I wrote a series of Markov chain chat simulators as a teenager. In fact I independently "invented" the concept, and I was crushed to discover I'd been scooped

I wrote a series of Markov chain chat simulators as a teenager. In fact I independently "invented" the concept, and I was crushed to discover I'd been scooped by several decades. I figured out a lot of tricks for making the text more coherent, but I abandoned the project when I decided the road to an artificial general intelligence did not pass through here: it's definitely attached to the main trunk, but represents a side-offshoot. My last thinking on the matter / current thinking in hindsight is that the apparent and peculiar power of Markov chains to model human language is simply a necessary consequence of the constraints of serialization, in much the same way that different libraries for serializing complex data structures might convergently evolve. Also, an ability to predict the likely next symbol provides some redundancy and error-correction ability. It could simply be that languages which approximately satisfy the Markov property are such a cheap and effective way to get error resilience that there's little reason for a language not to evolve that way. I use the word "approximately" deliberately: it was the realization that a chat AI making and holding coherent thoughts is incompatible with the Markov property that caused me to abandon my work.