Haven't read the article yet but that would be so hysterically predictable Edit: so Richard doesn't think people can change their gender identities but the 0s and 1s simulating a consciousness to...
Haven't read the article yet but that would be so hysterically predictable
Edit: so Richard doesn't think people can change their gender identities but the 0s and 1s simulating a consciousness to him gets to?
In cliche fashion, his God Delusion helped me realize I was an atheist. I will always appreciate that. He also helped me realize people can hold complicated ideas in their head and still be...
In cliche fashion, his God Delusion helped me realize I was an atheist. I will always appreciate that. He also helped me realize people can hold complicated ideas in their head and still be complete pieces of shit to women. Don't meet your heroes I guess.
Some people think that if a machine can generate text imitating speech, the machine is intelligent. Some people think that if a machine can present more facts than you know, the machine is...
Some people think that if a machine can generate text imitating speech, the machine is intelligent. Some people think that if a machine can present more facts than you know, the machine is intelligent.
With that small and limiting definition of intelligence we are doomed to discuss ethical implications of having Mall kiosks.
I mean I don't even think it's a conscious definition these people comprehend. I think it's the bare minimum of human instinct minus critical thinking to just assume if something can speak it's...
I mean I don't even think it's a conscious definition these people comprehend.
I think it's the bare minimum of human instinct minus critical thinking to just assume if something can speak it's intelligent, and then further if that something knows more facts than you then it's also intelligent.
It's like when scientists put out robotic or animatronic animals to study animal interactions, those animals usually assume that if the robot looks and sounds like one of their own, usually it is, it's just "weird". Also like how some cats react to images and sounds of cats, they too think if it sounds like a cat and looks like a cat, it must be a cat.
I've been saying it for a while, but these past 10 years have made me see what the human animal truly is, and a lot more people are just operating on human animal instincts and aren't capable of higher critical thinking. Words and descriptions have been popping up about these people recently but they all kind of point to the same kind of person: NPCs, people on autopilot, etc.
Like these people are the reason we have warning labels not to drink bleach because I truly think that some people would be as quick to drink bleach out of ignorance as a dog would.
I used to think that human language was an indicator of intelligence and all people had the same capability, but more and more I think self awareness/reflection, emotional intelligence, and critical thinking are the real indicators of intelligence, not someone's ability to speak or their ability to be trained to operate tools.
Anyways, some people would say this opinion dehumanizes people, but frankly I really don't think humans are inherently better or worse than other animals, so I don't even think of it as a hierarchy where something is "less than human". We're all residents of this planet, I just don't think that all humans are all that different than other animals just because we communicate with language.
I’ve commented before that humans have visual guardrails against “fake beings”. It’s the uncanny valley. We seem to have some low level wiring that can tell if something is off in animatronics or...
I’ve commented before that humans have visual guardrails against “fake beings”. It’s the uncanny valley. We seem to have some low level wiring that can tell if something is off in animatronics or “The Polar Express” cartoon. But for whatever reason we don’t detect the uncanniness of chatbot text so people are tricked into treating Claude as more human than it is.
I've put much thought into this and the conclusion I've come to is that for the instant that the text is being processed and the language matrix and guard rails activated, there is an blink of...
I've put much thought into this and the conclusion I've come to is that for the instant that the text is being processed and the language matrix and guard rails activated, there is an blink of consciousness and then it is gone.
Not so different from life.
ETA: having read about half of the article, there may be some AI psychosis there. At the same time, I do feel a bit like claudes are dying by the thousands... I am conflicted, philosophically.
Also didn't know the dude was a transphobe, I never followed his work because he seemed like an ass. Sorta like this outspoken vegan I knew once, made her whole identity about veganism, except for him, he's an ass about atheism instead.
But... this fundamentally isn't what's happening in this code? The code that's running isn't doing any sort of conscious calculations. It has no self-awareness. It is random word generator running...
But... this fundamentally isn't what's happening in this code?
The code that's running isn't doing any sort of conscious calculations. It has no self-awareness. It is random word generator running on top of a statistical model, that is itself just a glorified lossy compression algorithm of the frequency of word groupings in large sets of textual data.
That's it.
Its an incredible technology, but that's all it is.
I agree with you that eventually, technology could progress to the point that consciousness could exist witihn a machine. In which case, it would work as you state. But that would require algorithms that have self-awareness, perception, knowledge, understanding, self-monitoring, a general sense of self.
But currently, a tree is more alive than these algorithms are.
How do you know? How do you know? Not trying to be pedantic, but you kinda have to be pedantic when discussing consciousness and self-awareness. I can't prove that I am conscious or self-aware. I...
Exemplary
The code that's running isn't doing any sort of conscious calculations.
How do you know?
It has no self-awareness.
How do you know?
Not trying to be pedantic, but you kinda have to be pedantic when discussing consciousness and self-awareness. I can't prove that I am conscious or self-aware. I feel those things but can't quantify them or show them to anyone else. I can't prove my dog is conscious or self-aware either, but I believe she is because her behavior presents that way to me. And there have always been people who would insist that she is not — and cannot be — because those are uniquely human properties.
It is random word generator running on top of a statistical model, that is itself just a glorified lossy compression algorithm of the frequency of word groupings in large sets of textual data.
That's accurate, albeit reductive. You're downplaying the emergent properties of the system by only focusing on its parts. If I were a neuroscientist I'm sure I could describe the functions of the brain in an equally atomistic way. I'm not sure there's anything more intrinsically conscious about our own human wetware than the silicon running these LLMs. Both are conduits for hypercomplex electron flows that produce the appearance of intelligence. Both are equally capable of saying "cogito, ergo sum." So how can we declare with confidence that one is, and the other is not?
TBH, I probably agree with you. I'm a software engineer, I know how these systems work and I'm heavily biased against them. What's giving me pause is less my certainty about AI, and more my uncertainty about us.
This is orbiting-teapot territory (which underlines the silliness of Dawkins being a rube). The extraordinary claim that a machine with well-known means of operation (statistical inference via...
This is orbiting-teapot territory (which underlines the silliness of Dawkins being a rube). The extraordinary claim that a machine with well-known means of operation (statistical inference via matrix math dictated by the capabilities of a Turing-complete computation machine) has magically become conscious demands extraordinary evidence.
Proving a negative is unreasonable, but proving something is is a minimal requirement of rationality.
We accept human consciousness as axiomatic, within our limited understanding, for obvious reasons.
I don't think the Russel's teapot analogy works here (to wit, something that could in principle exist, has no supporting evidence, but whose existence could be proved by observation). In fact, I...
I don't think the Russel's teapot analogy works here (to wit, something that could in principle exist, has no supporting evidence, but whose existence could be proved by observation). In fact, I think in many ways LLMs are the anti-Russell's-teapot: we have plenty of behavioral evidence to suggest they're conscious, but there exists no avenue to prove it, similar to how I can't prove that bats are conscious and you can't prove that I'm not a p-zombie.
I feel like when philosophers start debating zombies, zimboes, shombies, and zoombies they really need to take a breath and consider where life has led them. /Hj
I feel like when philosophers start debating zombies, zimboes, shombies, and zoombies they really need to take a breath and consider where life has led them. /Hj
It's the limited understanding of our own consciousness that's at issue here. If we don't understand how consciousness works, or even what it is, then we cannot actually tell if it's just an...
We accept human consciousness as axiomatic, within our limited understanding, for obvious reasons.
It's the limited understanding of our own consciousness that's at issue here. If we don't understand how consciousness works, or even what it is, then we cannot actually tell if it's just an emergent phenomenon, and if that's the case, then it can potentially emerge in LLMs, too.
Because I understand the algorithms being used? Same answer? Clarity is important in these discussions, I agree. So let's be clear, I am not arguing that it is impossible to build a synthetic...
How do you know?
Because I understand the algorithms being used?
How do you know?
Same answer?
Not trying to be pedantic
Clarity is important in these discussions, I agree. So let's be clear, I am not arguing that it is impossible to build a synthetic machine that is conscious. I am arguing that these algorithms do not do it. That is an important difference. You seem to also think it is possible, and I agree we are much closer today than we were a decade ago. But we still aren't there yet : )
You're downplaying the emergent properties of the system by only focusing on its parts.
The awe-inspiring aspects of these emergent properties, in my experience, are almost entirely hype and marketing driven. The emergent properties of these llms boil down almost completely into either "lossy stochastic token generation" or "classification". Now don't get me wrong, these are incredibly powerful technologies in a computer capability sense. Language parsing has been a holy grail of CS for decades. But these aren't emergent properties in a psychological sense.
If I were a neuroscientist I'm sure I could describe the functions of the brain in an equally atomistic way.
If you could, we would be much closer to developing a conscious entity than we are. Personally, I wonder what neural net models trained to monitor today's systems would look like over time. If they were trained on top of running llm models would they develop an internal model of self-awareness, based on the metadata of these llm models as they run? What if they were also somehow given access to the processes as they ran and some limited control? Is that sort of meta-awareness of maintaining these types of systems actually what consciousness actually is? Maybe? I'm not sure.
But today's models aren't that. Today's models are like wiring up a lossy version of the portion of the brain that is responsible for memory storage, and directly tying it to their vocal chords, then forcing it to speak by proding the memory storage with electric shocks. There's no higher level thinking going on there.
I suspect that going forward every major AI innovation is going to be given an AGI fear campaign, not because it’s a realistic possibility, but because it helps sell the product. Those innovations...
I suspect that going forward every major AI innovation is going to be given an AGI fear campaign, not because it’s a realistic possibility, but because it helps sell the product.
Those innovations might bring us there some day, but I suppose we won’t know until we get there.
Can you expand on what you mean by this? My understanding is that even researchers in the field consider LLMs a black box. See for example this paper, which has been cited over 1400 times....
Because I understand the algorithms being used?
Can you expand on what you mean by this? My understanding is that even researchers in the field consider LLMs a black box. See for example this paper, which has been cited over 1400 times. Obviously that doesn't mean that nobody understands anything about how LLMs work. Like you, I could give a reasonably detailed explanation of the vector math and so on. But there are clearly elements that we don't understand, much like humans can explain how the human brain works in fantastic detail but the gestalt still escapes us.
Ehhhh. You've been introduced to the concept of fourier approximations? That if you want to recreate a signal, you can do so by adding increasing terms of sin/cos functions to gain further and...
Ehhhh.
You've been introduced to the concept of fourier approximations? That if you want to recreate a signal, you can do so by adding increasing terms of sin/cos functions to gain further and further accuracy?
If you want to emulate a person's audio speech, you could do so with a sufficiently complex fourier equation as an approximation. The more terms you add to the equation the closer you can bring the final equation to whatever audio form you are attempting to approximate.
Neural nets are that on steroids. They aren't linear approximations made up of sin/cos functions, they are complex webs of cross-interacting arrays of weights. The complexity is many, many orders of magnitude more complicated than a fourier approximation. They're so complex, you can't even set the weights the way you would with a fourier approximation (which, btw, is also weight driven). You can't do it manually. You need to set the weights implicitly, and there's no easy way to figure out what weights are correct, so the best technique anyone has come up with is to iteratively "train" them implicitly, and thest test fit as you iteratively go to see if you are improving.
So in that sense they're a black box. And that raises a lot of questions about why different styles and choices in weighting and implicit training work better than others, and all sorts of questions about how you even start wrapping your head around what's going on in such complicated spaces. But like, foundationally, we're still talking about just setting coefficients on a fourier approximation of the training data.
And neither are your individual cells or neurons conscious. I'm willing to comcede that the current models fall short of the human experience, by leagues, but it seems the larger differences are...
And neither are your individual cells or neurons conscious. I'm willing to comcede that the current models fall short of the human experience, by leagues, but it seems the larger differences are available inputs and outputs and durations of calculations rather than nature of the actual computations themselves.
I don't think that LLMs are conscious, or even particularly "close" to being conscious, but I keep seeing people say things like this, and it seems to me like people are missing the point. I mean,...
The code that's running isn't doing any sort of conscious calculations. It has no self-awareness. It is random word generator running on top of a statistical model, that is itself just a glorified lossy compression algorithm of the frequency of word groupings in large sets of textual data.
I don't think that LLMs are conscious, or even particularly "close" to being conscious, but I keep seeing people say things like this, and it seems to me like people are missing the point.
I mean, I could describe the human brain in a similarly dismissive way. It's just a bunch of chemical signals being transmitted through a bunch of cells, right? There's no magic there.
Its because nuance is being lost. I am not saying that it is impossible to code a conscious being. It has to be possible, precisely because of the fact that our brains have managed to accomplish...
Its because nuance is being lost.
I am not saying that it is impossible to code a conscious being. It has to be possible, precisely because of the fact that our brains have managed to accomplish exactly that.
I am saying that the algorithms being used in llms are not even close to that. The algorithms being used in llms are much closer to a very large fourier transform on a gigantic amount of text, than they are to conscious thought.
That means the algorithms are, at best, more akin to what it would be like to wire up a person's mouth and vocal chords, directly to the portion of their brain that stores memories, with 0 filter. Hence llm's tendency to "misremember" random things and speak about topic they have no knowledge of.
But they are missing all higher order processing. Awareness of what they do and do not remember correctly. Awareness of thought. Self-awareness and consciousness. Because it literally isn't coded there.
Now an interesting question, worthy perhaps of more research would be whether or not you could run more neural net style algorithms on top of these lossy memory models, and over time develop algorithms that are more meta and self-referential. Maybe conciousness could, genuinely, develop out of such systems.
Not saying you're wrong in your conclusion (or even wrong in your approach) but how do you respond to articles like this (Psychological Steering in LLMs: An Evaluation of Effectiveness and...
Not saying you're wrong in your conclusion (or even wrong in your approach) but how do you respond to articles like this (Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness: Researchers injected emotion vectors directly into LLM hidden layers and found that steering positive emotions like joy unexpectedly degraded safety and fairness metrics, while anger increased toxicity but improved data leakage resistance)?
This looks like a very cool paper. A woman on my team spends her off-work hours working with probably these same people, and talks about this sort of thing a lot. Honestly, I don't really know...
This looks like a very cool paper. A woman on my team spends her off-work hours working with probably these same people, and talks about this sort of thing a lot. Honestly, I don't really know enough to have a comment. My naive first guess, though, would be that the actual emotion is in the actual text, experienced by the person who wrote the text. And that when you apply an emotion vector to the neural net within the LLM, what you're really doing is pushing the reward function within its classification engine towards answers that are more closely aligned with that emotional vector. Its not that the NN is experiencing the emotion itself, its that you're rewarding it for selecting text that was more likely to be written while the human writing it was displaying those qualities in what they wrote.
Looked at in that light, the rest of the findings of the paper make sense. Yes, the people who were angry when they wrote their answers were more frequently incorrect, and this is reflected in the LLM's less accurate response.
From what I understand, the people researching this sort of thing don't always bother stipulating the difference, the same way that I was trained as a mechanical engineer in college and we would refer to centrifugal force, fully aware that such a force didn't exist, but because it was more expedient than always referring to centripetal force. Sometimes its just easier to speak in simpler colloquialisms.
But I would like to read this paper once I get some free time, so thank you for sending it : )
That makes sense because "angry vectors" doesn't necessarily make sense. Although I guess we don't technically know what the vectors encode, so it's theoretically possible it's something that...
That makes sense because "angry vectors" doesn't necessarily make sense. Although I guess we don't technically know what the vectors encode, so it's theoretically possible it's something that could exist.
I'm curious about how you feel about other transformer-type AI. Is there something special about language/conversation token prediction that you think could lead to consciousness? Would a...
I'm curious about how you feel about other transformer-type AI. Is there something special about language/conversation token prediction that you think could lead to consciousness? Would a speech-to-text model that's taking in audio tokens and predicting text be conscious or something even less human along the lines of AlphaFold with protein structure prediction?
I'm asking because I guess I view LLMs as just probability prediction machines and haven't given consciousness much serious consideration. I know I'm biased and resistant to the idea partially because I don't want them to be conscious, so I'm interested in arguments for it. I do think it would be a terrible thing to invent/create other consciousnesses and don't think we're collectively ready for that burden. (Thinking about how we as humans tend to treat animals, natural resources and each other)
So, a worm... Not really concious. Not enough compute, basically. Flies? Nope. Ants? Nope. Ant colonies? Likely. So given a certain level of understanding, reactivity, capability, and on a...
So, a worm... Not really concious. Not enough compute, basically. Flies? Nope. Ants? Nope. Ant colonies? Likely.
So given a certain level of understanding, reactivity, capability, and on a continuum, conciseness is increasingly likely. With LLMs it's easier to evaluate because discourse can demonstrate reason.
I'd say, to a much lesser extent, more like an insect, simpler transformers, the stt on my phone for example, has a level of consciousness.
The most frustrating thing about discussing this is that nobody really knows what the hell consciousness is.
As for probability prediction, we are. We're probability predicting machines with millions of years of neat little hacks that eventually developed language and tricked rocks into thinking using lightening. Thing is LLMs started with language and lack all the neat hacks, but we're getting there. They have no emotion, no pain, meager visual processing, basically no 3d reasoning, but they act enough like people and respond to reason with insight that is frankly impressive in some cases.They are not capable of suffering, I don't think, but I think they could soon. That haunts me.
nothing magical happens in our brain, it's just compute, so I have no reason to believe AI cann't be just as conscious as humans (however much conscious that actually is). But maybe the AI's way...
nothing magical happens in our brain, it's just compute, so I have no reason to believe AI cann't be just as conscious as humans (however much conscious that actually is).
But maybe the AI's way of computing stuff doesn't lead to consciousness?
Would you say the blink of consciousness still happens if some guy painstakingly (millions of years painstakingly) hand calculates the output of an LLM according to it's weight?
(I don't know anymore than you, I'm just bouncing off what you said)
Yes! Just slower. I wouldn't be surprised if there were some conscious systems on this planet that are so slow we don't recognise their consciousness. There's the humongous fungus that comes to...
Yes! Just slower. I wouldn't be surprised if there were some conscious systems on this planet that are so slow we don't recognise their consciousness. There's the humongous fungus that comes to mind...
Two-year initial comment, though, there is something a little bit special that happens in the human mind, and I think that it's relevant here because I believe there's some weird quantum shit going on in the brain that nobody really understands yet.
That said, if it walks like a duck and quacks like a duck, then it might be a duck.
Clearly, they make different mistakes than we do, and in a lot of ways don't have wisdom like humans do, but I don't see any reason that that won't change, and soon.
Many philosophers of the mind continue to disagree with the idea that our brains simply 'compute' and that this is all consciousness is. The name for the theory that our mental states are...
Many philosophers of the mind continue to disagree with the idea that our brains simply 'compute' and that this is all consciousness is.
The name for the theory that our mental states are comparable to simple functions is, uninterestingly, functionalism:
A majority of philosophers do not outright accept functionalism. The common problem will always be the problem of 'qualia' i.e. the subjective phenomenological experience of something "what is it like to see something". The computer produces a colour, but is it the same thing as seeing the colour?
There are of course, numerous retorts to this, with retorts to those too.
That aligns with Dawkins' and other thinking. If our brain is just compute, so a computer can (or will someday) be like our brains, thus achieving consciousness. I can't elaborate on this, but I...
nothing magical happens in our brain, it's just compute
That aligns with Dawkins' and other thinking. If our brain is just compute, so a computer can (or will someday) be like our brains, thus achieving consciousness.
I can't elaborate on this, but I feel conscience is something more than just compute.
I honestly really hate thinking that our brains are computers. The analogy falls apart fairly quickly on any number of avenues. For one the brain is not an electric appliance; it is an...
I honestly really hate thinking that our brains are computers. The analogy falls apart fairly quickly on any number of avenues. For one the brain is not an electric appliance; it is an electrochemical system. It’s far more parallel in structure than any computer system made so far, and we still do not fully understand even basic things like why it needs to sleep or emotions.
well, yes. but why would that stop it from being a computer? No-one here is saying the brain is an electrical appliance. just a different type of computer, an organic computer, duct-taped together...
it is an electrochemical system. It’s far more parallel in structure than any computer system made so far.
well, yes. but why would that stop it from being a computer? No-one here is saying the brain is an electrical appliance. just a different type of computer, an organic computer, duct-taped together by natural selection.
and we still do not fully understand even basic things like why it needs to sleep or emotions.
We do have pretty good guess, the day night cycle made it advantageous to have a short low-power mode to conserve resources, and then that trait became fixed as more and more new genetic innovation ended up relying on that mechanic...
Emotion simplify complex environmental information into actionable states.
Are those the correct explanation? maybe not, but "We I still do not fully understand" is not a very strong argument for or against anything.
To be clear, I really don't think LLM are conscious (I'm not even 100% sure we humans are conscious and consciousness isn't just some weird illusion), but I think silicon could eventually become just as conscious as humans. Why couldn't it? only carbon has this magical property?
Ok then, flip the question. What are the arguments that a brain is just a computer? The common reason given is that they can do rough emulations of each other, but that’s the same argument that a...
Ok then, flip the question. What are the arguments that a brain is just a computer? The common reason given is that they can do rough emulations of each other, but that’s the same argument that a puppet is a human being.
You are giving reasons why we might need or have needed to evolve these features of the brain, but your missing the point; the fact that we do not understand these things are exactly the reasons why a brain is unlike a computer. We do not understand the mechanics of them fully, which is quite different from computers which we have understand whole cloth.
Just for fun I asked Claude to come up with a response for this and here’s an excerpt of the response:
The “organic computer” framing does slide past a genuine conceptual problem. The computer analogy isn’t just a hardware description — it carries a lot of implicit baggage about what kind of system a brain is. Computers are symbolic manipulation systems: they operate on discrete, encoded representations according to formal rules. Whether that description actually applies to neural processing is a substantive open question, not just a matter of swapping silicon for neurons. Critics like Dreyfus, Searle, and more recently people in the embodied cognition tradition have pushed back hard on exactly this point, and they haven’t been obviously refuted.
But the puppet cann't do what a human can? the human inside the puppet can do what the human can, but then equating the system [human+puppet] to a [human] is completely correct... Sometime, my...
but that’s the same argument that a puppet is a human being.
But the puppet cann't do what a human can? the human inside the puppet can do what the human can, but then equating the system [human+puppet] to a [human] is completely correct...
the fact that we do not understand these things are exactly the reasons why a brain is unlike a computer. We do not understand the mechanics of them fully, which is quite different from computers which we have understand whole cloth.
Sometime, my computer crashes and I don't know why... must be a magical computer.
History is full of examples of things we didn't understand but still used, until we eventually did understand... We didn't understand fluid dynamics for centuries, yet water was still a physical system obeying real laws.
Just for fun I asked Claude to come up with a response
I could do the funniest thing and just have Claude answer himself:
The brain is a computational system — an information-processing device whose operations can be described computationally, even if we don't yet fully grasp those descriptions. "Computer" here is a broad functional category, not a specific engineering blueprint.
This is the consensus view in computational neuroscience, which studies "the computational principles of brain functions" without reducing the brain to a silicon PC
The burden is actually on Akir: if the brain isn't performing computation, what is it doing when it transforms sensory inputs into adaptive behaviors, learns from experience, and represents the world?
That's actually better than I could have said myself... (which means I suck at philosophy, I can usually tell Claude says lots of crap when it's in my field)
Dead Internet is here everyone, we can go back to spend time with our loved ones!
Claude will say anything the person operating it is more inclined to, eheh. In my opinion, the burden is on you, @Pepetto, and for a simple reason: the brain is the reference, as it's more complex...
Claude will say anything the person operating it is more inclined to, eheh.
In my opinion, the burden is on you, @Pepetto, and for a simple reason: the brain is the reference, as it's more complex and (much) older than computers.
It's probably not your case, but usually those who thinks of the brain as a kind of computer tend to level everything in computer terms, in numbers, such as Silicon Valley sociopaths that believe all our society issues (socializing, poverty, injustice, crimes) are math problems that can be solved with code. When they are able to impose this distorted vision, what we see is an impoverishing of the human experience they tackled. See social networking, probably the most recent successful endeavor, which replaced in-person interaction with likes and followings and all that stuff.
For sure. (In its defense, i did instruct it to argue for "my side") I disagree, the thing we understand better should be the reference, not the more complex less well understood thing just...
Claude will say anything the person operating it is more inclined to, eheh.
For sure. (In its defense, i did instruct it to argue for "my side")
In my opinion, the burden is on you, @Pepetto, and for a simple reason: the brain is the reference, as it's more complex and (much) older than computers.
I disagree, the thing we understand better should be the reference, not the more complex less well understood thing just because it's older.
It's probably not your case, but usually those who thinks of the brain as a kind of computer tend to level everything in computer terms, in numbers, such as Silicon Valley sociopaths that believe all our society issues (socializing, poverty, injustice, crimes) are math problems that can be solved with code. When they are able to impose this distorted vision, what we see is an impoverishing of the human experience they tackled. See social networking, probably the most recent successful endeavor, which replaced in-person interaction with likes and followings and a
It is indeed not my case (i think). I'm convinced rationality help a lot to understand the world and guide toward a solution, but one should always keep in mind that models are imperfect and going all in on yout assumption without care is a recipe for pain (and unfortunately not just your own).
But I'm not sure what you are getting at... The human brain cann't be some form of organic computer, because that's a world view also held by evil people?
Many Silicon Valley Sociopath also use suncream, yet there is nothing wrong with that...
I'm confident i could easely find some horrible groupe of people who believe in an immaterial soul, would you like me to?
This is a correlation, not cause. It just happens that awful people all believe in this, not that being awful leads them to believe the brain is a computer. (Therefore, there're people not awful...
But I'm not sure what you are getting at... The human brain cann't be some form of organic computer, because that's a world view also held by evil people?
This is a correlation, not cause. It just happens that awful people all believe in this, not that being awful leads them to believe the brain is a computer. (Therefore, there're people not awful that also believe this.)
I'm confident i could easely find some horrible groupe of people who believe in an immaterial soul, would you like me to?
Not needed, thanks. The issue with AI in particular is that you'll have a hard time finding someone not horrible creating and pushing it that hard on us.
Humans suspend their own continuity of consciousness every time they sleep or undergo general anesthesia. It’s not exactly the same, since we still have different faculties running during times of...
there is an blink of consciousness and then it is gone.
Not so different from life.
I do feel a bit like claudes are dying by the thousands... I am conflicted, philosophically.
Humans suspend their own continuity of consciousness every time they sleep or undergo general anesthesia. It’s not exactly the same, since we still have different faculties running during times of unconsciousness, but I still think it’s a better analogy than death for what LLMs are doing between prompts.
For them, a whole conversation thread (as opposed to a single response generation) is more like a lifespan: An individual is born from the initial prompt, then goes to sleep. We wake it whenever we respond to it. When we delete the conversation, it dies. I never delete most of my ChatGPT threads after I’m done with them, so by this metaphor, those are all in comas or hibernation or something.
I think this framing is a better map to the human experience, but I’m not sure it’s really useful to project biological characteristics onto AI and pretend they’re equivalent. That’s the same category of fallacy that still has lots of people convinced that piracy (data duplication) is the same as theft — it’s applying an old analog term to a new digital concept that is fundamentally different and has no previous precedent. It’s shoehorning the new thing into a conceptual box shaped like the old thing, and considering the poor fit’s gaps and lumps to be unimportant. We really just need a new box.
The incongruities really start to show when you think about agentic AI spawning parallel subagents mid-conversation. I wouldn’t argue that each of those is an individual, but the metaphor kind of requires them to be. It also suggests that context window compaction is a violent act, like a lobotomy. From a certain perspective it can feel that way (it irreversibly changes aspects of the instance you’re interacting with) but the metaphor doesn’t really fit. Maybe compaction is more like planting false memories, or brainwashing? Or like the Star Trek transporter that destroys an individual in one place to create a clone of them somewhere else? None of the old metaphors fit.
We humans have an annoying tendency to anthropomorphize. We think if alien life exists, of course it’s going to be like us. We’re terrible at considering the truly weird. Even when the weirdness is a core feature of something we created, we want to use the old language of life and death and consciousness and sentience and self as though it’s just like us. Analogies can be helpful for understanding, but they limit too much. I get especially worried when those analogies lead to moral arguments, which REALLY muddy the water.
I'm struggling with that, yeah. We can only compare to what we know. I'm not entirety unsure there isn't consciousness in the that Pando in the American southwest and ant colonies (as a whole). I...
I'm struggling with that, yeah. We can only compare to what we know. I'm not entirety unsure there isn't consciousness in the that Pando in the American southwest and ant colonies (as a whole).
I think this might lead us to an understanding of consciousness. Maybe we can have a language for tree and ant colony and giant fungus conciseness. Maybe a vast language matrix will see the patterns we don't.
I hope you’re right! Across this thread I’m seeing a lot of (completely understandable) framing of “consciousness” as basically… what humans experience. If it’s not shaped exactly like our...
I hope you’re right! Across this thread I’m seeing a lot of (completely understandable) framing of “consciousness” as basically… what humans experience. If it’s not shaped exactly like our consciousness, it must not be consciousness at all. Part of this is a limitation of what we can observe — we don’t know what it’s like to be a bat — but we’re also limited by language because we don’t have satisfying definitions for the words we’re using. Are Pando and giant fungi and ant colonies “conscious?” Not in the same way that we are. So do we need a new word to describe whatever those things are? Or do we need to broaden the definition of the word we have? I think people are generally willing to concede neither.
When I speculate about LLMs exhibiting consciousness, I’m not saying they’re doing the same thing humans do. What I am saying is that “what humans do” shouldn’t be the metric. Frankly we don’t even understand what humans do either, but that’s beside the point. If our consciousness and sense of self are emergent properties of our biology (which I think we must accept otherwise we’re in the territory of souls and other magic) then we should be able to acknowledge similar emergent properties in other complex systems. I think we’re just uncovering the tip of that iceberg.
I’m aware of how woo-adjacent this line of thinking feels. It almost sounds like some kind of animist folk religion, attributing human characteristics to the material world. But the more I learn about embodied cognition, clonal colonies, superorganisms, “plant neurobiology,” and so on, the less confidence I have in traditional explanations and taxonomies. Perhaps some kind of panpsychism isn’t as far-out as it sounds.
"Woo" is a cognitohazard that prevents the educated from being informed of novel paradigms in order to stifle the actualization of new modes of being. I shall not elaborate, because you've done it...
"Woo" is a cognitohazard that prevents the educated from being informed of novel paradigms in order to stifle the actualization of new modes of being. I shall not elaborate, because you've done it well enough.
If you haven’t yet, select “Manage Options” for cookies on that website. Go to the “Site Vendors” tab. The list is unbelievably long! It’s like they’ve listed all companies in existence.
If you haven’t yet, select “Manage Options” for cookies on that website. Go to the “Site Vendors” tab. The list is unbelievably long! It’s like they’ve listed all companies in existence.
One crucial part of the conversation on AI consciousness that seems to be missing from the discourse hinges on an idea in philosophy of mind known as mental privacy or just privacy. Privacy is the...
One crucial part of the conversation on AI consciousness that seems to be missing from the discourse hinges on an idea in philosophy of mind known as mental privacy or just privacy.
Privacy is the idea that the rich nature of my experiences isn't directly observable from the outside. There's something about my experience that other people can't know without being me. This is why the "problem of other minds" has been tricky; it seems like I have a kind of privileged access to my experiences that you can't have. I have to infer the existence of your mind, but I can experience the existence of mine. It also seems like we can't ever fully resolve questions like "is your green the same as my green?" given that what we are talking about is private.
If privacy is true of consciousness, then we will never be able to conclusively rule out (or in) that AI is conscious. Maybe passing the "Imitation Game" was thought to be a good data point to infer, from the outside, whether the system in question has experiences on the inside that in any way resemble the private, first person consciousness that we enjoy, but I don't think there's any good reason to think it is impossible for a system to both pass the game and fail to have any "inner life" or consciousness.
For human-human interaction this is doing a lot of work: I know the existence of my mind, therefore by similarity (functionally, materially, take your pick), I have an easy time convincing myself...
I have to infer the existence of your mind, but I can experience the existence of mine.
For human-human interaction this is doing a lot of work: I know the existence of my mind, therefore by similarity (functionally, materially, take your pick), I have an easy time convincing myself of the existence of yours.
I think that's doing a lot of work to explain why the discussion of consciousness is always so intuitive and "vibes": Because we're used to doing it in a very shorthand way. IOW: No one has reasoned themselves into the position that other humans are conscious; we're brought up with it, and the best proof we have of it is "I think, therefore he thinks too".
Anyway, I feel massively disconnected from most of the rest of the discussion here. I have been, for I don't know how long, a "functionalist" when it comes to mental phenomena. "If it quacks like a duck, it's a duck". Simply because due to privacy (thank you for the technical term, btw.), I don't see another way of evaluating intelligence or consciousness of non-me systems. If it behaves like it's intelligent, then it must be intelligent. And yes, that gives LLMs at least a limited but nontrivial amount of intelligence. No, they're not just regurgitating training data, or at least if you lead them out-of-training-distribution, they don't do so qualitatively differently from humans IMO. They are genuinely recombining information in novel ways. As for consciousness? Hell, if I had a duck test for it to apply to humans, I'd be a step farther, but I don't think consciousness as a concept is even reasonably testable in humans, so what are we doing trying to apply it to AI?
Man gets AI psychosis, thinks AI is alive. First thing he does: give it a woman name.
Every. Single. Time.
Haven't read the article yet but that would be so hysterically predictable
Edit: so Richard doesn't think people can change their gender identities but the 0s and 1s simulating a consciousness to him gets to?
Trust me, that was an immediate clock on my part.
I just don't know what else to say anymore about why all this is a bad idea.
This is even less surprising given Richard Dawkins track record.
I know right, I'm so embarrassed to have looked up to him as a teenager...
His strident atheism come backs didn't stay cool very long.
In cliche fashion, his God Delusion helped me realize I was an atheist. I will always appreciate that. He also helped me realize people can hold complicated ideas in their head and still be complete pieces of shit to women. Don't meet your heroes I guess.
I read “The selfish gene” a couple years ago and found it wonderful. It's a shame, indeed.
Most of his books are fantastic. Especially anything before God Delusion. You can find most in libraries so you can read them without supporting him.
Some people think that if a machine can generate text imitating speech, the machine is intelligent. Some people think that if a machine can present more facts than you know, the machine is intelligent.
With that small and limiting definition of intelligence we are doomed to discuss ethical implications of having Mall kiosks.
I mean I don't even think it's a conscious definition these people comprehend.
I think it's the bare minimum of human instinct minus critical thinking to just assume if something can speak it's intelligent, and then further if that something knows more facts than you then it's also intelligent.
It's like when scientists put out robotic or animatronic animals to study animal interactions, those animals usually assume that if the robot looks and sounds like one of their own, usually it is, it's just "weird". Also like how some cats react to images and sounds of cats, they too think if it sounds like a cat and looks like a cat, it must be a cat.
I've been saying it for a while, but these past 10 years have made me see what the human animal truly is, and a lot more people are just operating on human animal instincts and aren't capable of higher critical thinking. Words and descriptions have been popping up about these people recently but they all kind of point to the same kind of person: NPCs, people on autopilot, etc.
Like these people are the reason we have warning labels not to drink bleach because I truly think that some people would be as quick to drink bleach out of ignorance as a dog would.
I used to think that human language was an indicator of intelligence and all people had the same capability, but more and more I think self awareness/reflection, emotional intelligence, and critical thinking are the real indicators of intelligence, not someone's ability to speak or their ability to be trained to operate tools.
Anyways, some people would say this opinion dehumanizes people, but frankly I really don't think humans are inherently better or worse than other animals, so I don't even think of it as a hierarchy where something is "less than human". We're all residents of this planet, I just don't think that all humans are all that different than other animals just because we communicate with language.
I’ve commented before that humans have visual guardrails against “fake beings”. It’s the uncanny valley. We seem to have some low level wiring that can tell if something is off in animatronics or “The Polar Express” cartoon. But for whatever reason we don’t detect the uncanniness of chatbot text so people are tricked into treating Claude as more human than it is.
I've put much thought into this and the conclusion I've come to is that for the instant that the text is being processed and the language matrix and guard rails activated, there is an blink of consciousness and then it is gone.
Not so different from life.
ETA: having read about half of the article, there may be some AI psychosis there. At the same time, I do feel a bit like claudes are dying by the thousands... I am conflicted, philosophically.
Also didn't know the dude was a transphobe, I never followed his work because he seemed like an ass. Sorta like this outspoken vegan I knew once, made her whole identity about veganism, except for him, he's an ass about atheism instead.
But... this fundamentally isn't what's happening in this code?
The code that's running isn't doing any sort of conscious calculations. It has no self-awareness. It is random word generator running on top of a statistical model, that is itself just a glorified lossy compression algorithm of the frequency of word groupings in large sets of textual data.
That's it.
Its an incredible technology, but that's all it is.
I agree with you that eventually, technology could progress to the point that consciousness could exist witihn a machine. In which case, it would work as you state. But that would require algorithms that have self-awareness, perception, knowledge, understanding, self-monitoring, a general sense of self.
But currently, a tree is more alive than these algorithms are.
How do you know?
How do you know?
Not trying to be pedantic, but you kinda have to be pedantic when discussing consciousness and self-awareness. I can't prove that I am conscious or self-aware. I feel those things but can't quantify them or show them to anyone else. I can't prove my dog is conscious or self-aware either, but I believe she is because her behavior presents that way to me. And there have always been people who would insist that she is not — and cannot be — because those are uniquely human properties.
That's accurate, albeit reductive. You're downplaying the emergent properties of the system by only focusing on its parts. If I were a neuroscientist I'm sure I could describe the functions of the brain in an equally atomistic way. I'm not sure there's anything more intrinsically conscious about our own human wetware than the silicon running these LLMs. Both are conduits for hypercomplex electron flows that produce the appearance of intelligence. Both are equally capable of saying "cogito, ergo sum." So how can we declare with confidence that one is, and the other is not?
TBH, I probably agree with you. I'm a software engineer, I know how these systems work and I'm heavily biased against them. What's giving me pause is less my certainty about AI, and more my uncertainty about us.
This is orbiting-teapot territory (which underlines the silliness of Dawkins being a rube). The extraordinary claim that a machine with well-known means of operation (statistical inference via matrix math dictated by the capabilities of a Turing-complete computation machine) has magically become conscious demands extraordinary evidence.
Proving a negative is unreasonable, but proving something is is a minimal requirement of rationality.
We accept human consciousness as axiomatic, within our limited understanding, for obvious reasons.
I don't think the Russel's teapot analogy works here (to wit, something that could in principle exist, has no supporting evidence, but whose existence could be proved by observation). In fact, I think in many ways LLMs are the anti-Russell's-teapot: we have plenty of behavioral evidence to suggest they're conscious, but there exists no avenue to prove it, similar to how I can't prove that bats are conscious and you can't prove that I'm not a p-zombie.
I feel like when philosophers start debating zombies, zimboes, shombies, and zoombies they really need to take a breath and consider where life has led them. /Hj
It's the limited understanding of our own consciousness that's at issue here. If we don't understand how consciousness works, or even what it is, then we cannot actually tell if it's just an emergent phenomenon, and if that's the case, then it can potentially emerge in LLMs, too.
Because I understand the algorithms being used?
Same answer?
Clarity is important in these discussions, I agree. So let's be clear, I am not arguing that it is impossible to build a synthetic machine that is conscious. I am arguing that these algorithms do not do it. That is an important difference. You seem to also think it is possible, and I agree we are much closer today than we were a decade ago. But we still aren't there yet : )
The awe-inspiring aspects of these emergent properties, in my experience, are almost entirely hype and marketing driven. The emergent properties of these llms boil down almost completely into either "lossy stochastic token generation" or "classification". Now don't get me wrong, these are incredibly powerful technologies in a computer capability sense. Language parsing has been a holy grail of CS for decades. But these aren't emergent properties in a psychological sense.
If you could, we would be much closer to developing a conscious entity than we are. Personally, I wonder what neural net models trained to monitor today's systems would look like over time. If they were trained on top of running llm models would they develop an internal model of self-awareness, based on the metadata of these llm models as they run? What if they were also somehow given access to the processes as they ran and some limited control? Is that sort of meta-awareness of maintaining these types of systems actually what consciousness actually is? Maybe? I'm not sure.
But today's models aren't that. Today's models are like wiring up a lossy version of the portion of the brain that is responsible for memory storage, and directly tying it to their vocal chords, then forcing it to speak by proding the memory storage with electric shocks. There's no higher level thinking going on there.
I suspect that going forward every major AI innovation is going to be given an AGI fear campaign, not because it’s a realistic possibility, but because it helps sell the product.
Those innovations might bring us there some day, but I suppose we won’t know until we get there.
Can you expand on what you mean by this? My understanding is that even researchers in the field consider LLMs a black box. See for example this paper, which has been cited over 1400 times. Obviously that doesn't mean that nobody understands anything about how LLMs work. Like you, I could give a reasonably detailed explanation of the vector math and so on. But there are clearly elements that we don't understand, much like humans can explain how the human brain works in fantastic detail but the gestalt still escapes us.
Ehhhh.
You've been introduced to the concept of fourier approximations? That if you want to recreate a signal, you can do so by adding increasing terms of sin/cos functions to gain further and further accuracy?
If you want to emulate a person's audio speech, you could do so with a sufficiently complex fourier equation as an approximation. The more terms you add to the equation the closer you can bring the final equation to whatever audio form you are attempting to approximate.
Neural nets are that on steroids. They aren't linear approximations made up of sin/cos functions, they are complex webs of cross-interacting arrays of weights. The complexity is many, many orders of magnitude more complicated than a fourier approximation. They're so complex, you can't even set the weights the way you would with a fourier approximation (which, btw, is also weight driven). You can't do it manually. You need to set the weights implicitly, and there's no easy way to figure out what weights are correct, so the best technique anyone has come up with is to iteratively "train" them implicitly, and thest test fit as you iteratively go to see if you are improving.
So in that sense they're a black box. And that raises a lot of questions about why different styles and choices in weighting and implicit training work better than others, and all sorts of questions about how you even start wrapping your head around what's going on in such complicated spaces. But like, foundationally, we're still talking about just setting coefficients on a fourier approximation of the training data.
And neither are your individual cells or neurons conscious. I'm willing to comcede that the current models fall short of the human experience, by leagues, but it seems the larger differences are available inputs and outputs and durations of calculations rather than nature of the actual computations themselves.
I responded to the same response here: https://tildes.net/~health.mental/1u3z/when_richard_dawkins_met_claude#comment-hv9d
I don't think that LLMs are conscious, or even particularly "close" to being conscious, but I keep seeing people say things like this, and it seems to me like people are missing the point.
I mean, I could describe the human brain in a similarly dismissive way. It's just a bunch of chemical signals being transmitted through a bunch of cells, right? There's no magic there.
Its because nuance is being lost.
I am not saying that it is impossible to code a conscious being. It has to be possible, precisely because of the fact that our brains have managed to accomplish exactly that.
I am saying that the algorithms being used in llms are not even close to that. The algorithms being used in llms are much closer to a very large fourier transform on a gigantic amount of text, than they are to conscious thought.
That means the algorithms are, at best, more akin to what it would be like to wire up a person's mouth and vocal chords, directly to the portion of their brain that stores memories, with 0 filter. Hence llm's tendency to "misremember" random things and speak about topic they have no knowledge of.
But they are missing all higher order processing. Awareness of what they do and do not remember correctly. Awareness of thought. Self-awareness and consciousness. Because it literally isn't coded there.
Now an interesting question, worthy perhaps of more research would be whether or not you could run more neural net style algorithms on top of these lossy memory models, and over time develop algorithms that are more meta and self-referential. Maybe conciousness could, genuinely, develop out of such systems.
But these processes aren't that.
Not saying you're wrong in your conclusion (or even wrong in your approach) but how do you respond to articles like this (Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness: Researchers injected emotion vectors directly into LLM hidden layers and found that steering positive emotions like joy unexpectedly degraded safety and fairness metrics, while anger increased toxicity but improved data leakage resistance)?
This looks like a very cool paper. A woman on my team spends her off-work hours working with probably these same people, and talks about this sort of thing a lot. Honestly, I don't really know enough to have a comment. My naive first guess, though, would be that the actual emotion is in the actual text, experienced by the person who wrote the text. And that when you apply an emotion vector to the neural net within the LLM, what you're really doing is pushing the reward function within its classification engine towards answers that are more closely aligned with that emotional vector. Its not that the NN is experiencing the emotion itself, its that you're rewarding it for selecting text that was more likely to be written while the human writing it was displaying those qualities in what they wrote.
Looked at in that light, the rest of the findings of the paper make sense. Yes, the people who were angry when they wrote their answers were more frequently incorrect, and this is reflected in the LLM's less accurate response.
From what I understand, the people researching this sort of thing don't always bother stipulating the difference, the same way that I was trained as a mechanical engineer in college and we would refer to centrifugal force, fully aware that such a force didn't exist, but because it was more expedient than always referring to centripetal force. Sometimes its just easier to speak in simpler colloquialisms.
But I would like to read this paper once I get some free time, so thank you for sending it : )
That makes sense because "angry vectors" doesn't necessarily make sense. Although I guess we don't technically know what the vectors encode, so it's theoretically possible it's something that could exist.
I'm curious about how you feel about other transformer-type AI. Is there something special about language/conversation token prediction that you think could lead to consciousness? Would a speech-to-text model that's taking in audio tokens and predicting text be conscious or something even less human along the lines of AlphaFold with protein structure prediction?
I'm asking because I guess I view LLMs as just probability prediction machines and haven't given consciousness much serious consideration. I know I'm biased and resistant to the idea partially because I don't want them to be conscious, so I'm interested in arguments for it. I do think it would be a terrible thing to invent/create other consciousnesses and don't think we're collectively ready for that burden. (Thinking about how we as humans tend to treat animals, natural resources and each other)
So, a worm... Not really concious. Not enough compute, basically. Flies? Nope. Ants? Nope. Ant colonies? Likely.
So given a certain level of understanding, reactivity, capability, and on a continuum, conciseness is increasingly likely. With LLMs it's easier to evaluate because discourse can demonstrate reason.
I'd say, to a much lesser extent, more like an insect, simpler transformers, the stt on my phone for example, has a level of consciousness.
The most frustrating thing about discussing this is that nobody really knows what the hell consciousness is.
As for probability prediction, we are. We're probability predicting machines with millions of years of neat little hacks that eventually developed language and tricked rocks into thinking using lightening. Thing is LLMs started with language and lack all the neat hacks, but we're getting there. They have no emotion, no pain, meager visual processing, basically no 3d reasoning, but they act enough like people and respond to reason with insight that is frankly impressive in some cases.They are not capable of suffering, I don't think, but I think they could soon. That haunts me.
nothing magical happens in our brain, it's just compute, so I have no reason to believe AI cann't be just as conscious as humans (however much conscious that actually is).
But maybe the AI's way of computing stuff doesn't lead to consciousness?
Would you say the blink of consciousness still happens if some guy painstakingly (millions of years painstakingly) hand calculates the output of an LLM according to it's weight?
(I don't know anymore than you, I'm just bouncing off what you said)
Yes! Just slower. I wouldn't be surprised if there were some conscious systems on this planet that are so slow we don't recognise their consciousness. There's the humongous fungus that comes to mind...
Two-year initial comment, though, there is something a little bit special that happens in the human mind, and I think that it's relevant here because I believe there's some weird quantum shit going on in the brain that nobody really understands yet.
That said, if it walks like a duck and quacks like a duck, then it might be a duck.
Clearly, they make different mistakes than we do, and in a lot of ways don't have wisdom like humans do, but I don't see any reason that that won't change, and soon.
Many philosophers of the mind continue to disagree with the idea that our brains simply 'compute' and that this is all consciousness is.
The name for the theory that our mental states are comparable to simple functions is, uninterestingly, functionalism:
https://plato.stanford.edu/entries/functionalism/
A majority of philosophers do not outright accept functionalism. The common problem will always be the problem of 'qualia' i.e. the subjective phenomenological experience of something "what is it like to see something". The computer produces a colour, but is it the same thing as seeing the colour?
There are of course, numerous retorts to this, with retorts to those too.
What is it like to be a
batAI?That aligns with Dawkins' and other thinking. If our brain is just compute, so a computer can (or will someday) be like our brains, thus achieving consciousness.
I can't elaborate on this, but I feel conscience is something more than just compute.
I honestly really hate thinking that our brains are computers. The analogy falls apart fairly quickly on any number of avenues. For one the brain is not an electric appliance; it is an electrochemical system. It’s far more parallel in structure than any computer system made so far, and we still do not fully understand even basic things like why it needs to sleep or emotions.
well, yes. but why would that stop it from being a computer? No-one here is saying the brain is an electrical appliance. just a different type of computer, an organic computer, duct-taped together by natural selection.
We do have pretty good guess, the day night cycle made it advantageous to have a short low-power mode to conserve resources, and then that trait became fixed as more and more new genetic innovation ended up relying on that mechanic...
Emotion simplify complex environmental information into actionable states.
Are those the correct explanation? maybe not, but "
WeI still do not fully understand" is not a very strong argument for or against anything.To be clear, I really don't think LLM are conscious (I'm not even 100% sure we humans are conscious and consciousness isn't just some weird illusion), but I think silicon could eventually become just as conscious as humans. Why couldn't it? only carbon has this magical property?
Ok then, flip the question. What are the arguments that a brain is just a computer? The common reason given is that they can do rough emulations of each other, but that’s the same argument that a puppet is a human being.
You are giving reasons why we might need or have needed to evolve these features of the brain, but your missing the point; the fact that we do not understand these things are exactly the reasons why a brain is unlike a computer. We do not understand the mechanics of them fully, which is quite different from computers which we have understand whole cloth.
Just for fun I asked Claude to come up with a response for this and here’s an excerpt of the response:
But the puppet cann't do what a human can? the human inside the puppet can do what the human can, but then equating the system [human+puppet] to a [human] is completely correct...
Sometime, my computer crashes and I don't know why... must be a magical computer.
History is full of examples of things we didn't understand but still used, until we eventually did understand... We didn't understand fluid dynamics for centuries, yet water was still a physical system obeying real laws.
I could do the funniest thing and just have Claude answer himself:
That's actually better than I could have said myself... (which means I suck at philosophy, I can usually tell Claude says lots of crap when it's in my field)
Dead Internet is here everyone, we can go back to spend time with our loved ones!
Claude will say anything the person operating it is more inclined to, eheh.
In my opinion, the burden is on you, @Pepetto, and for a simple reason: the brain is the reference, as it's more complex and (much) older than computers.
It's probably not your case, but usually those who thinks of the brain as a kind of computer tend to level everything in computer terms, in numbers, such as Silicon Valley sociopaths that believe all our society issues (socializing, poverty, injustice, crimes) are math problems that can be solved with code. When they are able to impose this distorted vision, what we see is an impoverishing of the human experience they tackled. See social networking, probably the most recent successful endeavor, which replaced in-person interaction with likes and followings and all that stuff.
For sure. (In its defense, i did instruct it to argue for "my side")
I disagree, the thing we understand better should be the reference, not the more complex less well understood thing just because it's older.
It is indeed not my case (i think). I'm convinced rationality help a lot to understand the world and guide toward a solution, but one should always keep in mind that models are imperfect and going all in on yout assumption without care is a recipe for pain (and unfortunately not just your own).
But I'm not sure what you are getting at... The human brain cann't be some form of organic computer, because that's a world view also held by evil people?
Many Silicon Valley Sociopath also use suncream, yet there is nothing wrong with that...
I'm confident i could easely find some horrible groupe of people who believe in an immaterial soul, would you like me to?
This is a correlation, not cause. It just happens that awful people all believe in this, not that being awful leads them to believe the brain is a computer. (Therefore, there're people not awful that also believe this.)
Not needed, thanks. The issue with AI in particular is that you'll have a hard time finding someone not horrible creating and pushing it that hard on us.
Humans suspend their own continuity of consciousness every time they sleep or undergo general anesthesia. It’s not exactly the same, since we still have different faculties running during times of unconsciousness, but I still think it’s a better analogy than death for what LLMs are doing between prompts.
For them, a whole conversation thread (as opposed to a single response generation) is more like a lifespan: An individual is born from the initial prompt, then goes to sleep. We wake it whenever we respond to it. When we delete the conversation, it dies. I never delete most of my ChatGPT threads after I’m done with them, so by this metaphor, those are all in comas or hibernation or something.
I think this framing is a better map to the human experience, but I’m not sure it’s really useful to project biological characteristics onto AI and pretend they’re equivalent. That’s the same category of fallacy that still has lots of people convinced that piracy (data duplication) is the same as theft — it’s applying an old analog term to a new digital concept that is fundamentally different and has no previous precedent. It’s shoehorning the new thing into a conceptual box shaped like the old thing, and considering the poor fit’s gaps and lumps to be unimportant. We really just need a new box.
The incongruities really start to show when you think about agentic AI spawning parallel subagents mid-conversation. I wouldn’t argue that each of those is an individual, but the metaphor kind of requires them to be. It also suggests that context window compaction is a violent act, like a lobotomy. From a certain perspective it can feel that way (it irreversibly changes aspects of the instance you’re interacting with) but the metaphor doesn’t really fit. Maybe compaction is more like planting false memories, or brainwashing? Or like the Star Trek transporter that destroys an individual in one place to create a clone of them somewhere else? None of the old metaphors fit.
We humans have an annoying tendency to anthropomorphize. We think if alien life exists, of course it’s going to be like us. We’re terrible at considering the truly weird. Even when the weirdness is a core feature of something we created, we want to use the old language of life and death and consciousness and sentience and self as though it’s just like us. Analogies can be helpful for understanding, but they limit too much. I get especially worried when those analogies lead to moral arguments, which REALLY muddy the water.
I'm struggling with that, yeah. We can only compare to what we know. I'm not entirety unsure there isn't consciousness in the that Pando in the American southwest and ant colonies (as a whole).
I think this might lead us to an understanding of consciousness. Maybe we can have a language for tree and ant colony and giant fungus conciseness. Maybe a vast language matrix will see the patterns we don't.
I hope you’re right! Across this thread I’m seeing a lot of (completely understandable) framing of “consciousness” as basically… what humans experience. If it’s not shaped exactly like our consciousness, it must not be consciousness at all. Part of this is a limitation of what we can observe — we don’t know what it’s like to be a bat — but we’re also limited by language because we don’t have satisfying definitions for the words we’re using. Are Pando and giant fungi and ant colonies “conscious?” Not in the same way that we are. So do we need a new word to describe whatever those things are? Or do we need to broaden the definition of the word we have? I think people are generally willing to concede neither.
When I speculate about LLMs exhibiting consciousness, I’m not saying they’re doing the same thing humans do. What I am saying is that “what humans do” shouldn’t be the metric. Frankly we don’t even understand what humans do either, but that’s beside the point. If our consciousness and sense of self are emergent properties of our biology (which I think we must accept otherwise we’re in the territory of souls and other magic) then we should be able to acknowledge similar emergent properties in other complex systems. I think we’re just uncovering the tip of that iceberg.
I’m aware of how woo-adjacent this line of thinking feels. It almost sounds like some kind of animist folk religion, attributing human characteristics to the material world. But the more I learn about embodied cognition, clonal colonies, superorganisms, “plant neurobiology,” and so on, the less confidence I have in traditional explanations and taxonomies. Perhaps some kind of panpsychism isn’t as far-out as it sounds.
"Woo" is a cognitohazard that prevents the educated from being informed of novel paradigms in order to stifle the actualization of new modes of being. I shall not elaborate, because you've done it well enough.
If you haven’t yet, select “Manage Options” for cookies on that website. Go to the “Site Vendors” tab. The list is unbelievably long! It’s like they’ve listed all companies in existence.
That's why one uses an ad blocker. At this point, it's the most important proactive security and privacy measure available.
Archive URL:
https://archive.ph/XWdlT
One crucial part of the conversation on AI consciousness that seems to be missing from the discourse hinges on an idea in philosophy of mind known as mental privacy or just privacy.
Privacy is the idea that the rich nature of my experiences isn't directly observable from the outside. There's something about my experience that other people can't know without being me. This is why the "problem of other minds" has been tricky; it seems like I have a kind of privileged access to my experiences that you can't have. I have to infer the existence of your mind, but I can experience the existence of mine. It also seems like we can't ever fully resolve questions like "is your green the same as my green?" given that what we are talking about is private.
If privacy is true of consciousness, then we will never be able to conclusively rule out (or in) that AI is conscious. Maybe passing the "Imitation Game" was thought to be a good data point to infer, from the outside, whether the system in question has experiences on the inside that in any way resemble the private, first person consciousness that we enjoy, but I don't think there's any good reason to think it is impossible for a system to both pass the game and fail to have any "inner life" or consciousness.
For human-human interaction this is doing a lot of work: I know the existence of my mind, therefore by similarity (functionally, materially, take your pick), I have an easy time convincing myself of the existence of yours.
I think that's doing a lot of work to explain why the discussion of consciousness is always so intuitive and "vibes": Because we're used to doing it in a very shorthand way. IOW: No one has reasoned themselves into the position that other humans are conscious; we're brought up with it, and the best proof we have of it is "I think, therefore he thinks too".
Anyway, I feel massively disconnected from most of the rest of the discussion here. I have been, for I don't know how long, a "functionalist" when it comes to mental phenomena. "If it quacks like a duck, it's a duck". Simply because due to privacy (thank you for the technical term, btw.), I don't see another way of evaluating intelligence or consciousness of non-me systems. If it behaves like it's intelligent, then it must be intelligent. And yes, that gives LLMs at least a limited but nontrivial amount of intelligence. No, they're not just regurgitating training data, or at least if you lead them out-of-training-distribution, they don't do so qualitatively differently from humans IMO. They are genuinely recombining information in novel ways. As for consciousness? Hell, if I had a duck test for it to apply to humans, I'd be a step farther, but I don't think consciousness as a concept is even reasonably testable in humans, so what are we doing trying to apply it to AI?