the beautiful code
-
You say they have no knowledge and are only good for boilerplate. So you're contradicting yourself there.
I didn't say they have no knowledge, quite the opposite. Here a quote from the comment you answered:
LLMs are extremely knowledgeable (as in they "know" a lot) but are completely dumb.
There is a subtle difference between intelligent and knowledgeable. LLM know a lot in that sense that they can remember a lot of things, but they are dumb in that sense that they are completely unable to draw conclusions and put that knowledge into action in any other means besides spitting out again what they once learned.
That's why LLMs can tell you a lot about about all different kinds of game theory about tic tac toe but can't draw/win that game consistently.
So knowing a lot and still being dumb is not a contradiction.
-
Is there a chance that's right around the time the code no longer fits into the LLMs input window of tokens? The basic technology doesn't actually have a long term memory of any kind (at least outside of the training phase).
Was my first thought as well. These things really need to find a way to store a larger context without ballooning past the vram limit
-
Was my first thought as well. These things really need to find a way to store a larger context without ballooning past the vram limit
The thing being, it's kind of an inflexible blackbox technology, and that's easier said than done. In one fell swoop we've gotten all that soft, fuzzy common sense stuff that people were chasing for decades inside a computer, but it's ironically still beyond our reach to fully use.
From here, I either expect that steady progress will be made in finding more clever and constrained ways of using the raw neural net output, or we're back to an AI winter. I suppose it's possible a new architecture and/or training scheme will come along, but it doesn't seem imminent.
-
The thing being, it's kind of an inflexible blackbox technology, and that's easier said than done. In one fell swoop we've gotten all that soft, fuzzy common sense stuff that people were chasing for decades inside a computer, but it's ironically still beyond our reach to fully use.
From here, I either expect that steady progress will be made in finding more clever and constrained ways of using the raw neural net output, or we're back to an AI winter. I suppose it's possible a new architecture and/or training scheme will come along, but it doesn't seem imminent.
I fell like the way investments are currently made, coming up with something new is made almost impossible. Most of the hardware is designed with LLMs in mind
-
I actually was going to link the same one I always do, which I think I heard about through a blog or talk. If that's not good enough, it's easy to devise your own test and put it to an LLM. The way you phrased that makes it sound like you're more interested in ignoring any empirical evidence, though.
That’s unreal. No, you cannot come up with your own scientific test to determine a language model’s capacity for understanding. You don’t even have access to the “thinking” side of the LLM.
-
That’s unreal. No, you cannot come up with your own scientific test to determine a language model’s capacity for understanding. You don’t even have access to the “thinking” side of the LLM.
wrote on last edited by [email protected]You can devise a task it couldn't have seen in the training data, I mean. Building a comprehensive argument out of them requires a lot more work and time.
You don’t even have access to the “thinking” side of the LLM.
Obviously, that goes for the natural intelligences too, so it's not really a fair thing to require.
