10-04-2023, 11:20 PM
And to clear up one anticipated source of confusion: I am not claiming that we will be able to get an LLM (at least, any of the current generation ones) to answer a query like, "Translate the following Voynich string into English, or if no translation is possible, explain why." Sadly, that's not how this works. If Bing AI or some other LLM "understands" Voynichese, it understands Voynichese in the same way that the person in the "Chinese Room Thought Experiment" understands Chinese. It will only be able to map Voynichese onto English if it finds that the best way to model Voynichese is to model the human or process that created Voynichese as attempting to encode English according to certain parameters.
Indeed, don't think of these LLMs as understanding "English" or "Japanese." Think of them as understanding "Internetese" or "Reality-ese." As far as they know, "reality" is a language game that they are incredibly good at playing, so much so that they know that, under the rules of "reality-ese," when someone asks, "what color is the sky," the LLM is supposed to answer "Blue, except at night when it is black, and at dusk when it can be pink or orange depending on the weather..." Even though the LLM has never seen the sky.
Indeed, don't think of these LLMs as understanding "English" or "Japanese." Think of them as understanding "Internetese" or "Reality-ese." As far as they know, "reality" is a language game that they are incredibly good at playing, so much so that they know that, under the rules of "reality-ese," when someone asks, "what color is the sky," the LLM is supposed to answer "Blue, except at night when it is black, and at dusk when it can be pink or orange depending on the weather..." Even though the LLM has never seen the sky.