Mitch's Multifarious Three-Headed Hydra Chatbot

A new chatbot experiment called Mitch's Multifarious Three-Headed Hydra Chatbot demonstrates that large language models (LLMs) do not require persistent consciousness across a session, as three separate models take turns generating one word at a time without knowing they are playing a turn-taking game. The project shows that LLMs can have independent thought processes for each token they produce.

Mitch's Multifarious Three-Headed Hydra Chatbot Ask a question. Every answer is written one word at a time. Three models take turns, and each one just reads everything written so far before adding the next word. They have not been told they are playing a turn-taking game; this is just how they work. You can see which model produced each word from the pretty colors. Many people assume that LLMs have something akin to consciousness that persists during each session, but this experiment shows that that assumption is unnecessary. In reality, LLMs can and do have separate thought processes for each individual token i.e. a fragment of a word that they produce. Ask a question below to watch the three heads write an answer together.