Comment by demosthanos 21 hours ago
This is wrong. I write toy languages and frameworks for fun. These are APIs that simply don't exist outside of my code base, and LLMs are consistently able to:
* Read the signatures of the functions.
* Use the code correctly.
* Answer questions about the behavior of the underlying API by consulting the code.
Of course they're just guessing if they go beyond what's in their context window, but don't underestimate context window!
This is moving goalposts vs the original claim upthread that LLMs are just regurgitating human-authored stackoverflow answers and without those answers it would be useless.
It’s silly to say that something LLMs can reliably do is impossible and every time it happens it’s “dumb luck”.
If that's the distinction you're drawing then it's totally meaningless in the context of the question of where the information is going to come from if not Stack Overflow. We're never in a situation where we're using an open source library that has zero information about it: The code is by definition available to be put in the context window.
As they say, it sounds like you're technically correct, which is the best kind of correct. You're correct within the extremely artificial parameters that you created for yourself, but not in any real world context that matters when it comes to real people using these tools.
The argument is futile as the goal posts move constantly. In one moment the assertion is it’s just megacopy paste, then the next when evidence is shown that it’s able to one shot construct seemingly novel and correct answers from an api spec or grammar never seen before, the goal posts move to “it’s unable to produce results on things it’s never been trained on or in its context” - as if making up a fake language and asking it write code in it and its inability to do so without a grammar is an indication of literally anything.
To anyone who has used these tools in anger it’s remarkable given they’re only trained on large corpuses of language and feedback they’re able to produce what they do. I don’t claim they exist outside their weights, that’s absurd. But the entire point of non linear function activations with many layers and parameters is to learn highly complex non linear relationships. The fact they can be trained as much as they are with as much data as they have without overfitting or gradient explosions means the very nature of language contains immense information in its encoding and structure, and the network by definition of how it works and is trained does -not- just return what it was trained on. It’s able to curve fit complex functions that inter relate semantic concepts that are clearly not understood as we understand them, but in some ways it represents an “understanding” that’s sometimes perhaps more complex and nuanced than even we can.
Anyway the stochastic parrot euphemism misses the point that parrots are incredibly intelligent animals - which is apt since those who use that phrase are missing the point.
So, you're saying you provided examples of the code and APIs and more, in the context window, and it succeeds? That sounds very much unlike the post I responded to, which claimed "no knowledge". You're also seemingly missing this:
"If you're getting answers, it has seen it elsewhere"
The context window is 'elsewhere'.