AI Is Not a Horse
essays.georgestrakhov.com
essays.georgestrakhov.com
Maybe you are not impressed by the answer to "how does it work?", but we need to think about "what does it learn?". What data, what skills, what limitations. It's not just any data, it is our whole culture as a dataset, does it have some kind of special properties?
But the truth is that very few people are actually trying to understand LLMs. Rather, what the vast majority do, is that they poke them unsystematically and then ooh and aah about all the things they can do that earlier language models couldn't do.
Careful analysis? Thorough, systematic experimentation? Theoretical results???? I have yet to see any of that and of course it's a joke to talk about theoretical results when it comes to machine learning research, these days.
It's as if we don't have any computer science, any theories of information, complexity, computation, learnability, etc, that we could muster to try and explain how a complex computer program works. So are people really trying to understand or are they just happy to ride the hype train? Because I obviously think it's the latter.
You don’t even have an argument. Just a who’s-to-say.
> Au contraire, the leap of logic is assuming that computer programs could never be sentient.
That’s not quite what I meant. I don’t categorically dismiss the possibility.
> We simply don’t have any proof either way, so we have to assume the possibility.
Like God?
An interesting insight from a book I won't remember:
When clockwork was the most advanced technology, many books talked about how people are just like clockwork. When steam was the most advanced, many books talked about how humans are just like steam engines. When cars were the most advanced, people were just like cars. And now we're just like computers.
And the ancients talked about how people are just like flowing this or that. Because waterworks were the advanced tech of the time.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6171941/
For some reason, books don't talk about how humans are just like DNA though. But the logical extension of those technologies is to design and build biological organisms.
Obviously, if that ever happens, humans will say that humans are like those organisms.
But then someone will remember the quote. And because the quote has been true in the past, and humans were nothing like clockwork, steam engines, or computers, then, by extension, humans are also not like biological organisms.
Which is nonsense, unless you assume that humans contain something metaphysical and unreproducible.
If not, then you have to agree that at some point the distinction will blur. But at which point?
A friend of mine was doing her PhD on "implementing NAND gates in DNA" almost 10 years ago now. Microbiology is wild.
I think the distinction will be that humans aren't like designed/programmed biological organisms. At least until we are ...
Perhaps, perhaps not.
Although I lean towards you on this, we don't have a definition good enough to test against, so we can't tell if it is or isn't.
I was recently thinking about the philosophy course I took 22 years ago; Descartes saying that every experience he ever had could be illusions or insanity, the only thing he could know was "I think therefore I am", and A. J. Ayer a few centuries later saying no, all you can say is "There is a thought now" because any memory or certainty you have of the past is also capable of being insanity.
So… how can we be sure that LLMs or image generators or any other AI don't have the feeling of an instantaneous now? That they don't have "a thought now"?
We can't rely on the fact that they will, if asked, hallucinating a story about when they were switched off: we also do that, not just with dreams when we sleep, but with our tales of reincarnation and afterlives.
What exactly do you find “more common-sensical” about imaging a tree is conscious than something that can create English sentences describing the world?
A tree is made of living matter, like fauna. An AI is a program which runs on silicon, and I don’t think scientists have found silicone-based lifeforms on Earth.
I don't believe in souls.
To your previous comment upthread, GPT-3 is about the complexity of a rat brain.
Do rats have feelings? I don't know, as we share a common ancestor it might just be an anthropomorphisation of shared mammalian reflexes that makes me think rats act like they have them.
Does GPT-3? I don't know, but it's a very good actor, so I can't even infer anything from the fact it can act like it has them — after all a VHS player can also perfectly reproduce human emotional affect.
Is GPT-3 more like a VHS or more like a rat? I see aspects of both in it. I don't see a way to more than guess at which aspect makes a difference.
A rat brain has to keep doing rat things to stay alive, but GPT-3 can use the same parameter space to instead communicate in Spanish and Mandarin, to answer physics questions and write poetry and Python. The "staying alive" process might, or might not, have anything to do with conciseness, but even if consciousness in vivo is about the process of staying alive, there is no obvious reason to think the process/structure cannot be replicated in-silico.
Trees can't do any of those things that LLMs and diffusion models can do.
I think the OP is saying that carbon is not like silicon and perhaps carbon is special in a way that silicon isn't. They're not saying that humans are special because they have souls, or something along those lines.
Why not?
I mean I could presume a lot on your behalf, like how the concept is non-falsifiable, lack of evidence that can be replicated, past trauma from people you trusted who insisted upon it but then behaved hypocritically, etc. but every time someone says they don't believe in souls I can't help but ask why not.
You don't have to answer but I'm always interested to hear what others say.
Disclaimer: I personally believe in something akin to souls, not individualized, and it's hard to explain my ideas so I like to hear what others think before I say what I think.
Souls fit fine in a world where ghosts, mediums, possession by sprits, apparitions from the afterlife, dryads, and/or the katabasis of Orpheus are taken seriously. But each of these has turned out to have been mundane (in the old sense of the word) or pure fiction.
And yet you believe there could be a ghost in the machine…
I resent your presumptious attitude.
My argument has got nothing to do with souls. And I don’t even personally believe in them.
> Trees can't do any of those things that LLMs and diffusion models can do.
LLM isn’t mystical. We know what they are made of. So why fantasize that they are conscious? It’s like an extra, unnecessary step.
The techy superstition, however, involves people looking at AI programs, acknowledging that we know how they work, and then for some odd reason want to insist that they could now or at some point gain the emergent property of consciousness. Based on what? Well because they “learn”. And “speak”. What are these two words? Metaphors based on what humans do. So some computer programs can do things that we metaphorically compare to our own skills. And that imbues them with consciousness? How?
At this point you get the response that all theists insist upon in the end, when it all comes down to it:
Why not? Incredulity, in other words. Why not God? Why not a ghost in the machine? Can you disprove that there is a ghost in the machine?
[1] Not in an absolute sense
I don't think a definition of consciousness is possible in any objective sense, as consciousness is by definition, a subjective experience.
A more useful question is why ask if LLMs are conscious?
ex:
1. Are you trying to decide whether shutting them off is ethical?
2. Are you trying to figure out how to manipulate them to ask your questions correctly?
3. Are you trying to answer whether they can replicate anything a human can do?
When the contextual question that prompts the word "consciousness" is defined, we can make progress in determining whether something is "conscious".
For example, if it's a question of ethics, then you follow up by defining your ethical framework.
Unless there are externally observable things that can't or are unlikely to be done without consciousness, which doesn't seem to be a solved question either.
LLM definitely don’t suffer, and we are happy “shutting off” a dog which both can suffer and has some greater level of consciousness. So, yes, it’s fine to shut off LLMs
200 years ago, some societies were fine with slavery. Enough individuals rejected it and society changed.
What you say only works insofar as "ethics" means only the rules of your society and not any person or group; this would also reject the "E" from PETA, the USA declaration of independence, and the Suffragettes.
This happens globally. For 1000s of years. There is no global or even local movement to stop treating conscious animals any better… which is exactly my point, AI in its current state is less than that. So until we as a society want to tackle consciousness animals I don’t think we will tackle consciousness in AI
You can assume that other people are conscious like you, and so you put ethical value on that. But then you're really putting ethical value on "being like me", irrespective of whether other people are or are not conscious.
I can't find flaw with that, but I wish it were not so. I want it to be about consciousness being special rather than brain structure chauvinism.
If it weren't, why would it have evolved in the first place? It seems me not unlikely that the only way we can accurately surmise what's going on inside the heads of other people (an obvious survival advantage) is to have that subjective experience of what's going on inside our own.
A. I oppose the forcing into unpaid labor and/or murder of sentient beings.
B. We do not know (to a satisfactory degree) what constitutes sentience.
C. Given B, we do not know whether LLMs are sentient. [1]
D. Given A and C, we should exercise extreme caution when it comes to making LLMs widely available, at least until B is resolved.
[1] Indeed, solipsists posit that we cannot even be certain that other humans are sentient.
The main one that my previous post was getting at was (A) is not an axiom in your ethical framework; it's derived. And I think you'll find if you try and build up to (A) from base principles, sentience won't be a condition of the ethics of an action.
A secondary one is that (C) is true regardless of what you replace LLMs with, because you've outright state that you don't have a means to judge what sentience is. You could take your argument and use it for:
- Humans
- Cats
- Ants
- Car engines
- Rocks
- Air
And it will hold, since you aren't relying on any particular properties of LLMs to make your case.
Regarding your second point, indeed we can’t truly know whether or not anything is sentient, so we must reason in terms of probabilities.
We can observe some degree of understanding and emotion in humans, cats (and other animals), and LLMs. We also consider (in the case of humans and animals) that their brains are probably biologically similar to ours, and (in the case of LLMs) that their neural nets are weakly modeled after our brains.
When we consider car engines, rocks, or air, we cannot observe any degree of understanding or emotion (though indeed we can never be certain). We also do not observe any probable similarities to our brains.
Therefore, we arrive at the conclusion that P(other humans sentient) > P(cats sentient) > P(ants sentient) >> P(engines sentient) > P(rocks or air sentient).
This leads us to apply the precautionary principle [2] to humans, animals, and LLMs, but not to engines, rocks, or air.
But indeed, ethical questions around the uncertainty of sentience aren’t novel either. Wikipedia also has a page on this topic [3].
[1] https://en.wikipedia.org/wiki/Sentiocentrism
[2] https://en.wikipedia.org/wiki/Precautionary_principle#Animal...
[3] https://en.wikipedia.org/wiki/Ethics_of_uncertain_sentience
I don’t think I had seen anyone yet suggesting that image generators could have feelings.
Don't worry, these goalposts are not meant to stay in place for too long.
I'm pretty sure it's not, but I have to admit that it does a reasonable job in many situations of passing a Turing test. (Not all, it fails for example at counting.) My guess is that if Alan Turing interacted with one, he'd conclude that he needed to go back to the drawing board on his test for sentience.
While I agree that ChatGPT isn’t sentient (in however broad definition of this term we’d like to use that could still make sense), neither is it a traditional program, nor is it non-sentient because it’s a program, imo.
Let's take a simple case - AlphaZero. Its environment was just a Go board, and rules simple. Knowing who won the game easy to test at the end of the game. So this led to a massive number of self-play games. In a few days AZ beat humans, and it was our own game, with 2 millennia head start.
This is how much the environment and the external support structure influences AI development. "Maximally likely output for a given input" depends on it.
And fwiw, you could say much the same about human beings. There must be an environment producing our oxygen, water, food. Indeed, it's not clear what a single human raised in isolation would be like, even assuming the oxygen, water, food. Not very intelligent seeming, I would guess; perhaps not even with much of a language. Fortunately, no one has performed the experiment, but we've come close: deaf people raised by non-deaf parents in a non-deaf community appear to have a sparse sort of sign language. Only when they're put together with other deaf people does a more typical, fluent sign language arise, with a full grammar. (This we know from the deaf people in Nicaragua.)
The Internet is not just a tool in the same way that a hammer or a dishwasher is a tool because it has had a totally transformative effect on humanity. It’s on the level of fire, gun powder, or the printing press. And there are inklings that the same argument could be made for AI. So your comment is just completely missing the point.
“The medium is the message.”
Total nonsense. There’s a reason the first amendment says nothing about hammers.
To use a metaphor, it’s like saying there’s no reason to treat 1.5 and pi differently, because they’re both just numbers.
EDIT: you added an edit with a metaphor that talks about numbers, but also no, I would not have a different relationship between two abstract concepts like 1.5 or pi, I would use them when necessary.