one of the oddest things i hear about ai is "they're just predicting the next token". this is said dismissively! it takes me some effort to model this position, because it is so far removed from my own -- not (just) in terms of opinions/conscious beliefs, but in terms of the felt experience of existing it reflects. regardless of what llms are or are not doing, "token prediction" gestures at something so familiar to me that i debate if calling it a metaphor for some of my own cognitive processes fails to go far enough.
remember in the misty dawns of time (2022) when people were calling chatgpt 3 a "bullshit machine"? as i recall this was typically pointing at how the truth value of the output was unreliable, but it also pointed to a similarity between the llm and human thought. when i am bullshitting--not telling a specific deliberate lie, but speaking or writing with unfounded confidence--my internal experience is of holding a certain posture and picking the next thing to say that sounds the most plausible given what i've just said, without knowing where even the current sentence is going to end. "token prediction" certainly feels like an accurate description of what i'm doing.
(there's an extent to which this feels partially true of all my speech and writing, but that's a much stronger claim in which i have less confidence)
however, the place where "predicting the next token" gave me a way to describe my internal experience that i had previously lacked was not in the production of language, but the processing of it. i am not hard of hearing, but i do struggle to understand the spoken word, especially if it's recorded. this is not a major issue so i haven't bothered to figure out exactly what causes it, and given that it provides some natural immunity from short form video, it's starting to feel like something of a blessing. it does mean that listening is an effortful act, and one i reflect on frequently. over the years, i've noticed that while i can adjust to most "how your vowels and consonants sound" accents pretty quickly, grammatical accents trip me up a lot. focusing on how i listened in those cases, i found that i was trying to fill in what i expected to hear, and that wasn't matching the actually sensory input. having seen it once, i found this pattern was generally present in how i processed speech. it was like i had "3 c's instead of phonics" illiteracy, but for the spoken rather than written word. like a reader struggling through an unknown word with limited understanding of the characters themselves, i was filling in the sounds that seemed most likely--predicting the next token to try to keep up.
i do not believe that the material basis of cognition can be found through introspection. experientially, however, a lot of it sure does feel like predicting the next token