Fork me on GitHub
#community-development
<
2022-12-07
>
respatialized15:12:26

I want to raise the question of ChatGPT in this Slack. Most of the people posting responses from it are mentioning fun stuff in #off-topic and are clearly labeled (or at least ChatGPT is mentioned), which is in line with ChatGPT’s attribution policy. I think #off-topic is a great place for stuff like that. Where it gets thornier is when people ask real questions. I find myself in agreement with https://meta.stackoverflow.com/questions/421831/temporary-policy-chatgpt-is-banned to avoid people spamming unverified answers to questions; I think if people actually want what ChatGPT suggests is a solution they can use ChatGPT rather than relying on someone copying and pasting its responses here. So I think it’s a good idea to put a policy in place discouraging ChatGPT answers in “on topic” channels. I am making my point of view clear here, but I’m not trying to be prescriptive; I’d rather this be the starting point for a discussion about ChatGPT (and perhaps also Copilot and other ML tools)-generated content on this slack. So I’d like to hear other points of view as well.

eggsyntax15:12:13

Can you give some examples where that's happening here?

respatialized15:12:13

I haven’t seen it yet. But I’m trying to be proactive about the possibility - I think it makes sense to discuss before it becomes widespread.

p-himik15:12:12

I wouldn't be proactive with prohibitions. And I sincerely doubt that it will be widespread here. This platform has no shiny points and karma or whatever it is SO uses - there's almost 0 reason for people to be disingenuous about their answers.

☝️ 3
eggsyntax15:12:15

My first instinct as a moderator is to wait and see if there's a problem before deciding how to address it. As @U2FRKM4TW says, Clojurians seems like a less likely place for that to be an issue. Not an Official Statement or anything, just my immediate intuition 🙂

respatialized15:12:56

I have seen at least one post where the reverse is happening: someone asking questions about something ChatGPT suggested they do. Not sure whether that’s different from an ordinary question.

p-himik15:12:24

I'd say it's definitely different. And I wouldn't discourage such questions.

respatialized15:12:30

And I do agree that this Slack is smaller and less susceptible to the factors that motivated the StackOverflow ban.

pez15:12:18

I'm fighting the urge to post ChatGPT's answer to this post. 😃

😂 5
quoll16:12:06

This will change in the future, but for now ChatGPT is a terrible Clojure programmer. It takes a lot of work to coach it into writing correct code, and it does not seem to understand instructions like, “The number of opening parentheses must be the same as the number of closing parentheses”

quoll16:12:36

So I think it’s a while before anyone would attempt to provide an answer that comes from such a system!

seancorfield16:12:57

As a moderator, I would certainly hate to see channels here peppered with AI-generated responses -- it's bad enough that everyone's posting enthusiastic messages about their interactions with it (at least I can filter a lot of it out on Mastodon!). If experience reports about the AI-bots stay in #C03RZGPG3 I'll grumble and put up with it, but if it starts to creep into other channels here, I'm very likely to push for a ban on it.

❤️ 1
eggsyntax16:12:41

I think also that chatgpt is this week's shiny new toy, and so it's unsurprising that there's lots of interest in it, but I suspect a lot of that will slow down over the next week or two.

pez16:12:45

I have a hard time imagining who would start post ChatGPT answers on stuff that we usually talk about here. We'll see what ensues, I guess.

1
quoll16:12:32

I posted in #C03RZGPG3 an interaction I had where I gave it an unmodified 4clojure question. I was impressed at how close it got, and I was frustrated at how hard it was to shift it to a correct answer

👍 1
1
Rupert (All Street)16:12:00

@UFTRLDZEW a very worthwhile discussion to have (for many online communities), although as @U077BEWNQ suggests it might be too soon (as the hype may settle down soon). Having said that - I do agree with @U04V70XH6 that I wouldn't like to Clojurians Slack filled with AI responses.

eggsyntax16:12:07

I think lots of people are very reasonably interested in its capabilities -- unlike Sean I'm not at all grumbly about people posting about it in #C03RZGPG3. It's legitimately interesting! Let a thousand strange #C03RZGPG3 flowers bloom 😁

seancorfield16:12:47

(I will note that I am equally grumbly about AoC stuff too -- which is why I have that all filtered out on Mastodon and why I "strongly encourage" anyone who mentions AoC here to join #C0GLTDB2T and keep it all in that channel where the rest of us don't have to see it 😁 )

quoll16:12:51

If someone mischaracterized an AI response as coming from them, then that would be a huge problem. I would advocate a 2 strikes policy on such things. (i.e. 1 warning, then they banned).

Rupert (All Street)17:12:11

One problem is that their is no objective way to prove that a response is AI - and so a banning policy may not be implemented fairly.

quoll17:12:53

There will be cases where it’s clear, but many when it’s not

👍 1
eggsyntax17:12:09

> One problem is that their is no objective way to prove that a response is AI Quantum computing theorist Scott Aaronson is doing some interesting work currently on watermarking AI output -- see the 'My Projects at OpenAI' section of https://scottaaronson.blog/?p=6823.

👀 1
eggsyntax17:12:36

(which I have mixed feelings about, since it would eg enable OpenAI to attempt to assert ownership over all output from GPT-n)

Rupert (All Street)19:12:53

It's possible that self hosted, open source AI similar to GPT will be good enough in future - and these are unlikely to have watermarking (or automatic submission to plagiarism database).

eggsyntax19:12:00

I would hope so! There's already a good open source text-to-image AI out there (https://stablediffusionweb.com/) and it'd be surprising if we didn't get a good LLM as well. Maybe the closed ones will always have the very latest state of the art but those innovations will hopefully spread to the open source ones fairly quickly. Which obviously raises lots of different social issues...but I'd personally rather err on the side of putting tools in the hands of everyone than keeping them in the hands of a few organizations (with some possible AI-safety caveats about AGI and clearly viable paths to AGI).

👍 1
mal18:12:48

I, for one, welcome our ChatGPT overlords. Attribution is going to be hard in a system where the answers can change continually and maybe aren’t even “stored.”

Rupert (All Street)23:12:49

Given that ChatGPT can be made to say practically anything and is well known for hallucinating, I don't think attributions to ChatGPT would be worth much. Better would be to attribute the original sources of information (but be aware that ChaptGPT is quite good at hallucinating sources too).

respatialized23:12:41

They are absolutely worth it as a line of demarcation between something created by a human author and something created by a probabilistic speech model. It's also part of the terms of service for OpenAI accounts, I believe. But I agree partially; works consisting mostly of edited ChatGPT responses are likely to be of low value as compared with the original corpus. As I suggested above, if I care what responses ChatGPT offers for a given prompt, I will just use ChatGPT directly.