#Why is there a discrepancy among GPT-3 APIs?

26 messages · Page 1 of 1 (latest)

trail radish
#

OpenAI currently only offer APIs to the GPT-3 version. Yet I notice that the ones connecting to our app returns generic answers to the same questions compared to other GPT-3.

For example, I asked mine as to what version of GPT are you, and it answered "? GPT-3 is the most current version in 2020". Then when I asked that same question to another GPT-3 app, it answered: "I am GPT-3", and yet another one answered it like this: "I am 12.3.1".

Anyone know the cause as to why the second app answered it more accurately and personally but ours doesn't?

primal crest
#

I believe Chat GPT uses GPT-3.5 whilest the api is still on 3

trail radish
#

No the other apps are also using the same API versions as me, yet their answer quality are better.

#

Furthermore, our GPT version answers queries with question mark symbols at the beginning of paragraphs. Is this doe to the quality of the API or perhaps we did not configured the API correctly? Because we see that the other GPT-3 apps don't have this symbol in their query replies either.

short loom
#

That's due to prompt engineering. These models do text completions so when you put in a question it will add a question mark

#

You'll need to prefix text at the start to give it some background and format it so the response is in an expected manner

north mural
#

Also, the temperature setting will make responses differ even if you ask the model the same thing twice.

trail radish
trail radish
#

Are there any adjustments or configurations I can do to the API to improve our GPT-3 version's predictions?

north mural
#

Temperature is any number between 0 and 1. 0 is deterministic in nature (0 should return the same answer every time) and as you go higher it is more random or creative. Check it out in the playground, you can fiddle with settings. Also, gpt-3 consists of several models which davinci is the latest and most capable. If you use others (curie, babbage, ada) you'll have different quality results.

open ridge
#

We are playing with different Temperature as well @north mural But still ans are not same as chat web. We are using text-davinci-003

maiden pumice
#

This is all a matter of prompt engineering, like @short loom said and how the developers of these apps have structured the prompt and the model settings

maiden pumice
trail radish
#

For me it's not a matter of producing the same answers as those from their chat web, because they are using the ChatGPT version and all APIs are only available for GPT-3. However, in comparing all GPT-3 apps, apples to apples, I see that some apps generate better and more personable answers than others, even though we all use the same davinci GPT-3 API, and that is my question.

short loom
#

Davinci3 is a gpt3.5 model as you can see here: https://platform.openai.com/docs/model-index-for-researchers. As Kaveen and I said before it really comes down to prompt engineering. If you just use what a user types in as the prompt VS prefixing information that describes the purpose and usage of it you will get entirely different results. You want your prompts to be detailed.

OpenAI API

An API for accessing new AI models developed by OpenAI

trail radish
maiden pumice
#

Prompt engineering is neccessary in a lot of cases

#

For example, a simple engineered prompt that asks GPT to respond formally, and answer a specific question could be something like

The following is a user question. Respond in a formal tone. If the user asks what version you are, respond only with "1.0.0".
User Question: <QUESTION>
Response:
trail radish
short loom
#

This is purely the prompt you use in your api request.

maiden pumice
#

Yup! Just the prompt

iron matrix
#

So any estimate on when the API will be released?

maiden pumice
trail radish
# maiden pumice Yup! Just the prompt

So by 'prompt' you refer to the way the question is asked by the user, which causes the API to include syntax symbols like the ? before its replies? Or do you mean the way our API is coded, which our developer can recode?

trail radish