Last night, I live streamed myself coding this Llama 2 Agent on a Single GPU (Colab). After 6 hours it actually has some good results.
How it works is it takes in your intuition (e.g. "I think x would be cool") and develops a business idea (with a name and branding colors) and a business plan. After the business plan is developed, it criticizes this plan recursively until the "Investor" prompt is satisfied with the plan. After all this it will generate the final MVP idea and pass it to a the React Engineer Agent I live coded 2 days ago (https://github.com/jawerty/10x-react-engineer). Finally once the code is generated you have an option to pivot.
As for the AI's memory, I have a small vector db for history of good intuition+idea examples to use as context (builds as you generate) and a simple backlog of business idea criticisms.
Overall, this experiment showed me that you definitely don't need OpenAI keys to have an effective LLM Agent for you. This is only using the Llama 13b parameter chat model and doing a wonderful job. I spend maybe 5 dollars of Colab compute building this for 6 hours which is way more than it would be to mess around with this.
One caveat is the final react code base often needs some modifications to the package.json/webpack setup. However, It's generating the components/folder surprisingly well starting from the entrypoint file. I addressed most of these concerns in this video but still going to fine tune a model that's way better in my next video or just move to a fine-tuned starcoder for this portion.
Let me know what you think and play around with the Colab if you can!
The live stream: https://youtube.com/live/ZCyxDRSw0wE The 10X-React-Engineer: https://github.com/jawerty/10x-react-engineer
If you want to see me live code more let me know any ideas you'd like to see built. I've been streaming projects like these most days this past month and don't feel like stopping haha.