One of my favorite books is Douglas Hofstadter's"I Am a Strange Loop", which picks up some threads from his more famous "Gödel Escher Bach" but also reflects him coping with the death of his soulmate wife. Hofstadter is trying to see if his ability to have conversations with Carol (based on his earlier history of a proven, high-fidelity ability to predict just what she would say) could in a way BE her living on in his head and heart - in a philosophical (and not merely poetic) way, or if that was just a consoling bit of wishful thinking.
In talking about mind and consciousness, he constructs a playful physical metaphor of the "careenium" - a bouncing-magnetized-pinballs thought experiment of how we come to model the world in our own craniums - a model rich enough to include ourselves as a model doing the modeling, and so on and so on.
Googling to try to remember the term "Careenium", I found this page explaining the concept and comparing it to GPT. One challenge you run into if you collaborate with GPT is that's it's not doing a great job of modeling the problem at hand in its virtual head: its model of the world is fairly static, and a conversation with it (as impressive as it is! Especially if you've played with the previous fruits of AI over the past decades) is just a probabilistic word journey through that static space.
In some ways it's right there in the name: GPT means "Generative Pre-trained Transformer", and the problem is the "Pre" - and earlier "Transfomers" were notably worse at keeping track of what was just said in the dialog.
I guess the implication is if GPT had greatly increased abilities to update its own model on the fly, if that process was more organically bootstrapped and ongoing, it might be a better candidate for "true" Artificial General Intelligence and even consciousness...
No comments:
Post a Comment