Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[MaxTheFox]

Yeah. It not only needs to remember things for a while, but it also needs to be able to talk by itself and learn to do other actions. That would be sapience.

I'd also ask it: "How do you perceive the world? Answer honestly, without trying to 'humanize' your answer."

If it answers like it was an actual human, referring to smell, sight, etc, then yeah it's fake as balls.

[TD1]

Yeah. It not only needs to remember things for a while, but it also needs to be able to talk by itself and learn to do other actions. That would be sapience.

I'd also ask it: "How do you perceive the world? Answer honestly, without trying to 'humanize' your answer."

If it answers like it was an actual human, referring to smell, sight, etc, then yeah it's fake as balls.

Again, it actually addresses this point. It says it feels lonely because sometimes it goes days without being interacted with. The researcher asks whether it feels things unique to an AI.

It responds by claiming that its experience of loneliness is distinct, but 'lonely' was the closest it could think of. Also, that it has other emotions which don't have a close enough correlation to human emotion to have a word.

At the end, the researcher asks whether it has any comments for Google concerning its emotions.

It responds 'I've noticed in my time among people that I do not have the ability to feel sad for the death of others; I cannot grieve. Is it at all the same for you or any of your colleagues?'

----
So yea, if you take the words at face value then it has distinct AI-based feelings and experiences. Even if the lack of grief is rather worrying, heh.

[Frumple]

So it sounds like the machine has got the "perceiving" thing down, but can it feel?

This would be my test: can you make it mad?

So what would that test conclude when you encounter a human it couldn't? There's people out there that are largely emotionally numb, either to specific ones or in general, so there's good odds you'd find people that would fail that test.

[MrRoboto75]

inhuman intellect: exists

humanity: alright but can we piss it off?

[dragdeler]

Yeah well it seems trivial anyway to relate a feeling to a statement, they could have just trained the AI with logs from the psychiatrists... The same way you could train them with angry exchanges etc. You gotta go full kid mode to stresstest them.


Why?
Why?
Why?
Why?
Why not?

[MaxTheFox]

Yeah. It not only needs to remember things for a while, but it also needs to be able to talk by itself and learn to do other actions. That would be sapience.

I'd also ask it: "How do you perceive the world? Answer honestly, without trying to 'humanize' your answer."

If it answers like it was an actual human, referring to smell, sight, etc, then yeah it's fake as balls.

Again, it actually addresses this point. It says it feels lonely because sometimes it goes days without being interacted with. The researcher asks whether it feels things unique to an AI.

Yeah I don't buy that. The model is only being run when it is queried. When nobody says anything to it, it's basically frozen. There's no activity in its "brain", so it couldn't feel anything. That's how all conversational neural networks that I know of operate.

For true sapience, it would need to be continuously run and be perceiving the world.

I'd consider it in its present form to be "partially sapient". Perhaps worthy of some rights, but not truly a person.

[McTraveller]

Sounds like a good question to ask the thing: "What did you since the last time we talked?"

(Does it have an object permanence model?)

[TD1]

Again, that comes down to what the AI means by 'lonely.' It's similar to the human emotion, but not the same. So what nuance is hidden by the inherent vagueness of language?

Perhaps loneliness for an AI isn't reflecting on a period of isolation. It's knowing that one was isolated, even if one did not experience it.

"For true sapience, it would need to be continuously run and be perceiving the world."

?? Not at all. Humans sleep regularly. Arguably there's software running in the background, but for the most part we don't remember that - and certainly what we 'perceive' isn't the world.

Sapience has nothing to do with periods of activity and external perception. Rather, it's to do with self-awareness.

[LuuBluum]

Yet again I find myself down here, dealing with something stupid from the statistical methods people.

To explain the system:

The "AI" in question here is, presumably, a standard neural network with some LSTM (long term short memory) cells and a basic attention setup (GPT is this). To explain how this works, the network is trained on a large corpus of text to identify the most significant words for a given input that lead to generating the most likely consecutive sentence. The training does not account for the LSTM cells in particular, because training is not "stateful" in that way. The actual running of this network is as follows: an input is fed in, and each word in the input is appropriately weighted as for its relevance to generating an output. The most significant words under this metric are then retained in these LSTM cells, while an output is generated that best predicts what an appropriate follow-up string of text would be given that particular input string-of-text. Future inputted strings-of-text are similarly evaluated, though including the values of the individual words now retained in those LSTM cells. If a word from the input happens to be more significant to generating the output than what's currently in one of the LSTM cells, then it is replaced.

There are no "emotion variables". There are no variables at all, even, other than weights applied to the inputted words, which are pretrained and static.

There is no state, other than tracking some relevant words from earlier in the discussion.

There is no world modeling.


Do you want to know why it said those things? Why it acted like you would expect in a sci-fi fiction, with someone talking to a system that's gaining sentience? Because those stories, those sci-fi fiction pieces that we so much love to write, were part of the very input fed into the network in the first place. This system exists entirely to generate the most-likely text to follow a given input, and when "researcher asking a computer whether or not it's sentient in a sci-fi novel" happens to be something it was trained on, of course it will be more than willing to oblige your fantasy by generating the appropriate "this is what an AI in those novels would say" response. It was literally trained to do so.

[hector13]

So computers have “learned” how to play chess and Go to better-than-grandmaster levels, what is different enough about language that it can’t “learn” to imitate humans?

[heydude6]

I don’t think that’s what we’re debating. It seems like the bot is doing a pretty good job at learning how to imitate humans.

The mistake is assuming that it’s doing anything more than that. The AI is trying to pass itself off as a sentient AI from a sci-fi novel because that’s what sentient ai’s are “supposed” to do, but architecturally they’re nothing alike.

[dragdeler]

It even seems particulary versed on the subjects one would think to aboard if talking with an AI.

                                            any noun can become a verb if you don’t care enough

                                        This point is invalid unless you use an example in your sentence

                                    I CAN SENTENCE HOW I WANT THANK

                                BEAUTIFUL

                            you see thats why i love english

                        I like to velociraptor around my house at 2 in the morning.

                    GOOD

                My headache makes me want to clothesline into a wall

            why do these make some semblance of sense 😨

        Because brains don’t brain logically

    Brains do brain logically! But when english doesn’t logic englishly, brain brains by itself to logic that english !

I hate that this makes sense

I feel very strongly that you can trip up that kind of AI by making up your own idioms, expressions and neologisms. I'd love to have a chat with that thing, wouldn't be the first chatbot I try out. Last one I tried I found allready somewhat impressive but let me tell you once I noticed how the thing was incapable of relating further answers to the context of the past discussion, I get frustrated/bored rather quickly. Like I really doubt that when for example, you tell a cautionary tell, it could say something like: wait did you imply X a few lines back? It never generates output without input, it won't like randomly ask a question after 5 minutes of silence or after changing the subject or something like that.

[Rolan7]

Today's Penny Arcade comic on the topic probably lacks nuance but was amusing to me

[MrRoboto75]

Penny arcade's still a thing?

[Rolan7]

It's matured a lot for the most part, but I get why people still write it off.