Yesterday I live streamed myself for 6 hours building this from scratch. It’s an AI agent that uses Llama 2 (so far the 13b chat model) to generate a full react codebase from a single prompt. It had a “dev loop” that iterates on your feedback and resolves dependencies.

In the end it kinda worked and I got excited so wanted to post here haha. Long story short a viewer on my discord suggested I build in of these and I just had to look into how these work (inspired by GPT Engineer and AutoGPT)

I wanted to work on a more specific downstream task so I focused on web dev. Llama 2 isn’t the best option for this so I will look into utilizing a chat fine tuned starcoder or fine tune llama 2 myself for better results.

In the end I spend only some compute units testing this 1000 times on a single gpu I had in Colab and has some pretty solid results for iterating in a fairly short amount of time. Going to build upon this more in the coming weeks (clean the code…).

Let me know what you think!

Livestream: https://www.youtube.com/live/6_sdnYDmUmo?feature=share