Language and Thought

Recurrent interactions by LLMs collapse into nonsense. https://www.nature.com/articles/d41586-024-02355-z

AI models collapse when trained on recursively generated data - Nature

Interestingly, when people communicate they add something on each interaction. This has something to do with language being only approximate in how it maps onto thought. And yes, SOME interactions have exactly this collapsing structure …. like in politics when people don’t listen but go on repeating phrases …, and in other situations when people exchange words without thinking. But while in general language can be used as an exchange of clichés, and it certainly can be used to hide your own thoughts and express fictive content (aka as lying), its main use as a tool is not communication, but rather as a way to structure and order, to formalize, and retrieve memorized content.

And here the properties of “natural language” - as it was called since we invented programming languages — come into play. (We invented formal notations a long time before that, and used them in mathematics, but no one seemed to want to confuse them with the way humans spoke and wrote, and therefore it was sufficient to use just “language” or “speech”, and possibly “human language” to distinguish it from the communication systems of animals. ) These linguistic properties are derived from human thought and…