Agent-0: Replicating O1's Chain of Thought Reasoning

Watch the Video Overview
For a detailed walkthrough of how this project works and what to expect, check out the video linked above.

Project Description

This project is a proof of concept that aims to replicate the reasoning capabilities of OpenAI's newly released O1 model. O1 uses chain-of-thought prompting and reinforcement learning to reflect on its solutions, improving responses through iterative reasoning. Our goal is to mimic this behavior using alternative models.

In this implementation, we use a sequential agent-based system powered by the Gemini API (or any model with function-calling capabilities). The system proposes solutions to coding-related problems and iteratively refines them using chain-of-thought and reflection techniques at each stage. The Gemini API, with its code execution abilities, is ideal for this project. While it works with Gemini Flash, we recommend using the Pro version to avoid issues with external package dependencies, as the Pro version generally sticks to Python's standard library.

Important Note

This is a very early version and was created as a weekend hack, so expect it to fail in various scenarios. It currently works best for problems that can be solved through coding. We encourage you to give it a try and report any bugs or issues you encounter.

How to Run

1. Set Environment Variable

You need to set an environment variable for your Google API key:

export GOOGLE_API_KEY=<your_api_key>

2. Run the Script

python main.py

2. Create a Conda Virtual Environment

It's recommended to use a Conda environment for this project. To create and activate a new Conda environment:

conda create -n agent-0 python=3.10
conda activate agent-0

3. Install Dependencies

The only dependency required for this project is google-generativeai. Install it using pip:

pip install google-generativeai

4. Run the Script

python main.py

Give it a try and let us know what you think! Make sure to give it a star if you enjoyed it.