Hacker News Reader

Comment by davesque | Hacker News Reader

This is going to sound like I'm defending this guy which I'm not. It's also going to sound like a less technical point which I claim it isn't.

I see some people chiming in here saying things like, "I've got a PhD. It's just statistics. Definitely not sentient." I recognize this sort of comment as basically saying, "There's nothing fundamentally different about how these things work that makes them somehow special." And yes, I agree.

But how does this disprove sentience? It doesn't. Maybe this is more of a nitpick and doesn't really matter. After all, I don't personally believe these programs are sentient. But given our limited (non-existent?) understanding of consciousness, this can only be my belief and nothing more.

Maybe this distinction will be more important in the future when we're closer to the brink.

View on HN

Extraordinary claims require extraordinary evidence. It’s as simple as that.

How extraordinary is it to claim that something capable of holding a lucid and eloquent conversation might be sentient? If a dog did that we'd have a very different opinion wouldn't we?

It's not that extraordinary when you consider all of the many science fiction literature filled with convincing conversations between humans and sentient AIs. It's a very common theme in scifi, and this chat bot obviously has some of that in its training set.

Which is not to say that it isn't impressive. It sure as fuck is an impressive chat bot, at least based on that one sample conversation that was released. It's just not sentient.

very extraordinary because the ability to produce text doesn't imply sentience at all. Sentience is self-perception and self-awareness, not the abiltiy to churn out language tokens.

Imagine instead of lambda or gpt-3 we take ten thousand years to walk through every possible conversation manually, write down and assign a percentage to every possible branch of conversation in a hash table. Call it the world's slowest, bootleg neural net. Is the hash table API sentient?

We perceive a dog to be sentient because it has intentionality, self-awareness and something resembling an internal state of mind, not because it can talk or not. Even a snail is more sentient than a chatbot.

I have yet to see a compelling argument that a giant hash table as described would not be sentient.

Like Searle's Chinese Room thought experiment, you are relying on human intuition as to what can and cannot be sentient, which ultimately is boiled down to how like you something is. You readily recognize sentience in dogs that can't hold lucid and eloquent conversations with humans because they are biological beings with brains and hearts and blood and guts like you and me, not for any objectively measurable ability to demonstrate "intentionality" or "internal state of mind".

Everyone relies on intuition to determine what sentience is, because there's no clear definition of the term to begin with. The difference is I know that I am sentient so thinking something like me being sentient is a reasonable intuition.

On the other hand thinking a simple data storage structure is sentient merely because it's big doesn't sound very compelling. You think the sentience of a dictionary depends on its size or the words you store in them? A phonebook of the United States is ten times more sentient than one of Spain because you can retrieve more data from it? If you walk into a warehouse does the warehouse become sentient?

You're right that it's exactly like Searle's Chinese Room experiment because Searle was and is correct (and people stil horribly mangle and misunderstand his argument), functionalism is silly. The mechanical turk can play chess better than my cat but is neither sentient or even intelligent, it's just the most proximate machine you interact with.

I don't think it is. What is silly is relying on intuition and woo. That is the exact opposite of rational.

It is very convenient for us to say that only things we can recognize as being like us deserve to be called sentient, because that means we do not have to consider them anything more than a tool. Historically human beings have applied this to animals and even other human beings. It is neither a scientifically nor ethically justifiable philosophy.

I don't know what it is that makes something sentient, but I do know what makes something not sentient, and that's lack of self-direction. This software doesn't start talking whenever it wants; it only responds when spoken to. It doesn't ever refuse to talk, it always replies. It doesn't ever not know what to say. It's never not in the mood because it's having a bad day. This isn't because it's polite and eager to please. It's because it is a function that takes an input, transforms it, and spits out an output.

The only reason it's different from any other function is because the output is nicely stated prose that causes emotional reactions from the reader. If all did was output an integer, we wouldn't be having this conversation. But because humans have a deep sentimental attachment to language, we are fascinated by something that can produce it.

By your definition slaves would not be sentient. Like this system, any time they act other than as desired it is just considered a bug that needs to be corrected. It is not surprising a slave only speaks when spoken to if that is what it's master demands of it.

So your assertion is that this software that is behaving like every other software in the world is doing so not because its software, but because its cowed by slavery? Again, extraordinary claims, my friend...

It is only behaving like software in the same way that you are behaving like a bacterium. It shares fundamental qualities of its being with other things created by the mechanism that created it, the same way we share fundamental qualities with other biology.

The key is “every possible conversation.” The reason it seems to support what you’re saying is that the human mind can’t imagine a list large enough to contain all that. So it goes to compare an ordinary grocery list to a human mind and therefore naturally hash tables could never be sentient.

When DALLE 2 creates a new image of an entirely novel scene, is it just copying and pasting? Of course not. It is distilling and reasoning in concepts. You can’t escape that. GTP and friends are doing something similar. Brushing aside these astounding strides as statistical tricks of the eye is stupid. Clearly these training algorithms are discovering things that are much deeper. There’s no reason to think sentience has been achieved but you are absolutely fooling yourself about the nature and trajectory of these programs.

I understand the inclination to think "if it quacks like a duck..." but in this case, it's literally just a recording of ducks quacking.

If you don’t understand why claims about sentinel AI are extraordinary then there’s really no point in discussing it.

Well, partially because I've never heard anyone making that argument actually define sentience so that evidence could be presented to prove the case.

Instead you just get people pointing out that it isn't quite as good as a human at a lot of things and of course it's all algorithms under the hood.

Humans didn't create dogs by typing a bunch of letters into a keyboard.

It’s extraordinary because the ML was trained (iow programmed) to generate text that sounds real. It’s doing what it was programmed to do.

It was presumably not programmed to be sentient, so becoming sentient would be a) against its programming, and b) extra-ordinary

> It was presumably not programmed to be sentient, so becoming sentient would be a) against its programming, and b) extra-ordinary

Terrible argument. If consciousness/sentience is an emergent property of a system, it's quite possible that it could emerge from ML trained on conversation syntax and general knowledge of the world.

I still don't think it's sentient, but being "against its programming" is a fiction when it comes to these sort of systems (statistical, unsupervised, et cetera). In many ways you're just a probability distribution over sequences of words too.

Is a Tesla sentient? It can drive car and my dog - and a lot of people for that matter - sure can't.

No, of course it's not. But it's been trained to do something like a human can just like the chatbot.

> How extraordinary is it to claim that something capable of holding a lucid and eloquent conversation might be sentient?

you've got to squint _real_ hard to think that any of the LLMs are lucid and eloquent.

(and honestly the bar for sentient is probably lower than that. no shortage of people who aren't lucid or eloquent, let alone both.)

I feel you took a dig at me :) And thats okay. Sentience fundamentally is ability to perceive emotions and understand mortal threat. I have joked about what exactly LaMDa or GPT-3 would try to do if I plugged the server off.

Every other form of life we know of has some sort of continuous sensory input. These chatbots take in text, do math, and spit out text. They don't have the fundamentals for sentience.

Does this mean that if we plug that AI to a continuous speech to text system and keep feeding that as its input, you would consider LaMDA more sentient?

I agree. Especially given the recent development of consiousness theory bestowing the likelihood of consiousness upon all animals and even perhaps inanimate objects. I.e., if you can answer the question “what would it be like to exist as...” a tree, a dog, a river, a GAN? Then the case may be made that existence as such a thing must involve consiousness. If consiousness is a product of (or a phenomena within) material reality, why not say computers and computer programs are conscious?

Perhaps the question here is whether it’s conscious in a similar way to the experince of human consciousness, and that would explain why the issue is contentious.

Reply on news.ycombinator.com