3 ms·
> So I clearly remember wondering "Am I just a kind of slightly more advanced Markov chain?" I think humans employ markov chains (MCs) all the time. (Back in
by randcraw 8y ago
> So I clearly remember wondering "Am I just a kind of slightly more advanced Markov chain?"
I think humans employ markov chains (MCs) all the time. (Back in the days of symbolic AI, these became popular as 'frames', AKA case-based reasoning.)
But it's clear that human cognition far exceeds the capacity of simple probabilistic devices like MCs, much as context-free grammars exceed finite state machines. Eventually it became clear that too much of intelligence cannot be modeled viably using simple probabilistic mechanisms like MCs or frames (like memory, logic, and learning).
I'm hopeful that work like GPT-2 will accomplish the same revelation for the limits inherent in probability-based models via deep nets. As long as AI models fail to model semantics explicitly, they will forever create only narrow savants or general morons.