PyBot: A Conversational Bot in Pure Python Without Neural Networks
The PyBot project by Hugging Face is an experiment in creating a conversational bot using pure Python alone. Unlike traditional language models, the solution does not use neural network weights, matrix multiplications, or transformers.
Development was carried out using a swarm of over 1,000 AI agents that applied if-else branching rules, regular expressions, and lookup tables. During the evolution process, more than 5,000 code versions were generated and evaluated.
The final codebase consists of 640,000 lines of working code and 550,000 lines of self-written tests. To run it, you only need a Python 3.8+ interpreter without using a GPU or external libraries.
Results from the SmolTalkBench benchmark show an increase in the deterministic bot's win rate from 0% to 26%. This brought PyBot to the performance level of the SmolLM-360M-Instruct neural network.
Key features of the project include complete absence of dependencies and instant startup via the python -m pybot command. The project is available open-source for study and testing.
- 640k lines of code and 550k lines of tests created by a swarm of 1,000+ agents
- 26% win rate on SmolTalkBench ensures parity with the SmolLM-360M-Instruct neural network
- Absence of weights and dependencies allows running the bot on any CPU










