
Description
Today's chatbots talk through model weights. Sixty years ago ELIZA held conversations by matching what you typed against patterns and filling in templates; its limit was that every rule had to be written by hand.
Hugging Face handed that job to a swarm of over 1,000 agents: build a chatbot in pure Python with only if-else branches, regex and lookup tables, no weights and no training. The agents wrote and evaluated over 5,000 versions, and the win rate on SmolTalkBench climbed from 0% to 26%, on par with SmolLM-360M-Instruct, once state of the art among models under 1B parameters.
The result is PyBot: around 640k lines of code that answer everyday questions, backed by 550k lines of tests the agents wrote themselves. It needs Python 3.8 or newer and nothing else, with no packages to install and no model to download.
Chat in the terminal: clone the repo and run python -m pybot; /reset starts over, /exit quits, and it remembers what you told it, like your name.
One-shot questions: -p asks one question and exits, or pipe messages in for one reply per line.
Try it online: a Hugging Face Space shows which of PyBot's modules produced each reply.
Zero dependencies: Python 3.8+ only; startup takes a few seconds to load 640k lines, then about a second per reply.
A visible experiment: the repo charts win rate and lines of code across commits, so you can watch the rules pile up.
Hugging Face handed that job to a swarm of over 1,000 agents: build a chatbot in pure Python with only if-else branches, regex and lookup tables, no weights and no training. The agents wrote and evaluated over 5,000 versions, and the win rate on SmolTalkBench climbed from 0% to 26%, on par with SmolLM-360M-Instruct, once state of the art among models under 1B parameters.
The result is PyBot: around 640k lines of code that answer everyday questions, backed by 550k lines of tests the agents wrote themselves. It needs Python 3.8 or newer and nothing else, with no packages to install and no model to download.
Features
Chat in the terminal: clone the repo and run python -m pybot; /reset starts over, /exit quits, and it remembers what you told it, like your name.
One-shot questions: -p asks one question and exits, or pipe messages in for one reply per line.
Try it online: a Hugging Face Space shows which of PyBot's modules produced each reply.
Zero dependencies: Python 3.8+ only; startup takes a few seconds to load 640k lines, then about a second per reply.
A visible experiment: the repo charts win rate and lines of code across commits, so you can watch the rules pile up.
Tags:Chatbot

