18 comments

[ 27.6 ms ] story [ 341 ms ] thread
“Learning” by prompting, calculating the loss against evals, and updating the prompt
(comment deleted)
Yeah I've been thinking about this lately.

LLMs come and go.

Prompt engineering techniques come and go.

But eval / labelled dataset is always useful once you built it.

In my experience, architecture stays. A nice modular codebase (where you can switch-out the various models and have control on the minutia of the pipeline from data to model) brings a lot of value and can fairly easily be kept up-to-date with best practices.
(comment deleted)
Isn’t this just a very naive implementation of what DsPY does?

https://github.com/stanfordnlp/dspy

I don’t understand what is exceptional here.

(author) It's very similar. I think the extension of this is that we could build a frontend gui for something like this for prompt management. Love what they're doing
Post Author: getting a lot of requests so scaling the backend. Standby.
You should add a very short description of what this is and how it works. I kind of get that it is a way to tune/generate prompts for specific input output test cases, But its not clear exactly what it does.
Shouldn't this be LoRA training?
can you please add more info on the page to show why it is important and how its helpful
Yeah I'll do that now!
It'd be great to get a little write up of the technique
Added it!
Sorry to ask, but would you mind expanding a little more? For example, when you state "our system iteratively rewrites the prompt to better fit the desired behavior", what is the use case for this? Is it to make personal datasets? Thank you.
Just yesterday, I wrote an article about FT and learned about services like Entry Point AI.

Seems like an awesome idea. I'm curious how long it will take to get a model on a reasonable level.

Phind is pretty good and also the fastest model I used recently, so I'd assume it's quite small, no?

(comment deleted)