I wrote 100% of my post myself. There’s no bragging going on. I’m warning people that I learned it from chatgpt so they know it may not be trustworthy. Still in my experience the vast majority of what I learn from ChatGPT/gemini/claude is correct, and in this kind of case where the knowledge is interesting but not something I’m going to need to rely on for anything, the benefits of quickly learning are so high for me that I thought someone else might appreciate it too. I’m sorry if I offended you.
> the vast majority of what I learn from ChatGPT/gemini/claude is correct
Roughly 90% according to benchmarks. Which means you're learning 10% bullshit.
> the benefits of quickly learning are so high for me that I thought someone else might appreciate it too
ChatGPT has been there for two and a half years already, of course everyone on this forum knows about it, you don't have to tell people how cool it is… Rather you can safely assume that if someone asks a question here, and not as a ChatGPT prompt, it means they doesn't want an answer from ChatGPT!
I didn’t give you an answer from chatgpt. I’m not telling anyone how cool it is. I’ve been here for years and I’m well aware of the level of exposure to LLMs there is for people on HN .
I did give out a warning to people that I had learned something from chatgpt, not from some other source, so they could take it for what it was worth.
I’m so tired of seeing people in other forums try to help someone and having someone give them a shit response. One thing I like about HN is the fact that that doesn’t happen much here. I still think what I wrote in response to your question was helpful and thoughtful. Back up a minute and just think it over, please.
I’d love to see which benchmarks you’re talking about, and if they apply to all three of those specific LLMs, and which version of their models, and if they differentiate topics. My guess is that in the case of LLM output that is a response to a purely factual question, or a response to a question asking a technological question where I present it with something I know occurs but want to know the means by which it occurs, the amount of hallucination is much less than 10%. And if you have some background in the area you are asking about, you can easily filter out some of what is hallucinated, making the learned material well about 90%. But I’m just guessing based on my experience of asking LLMs a lot of questions about both things I already know quite well and others that I don’t. So, which benchmarks are you talking about?
Why the hell are people doing that over again? Nobody asked you to copy-paste a plausible AI-generated response you have no idea about its accuracy.
That some people like do it is depressing, but the fact that you are bragging about that is truly beyond words.