Re:Bard is getting better at logic and reasoning
by hackefeller on Monday June 12, 2023 @01:51AM
(#36288926)
Do you think the "next word prediction" argument is so popular because we want to believe our intelligence is more complex than it is?
I think some people get caught up on the “next word prediction” point, because this is just the mechanism. For the next word prediction to work, the LLM has all sorts of internal representations of the world inside it which is where the capability comes from.
Human reasoning probably comes from evolution (genetic survival/replication), and then somehow thought was an emergent behaviour that unexpectedly came from that process. A thinking machine wasn’t designed, it just kind of came to be over millennia.
Seems to be kind of the same with AI, but the first example of these emergent behaviours seems to be coming out of the back of building a next-word-guesser. It’s a little unexpected, but a simple framework seems to be allowing a neural net to somehow build representations of the world inside it.