3 ms·
It can do arithmetic on large numbers that were certainly never in its training data, not flawlessly but much, much better than chance. I do not see how this i
by retrac 4y ago
It can do arithmetic on large numbers that were certainly never in its training data, not flawlessly but much, much better than chance. I do not see how this is possible without some kind of model of arithmetic.
If you train a transformer or neural network to play checkers from a database of game moves, there will be 192 nodes in the model, tightly coupled to each other in roughly a square, cube, or hypercube geometry. 64 nodes for each of black, red, or unoccupied. Seems like a model of the game to me.
There is also the proof that gradient descent can approximate any function with arbitrary accuracy, given infinite time and samples. I don't understand the proof itself, but if that is true, then implies that machine learning can, in theory, model anything that can be described mathematically.
Does anyone actually argue these systems don't create models with some abstraction? Or is this something of a strawman we're all just tempted to tear down for some reason?
- famouswaffles 4y agoYou'd be surprised how many people are on the "nothing to see here" bandwagon despite how obviously false that is. It feels so ridiculous when someone says GPT doesn't reason or understand because I and many others can clearly show output demonstrating this. When presented with this, the next thing is to double down. "Oh forget what you're seeing, it's not "true understanding" TM". This just makes the conversation even more bizzare. Either your dog fetches what you throw at it or it doesn't. The idea of "pretend fetching" is so obviously silly and it's exactly what these people argue. I'm supposed to believe in this mysterious special distinction that nobody can apparently properly define nevermind test for. General understanding isn't binary. You can understand some things and not understand other things. And it's not like everything encoded in language is encoded at the same level. GPT-4's spatial understanding is much better but that doesn't mean GPT-3 didn't understand anything at all. It just means it didn't understand space.
- retrac 4y agoI consider the fact that these systems appear to create a very complex human-like internal model to be one of the, quite suddenly, most pressing philosophical problems of our time. If we had introspection into ML models, would they be sensible, like how humans think and model in the abstract? Is there a "carry the one" weight somewhere, that's linked to numbers that triggers when GPT is given a pattern that invokes that subnetwork, it adds almost like humans do with pen and paper? Or is it entirely different? I'm not sure we can say whether it "understands" until some questions like that have a start of an answer. Perhaps it's possible to model and predict something, even abstractly reason, without understanding that can be related in a way a human would comprehend. That would be unsettling in its own way.
- famouswaffles 4y agoFor me, if the output can consistently demonstrate understanding then it understands. The reason LLMs are able to be plugged into everything under the sun is that general intelligence and understanding. It's not a parlor trick. At this point you can say, it's only mimicking understanding but that's just a circular argument. If mimicry understanding is such an important distinction over true understanding then it should be obvious. I should be able to test for it. If you can't test for it then it's just an arbitrary distinction you think is important but paradoxically isn't affecting the output at all. Not saying LLMs understand exactly like humans. Maybe they do. Maybe not. But they understand.