Thing that would be very useful: possibility of communicating with the assistant

Community
1

one thing that would be very useful is the possibility of communicating with the assistant to make him understand the requests better, for example:
user: shows the first 3 names in the address book, each with the last two messages, example:
John
message1, message2
assistant:
bill call me tonight, we could have dinner together
…
user: messages go under the name please

2

What you describe is multi-shot training. Providing input and output exchanges before the ultimate input.

It is far better not to correct the AI, but just to show the AI responding in the correct manner to an example user input.

Here for example is just one-shot. I show the AI an input and then demonstrate how the assistant should answer. The second user input is the actual one satisfied by AI, and thus employs the same answering format demonstrated:

image
image666×680 20.4 KB

Without the example, the AI is a bit confused about what type of output it should produce:

image

Unfortunately, “assistants” does not offer placement of “assistant” role messages in a thread, only user questions. (Likely because they can make the AI work the way a developer wants…)

1 Like
3

Can you please send me in text format as I don’t know English and have it translated by Google?

Thanks in advance🙂

4

you can just drag and drop the images into ChatGPT and ask for a translation

5

It’s true, I’m still not used to it​:grin::wink:

thanks!

1 Like
6

In the API Playground, I first write a user question that doesn’t have any instructions, it just has some common metaphors, simple sayings that one might not know the origin.

Then I write my own assistant reply, that shows the AI printing each metaphor, and then providing a definition for each one.

When the AI then gets another list of metaphors, it knows the output format that was expected of it from seeing it previously.