One-Shot-Learning on 3.5 using GPT-4o outputs

API
,
1

Hello all,

Just wanted to pass along some information here. While gpt-4o is much better at coding, it’s not truly better in every area. For anyone whom has seen maybe a litttllleee bit better responses for their use case, this post is for you.

I’ve discovered if you have a formatted summarization. Meaning a bot that is used for summarization of something or the like, you can basically just run a gpt-4o call, and copy and paste that response as a one shot example into 3.5 and change the model to 3.5 and you’ll just start getting responses with similar depth and clarity as gpt-4o

Don’t pay 10x the cost for your business when you can jerryrig a 3.5 model to get the same output as a 4o model.

In other words, don’t always believe the hype. Yes the multi modality is awesome and a huge step forward. But if you are using GPT in a way where you’re sending it data and expecting a formatted response each time, like data summarization (rather than a chat bot), you can indeed get 3.5 to respond similar to gpt-4o

Just passing this on hoping it saves someone 10x the money! :slight_smile:

8 Likes
2

I’ve never thoith about this.
But how it is compatible this approach with the smaller context wimdows of 3.5? Could you provide a practical example?

Thank you!

3

who is the best foot ball player

4

Can’t understand. What do you mean?

5

I mean that just all depends on what you’re trying to accomplish. If its too much for your context window you might be screwed, you could also feed less data into it?

The idea is to tell 3.5 with a one-shot prompt, sound similar to this ‘X’ where X is the gpt4o response you’ve already gotten.

So for instance if you want it to summarize a court room hearing or something - if you get that 4o summary, stick that into 3.5 say “sound like this” (obviously not that simple of a prompt but for the example in this message) and see if you get it to respond similarly to 4o