Restarting partially completed chat completion API calls

Interesting perspective. I found a point where gpt-3.5-turbo-0125 or gpt-4-turbo (0125) would start over in the penguin prose.

GPT-4-0613, no hiccup. gpt-3.5-turbo-0613, no problem.

Latest models broke completion (among other deoptimizations) where the deprecations guide specifically recommends chat as an edit replacement, and previously had gpt-4 pointed at also to replace completions.


I found the most performative against the new behavior is a user message “[continue AI completion]”

However, the messages being wrapped in a container for “ChatML”, and an unseen “assistant” prompt, means the flow is broken up and the AI is re-prompted.

1 Like