Item — HN Reader

Agreed, oftentimes the truly zealous AI pundits act as if our modern day LLMs are completely, 100% equivalent to humans, which I find utterly insane as a concep. For example any discussion about copyrighted works, which is a hot topic, will inevitably end up with someone equating an LLM "learning" to a human learning, as if the two are identical.

I think human languages and psyches just aren't built to cope with the concept of AI. Many words have loose meanings, like "learning" in the case of AIs, that can easily be twisted to mean one of dozens of definitions, depending on the stance of the person talking about it. It'll be interesting as the technology becomes more prevalent and mundane how people start treating it all. I'm hoping we get to realizing that a computer isn't a human regardless of the eloquence of its "speech" or whatever words we use to describe what it does, but I guess we'll see

6 comments

> oftentimes the truly zealous AI pundits act as if our modern day LLMs are completely, 100% equivalent to humans, which I find utterly insane as a concept

Never heard anyone say this and I know (and know of) a lot of doomers. Honestly, this entire line of discussion would be far less frustrating if it weren't for the endless strawmanning and name-calling.

I'm sure most would put me pretty far onto the doomer side. I can confirm that I'm not concerned with how LLMs compare to human equivalence today.

I'm concerned with whether they are artificial intelligence at all, what the bounding functions for AI's development are, how we could possibly solve the alignment problem, and whether we even understand human intelligence enough to recognize an artificial intelligence at all or predict how intelligence might work in something drastically more powerful than us.

> I'm hoping we get to realizing that a computer isn't a human regardless of the eloquence of its "speech" or whatever words we use to describe what it does, but I guess we'll see

My anecdotal experience with how people treat Alexa devices does not inspire confidence in me with regards to this. I can't even convince my wife not to gender it when referring to it.

To be fair, anthropomorphizing is kind of a built in feature for us — or more accurately, ascribing intentionality to things as a means of explanation. Magnets “want” to stick together, my vintage computer “thinks” it’s 1993, and the furious gods are hurling burning rocks from the skies because my neighbor ate leavened bread on the wrong day.

It’s not limited to AI or LLMs - theory of mind is a powerful explanatory tool that helps us navigate a complex world, and we misapply it all the time.

I don't think this is really down to people understanding what a computer is, but more down to how humans interact with nonhumans. We anthropomorphize animals and objects all the time. A computer program is ultimately a (complicated) object, that we often give all sorts of human trappings, like a human voice that expresses things in human languages.

If you can take pity on the final, dented avocado at the shops because it looks "sad", you will for sure end up calling Alexa 'she'. Avocados can't be sad, but they can look sad /to humans/, and a machine can't really have a gender or be polite in the human sense, but it can definitely sound like a polite lady.

I think humans will just fundamentally relate to anything they perceive socially as another human, even if we know full well they aren't human. Probably it's a lot less work for a human brain, than it is to try to engage with the true essence of being an avocado or an Alexa.

I agree, but I'm not sure why you don't think that is at least relevant if not partially responsible for people seeing humanity in a language model.

If we anthropomorphic animals and even vegetables regularly, what happens when something appears to talk back to us intelligently? I don't think these mechanisms are nearly as distinct as you make them out to be.

Perhaps I misinterpreted your original post, but I do think there's a difference between not intellectually 'getting' that a computer isn't thinking in the same way a human is thinking, and this anthropomorphizing that we do to all sorts of things. I'll try to be clearer:

Obviously making an object mimick human traits makes it a lot easier to anthropomorphize. Like if you put a little cowboy hat on a pear. But you wouldn't say that somebody doesn't "understand that fruits aren't human" if they assign human traits to them. Making a machine that talks back at us in a seemingly intelligent way is a much more intensely human trait for it to have. It's not down to our intellectual understanding of the thing, but how our brains decode our (social) world.

So the anthropomorphizing happens regardless of the level of technical understanding. We both know it's just a pear, but it's also a cowboy now.

Of course there's a level of playfulness involved with all of this too. It's kinda fun to backsass your GPS instructions, or tell Alexa she's being nosy, or ask your dog what he thinks of the presidential debate. But that's not the same as foolishly overestimating the political acumen of the dog, or the nature of Alexa's intelligence.

I think it's irrelevant what people "know", if that's not how they act. It doesn't matter that people know a fruit isn't human if they act like it is in ways, because how people interact with the world is the only thing that actually matters. If you always act like the bear with a hat has some level of feelings even if you don't believe it, that's functionally identical to actually believing it, so any difference is irrelevant. No matter how much they express they know it has no feelings, if they feel bad or feel bad for it if it's destroyed, what's more true, what they said or how they actually responded?

Most people don't act like things have feeling all the time, or to the same level as a living being might have feelings, but many people do it to a degree. Given that much of how we feel is immediate (and thus not reasoned), subconscious, and sometimes Pavlovian, I don't think it's a stretch to think that even acting like an inanimate object has feelings or human traits regularly might lead towards subconscious feelings about it that are not entirely rational or immediately understood by the person.

For a slightly different argument, consider why it's generally considered good advice not to name the animals meant for slaughter. Or how people treat cars or large pieces of equipment that aren't always reliable (are "temperamental"), or that are tightly linked to safety and livelihood, such as boats.

We are social animals. We bond to others easily by our nature, because it's beneficial for survival. I think we do it so easily that it extends towards animals, often to good effect, but sometimes even to inanimate objects that have a lot of significance. In the past this seems to even have been extended to weapons and armor, given the number of named items in history.

> For example any discussion about copyrighted works, which is a hot topic, will inevitably end up with someone equating an LLM "learning" to a human learning, as if the two are identical.

The argument is not that they are identical. LLMs and diffusion models learning is a new thing. We are all trying to come to terms what that means, and how we should regulate it. (And if at all we should regulate it) Do note, we are talking about here what the law should be, not what the law is.

And in doing so we compare this new thing to already existing things. "It is a bit similar to this in this regard" or "it is unlike that thing in this other regard".

I don't think it is controversial to say that if an LLM outputs byte-to-byte the text of a copyrighted work the work remains under copyright. The person who run the LLM does not magically gain rights by that. If you coax your model to output the text of the Harry Potter books you don't suddenly become able to publish it.

The question is what happens if the new work contains elements from copyrighted works. For example if it borrows the "magical school" setting from HP but mixes it with Nordic mythology and makes it as deadly as Game of Thrones. What then? Can they publish this new thing? Do they need to pay fees to J K Rowling, and George R. R. Martin?

It is generally permissible to publish that new work if it was created by a human. If it is sufficiently different from the other works you are free to write and publish it. Does this suddenly change just because the text was output by an LLM?

The argument is not that "human learning" and "machine learning" is the same. It is that they are similar enough that one has to argue why you think one can create new work and why the other can't.

Human learning and machine learning aren't even close to similar and pretending like they are is very stupid.

That is not a very well formed argument, is it? You basically state your opinion and then call people who have an oposite opinion names.

Let’s try to elevate the conversation: there is a newly published author MrX. His book combines themes from many copyrighted works in a novel way and is a critical and commercial success. There is no allegation of any copyright infringement around the work. Suddenly information comes out that MrX used LLMs heavily in the creation of the book and the LLM was trained on copyrighted works.

Do you think MrX owes licencing fees to those other authors? And which ones? All the ones in the training set? Just the ones which has thematic/stylistic similarity to his work?

How come those thematic/stylistic similarities only matter if MrX used an LLM and don’t matter if he used his own head?

Yes MrX owes licensing fees, full stop.

Die mad about it

Feel free to expound on the important differences

Well for one, human learning actually produces an understanding of the thing.

Understanding meaning what?

> For example any discussion about copyrighted works, which is a hot topic, will inevitably end up with someone equating an LLM "learning" to a human learning, as if the two are identical.

I don’t think the point is that the two are exactly identical or that humans and LLMs are equivalent, but that the processes are similar enough in the general level that any attempt to regulate LLM training in copyrighted material will inevitably have the same ramifications for human learning.

In pretty much all attempts I’ve seen to differentiate the two, it inevitably boils down to hand-waving about how human beings are “special”.

We've reached the inevitable part of the conversation! What distinction are you drawing between what an LLM does and what a human does? Because as far as I can see they are identical.

A human artist looks at a lot of different sources, builds up a black-box statistical model of how to create from that and can reproduce other styles on demand based on a few samples. Generative AI follows the same process. What distinction do you want to draw to say that they should be treated differently legally? And why would that even be desirable?

I'm pretty sure the artist is conscious and the AI isn't which means there's something reductive about your claim that they are both merely "applying a black-box statistical" model.

Even if it's true (which is debatable,) it doesn't appear to be more informative than saying "they are both made of atoms."

Well, my personal opinion is with the rise of neural nets we've basically proven that "consciousness" is an illusion and there is nothing there to find. But for the sake of argument, lets assume that there is something called consciousness, artists have it and neural nets don't.

How are you going to demonstrate that consciousness is responsible for what the artists are doing? We have undeniable proof that the art could be created by a statistical model, there is solid evidence that the brain creates art by simulating a mathematical neural network to achieve creative outcomes - the brain is full of relatively simple neurons linking together in a way that is logically similar to the way we're encoding information into these matrices.

So it is quite reasonable to believe that the artists are conscious but suspect that consciousness isn't involved in the process of creating a copyrighted work. How does that get dealt with?

I'll talk about fiction, since I've written some. If I write a ghost story it's because I enjoy ghost stories and want to take a crack at my own. While I don't know why ideas pop into my head, I do know that I pick the ones that are subjectively fun or to my taste. And if I do a clever job or a bad job I have a sense of it when I reread what I wrote.

These AI's aren't doing anything like that. They have no preference or intent. Their choices change depending on setting like temperature or prompts.

Or let's try a different example. Stephen King wrote a novel where he imagined the protagonist gets killed and eaten by the villain's pet pig (Misery). He struggled to come up with a different ending because he said nobody wants to read a whole novel just to see the main character die in the end. He thought about it and did a different ending.

Are you claiming Stephen King's conscious deliberation wasn't part of his writing process? I'd say it clearly was.

Also, I don't really understand the consciousness is an illusion argument. If none of us are conscious, why should I justify any copyright policy preference to you? That would be like justifying a copyright policy preference to a doorknob. But somehow I'm also a doorknob in this analogy???

Suppose Bob says he's conscious and Jim says he isn't and we believe them. Doesn't that suggest we would have different policy preferences on how they are treated? It would appear murdering Jim wouldn't be particularly harmful but murdering Bob would. I don't have to show how Jim and Bob's mind differ to prefer policies that benefit Bob over Jim.

> ...[w]hile I don't know why ideas pop into my head...

If you're trying to argue that you're doing something different from statistical sampling, not knowing how you're doing it isn't a very strong place to argue from. What you're experiencing is probably what it feels like for a sack of meat to take a statistical sample from an internal black-box model. Biologically that seems to be what is happening.

I have also done a fair amount of writing. The writing comes from a completely different part of my mind than the part that experiences the world. I see no reason to believe it is linked to consciousness, even allowing that consciousness does exist which is questionable in itself.

It is an unreasonable position to say that you don't know the process but it must be different from a known process that you also don't have experience using.

> Are you claiming Stephen King's conscious deliberation wasn't part of his writing process? I'd say it clearly was.

Unless you're claiming to have a psychic connection to Stephen King's consciousness, this is a remarkably weak claim. You have no idea how he was writing. Maybe he's even a philosophical zombie. Thanks to the rise of LLMs we know that philosophical zombies can write well.

And "clearly" is not so - I could spit out a lost Stephen King work in this comment that, depending on how good ChatGPT is these days, would be passable. It isn't obvious that it is the work of a conscious mind. It in fact would obviously be from a statistical model.

> If none of us are conscious, why should I justify any copyright policy preference to you?

I've been against copyright for more than a decade now. You tell me why it is justified even if consciousness is a factor. The edifice of copyright is culturally and economically destructive and also has been artistically devastating (there haven't been anywhere near as many great works in the last 50 years as there should have been in a culturally dynamic society).

I'm referring to how Stephen King discusses his writing process in On Writing. I doubt you actually believe Stephen King might be a p zombie and I'm skeptical you really think consciousness is an illusion. I think if i chained you to a bed and sawed off your leg (like what happens to the protagonist in Misery) you would insist you were a conscious actor who would prefer to not suffer. I don't even know what consciousness is an illusion is supposed to mean.

If I sawed off your leg would it have the moral consideration of removing the leg of a barbie doll's leg if you feel your consciousness is an illusion?

When I write a story my brain might be doing something you could refer to as a black box calculation if you squint a little, but how is it "statistics?" When I feel the desire to urinate, or post comments on hacker news, or admire a rainbow, or sleep am I also "doing statistics?"

You seem to be referring to what people traditionally call "thinking" or "cognition" and rebranding it as "statistics" in search of some rhetorical point.

My point is human beings have things called "personalities" and "preferences" that inform their decision makings, including what to write. In what sense is that "statistics"?

The idea that the human subconscious is not consciously accessible is not a new idea. Freud had a few things to say about that. I don't think it tells us much about AI. I do think my subconscious ideas are informed by my consious preferences. If I hate puns I'm not going to imagine story jdeas involving puns, for example.

Most authors would prefer copyright exists because they'd prefer book publishers, bookstore retailers and the like pay them royalties instead of selling the books they made without paying them. It's pretty simple conceptually, at least with traditional books.

Copyright existed far longer than the last 50 years so how is our 50 years of culture relevant? The U.S. has had copyright since 1790.

have you ever used or spoken to GPT2 or GPT3? Not chatgpt. The one before they did any RLHF, to train it to respond a certain way. If you asked it whether it would like to be hurt, it would beg you not to. If you asked it to move it's leg out of the way, it would apologize, and claim it obliged. It would claim to be conscious, to be aware.

Of course, these statement do not come from a place of considered action: everybody knows the machine is not conscious in the same way as a human, but the point is that an unfeeling machine even being able to make such claims makes us have to move the "consciousness" divider further and further back into the shadows, until it's just some nebulous vibe people insist must be somewhere. It's possible there is a clear and unambiguous way to define humans and intelligent animals conscious, but nobody has come up with a workable definition yet that lets us neatly divide them.

Another slight thing that gives me a little pause: you know how great our brain is at confabulating right? Have you ever done something, then had someone ask you why you did it? You generally tell them a story about how you thought about doing something, weighed the pros and cons etc, that isn't actually true when you think deep down. We like to think we are a single being that thinks carefully about everything it does. instead we're more like an explaining machine sitting on top of a big pile of confusing processes and making up stories why the processes do what they do. How exactly this last thought relates to the discussion I haven't figured out yet, its just something that comes to mind :)

> I doubt you actually believe Stephen King might be a p zombie and I'm skeptical you really think consciousness is an illusion.

Consciousness is an unobservable, undefinable thing which with LLMs in the mix we can theorise has no impact on reality; since we can reproduce all the important parts with matrices and a few basic functions. You can doubt facts all you want, but that is a pretty ironclad position as far as logic, evidence and rationality goes. Consciousnesses is going the way of the dodo in terms of importance.

> If I sawed off your leg would it have the moral consideration of removing the leg of a barbie doll's leg if you feel your consciousness is an illusion?

For sake of argument, lets say conclusive proof arises that Stephan King is a philosophical zombie. Do you believe that suddenly you can murder him? No; that'd be stupid and immoral. Morality isn't predicated on consciousness. I'm perfectly happy to argue about morality but consciousness isn't a thing that makes sense outside of talking about someone being knocked unconscious as a descriptive state.

> When I feel the desire to urinate, or post comments on hacker news, or admire a rainbow, or sleep am I also "doing statistics?"

No, you're responding to stimulus. But right now it looks extremely likely that the creative process is driven by statistics as has been revealed by the latest and greatest in AI. Unless you can think of a different mechanism - I'm happy to be surprised by other ideas at the moment it is the only serious explanation I know of.

> You seem to be referring to what people traditionally call "thinking" or "cognition" and rebranding it as "statistics" in search of some rhetorical point.

I don't think I've said anything about thinking or cognition. Although statistics will crack those too, but I'm expecting them to be more stateful processes than the current generation of AI techniques.

> Copyright existed far longer than the last 50 years so how is our 50 years of culture relevant? The U.S. has had copyright since 1790.

Yeah but the law has been continuously strengthened since then and as it's scope increases the damage gets worse. The last 50 years are where new works are effectively not going to enter the public domain before everyone who was around when they were created is dead.

> Well, my personal opinion is with the rise of neural nets we've basically proven that "consciousness" is an illusion and there is nothing there to find.

I keep seeing this claim being made but I never understand what people mean by it. Do you mean that the colors we see, the sounds we hear, the tastes, smells, feels, emotions, dreams, inner dialog are all illusions? Isn't an illusion an experience? You're saying that experience itself is an illusion and there is nothing to experience.

I can't make sense of that. At any rate, I see no reason to suppose LLMs have experiences. They don't have bodies, so what would they be experiencing? When you say an LLM is identical to a person, I can't make good sense of that either. There's a thousand things people do that language models don't. Just the simple fact that I have to eat on a regular basis to survive is meaningful in a way that it can't be for a language model.

If an LLM generates text about preparing a certain meal because it's hungry, I know that's not true in a way it can be true of a human. So right away, there's reasons we say things that go beyond the statistical black box reasoning of an LLM. They don't have any bodies to attend to.

I agree, it almost seems like some type of coping mechanism for that fact that after all the ability to get computers to generate art and coherent sentences, we're still completely none the wiser about understanding objective reality and consciousness and even knowing how to really enjoy the gift of having the experience of consciousness. So instead people create these type of "cop outs".

Mechanical drawing machines have existed forever, I loved, love, loved them when I was a kid and I used to hang the generated images on my wall. Never did I once look at those machines who could draw some pretty freaking awesome abstract art and think to myself, "well that's it, the vale of consciousness is so thin now, it's all an illusion", or "the machine is conscious".

As impressive as some of these models are at generating art, they are still drawing machines. They display the same amount of consciousness as a mechanical drawing machine.

I saw someone on Twitter ask ChatGPT-4 to draw a normal image, you know what it drew ? A picture of a suburban neighborhood, why might a conscious drawing machine do that?

The "it's an illusion" part is a piece of rhetorically toxic language that usually comes up in these discussions, as its a bit provocative. But it's equally anthropocentric to say that someone without a body can't have a conscious experience. When you're dreaming your body is shut off - but you can still be conscious (lucid dreaming or not). You can even have conscious experiences without most of your brain - when you hit your little toe on something, you have a few seconds of terror and pain that surely doesn't require most of your brain areas to experience. In fact, you can argue you won't even need your brain. Is that not a conscious experience? (I'm not really trying to argue against you, I just find this boundary interesting between what you'd call conscious experience and not)

Reading through this entire exchange and it would seem your argument inevitably falls back on this argument of consciousness, which is both arbitrary (why does that matter in the original context of the distinction in learning, particularly from copyrighted works?) and ill-defined (what even is consciousness? How do we determine if another being is conscious or not?).

> A human artist looks at a lot of different sources, builds up a black-box statistical model of how to create from that and can reproduce other styles on demand based on a few samples. Generative AI follows the same process.

The reason programs are different in front of the law is that humans can program programs to do whatever they like at scale, humans can't program humans to do whatever they like at scale.

So for example if it became legal to use images from an AI, then we would program an AI that basically copies images and it would be legal, because it is an AI. But at that point we know the law has been violated because programs that just copies images aren't a legal way to get copyrighted images for free.

You saying "But the AI is a black box" just means that you could have hidden anything in there, it doesn't prove anything. Legally it is the same as if you wrote a program to copy images.

It's this kind of thinking that is the only reason machines would ever become the primary life form on earth.

One distinction would be that every source a human looked at involved a payment for access to/a copy of each and every source inputted into their black-box.

Which is not the case with the current AI models (or rather, the companies profiting off them), to my understanding.

You mean all the landscape artists and portrait and figure and still life painters throughout all human history who just casually made art of whatever they saw around them?

Did you write this reply because it was the most typical way that the thread up until GP could have continued?

I don't mean that to sound flippant -- if this wasn't your intent, then obviously you understand the distinction.

I also only downvote comments containing things like "I know I'm going to get downvoted for this..." (tiny handful of my downvotes violate that rule). I don't like to disappoint expectations.

In what way is this argument any more productive than showing up to Pluto with a plucked chicken and yelling "BEHOLD! A MAN!"?

We can make all kinds of navel-gazing arguments about why LLMs and humans are identical but at the end of the day you'll go home and poop and the LLM will blissfully idle on a forgotten AWS compute instance burning a hole into someone's pocket and any idiot can see that you two are not the same in any sense.

Just because onthologies are leaky and fuzzy when applied to real world that doesn't mean there aren't tangible differences between things that are clearly different. I can construct a semantic argument for why deleting an LLM is akin to murder (and use it to either reason that deleting an LLM should be punished with the death penalty or that it should be okay to kill random people who annoy me) but that won't impress the judge or jury unless I'm trying to make a case for criminal insanity.

Making a machine do things humans can do but faster/cheaper/more has tangibly different implications than letting only humans do it. It should be treated differently for the same reasons that just because a public water fountain is free to use that doesn't mean you can put a hose on it and use it to water your lawn.

I'd be doing a lot better than expected if my arguments have the sort of historical pull that Diogenes' had. And he made a point accurately - they had to change the definition and he showed their approach up as not being able to accurately identify a man without making hilarious mistakes along the way.

Who the heck cares if AGI will 100% mimick humans and human minds, that's purely academic discussion.

What is a serious concern are overall capabilities (let's say penetration of networks and installing hacks across whole internet and further), combined with say malevolency.

Its trivial to judge mankind from whole internet as something to be removed from this planet for greater good or maybe managed tightly in some cozy concentration camps, just look at the freakin' news. Let's stop kidding ourselves, we are often very deeply flawed and literally nobody alive is or ever was perfect, despite what religions try to say.

I am concerned about some capable form of AI/AGI exactly because it will grok humanity based on purely data available. And lack of any significant control or even understanding of what's actually happening to those models and how they evolve their knowledge/opinions.

Even if the risk is 1%, that's existential risk we are running towards blindly. And I honestly think its way higher than 1%. Even if I will be proven wrong some proper caution is a smart approach.

But you can't expect when people's net wealth tightly coupled with moving as fast as possible and ignoring those concerns to do the best decisions for future of mankind, that's a pipe dream in a same vein as proper communism is. If that would be the case very few bright people would work at facebook for example (and many many other companies including parts of my own).