We love good ol’ completion models at Replit. We just replaced GPT-4 on a backend task with the new gpt-3.5-turbo with no accuracy hit. Faster + cheaper.
Madhav Singhal @madhavsinghal_
20 Sep 2023
Just ran an internal eval on a production task with gpt-3.5-turbo-instruct and it performs at GPT-4 level 👀

Sep 20, 2023 · 2:34 AM UTC

20
16
3
293
Replying to @amasad
This is interesting. Would you use gpt 35 turbo for coding as well?
1
Testing the new instruct rn
2
Replying to @amasad
what was the task. did it involve code gen?
1
8
Classification kinda. No
8
Replying to @amasad
Do you mean turbo or instruct?
Replying to @amasad
kind of shocking tbh fingers crossed we get gpt-4 instruct soon.
5
Replying to @amasad
rlhf really did a number to these models Been testing instruct 3.5 as well, for a variety of tasks and its reallly good
2
7
Replying to @amasad
How did you define accuracy? Or was this a fairly simple deterministic operation?
Replying to @amasad
Any chance you could share some examples of the kinds of prompts you're using here? I'm still trying to figure out what to use instruct for as opposed to regular turbo
1
8
Replying to @amasad
Will cover in #thursdai, definitely interesting from OAI
Load more