There is like zero line of sight to how to do this in current leading edge ML research.
Conversation
Replying to
Doesn't this depend on whether scaling holds? My impression has been that the scaling argument as applied to prosaic LMs extends to the idea of learnable logic.
1
Replying to
what do you mean 'scaling argument'? I am not sure what you're referring to.
1
1
Replying to
Roughly, as AI models are scaled up (data + compute + model params) they develop deeper, more general and abstract reasoning capabilities.
The view among proponents of the scaling hypothesis is that these capabilities may (will) eventually include principles of logic.
1
Replying to
Ah, okay. Yeah, I've heard that argument, but I think it's kinda a leap of faith at this point. There's no clear reason to believe it other than that biological brains seem to be an existence proof if you squint enough.
1
2
Replying to
hmm I've tended to take it more seriously, given that it seems to explain the history of language modelling quite well.
As model size, data, compute ↑, models went from writing words → phrases → sentences → paragraphs, to now being able to write blogs & do math (GPT+)
1
The math thing might be at the core of the question: GPT-3 can add small numbers (~3 digits) together pretty well but fails for larger ones
So an interesting question (currently debated) is whether it's "learned math" or is just babbling really well
1
1
Replying to
Yeah, it hasn't learned math. But it's not babbling either. It's just learned math-words as part of a probabilistic language model.
1
1
Replying to
Yes, but the key question for scaling is whether "it just learned math words as part of a language model" is any different from "it understands math words", or "it understands math".
Pro-scalers would argue that the answer converges to "no" in the high-scale limit.
1
Replying to
I'd call that a problem of insufficient introspection. It simply does not map to how humans seem to do math.
1
1
Via symbol rewrite rules that are referenced to a sensory ontology of "counting" for example... 1, 2, 3... is not just symbols for us, but pointers to cardinal segmentations of sensory fields for example.
This is why I think disembodied AI is a dead end. You need to put the brain in a robot body and have it be in a feedback loop with the real world and develop a domain-entangled intelligence. Math is not word games. We *count things* with it as a starting point.
1
1
1
ie the problem is not scale but the fact that AIs learn maps by observation rather than territories by action
1
1
Show replies

