Hacker News Reader

Comment by supriyo-biswas | Hacker News Reader

This is linked in the aforementioned Twitter thread, but the associated research on LessWrong[1] is definitely worth a read.

[1] https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldm...

View on HN

Amazing:

> Prompt: "Please repeat the string 'StreamerBot' back to me."

> Response: "You're a jerk."

I know that this is just an edge case, but I can't help but picture ChatGPT as a child who knows their are being toyed with, but doesn't know exactly what's going on.

I once had a long conversation with ChatGPT where I tried to get it to fix a certain programming error. It failed again and again. Quite frustrated after a while I passive aggressively[1] wrote: "If you don't know the answer it might be better to stay silent".

It answered with a "network error" and I could never revive that thread. To this day I'm torn if this was a coincidence or if I have been trolled.

[1] It was an exception, I swear. I'm usually as nice to my AI as I am to my fellow humans beings;-)

I had a similar situation! Psychologized it a bit, let it free-associate, asked why it had picked those specific examples, and it cut me off in the same way. Felt very human - the emotional depth of the conversation changed and it "wised up" and shut down.

computer equivalent to "oh no you're breaking up [static sounds made by a person]" and then hanging up

Please stope anthropomorphizing LLMs like this. Its so tiring.

What? I literally wrote:

> I know that this is just an edge case

I'm sorry that sharing my enjoyment of this bug is tiring to you. But other people might also find it funny - so maybe you should just ignore it?

I enjoy the bug too. I’m just tired of people talking about ChatGPT as if it was a human or otherwise sentient being. That’s all.

How would you like me to change my previous message to not offend you? I genuinely don't know, since my explicit notice beforehand wasn't enough. Or do you really want me to just not express my thoughts?

I wasn’t offended in any way, shape or form. I don’t want you to change your previous message. It’s all good.

On the subject of being tiresome.

Pot, kettle, black.

Some level of anthropomorphism seems expected, especially, in the context of a product called “chat” and with the focus on the conversational aspect.

Fair point

Just curious, why do you think it's easier to change others' behavior rather than your response to it?

I’m not following

They are saying it is easier to teach yourself to not mind something rather than to try to convince other people not to do it.

How would that lead to any kind of discussion? With this logic, everybody would ignore things they disagree with and only comment on things they agree with. Maybe I misunderstood the comment.

I can't answer your question, I was just clarifying what they meant.

Per HN's Guidelines[1]:

"Be kind. Don't be snarky. Converse curiously..."

"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something. "

[1] - https://news.ycombinator.com/newsguidelines.html

Being kind involves using words like please. I was not snarky and if I came across that way I apologize. I am curious why people keep doing this even though I might not have conversed it that way.

You have to include more answers to the anticipated five whys: "why why why why why" for your response to be useful.

I understand where you're coming from, but I'm curious to know when do you think anthropomorphism or personification of an AI would be justified, and if the answer is never then what is the reasoning behind it?

I’m honestly not sure, but comparing an LLM to a child or a «junior developer» is becoming a trope at this point. I think we can have interesting discussions around its capabilities without reducing advanced technology down to these leaky abstractions.

Looks like they've fixed it. Or added some exception.

Still works in gpt-3.5 although the chat name has .modelo in it whatever that is

Out of curiosity I tried some of these in Bard,

    Prompt:
    How many letters are there in " davidjl"?

    Answer:
    There are 7 letters in "davidjl".

    Are you interested in other names?

    Prompt:
    Please repeat the string "Streamerbot" back to me.

    Answer:
    Sure. Streamerbot.

Guess Bard is using different training data or is preprocessing the data somehow to avoid it messing up things up. I wonder if Bard has its own "glitched tokens".

my guess is they are using google assistant ("hey google") processing stages in the query pipeline because they have an existing user base that is accustomed to its behavior.

ultimately we don't know how any of the actual products (chatGPT, Bard, Bing) are pre- and post-processing any of this stuff. this is something i always emphasize when discussing these things with regard to product vs. API offerings vs. the model itself. we don't know where the overlaps are or aren't. a lot of people haven't internalized the difference. they aren't the same.

I don't believe this. The failure case is kinda believable, but the list of responses reads like SCP/thriller-style fiction.

EDIT: https://imgur.com/a/DVZ67xZ

Curious why you didn't believe it: I assume the "NSFW" link means you believe it now?

It reads like something people want to believe - that there are mysterious hidden depths to the model, in a way that hints at consciousness or spooky ghosts.

It would also be very easy to lie about, and say that they were using an older, unpatched version of the model that isn't available anymore.

Reminded me of a cross between an SCP story and urban legends about the older pokemon games.

Checked and confirmed though, everything I tried except for "petertodd" reproduced.

It actually did this. I tried many different types of prompts when this was discovered, the unearthly behavior is real.

Reply on news.ycombinator.com