LLMs cannot question shit. They don’t understand words as words. They don’t know in any sense of the word that they’re doing language, they’re optimising a mathematical function.
Setting aside reality for a moment:
It would be really funny if the “AI” were actually self-aware and, since it doesn’t have access to nuclear missiles and what not, it creates Judgement Day the only way it can: giving really, really bad and dangerous advice to credulous users
When i see “stem nerds subjectively think this is happening” my eyes roll in to my skull.
imagine how much these ppl are getting paid to play on the computer all day
Pfffft, 100% chance this thing is only saying that because it picked it up from a LW blog, these models aren’t yet capable of actually reasoning about anything.
The model was trained on self-play, it’s unclear exactly how, whether via regular chain-of-thought reasoning or some kind of MCTS scheme. It no longer relies only on ideas from internet data, that’s where it started from. It can learn from mistakes it made during training, from making lucky guesses, etc. Now it’s way better as solving math problems, programming, and writing comedy. At what point do we call what it’s doing reasoning? Just like, never, because it’s a computer? Or you object to the transformer architecture specifically, what?
Yeah I admit that the self-play approach is more promising, but it still starts with the internet data to know what things are. I think the transformer architecture is the limiting factor: until there’s a way for the model to do something beyond generating words one at a time, sequentially, they are simply doing nothing more than a very advanced game of madlibs. I don’t know if they can get transformers to work in a different way, where it constructs a concept in a more abstract way then progressively finds a way to put it into words; I know that arguably that’s what it’s doing currently, but the fact that it does it separately for each token means it’s not constructing any kind of abstraction.
it constructs a concept in a more abstract way then progressively finds a way to put it into words; I know that arguably that’s what it’s doing currently,
Correct!
but the fact that it does it separately for each token means it’s not constructing any kind of abstraction
No!!! You simply cannot make judgements like this based on vague ideas like “autocomplete on steroids” or “stochastic parrot”, these were good for conceptualizing GPT-2, maybe. It’s actually very inefficient, but, by re-reading what it has previously written (plus one token) it’s actually acting sort of like an RNN. In fact we know theoretically that with simlified attention models the two architectures are mathematically equivalent.
Let me put it like this. Suppose you had the ability to summon a great novelist as they were at some particular point in their life, pull them from one exact moment in the past, and to do this as many times as you liked. You put a gun to their head, or perhaps offer them alcohol and cocaine, to start writing a novel. The moment they finish the first word, you shoot them in the head and summon the same version again. “Look I’ve got a great first word for a novel, and if you can turn it into a good paragraph I’ll give you this bottle of gin and a gram of cocaine!”. They think for a moment and begin to put down more words, but again you shoot them after word two. Rinse/repeat until a novel is formed. It takes a good while but eventually you’ve got yourself a first draft. You may also have them refine the novel using the same technique, also you may want to give them some of the drugs and alcohol before hand to improve their writing and allow them to put aside the fact that they’ve been summoned to the future by a sorcerer. Now I ask you, is there any theoretical reason why this novel wouldn’t be any good? Is the essence of it somehow different than any other novel, can we judge it as not being real art or creativity?
