New Demo: Pawsitively – LLM content moderation API in 50 LOC, including the entire Phoenix app. Plus a LiveView front-end to play with it. Implemented and deployed entirely as a livebook! 🧵

Currently using mistral-large:123b, which is chonky, but smaller models can work with more effort. You should also check out llama-guard (8b) that was purpose built for detecting rule-breaking content. Could also combine llama-guard as first line, and larger LLM for feedback

With several examples of content that tries to trick the score evaluation, I haven't been able to get passed it. Of course LLMs are not perfect, and I did nothing more than give in a few shots of cheating. Feel free to try to trick it:

is 🤯. A phx app defined entirely in the notebook, with notebook code making requests to itself. Then you can just deploy the livemd file? Whaaat?! I don't think folks understand how far ahead we are in the notebook space. Will continue to accelerate :)

Shout out to @wojtekmach and @joao_lubien for putting Phoenix playground together. I think we can get this kind of thing into phoenix proper to make one off server scripts built-in supported. Of course it also uses req 😍

flycast docs and step-by-step LLM deployment video:

the livebook source (you'll need to set up your own LLM /chat/completions endpoint):

And as always, remember to keep the conversation positive 🐾

So basically, I can't use it

This is incredibly impressive. Recording an almost 6 minutes demo with low battery? Livebook and Playground are also cool I guess

Cool demo! Curious about with Livebook and Phoenix Playground if you were able to rerun the block with server? I tried and it complained about the port being already taken.