site banner

Friday Fun Thread for February 24, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

I've been testing OpenAI on some questions on slightly obscure Finnish history, many of which it gets egregiously wrong. I run some of these questions on Bing. It gets them more correct, evidently thanks to its web search capability, but still commits some fairly obvious flagrant errors. Perhaps more on this later.

I suppose this is the equivalent of doing the "rigorous research" of literally looking up the first Google result, but this failure mode worries me. I've already heard a bunch of people say "I don't even use Google anymore, I just go to ChatGPT", but when Google serves you an answer, you at least know who published it, you can take their bias into account, and decide whether you want to keep looking or not. If ChatGPT becomes good enough to be trusted 99.9% of times, and it's only this 0.1% of cases that it gets egregiously wrong (whether by design, or accident), will people be able to find indications that the information is inaccurate, or will they just accept the false information as true?

Again, the problem in itself is nothing new, but I think we're pushing things in the direction of making people even easier to manipulate by the powers that be (or by stupid mistakes on obscure topics).

Bing tries to provide references.

Chatgpt is great for a certain class of question where you know the answer is true once you see it. E.g. ask it how to do something with a certain JavaScript library and it will give you the code. It's then very easy to confirm the validity by cross checking the documentation.

Generally I treat chatgpt as a wise but crazy old man. It will say insightful things, give you reference and names and ideas, but it's up to you to go away and confirm and research and come back with followup questions. Using chatgpt has to be embedded with your own research process. It can't be used in isolation, like you say.