I wanted to ask a few questions regarding the new models and upcoming deprecations:
Can babbage-002 and davinci-002 models be used both with completions and chat APIs? It’s not entirely clear in the docs.
Is the completions API also going to be deprecated in January 2024? I noticed that it’s marked as legacy in the API reference. I wonder, how urgent it’s going to be to go away from this API then.
Are babbage-002 and davinci-002 trained specifically on chat data like gpt 3.5 turbo, so that we could expect better results from them using chat API rather than some arbitrary format through completions API with which these models won’t be initially familiar?
Thank you for the answers.
Hi and welcome to the developer forum!
In general, completions are being replaced with instruct models where you can instruct them to do a completion or and other task you desire.
To question 1, they are completion only, no “chat” version.
On point 2, I would be planning to replace any reliance on the completions endpoint by Jan 2024.
On point 3, Those models are not trained for conversational interaction.
Hi, thank you for the answer Spencer, it helps a lot!
Regarding point 2, why to move away from completions endpoint by Jan 2024, if the newly released models are completion only? I mean, these are very good models. Why to release them now, if their only API is not meant to be used in 5 months?
For some context, I would be fine with moving completely to chat API, but the smallest model there is GPT 3.5 turbo which feels like an overkill for some of the very simple tasks that are done just fine by the new babbage-002 at a great price.
The announcement said that they would be moving to an instruct model to replace completions, that’s why I advise to move from it.
Now I fully accept that the “new” models are davinchi-002 and babbage-002 and they are not set for depreciation, but… I see a consistent move towards instruct chat models based around 3.5-Turbo for mass use commercial applications, so I personally would move my solutions from the base models.