[danieka]

My question is, did you learn the truth from ChatGPT? It seems like it will always give plausible sounding answer, that may or may not be correct. For example I asked it today to enumerate the rights given to all christians as part of the priesthood of believers according to Luther. It enumerate a couple of plausible sounding rights. Unfortunately none of them were correct.

[musingsole]

It's less like finding a Wikipedia article and more like asking a gossipy cousin about a topic.

It tends to land in the right ballpark and capture the tone of a thing.

[visarga]

> It tends to land in the right ballpark and capture the tone of a thing.

One weird trick is to use this text to search the web. Because it is already in the ballpark it will be a very good query. The model would bullshit "The height of Everest is 8,230 m", you take this string and search it, find "height of Everest is 8,849 m" -> correct the original response.

Now the problem is, how do you decide what information to trust in search?

AI needs a big push to index and do consistency checks for all facts. Let's have the model write a billion wiki entries and knowledge base concepts. Check for support, consistency, competing explanations - everything should be in there, we don't decide what is true. Then a language model can say when a topic is controversial, or when it doesn't know something.

Of course the goal is to find the truth, but we know it is going to be tricky. In the meantime we can have models that know when they don't know, or when the information is not certain.

[MAGZine]

I definitely agree with this, but what I have found it useful for is for discovering 'unknown unknowns' that I can then research in more depth.

It's too bad that it produces such plausible-sounding answers while being subtly wrong. But it can certain help people to expand their knowledge. I don't think it's a replacement for a comprehensive education on a topic, but can help people get closer.

[rib3ye]

I am learning a JS game dev platform and ChatGPT has been a GREAT help, even though it’s only right 90% of the time.

It cooler if it were an “auto-answer” tool for Stack Overflow. That way I could still get instant answers, and the community could asynchronously validate the answers.

[Jcowell]

If you’re learning something, glow do you know if it’s correct or wrong. I can see it as a tool to remind you (like damn I forgot how to reverse a linked list) but not on a topic that you wouldn’t be able to remember or not since you didn’t know it in the first place.

[tester457]

I don't like asking it for real life facts. It's better at writing emails, simple code, bash scripts, grunt work, etc.