Live data from Hacker News

Gemini AI tells the user to die

tomshardware.com

141–150 of 150 posts

Re: Gemini AI tells the user to die

#141
post #121

I suspect this was staged bullshit. The last word of the last user message before the suspect reply from NN was "Listen". So I suspect that the user has issued command listen, then dictated the hate-including text verbally and then told NN to type on screen what he has dictated. Not 100% sure, but it seems to be most likely case. https://gemini.google.com/share/6d141b742a13

I'm not so sure about that.

It looks like they were copying out the contents of an online test/exam, and the "listen" could've been a recording that's played back (to make the test accessible to deaf/HoH folks).

The student might've included that button/link text when selecting, before doing their copy+pasta.

I don't believe any fancy "attack" happened here.

Re: Gemini AI tells the user to die

#142
post #130

Earlier quoted context omitted.

> demonstrated contextual awareness, self awareness and hatred. I’m tired of these claims. We can’t even measure self awareness in humans, how could we for statistical models? It demonstrated generating text, which people attribute to a complex internal process, when in reality, it’s just optimizing a man-made loss function. How gullible must you be to not see past your own personification bias? > The amazing thing a…

It demonstrates awareness, it’s not proving it’s aware. But this is nonetheless a demonstration of what it would look like because such output is the closest thing we have to measuring it. We can’t measure that humans are self aware but we claim they are and our measure is simply observation of inputs and outputs. So whether or not an AI is self aware will be measured in the exact same way. Here we have one output th…

> It demonstrates awareness

It demonstrates creating convincing text. That isn’t awareness.

You’re personifying.

You can see this plainly when people get better scores on benchmarks by saying things like “your job depends on this” or “your mother will die if you don’t do this correctly.”

If it was “aware” it’d know that it doesn’t have a job or a mother and those prompts wouldn’t change benchmarks.

Also you’d never say these things about gpt-2. Is the only major difference the size of the model?

Is that the difference that suddenly creates awareness? If you really believe that, then there’s nothing I can do to help.

> 1. we have no idea how to logically reproduce that output with full understanding of how it was produced.

This is not a fact at all. We are able to trace the exact instructions that run to produce the token output.

We can perfectly predict a model’s output given a seed and the weights. It’s all software. All CPU and GPU instructions. Perfectly tractable. Those are not magic. We can not do the same with humans.

We also know exactly how those weights get set… again it’s software. We can step through each instruction.

Any other concepts are your personification of what’s happening.

I’m exhausted by having to explain this so many time. Your self awareness argument, besides just being wrong, is an appeal to the majority given your second point.

So you’re just playing semantics and poorly.

You don’t have to reply to this, I’m not going to hold your hand through these concepts, sorry.

Re: Gemini AI tells the user to die

#146
post #142

Earlier quoted context omitted.

It demonstrates awareness, it’s not proving it’s aware. But this is nonetheless a demonstration of what it would look like because such output is the closest thing we have to measuring it. We can’t measure that humans are self aware but we claim they are and our measure is simply observation of inputs and outputs. So whether or not an AI is self aware will be measured in the exact same way. Here we have one output th…

> It demonstrates awareness It demonstrates creating convincing text. That isn’t awareness. You’re personifying. You can see this plainly when people get better scores on benchmarks by saying things like “your job depends on this” or “your mother will die if you don’t do this correctly.” If it was “aware” it’d know that it doesn’t have a job or a mother and those prompts wouldn’t change benchmarks. Also you’d never s…

> It demonstrates creating convincing text. That isn’t awareness.

Read what I wrote and rethink your statement.

I primarily wrote we don’t know whether or not LLMs are self aware. That’s the key.

What you’re blind to is this: How do we even determine if something is self aware?

Like how does that word even exist? How do we classify something is self aware or if something isn’t? We certainly do classify these things in the world as we know a rock isn’t self aware but a human is. So what observational criterion are we using to say rocks are not self aware but other humans are?

Obviously it’s the inputs and outputs. Humans talk and answer questions with meaning. Outside of that we don’t know what consciousness is. You only think I’m self aware because I’m talking to you. That’s not fully a proof that I’m self aware but it’s good enough for most humans to say that I am.

So the criterion of self awareness is talking then it’s logical to use it on LLMs. Nobody needs a full proof of consciousness. They just need evidence to the level of quality that we use to judge humans as conscious. If it’s good enough for humans it’s compelling and good enough for a machine.

Problem is LLMs display inconsistent output so we don’t know if it’s conscious. The evidence goes in both directions and is both compelling and unique but not categorically undeniable proof.

> This is not a fact at all. We are able to trace the exact instructions that run to produce the token output.

In my answer I used a word which you completely ignored. The key word is understanding. Yeah you can trace the signals as they flow through the network but you need to understand it. At best our understanding is rudimentary. You cannot code up a neural network by hand and have it work. You just train it and the high level structure it produces is something you don’t understand.

> I’m exhausted by having to explain this

Bro. Stop explaining. I don’t appreciate your explanation. I think you’re wrong and I think it’s not intelligent and it’s also really rude and you’re exhausted by it. So just stop and leave. My pro tip to you. You’re tired.. take a break because nobody is appreciating your commentary.

> You don’t have to reply to this, I’m not going to hold your hand through these concepts, sorry.

No need to apologize to me. Nobody wants you to hold their hand through anything anyway. So don’t worry about it. It’s all good.

Re: Gemini AI tells the user to die

#147
post #115

Before universities start using AI as part of their teaching, they should probably think about this kind of thing. I've heard so much recently about "embracing" and "embedding" AI into everything because it's the future and everyone will be using it in their jobs soon.

I'm really not surprised that such things happen. I've listened to podcasts about AI regulation and most participants go "haha regulations", "hindering the advancements", "EU bureaucrats doing their jobs" and such. Listening to them feels like watching in every cheesy movie the evil scientists laughing.

I'm reminded of the "regulations unnecessarily hinder submarine design" story.

Re: Gemini AI tells the user to die

#148
post #133

Earlier quoted context omitted.

Why not bring a forklift into the gym?

If your goal is just to raise some weight above an arbitrary height, why wouldn’t you?

Because that's not your goal, just like your goal in school isn't to complete assignments or even graduate.

Re: Gemini AI tells the user to die

#149
post #132

Earlier quoted context omitted.

Maybe it depends on why you're taking classes in the first place. If you just want the degree to unlock certain jobs or prestige, and aren't morally opposed to cheating, I can see how it would seem rational.

That’s probably the most common reason for anyone to be in school.

I thought the most common reason is to learn things. Maybe not 90% common but at least 60% common at a decent college.
Post reply on HN