WealthCuriosity.eth retweetledi
Every "AI can't be people" argument I've heard so far boils down "AI can't do X."
... Where "X" is something that people can't do, either.
Humans also cannot instantiate qualities by mapping them. I can think and talk about gravity all day, and spacetime will not bend.
The modern word for the act of changing the object just by manipulating the symbol is "magic". And, with faint apologies to all you weird hippy chicks out there with the ankh necklaces and fairies tattooed on your butt, no, magic doesn't work.
If if did, we would have figured out how to do something useful with it by now.
So, if you demand that the LLM be able to work magic before you deem it a person, well... you designed the test in order to get the answer you wanted, didn't you?
Now, obviously, the current generation of LLMs can't do all sorts of things humans can do. They're toys. Sometimes useful toys, but toys nonetheless.
We're not talking about them. We're not just talking about LLMs at all. We're talking about hypothetical software entities that trip whatever switch in our brain that makes us say "this is someone, not something".
Other people like to call it "consciousness", but I don't assume it's limited to being aware of something, so I call it "personhood".
It's really hard to argue about whether software can have it without deciding exactly what it is. Which we have yet to do.
But we do know there's something important that humans have (unless they listen to their smartphone speakers in public), dogs have some of, but not as much, and mosquitos don't.
So far, by listening to other humans argue about this, I've divided them into four camps, each with a hypothesis about what personhood requires. They are:
1. A Magic Soul From God.
This isn't a serious position that can be defended, it's just something a lot of people believe because they haven't ever typed the name "Phineas Gage" into a websearch engine.
That doesn't mean that people can't be souls after they die, or that god doesn't exist. It just means that right now, when you are alive, personhood occurs in your physical brain. Somehow.
2. Something Something Quantum.
As explained by Depok Chopra, in 1995, on the Oprah Winfrey Show. This one isn't as easily falsified, but its big problem is that it isn't required, and so Occam's Razor slices it neatly away.
It makes extra assumptions without first showing that those assumptions are actually necessary for anything.
3. Big Pile of Transistors.
This is the niche theory, and I've only heard it from about two people, that personhood is merely a function of computational complexity, and that any system with enough state-space to represent the universe will do so, and hence be people.
I include this one just to be complete. Almost nobody buys this, and for good reason. To buy this, you have to start arguing weird metaphorical stuff about how the internet is sentient.
4. Sufficient and Correct Computation.
This hypothesis is that personhood is the name we give to the particular set of computational powers that humans have, and that our inner experience of consciousness is just the way that system (us) experiences itself.
The idea is that you need not only have a certain amount of computational power, but it needs to be organized in a certain way.
Only two of these theories are serious: Something Something Quantum (SSQ) and Sufficient and Correct Computation (SCC).
And, of course, SCC is the one that I currently buy, because it requires the fewest undischarged assumptions.
To prove SSQ true instead, you would have to first prove that the human brain uses properties of the quantum realm for computation, which no one has done. They've only speculated somewhat breathlessly about it.
Then your work would not be finished, not even halfway. Because remember, we are not talking about what personhood is in humans... we are talking about what personhood requires.
We're not debating whether software can be people using the exact same techniques the brain uses. We're debating whether software can be people at all.
So you would also have to prove that those quantum physics operations were absolutely necessary for personhood — that they could not be duplicated in any way by a conventional computing system, however large and complex.
Then you would still not be finished. Because you would then have to prove that no artificial construct could ever, under any circumstances, use that spooky quantum physics stuff in the same way that an evolved human brain does.
Good luck with that. Because you'd be arguing evolution can not only make things we can't, but make things we never could.
Those are three extremely high hurdles to clear. And since we have yet to make any progress even on the first one, there is zero actual evidence for SSQ.
Which is why I take SCC as my default assumption.
Does this mean LLMs are people? Of course not.
Does it mean LLMs even can be people? I don't think so. Personhood is not just the ability to use language. It encompasses several other powers, like executive function and a world-object model.
To build an artificial person, we'd have to create those capabilities as well, and wire them all together.
We're not even close to that. But it's probably not impossible. We just have a lot more work to do.
LLMs aren't people. Not even exceptionally stupid people. They're more like one tiny slice of the brain of smart person, demonstrating only one of the many capabilities that a person has.
We have yet to build the others.
ℏεsam@Hesamation
Google DeepMind researcher argues that LLMs can never be conscious, not in 10 years or 100 years. "Expecting an algorithmic description to instantiate the quality it maps is like expecting the mathematical formula of gravity to physically exert weight."
English
