Note to Self: My Views This Afternoon on the Current State of MAMLM GPT LLM Pseudo-AGI
Behind the paywall because this is of interest only to those genuinely interested in closely following my MAMLM-interpretation & assessment journey. Current viewpoint: Pantomimes of thought: LLMs working so weirdly well except where they do not, as we traverse the landscape of letter, number, and code from kernel smoothing to Beelzebub and examine the pseudo-AGI, where its successes are illuminating and its failures perhaps even more so. See “AI” as a gigantic, overconfident parrot doing cold reading at scale, both the magic and the madness snap somewhat into focus—astonishingly good at pantomiming human thought, and alarmingly bad at keeping their stories straight, unless their code does not run, in which case they stamp their Clever Hans feet again and again until it does run…
To recapitulate:
At one level, what an MAMLM GPT LLM is is a roiling boil of linear algebra that does this:
It takes every single human conversation and piece of text it can get its hands on.
It then matches that block of text to what the next word in that conversation is.
So you then feed the LLM a piece of text.
It then picks a block of text out of its training data that it judges to be “close” to the piece of text you fed it. (How does it do that? Mathymath.)
It then responds to you with that next word.
And then it continues on by adding the word it just gave you to the block of text and doing the whole process again.
Now this is not “thinking.”