5 ms·
ChatGPT can provide the same answers SO can because it mined SO as part of its training data.
by jibal 11mo ago
ChatGPT can provide the same answers SO can because it mined SO as part of its training data.
- nomel 11mo agoFrom my experience, the majority of answers (accepted or highest voted) are wrong/outdated, with the second or third being the correct answer.
- Aurornis 11mo agoIronically, that's where LLMs seem to have an advantage. They're not regurgitating the top Stack Overflow answer. They've scraped the entire Stack Overflow page, including context in additional answers, and they've also been trained on the same question from countless other parts of the internet.
- jibal 11mo agoAn advantage over what? Reading an SO page "scrapes" the whole page with your eyeballs. Anyway, I didn't say anything about which is preferable to use, I merely commented on where the information comes from, so your response is non sequitur. If sources like SO go away, it will affect the quality of LLM responses. "the same question from countless other parts of the internet" is a figment of your imagination when it comes to the kind of content SO accumulated. I've been in this industry for a very long time (my name appears in RFC #57) and I know what it was like to try to get answers to technical questions before SO showed up.
- jibal 11mo agoThen your experience is extremely limited or you're misreporting it.
- nomel 11mo agoOr, the context of your use, for a site that covers many thousands, is different than mine, leading to a different experience. The most-being-wrong applies very reliably for any API, library, or language that isn't stable through time. You probably work C, or similar, where backwards breaking changes, or shifts in "the correct way", are rare. I'm jealous! The answer being wrong at any point in time is probably < 10%, in my experience.
- jibal 11mo agoYou've basically confirmed what I said.
- nh23423fefe 11mo agoYes. Giant disorganized piles of atomized information are less useful than optimized integrated data representations queryable by LLMs.