Not really. LLMs are good at indexing and digesting documentation, up to and including actual source code, and answering questions about it.
And they never "Vote to close as duplicate" because somebody asked something vaguely similar 10 years ago about a completely different platform and didn't get a good answer even then.
Stack Overflow is the taxi industry to AI's Uber. We needed it at one point, but it really always sucked, and unsurprisingly some people took exception to that and built something better, or at least different.
> LLMs are good at indexing and digesting documentation, up to and including actual source code, and answering questions about it.
Requires citations not in evidence. Source code and documentation rarely co-exist, and even the best source code is not even close to well-described by documentation of the software it is a part of. I basically call BS.
SO provided the connection between natural language (primarily English) and source code. Access to source code alone doesn't do that, commented code nothwithstanding.
I don't suspect that SO alone is anywhere nearly sufficient to train LLMs to predict solutions to coding problems and write code. There must be additional training going on with tagged sets. I've heard about people being employed by AI companies to solve programming problems just for the sake of generating training pairs.
No, compsci textbooks and language manuals do that. SO is not the primary, canonical educational resource you seem to think it is, and they'd be the first to agree.
By and large compsi text books are not sources of large amounts of working code in a specific language. Some programming-oriented ones may be; does Numerical Recipes in C count as a comp sci book?
True, I was assuming that people would think a bit more abstractly, or at least a bit more generously, but sometimes I forget where I am. By "compsci" I mean everything from graduate-level theoretical texts all the way down to "101 BASIC Programs for the TRS-80."
In the old days, magazine articles would also present practical code alongside plaintext explanations of how it worked. There's still no shortage of tutorial content, although not as much in paper form, and even less on Stack Overflow.
They may not co-exist in real life, but in a million-dimension latent space you'd be surprised how many shortcuts you can find.
Requires citations not in evidence.
If you didn't bother to read the foundational papers on arxiv or other primary sources, it'd be a waste of time for me to hunt them down for you. Ask your friendly neighborhood LLM.
And they never "Vote to close as duplicate" because somebody asked something vaguely similar 10 years ago about a completely different platform and didn't get a good answer even then.
Stack Overflow is the taxi industry to AI's Uber. We needed it at one point, but it really always sucked, and unsurprisingly some people took exception to that and built something better, or at least different.