Coding was livestreamed here -> https://youtube.com/live/ZCyxDRSw0wE
AutoStartup is a Llama 2 autonomous agent built around my previous https://github.com/jawerty/10x-react-engineer project. It divises a startup idea, business plans and generates React codebases all based on a simple user input "intuition". I built this on a live stream on August 16th, 2023 in 6 and a half hours. Have been challenging myself to build more and have fun doing it!
First, AutoStartup takes in an "intuition" from the user. An example being "I think a website for dogsitters would be cool". Then the Agent divises a business idea from this as well as a business plan. It will iterate on the business plan with "Criticisms" (similar to AutoGPT). After each criticism it'll adjust the business plan and "pitch" to an investor prompt. The investor will either approve or disapprove of the final plan. After this final approval it will generate the code base and perform "pivots" based on user feedback
- User intutiton (e.g. "I think a website for dogsitters would be cool")
- Idea Loop
- Develops initial business idea
- Writes business plan
- Criticize loop
- Criticizes business plan and generates new plan
- Uses new plan to ask "investor" for approval
- Upon approval continue
- Else: restart loop with new plan
- Generate a product idea from finalized business plan
- Generates React Codebase
- Utilize 10x-React-Engineer (a project I built on a stream a couple days before this one) to generate the codebase
- (optional) Make a pivot based on the user's feedback
- 100% Llama 2 Inference
- No OpenAI keys necessary
- React Codebase Generation
- Memory
- Vector search for historically successful ideas paired with intuitions.
- Previous criticisms are utilized for investor approvals
- Lean Startup concepts
- Utilizes lean startup methodologies (pivot, tight mvp build loop)
Please understand you need to have a quality GPU that can load the LLama 2 13b parameter chat model. Hpwever, I built this all on Google Colab and think this is a great way to get around the costs of AI Agents + GPT-4.
If you want to run from source run these commands. (After getting access to llama 2 and logging into huggingface-cli)
Install
$ pip3 install -r requirements.txt
Run the main loop
$ python3 main.py
Here's a Google Colab with the code for you to play with
- Fine-tune a llama 2 chat model to be better at coding react. So far it's ok but needs some bug fixing more often than not