The No Language Model Post


This is a small project written by Christopher De Vries to explore the current craze over Large Language Models like ChatGTP. ChatGPT is not magic, and it is not a general intelligence. The results it produces may seem extraordinary, but as we have recently seen the remarkable results that we get from ChatGPT have no guarantee of being accurate. This reminded me of some work I had done with a very simple predictive algorithm back in the 90s.

In contrast with Large Language Models like ChatGPT, this is a No Language Model. It is based on a Markov chain which is a process developed in 1906 by which we predict the next item in a list using the previous N items and a probability model based on existing sequences. In this case the existing sequence is just a sequence of letters from news articles about ChatGPT, and we predict the next letter based on the previous 6 letters.

Markov chains can produce articles which look surprisingly similar to English text, but they contain no vocabulary database, know nothing of grammar, and have no understanding of the topic about which they write. A Markov chain is just a simple algorithm. LLMs may be more complex, but they fundamentally have similar limitations. To say that they “hallucinate” or “dream” is plainly false. They are inherently uncreative and unreliable, generating random babbel using a large volume of human-produced text.

You can see how this site works by exploring the code on github.