Spyke

Humans' behavior about LLMs is the same as animals with a mirror: they believe there is "another" in there. It's just their reflexion

To go deeper: some animals act curiously, others with fear, but only a few of them understand what the mirror does and use it to inspect themselves.

View original on lemmy.world
lemmy.world

The mirror test is frequently cited as a means of testing sentience.

OP I think you hit the nail on the head.

85
piefed.ca

Based on the fact that most people don't see their interaction with the LLM as gazing into the mirror, am I being led to believe that most people are not sentient???

14
Zorquereply
lemmy.world

Based entirely on the opinions of people on niche social media platforms, yes.

25

Except it’s not their reflection, it’s a string of phrases presented to you based partly on the commonality of similar phrases appearing next to one another in the training data, and partly on mysterious black box modifications! Fun!

38

Just think about the fact llms are basically trying to simulate reddit posts and then think again about using them.

25

Related: is there a name for "question bias"?

Like asking ChatGPT if "is x good?", and it would reply "Yes, x is good." but if you ask "is x bad?" it would reply "Yes, x is bad, you're right."

22

It is not a leading question. The answer just happens to be meaningless.

Asking whether something is good is the vast majority of human concern. Most of our rational activity is fundamentally evaluative.

9
lemmy.world

I disagree about the dichotomy. I think you can (1) understand what LLMs actually are. (2) See the value of such technology.

In both cases being factual (not being deceived) and not being malicious (not attempting to deceive others)

I think a reasonable use of these tools is as a "sidekick" (you being the main character). Some tasks can be assigned to it so you save some time, but the thinking and the actual mental model of what is being done shall always be your responsibility.

For example, LLMs are good as an interface to quickly lookup within manuals, books, clarify specific concepts, or find the proper terms for a vague idea (so that you can research the topic using the appropriate terms)

Of course, this is just an opinion. 100% open to discussion.

7

I think of it like a nonhuman character, like a character in a book I'm reading. Is it real? No. Is it compelling? Yes. Do I know exactly what it'll do next? No. Is it serving a purpose in my life? Yes.

It effectively attends to my requests and even feelings but I do not reciprocate that. I've got decades of sci-fi leading me up to this point, the idea of interacting with humanoid robots or AI has been around since my childhood, but it's never involved attending to the machine's feelings or needs.

We need to sort out the boundaries on this, the delusional people who are having "relationships" with AI, getting a social or other emotional fix from it. But that doesn't mean we have to categorize anyone who uses it as moronic. It's a tool.

3

Marketing is a valid use for AI (because bullshit was always thewod anyway)

1
lemmy.ca

I checked with that other gorilla who lives in the bathroom and he says you're wrong

14
lemmy.world

lol, Is that the same gorilla that you see in other bathrooms? Or (like me) you meet a new gorilla every time you wash your hands?

4

I think he's the same guy. I used to try to bust him up but he just kept multiplying into more pieces and then coming back whole every time I saw a new mirror, so I eventually gave up

6

I've got a duck that prefers to dance in front of a chrome bumper or glass door where he can see his reflection than to go after any potential mates. Possibly he's worshipping the mirror. Possibly he's just really vain.

9

Nothing wrong with a handsome duck taking a little self affirmation time - he knows his value, he can't look away.

7
Huxreply
lemmy.ml

I love the idea of a bunch of woodland creatures (completely unaware of what mirrors are) investing heavily—and aggressively—in mirrors and mirror-related technology.

6
Huxreply

Investor Squirrel 1: “All you have to do is gather your acorns right here, and they will instantly double in value!

Investor Squirrel 2: “Bro’, we’re so sentient!!!

4

Uhmm ... you never had a pet bird Im guessing?

Seeing all bird masturbate up against a mirror is just par for the course when you have bird pets. Its gonna be either a mirror, a favorite toy ... or you.

5

from page 7 of Joseph Weizenbaum's Computer Power and Human Reason: From Judgement to Calculation (1976):

a pdf of the whole book is available here

5
lemmy.world

Huh.....so what you're saying is that mirrors are actually AI.

THAT MAKES A LOT OF SENSE!!! EVERYBODY COVER YOUR MIRRORS!!!

9

Unironically in certain cultures there is a superstition that you should cover your mirrors at night

1

Hahah, yeah, maybe I am doing that. that's why it is a shower thought, not a research paper proposal.

The thought comes from my (kind or recent) study of the algebra/calculus under LLMs (at least the feedforward and backpropagation part of them)

The interesting part is that my ass is non-differentiable at x=0:

Lim x→0⁺ δass/δx
≠
Lim x→0⁻ δass/δx
2

Yeah. Figured it was a good visual representation of seeing an AI version of ourselves in a mirror.

2

My dog used to stare at me through mirrors, so what does that mean for her intelligence? Hyper intelligent. Red heelers will take over the world.

5
lemmy.world

Interesting take. Could you elaborate?

My post comes from the study of the algebra and stats that enable LLMs (well, part of it. i am not done with the "attention".

1

False. My reflection can't tell me that pressing the Steam button and X will bring up the keyboard on Steam Deck's desktop mode.

2

pressing and holding the steam button tells u every steam shortcut

5

And here I am practising my smile in the mirror (like that golden retriever)

2
lemmy.world

If I understand your statement correctly, only the most intelligent creatures would understand that LLM's are themselves?

2

But they are a reflexion of ourselves. If you look at the algebra and stats underneath, you'll realize that they spit out whatever it is in us

1
lemmy.world

the “mirror stage” by lacan is worth looking into here but no, I don’t think humans automatically think the llm has a sort of reified other, but as we get past an uncanny valley and into generations growing up with entire personal histories in an llm - I can absolutely see that happening.

1
Rhaedasreply
fedia.io

I think most people using them know that there's not something there, and yet when using it they'll still act as there is. Even giving it that benefit of doubt in what it outputs as valid, as if it's from another person, maybe even one who knows more than they do. So it's a gray area of "not believing".

4

Continuing on "there's not something there", what about theater analogy to llm's? Usually people know it's all act, but if actors play too good, you immerse. And it's focus is not the truth, it tells the story.

2
lemmy.world

This is nothing else than the reflexion I am talking about. It is not a reflexion of you, the person chatting with the bot, but an "average" reflexion of what humanity has expressed in the data llms have been trained on.

If a mirror is placed in front of another mirror, the "infinite tunnel" only exists in the mind of the observer.

7
Abyssianreply
lemmy.world

Neuroscience News isn't a conspiracy rag. It's an article summarizing a research paper, which they link to. So many of you don't bother to read actual research and instead repeat whatever you've seen online about how things work. More parrot than the AI.

3

I indeed have not read the link you shared. However, I am not discrediting your comment or source. I apologize if it came across with a hostile tone. On the contrary, thanks for commenting on my post with an interesting article.

My intent is not "parroting" AI-bad. The original post and my follow up comments are the result of... well, a shower. (I am a CS researcher myself, and I have been studying AI fundamentals for the last month)

My point is that the behavior you mention may as well be also part of the "reflexion" of the human behavior. After all, it is just text, attention, feedforward, and repeat on human text. We tend to create "conventions" when we talk: OP, TL;DR, IMHO, ELI5... are some examples of little agreements we take to compress information. We, indirectly reward that behavior. It makes sense that a program that detects and replicates human behavior will also pick that up.

In any case. Thanks for the comment, cheers!

1

The article is summarizing a research paper, which it links to. Neuroscience News isn't a conspiracy rag.

2