PyBot’s creators said on November 16 that roughly 1,000 AI agents spent a week building a chatbot that answers through Python rules rather than an LLM. Source Source
According to the project’s README, the result contains about 640,000 lines of code across 528 Python files. Its conversational machinery uses conditional branches, regular-expression patterns for matching text, and lookup tables—not neural-network weights. “No LLM” describes how PyBot answers, not how it was built. The construction crew was very much AI-powered. Source Source
In the account posted by _lewtun, each construction agent was an instance of DeepSeek-V4.1-Flash, working on capabilities including understanding, answers, memory and routing. The agents coordinated through a Slack-like message board, exchanging about 175,000 messages and merging roughly 8,000 pull requests into a shared repository. Apparently, removing the LLM from a chatbot still leaves room for an extremely busy group chat. Source
How the rulebook answers
The README describes a system in which conversational modules can each propose a reply on every turn. An arbiter selects the answer the user receives. The modules draw on shared components for understanding messages, remembering the conversation, looking up knowledge and composing responses. Source
The browser demo’s interface invites users to see which files produced a reply and identifies the winning module. The page says PyBot runs in the browser using Pyodide. Source
What “on par” actually means
The README says the agents wrote and evaluated more than 5,000 versions, raising PyBot’s SmolTalkBench win rate from 0% to a rounded 26%. In that evaluation, an LLM judge compares each reply with a reference answer. The project says the result puts PyBot on par with SmolLM-360M-Instruct. Source
The reported 26% is a win rate against reference answers, not a factual-accuracy score or a head-to-head win rate against SmolLM. The supplied account doesn’t include independent testing or say how the judge was set up. It also doesn’t establish whether PyBot encountered the test questions during development, leaving its performance on unfamiliar questions unclear. Matching SmolLM on this benchmark isn’t an eviction notice for LLMs. Source
For readers who want to inspect the rulebook, the project links its code and a browser demo. The README documents a terminal route requiring Python 3.8 or newer, with no external Python packages or model download: Source
git clone https://github.com/huggingface/pybot
cd pybot
python -m pybot
Comments