Repeated responses from gpt-4-0613?

Is anyone else seeing repeated responses coming back from gpt-4-0613 right now?

I just noticed this start happening, but I don’t see an incident open. I’m still investigating my application, but this started happening seemingly with no changes on my side. The repetitious responses seem to be coming back directly from the API streaming response.

Just want to know if anyone else is experiencing this?

4 Likes

Yes, we are seeing previous messages being redundantly appended chunk-by-chunk via the streaming completion API. We’re getting a verbatim copy of a previous message chunk-by-chunk appended after the expected API response. The chunks from the correct initial response have the same id as the redundant chunks, and they stream through exactly as if they were part of the normal response.

It’s not every completion, maybe something like 10-20% of the time.

2 Likes

Yes, the behavior I’m seeing matches your exact description.

We’re also seeing this – older messages are being appended verbatim into the latest streaming completion. We’re using gpt-4-0125-preview, and this seems to have only started in the last 24 hours or so.

2 Likes

We’re seeing the issue while using the official OpenAI python library v1.3.7, rebuilding with 1.14.3 now to see if the issue is still there.

Yes, we’re having repetition also in gpt-4-0125-preview, and function problems.

We send in a User message, and get back an assistant response. Normally, our assistant is asking a question like “How old are you?” and get the user’s response, which makes the Tools call to save the response. Instead we are getting a single response like:

role:assistant
content: “How old are you?”
tools: (function - save_age(35 years old))

It’s hallucinating - calling functions before the user even responds to the question.

1 Like

We are also experiencing the same issue, please let me know if you find a solution

1 Like

I’ve replicated this with basically all models. It’s happening when I stream the responses back. I started noticing weird behavior sometime yesterday.

1 Like

Same here, happening with gpt 4 turbo the same issue.

1 Like

Thanks for highlighting this - we’re investigating.

3 Likes

Thanks for bringing this to our attention! We’ve looked into the issue and found that there was a glitch causing repeated responses; our team has just implemented a fix for this.

You can keep an eye on the current status of this fix on our status page here. We’re monitoring the situation closely and will post updates here.

If you’re still noticing any weird behaviors or repeated responses, please do let us know!

4 Likes

Amazing. Really appreciate the prompt attention. Thank you!

We are still experiencing some repeated responses. Our use case is a LangChain Agent using GPT4-turbo-preview. Temperature 1. Max Tokens 4096.