91–100 of 143 posts
I think it could work - but you'd have to find a way to let ChatGPT exclusively deal with the sort of repeat questions that are rather annoying to human posters - while all the actually interesting, new questions (which are also more likely to cause ChatGPT to create a flawed answer), are reserved for humans. (Also don't let ChatGPT pose as a human - give it it's own user name and automate it.)
I think the ban on ChatGPT answers on Stack Overflow is a mistake. Sure, some of the answers might not be correct, but that's true for any tool. And the fact that ChatGPT answers are easy to produce means that more people can contribute, which can only be a good thing for the community. Plus, it's not like the moderators have to accept every answer - they can still weed out the bad ones. Let's not punish everyone bec…
Can't agree. When writing code I actually pretty sure that 1 + 1 = 2, and 100 + 100 = 200. Of course, of course it could be not true for all languages and all environments, but in one language and in one environment 1 + 1 = 2 is true is always, and not with 99.9999999% chance.
Thats my issue with ChatGPT answers - I can't trust it. ChatGPT shifts complexity of reading, analyzing, verifying correctness of answer to human.
I tried ChatGPT with a few SO questions as input, it was completely missing the point of the questions more often than not. I got 3 bad answers and one correct one out of it. The scary part was just how confidently incorrect it was. The text looked very good, but there were big errors in there. I also tried to ask it a scientific question on a pretty niche field, and it produced a very reasonably looking answer. Most…
Scientific questions are, without being disparaging, mostly trivia. ChatGPT knows all of Wikipedia. If an answer could be found by extracting a passage from a Wikipedia article, ChatGPT will do a good job of answering it. Ask ChatGPT to do any engineering work, and it falls flat. It's poor mathematical skills and lack of ability to create complex plans mean that it is relying purely on it's coherent-sounding dialog.
I can do the complex plans myself, but automation is quite helpful when looking up, for example, what is the idiomatic way to do datetime manipulation in some language which I haven't used in years.
I tried to create openai account, but seems e-mail is not enough for them and they require also phone number, not sure why, so I gave up, I provide my phone number only where it's essential.
It is to reduce abuse of their API because these requests are very expensive to serve as it requires 100s of gigabytes of VRAM to serve a request, requests take a few seconds at a time, and they are handling a ton of free requests. If you were paying for their API you could be paying up to about $0.10 per text generation.
Earlier quoted context omitted.
An intended double-loop might be interesting: Train one GPT on the output of another.
OpenAI isn't giving out this service from the goodness of their hearts just for the fun of it. I bet they've gotten it to a place where they can feed its output back into itself for self improvement.
I agree OpenAI have something going on here, I'd assumed it was that they wanted to know the strategies we came up with for bypassing their filters, and to get into a cat and mouse game with us so that they could improve them to the point of being productized. Also, marketing, both sales, recruitment, and mindshare.
ETA: I think I underestimated before the extent to which the output is often just your own words reflected back to you, in which case, it may well be adding more signal then noise.
https://news.ycombinator.com/item?id=33863413
Still, some prompts return a walk of text from a single sentence, and I have to imagine those are worse then useless for training.
I think the ban on ChatGPT answers on Stack Overflow is a mistake. Sure, some of the answers might not be correct, but that's true for any tool. And the fact that ChatGPT answers are easy to produce means that more people can contribute, which can only be a good thing for the community. Plus, it's not like the moderators have to accept every answer - they can still weed out the bad ones. Let's not punish everyone bec…
After the abuse: intelligent experts using their time to check the vaticinations of a delirious toy.
I think the ban on ChatGPT answers on Stack Overflow is a mistake. Sure, some of the answers might not be correct, but that's true for any tool. And the fact that ChatGPT answers are easy to produce means that more people can contribute, which can only be a good thing for the community. Plus, it's not like the moderators have to accept every answer - they can still weed out the bad ones. Let's not punish everyone bec…
Going to be blunt... I'm kinda sick of this already This pattern... It was cute for like the first couple of times... Clever even... But now the novelty is wearing thin and all I see are the fallacies inherent in trying to put this forward as some sort of argument/comment in favour of a viewpoint... because I'm still giving the commenter the benefit of the doubt and assuming they are trying to make a substantive cont…
Just finished Neil Stephenson's Anethem. In it (I'm paraphrasing here) the open internet is destroyed by media companies who flood the internet with "good garbage", that is stuff that is _almost_ correct. It forced people to pay for curated accurate content (that is, pay the same people who generated the good garbage in the first place).
I tried to create openai account, but seems e-mail is not enough for them and they require also phone number, not sure why, so I gave up, I provide my phone number only where it's essential.
They send an SMS code to sign up with.