I sometimes need to copy information from one chat instance to another.

balmy frigate
#

same thing happened here, we need to get an option of switching back

cunning tree
#

I'm getting this too which is very frustrating. I need to stay in 4.0 as I'm working on code and its way more accurate in 4.0. Instead of it switching I would much rather have an option to provide longer prompts that use up more tokens.

Showing my age but reminds me of the early twitter character limits 🥴