We don’t know that in the same way we don’t know the Kool-Aid Man won’t burst through the wall at any given moment and start singing Freebird. Technically the chances aren’t 0, but they may as well be.
I am a software engineer. I deal with LLMs and their implementation every day. I’m not a specialist, but I have a good enough idea of how they work. Their core functionality is fundamentally no different than the predictive text function on your phone. Even setting aside the fact that sentience is a complex and ill-defined concept, certainly you can agree that your autocorrect is nowhere in the realm of sentient.
Not the OP, but any software developer who has spent some time with the inner workings of LLMs can tell you that. Open weight models are a good opportunity to explore how LLMs work on the inside and to experiment with them a lot more than you can do with the so called Frontier models. The technology however is the same. These are just very fancy statistical prediction machines, and the reason why you are so impressed with them is that the statistical average of what sounds reasonable in a conversation is exactly what people expect to hear on average, given a specific context. It’s a neat little trick that works only only as long as the context fits (contentwise and spacewise), and once you realize that you’ll see LLMs as what they truly are. There is no magic, and the “intelligence” is the intelligence of the people who built those systems, the systems themselves are not intelligent in any commonly used sense of the word.
I am of the opinion that there’s no magic in sapience in the first place. Our brains are just pattern matching, statistical prediction machines that have had trillions of iterations of improvements over time. The intelligence we flatter ourselves with is only a small part of what we have evolved in our brains. For eons the problem space was all physical, but with AI, the starting point is the “intelligence” piece.
While that is partly true, it oversimplifies a few things. And these are the things that make LLMs a dead end for intelligence:
Evolution means constant learning and adaption. LLMs can’t be adapted on the go, they are stuck with their initial training data and retraining is a slow and externally controlled process. Apart from training, every additional information you give it while running must be loaded into the context, the space for context is limited and must be flushed before supplying different context. That’s a hard limit for evolution, at least with current technology.
The physical world is more important than you might think because everything hinges on this realm, including the electricity, cabling and hardware involved in running a model. The idea of an intelligent entity that exists purely as an ethereal cloud of logic unbound by physics is hard to justify, because the shape of logic itself is defined by the physical ground we are standing on. Therefore, an AI would need a first-hand access and understanding of the physical world to be successful. LLMs can’t have any of that, because they are purely language-bound.
I personally think Yann LeCun (a pioneer of machine learning and AI) is on the right track with his research and his opinion that LLMs are a dead end for artificial intelligence. The reasons are more on a technical than a philosophical side.
The problem is that despite being a literally true statement, it relies on the implicit assumption that a system founded in statistics necessarily has limitations that preclude sentience (can generalize this to other capabilities, the same argument is often used), when whether that is the case is the entire question at issue. It’s totally meaningless if someone just feels intuitively that this must be true, that’s not a real argument.
Worse, it gestures at an idea that is demonstrably false, that this technology is equivalent in all important ways to processing inputs through a math function someone wrote, when the whole innovation here is that the training process produces novel mechanisms for solving the given problem, mechanisms no one specifically intended or predicted, or in most cases even understands. And those mechanisms can do some seriously impressive things that were previously impossible.
A good argument here would at minimum need to actually explain why this is insufficient.
The problem is that despite being a literally true statement, it relies on the implicit assumption that a system founded in statistics necessarily has limitations that preclude sentience
Nope. We do not assume any system can produce sentience just by being complex. We have no reason to begin with that assumption so there’s no assumption to be made in dismissing it.
I am not making the claim that it can produce sentience (personally I don’t see sentience as a very meaningful concept anyway), I’m saying the arguments being made here that it cannot are deficient. Reminder of the context: https://lemmy.dbzer0.com/post/77644247/28876812
Play semantic games all you like, but nobody who actually understands how this shit works (and doesn’t have a vested interest in lying about it to the public) would suggest it’s capable of gaining sentience. It’s something you shouldn’t let people suggest as a serious possibility at this point.
but nobody who actually understands how this shit works (and doesn’t have a vested interest in lying about it to the public) would suggest it’s capable of gaining sentience.
Why is that? All I see is arguments misrepresenting how it works (see my earlier comment and the linked source) to try to say that it is certainly not capable of that without necessary elaboration.
It’s something you shouldn’t let people suggest as a serious possibility at this point.
You’re saying this as though it has been conclusively debunked, which backs up what I’m saying wrt what the goalposts are here: there is no such certainty. Arguing that the issue is uncertain is not a semantic game.
Edit: The crux of this is that there has been a huge amount of progress with transformer models, with constant claims that the line of where the limits are and what capabilities will never be achieved being made and disproved in recent years. Given that, I think it’s fair to say that any claims that a particular capability is forever out of reach call for a really strong justification.
Are whales sentient? Apes? Dogs? Squid? Alligators? Spiders? Fleas? What’s the test? (Yeah, the mirror test - how well do you think that really works?)
H. sapiens has more in common with fleas than LLMs - the LLMs only seem more familiar because they’re trained on our text-based interactions. Whatever sentience LLMs do or do not develop, we will be debating how to define it probably until they start reproducing themselves and force us to a sit-down negotiation over how we will divide the Earth’s resources - or maybe they won’t bother with negotiation.
So, there’s a difference between sentience and sapience, being self aware, or having consciousness.
Sentient just means the ability to experience feelings.
Pretty much all the animals you mentioned are indeed sentient.
LLMs are not conscious or self aware. They don’t have the ability to feel, so they’re not sentient. There’s an argument that they’re reaching sapience or wisdom, which is the ability to apply experience and tacit knowledge to new situations and complex questions.
Which is the issue. They can become sapient before sentient, meaning they can apply experience to new situations without the ability to use feeling as part of that. That’s pretty much the science fiction AI nightmare. A cold, unemotional, unfeeling computer that can apply logic and experience to problems.
Yet another slippery fish to try to wrestle with bare hands. What constitutes complexity? Some relatively simple looking questions turn out to be highly complex: Are there any whole-number solutions to (a^n + b^n = c^n) if (n) is greater than 2?
Then, any number of complex looking questions turn out to be quite simple with a little insight: such as any large product which has a factor that reduces to 0.
What’s simple for some people is hard for others, and I would posit that is greatly amplified when you stop comparing humans to humans and start comparing humans to animals or machines…
they can apply experience to new situations without the ability to use feeling as part of that.
We put unfeeling psychopaths in charge of our largest corporations, to maximize profits - a machine might be less scary in the long run.
Well, that’s not really a philosophical question. Simple questions are ones that only deal with one question at a time and have single answers.
Complex question involve multiple implicit questions and require inference to answer and there are multiple possible correct answers.
“Are there any whole-number solutions to (a^n + b^n =c^n) where n>2” isn’t complex. That’s actually extremely easy to answer. You could just keep brute forcing through every whole number in known existence if you had the compute power. Like trying to find the next digit of pi. But that’s not Fermat’s last theorem. The actual problem is “there are no whole-number solutions to (a^n + b^n =c^n) where n>2, prove it“ which is much more complex because it requires answering multiple questions along the way. The implicit questions that arise are simple “why” questions that have to be answered.
The answer to your question is “no”. But then the follow is “well why”. Wrapping it all up in a nice equation that explains all the “well why” questions before they’re even asked is the complex part.
AI is already getting to that point. LLMs are not synonymous with AI. They’re a type of AI but there are far more models out there than just LLMs. They’re just a subset of generative AI.
“Easy” and “hard” are NOT the same thing as “simple” and “complex”.
Choose your field… two rabbis (dirty old men, of course) having agreed that abortion is a sin: debating whether it is a sin for a woman to perform feminine hygeine after rape to possibly prevent pregnancy from that rape. Where on the continuum from Monty Python’s “every sin is sacred” up to late term does abortion become a sin? Do the circumstances of the insemination come into play or not, and if so… how? What considerations are given to alienating the woman and perhaps her family from the synagogue if too hard-line of an interpretation is taught? Should the considerations be drawn up mathematically with weights and formulas if possible? Like physics, if you have a formula you can check it for correctness by putting in test values and seeing if the results are reasonable.
“Easy” and “hard” are NOT the same thing as “simple” and “complex”.
No, but they rhyme - so to speak. Hard problems become easier when their apparent complexities are simplified.
Well that first one is very much a complex question with no correct answer, that also requires sentience, not just sapience. Those are the questions where the disconnect becomes dangerous, because now you bring in morals, and moral questions are by definition emotional. There’s no such thing as a concrete standard of morals.
The rabbis agree that abortion is sin. Do they agree that preventing pregnancy from the start counts as abortion? Given that they’re rabbis, the answer would probably be no, because even the most Orthodox Jews believe women are allowed contraception.
But that’s not really an answer to the real question, because the real question of when does abortion become a sin requires answering the next question of when does it become abortion and not prevention? Is it the Catholic view where preventing a pregnancy is the same as abortion, the biological point where it becomes a living organism with its own DNA, or the medical view where it becomes a pregnancy when it reaches the implantation stage?
Again, that’s not a philosophical issue. That question is complex because it requires answering MANY other questions along the way to reach a final answer to the original question. It’s an extremely complex question because each of those questions that arise also have a million questions to answer to get the answer to them. And all of them being “right” or “wrong” rely on large scale social acceptance of the answers to each other questions, as well as the ultimate question of if there is even such a thing as a concrete “right” or “wrong” answer to any ethical question.
Those answers are always based on emotion. All moral and ethical questions are based on emotion because morals and ethics are emotional constructs.
That’s where AI fails. AI doesn’t have sentience. It can’t feel emotion and therefore can’t reason using emotion. It can only go by what it is trained on as concretes.
And, kinda? They kinda rhyme. There are PLENTY of things that are simple but hard, and plenty of things that are complex but easy. Like losing weight. Losing weight is extremely simple. It’s basic thermodynamics. Consume less energy than you expend, and you lose weight. Period. No exception. Now that can be extremely HARD for most people to actually do. Tying your shoes is relatively complex. It’s multiple steps that all depend on the previous step being completed properly. But it’s easy enough to do that most people can master it by the time they’re 8.
There’s no such thing as a concrete standard of morals.
My sociology professor used to say: "if you look far enough in human societies, you will find examples of just about everything imaginable being accepted, or restricted except for incest and cannibalism, those two have clearly observable problems and generally societies will shun them eventually - with notable exceptions such as the Hapsburgs of Europe and the odd cannibal tribe in the Pacific.
It’s an extremely complex question because each of those questions that arise also have a million questions to answer to get the answer to them. And all of them being “right” or “wrong” rely on large scale social acceptance of the answers to each other questions
For an AI to participate in such discussions, it - like a human - would need to learn the locally accepted social mores and synthesize any gaps from the available information. The final result for this and similar questions will depend heavily on the society it is learning from - uncontacted tribes in the Amazon will have different answers from tribes in equatorial West Africa, and of course different answers still from Christian, Muslim, Hindu etc. etc. (each of those having dozens to hundreds of sects…) There is no global answer.
AI doesn’t have sentience. It can’t feel emotion and therefore can’t reason using emotion.
To the extent that emotion is expressed in lexical tokens it can process, AI can process those tokens in context and “feel” to the fidelity that writers can express what emotion is and how it should / should not, does / does not impact decisions.
…losing weight is simple… that can be extremely HARD for most people to actually do.
Knowing the physics doesn’t make self-control any easier. Losing weight is a behavioral thing and many of the factors in the behavioral choices aren’t under rational control - the decision to eat is strongly driven by all kinds of biological factors that don’t care about your reason and philosophies. I wonder what similar problems AI may develop in the future as an emergent behavior of other things we are building into it.
> The large language model chatbots use is not a path to sentience.
We don’t know that and won’t know that until there are sentient machines.
You’re right, in the sense that we wouldn’t know what is a path to sapience until it happens, but we can only work with what we know, and what we do know is that the transformer framework language models use is not what a brain is, and a brain is, as far as we know, the only path. The artificial neural network technology transformers are based on is rooted on a mathematical representation of an extremely simplified model of how neurons work. Human minds don’t work in tokens, and brains have chemical signals and things that also have an impact. The physical wiring is part of the story.
We’d need to objectively define what sentience/sapience is to begin with, anyway. We can’t even do that for people and animals, expanding it to machines seems an exercise in futility. It wasn’t that long ago that babies and animals were considered to be little more than automata as far as consciousness goes.
It would be amazing to be proven wrong, because the implications for philosophy and neurology would be staggering, but I digress.
We don’t know that and won’t know that until there are sentient machines.
We don’t know that in the same way we don’t know the Kool-Aid Man won’t burst through the wall at any given moment and start singing Freebird. Technically the chances aren’t 0, but they may as well be.
What is your area of expertise that makes you so certain about how sentience can be attained?
I am a software engineer. I deal with LLMs and their implementation every day. I’m not a specialist, but I have a good enough idea of how they work. Their core functionality is fundamentally no different than the predictive text function on your phone. Even setting aside the fact that sentience is a complex and ill-defined concept, certainly you can agree that your autocorrect is nowhere in the realm of sentient.
Not the OP, but any software developer who has spent some time with the inner workings of LLMs can tell you that. Open weight models are a good opportunity to explore how LLMs work on the inside and to experiment with them a lot more than you can do with the so called Frontier models. The technology however is the same. These are just very fancy statistical prediction machines, and the reason why you are so impressed with them is that the statistical average of what sounds reasonable in a conversation is exactly what people expect to hear on average, given a specific context. It’s a neat little trick that works only only as long as the context fits (contentwise and spacewise), and once you realize that you’ll see LLMs as what they truly are. There is no magic, and the “intelligence” is the intelligence of the people who built those systems, the systems themselves are not intelligent in any commonly used sense of the word.
I am of the opinion that there’s no magic in sapience in the first place. Our brains are just pattern matching, statistical prediction machines that have had trillions of iterations of improvements over time. The intelligence we flatter ourselves with is only a small part of what we have evolved in our brains. For eons the problem space was all physical, but with AI, the starting point is the “intelligence” piece.
While that is partly true, it oversimplifies a few things. And these are the things that make LLMs a dead end for intelligence:
I personally think Yann LeCun (a pioneer of machine learning and AI) is on the right track with his research and his opinion that LLMs are a dead end for artificial intelligence. The reasons are more on a technical than a philosophical side.
LLMs are simply a statistics model, nothing more, nothing less
Sure, when you engage in reductionism, you can say the same thing about our brains.
You could yes, you’d be wrong, but you could say that.
No you can’t you dumbass.
You’re a charming fellow.
Reductionism is the tendency to make inappropriate reductions. This is a perfectly appropriate reduction.
The problem is that despite being a literally true statement, it relies on the implicit assumption that a system founded in statistics necessarily has limitations that preclude sentience (can generalize this to other capabilities, the same argument is often used), when whether that is the case is the entire question at issue. It’s totally meaningless if someone just feels intuitively that this must be true, that’s not a real argument.
Worse, it gestures at an idea that is demonstrably false, that this technology is equivalent in all important ways to processing inputs through a math function someone wrote, when the whole innovation here is that the training process produces novel mechanisms for solving the given problem, mechanisms no one specifically intended or predicted, or in most cases even understands. And those mechanisms can do some seriously impressive things that were previously impossible.
A good argument here would at minimum need to actually explain why this is insufficient.
Nope. We do not assume any system can produce sentience just by being complex. We have no reason to begin with that assumption so there’s no assumption to be made in dismissing it.
I am not making the claim that it can produce sentience (personally I don’t see sentience as a very meaningful concept anyway), I’m saying the arguments being made here that it cannot are deficient. Reminder of the context: https://lemmy.dbzer0.com/post/77644247/28876812
Play semantic games all you like, but nobody who actually understands how this shit works (and doesn’t have a vested interest in lying about it to the public) would suggest it’s capable of gaining sentience. It’s something you shouldn’t let people suggest as a serious possibility at this point.
Why is that? All I see is arguments misrepresenting how it works (see my earlier comment and the linked source) to try to say that it is certainly not capable of that without necessary elaboration.
You’re saying this as though it has been conclusively debunked, which backs up what I’m saying wrt what the goalposts are here: there is no such certainty. Arguing that the issue is uncertain is not a semantic game.
Edit: The crux of this is that there has been a huge amount of progress with transformer models, with constant claims that the line of where the limits are and what capabilities will never be achieved being made and disproved in recent years. Given that, I think it’s fair to say that any claims that a particular capability is forever out of reach call for a really strong justification.
Are whales sentient? Apes? Dogs? Squid? Alligators? Spiders? Fleas? What’s the test? (Yeah, the mirror test - how well do you think that really works?)
H. sapiens has more in common with fleas than LLMs - the LLMs only seem more familiar because they’re trained on our text-based interactions. Whatever sentience LLMs do or do not develop, we will be debating how to define it probably until they start reproducing themselves and force us to a sit-down negotiation over how we will divide the Earth’s resources - or maybe they won’t bother with negotiation.
So, there’s a difference between sentience and sapience, being self aware, or having consciousness.
Sentient just means the ability to experience feelings.
Pretty much all the animals you mentioned are indeed sentient.
LLMs are not conscious or self aware. They don’t have the ability to feel, so they’re not sentient. There’s an argument that they’re reaching sapience or wisdom, which is the ability to apply experience and tacit knowledge to new situations and complex questions.
Which is the issue. They can become sapient before sentient, meaning they can apply experience to new situations without the ability to use feeling as part of that. That’s pretty much the science fiction AI nightmare. A cold, unemotional, unfeeling computer that can apply logic and experience to problems.
Yet another slippery fish to try to wrestle with bare hands. What constitutes complexity? Some relatively simple looking questions turn out to be highly complex: Are there any whole-number solutions to (a^n + b^n = c^n) if (n) is greater than 2?
Then, any number of complex looking questions turn out to be quite simple with a little insight: such as any large product which has a factor that reduces to 0.
What’s simple for some people is hard for others, and I would posit that is greatly amplified when you stop comparing humans to humans and start comparing humans to animals or machines…
We put unfeeling psychopaths in charge of our largest corporations, to maximize profits - a machine might be less scary in the long run.
Well, that’s not really a philosophical question. Simple questions are ones that only deal with one question at a time and have single answers.
Complex question involve multiple implicit questions and require inference to answer and there are multiple possible correct answers.
“Are there any whole-number solutions to (a^n + b^n =c^n) where n>2” isn’t complex. That’s actually extremely easy to answer. You could just keep brute forcing through every whole number in known existence if you had the compute power. Like trying to find the next digit of pi. But that’s not Fermat’s last theorem. The actual problem is “there are no whole-number solutions to (a^n + b^n =c^n) where n>2, prove it“ which is much more complex because it requires answering multiple questions along the way. The implicit questions that arise are simple “why” questions that have to be answered.
The answer to your question is “no”. But then the follow is “well why”. Wrapping it all up in a nice equation that explains all the “well why” questions before they’re even asked is the complex part.
AI is already getting to that point. LLMs are not synonymous with AI. They’re a type of AI but there are far more models out there than just LLMs. They’re just a subset of generative AI.
“Easy” and “hard” are NOT the same thing as “simple” and “complex”.
Choose your field… two rabbis (dirty old men, of course) having agreed that abortion is a sin: debating whether it is a sin for a woman to perform feminine hygeine after rape to possibly prevent pregnancy from that rape. Where on the continuum from Monty Python’s “every sin is sacred” up to late term does abortion become a sin? Do the circumstances of the insemination come into play or not, and if so… how? What considerations are given to alienating the woman and perhaps her family from the synagogue if too hard-line of an interpretation is taught? Should the considerations be drawn up mathematically with weights and formulas if possible? Like physics, if you have a formula you can check it for correctness by putting in test values and seeing if the results are reasonable.
No, but they rhyme - so to speak. Hard problems become easier when their apparent complexities are simplified.
Well that first one is very much a complex question with no correct answer, that also requires sentience, not just sapience. Those are the questions where the disconnect becomes dangerous, because now you bring in morals, and moral questions are by definition emotional. There’s no such thing as a concrete standard of morals.
The rabbis agree that abortion is sin. Do they agree that preventing pregnancy from the start counts as abortion? Given that they’re rabbis, the answer would probably be no, because even the most Orthodox Jews believe women are allowed contraception.
But that’s not really an answer to the real question, because the real question of when does abortion become a sin requires answering the next question of when does it become abortion and not prevention? Is it the Catholic view where preventing a pregnancy is the same as abortion, the biological point where it becomes a living organism with its own DNA, or the medical view where it becomes a pregnancy when it reaches the implantation stage?
Again, that’s not a philosophical issue. That question is complex because it requires answering MANY other questions along the way to reach a final answer to the original question. It’s an extremely complex question because each of those questions that arise also have a million questions to answer to get the answer to them. And all of them being “right” or “wrong” rely on large scale social acceptance of the answers to each other questions, as well as the ultimate question of if there is even such a thing as a concrete “right” or “wrong” answer to any ethical question.
Those answers are always based on emotion. All moral and ethical questions are based on emotion because morals and ethics are emotional constructs.
That’s where AI fails. AI doesn’t have sentience. It can’t feel emotion and therefore can’t reason using emotion. It can only go by what it is trained on as concretes.
And, kinda? They kinda rhyme. There are PLENTY of things that are simple but hard, and plenty of things that are complex but easy. Like losing weight. Losing weight is extremely simple. It’s basic thermodynamics. Consume less energy than you expend, and you lose weight. Period. No exception. Now that can be extremely HARD for most people to actually do. Tying your shoes is relatively complex. It’s multiple steps that all depend on the previous step being completed properly. But it’s easy enough to do that most people can master it by the time they’re 8.
My sociology professor used to say: "if you look far enough in human societies, you will find examples of just about everything imaginable being accepted, or restricted except for incest and cannibalism, those two have clearly observable problems and generally societies will shun them eventually - with notable exceptions such as the Hapsburgs of Europe and the odd cannibal tribe in the Pacific.
For an AI to participate in such discussions, it - like a human - would need to learn the locally accepted social mores and synthesize any gaps from the available information. The final result for this and similar questions will depend heavily on the society it is learning from - uncontacted tribes in the Amazon will have different answers from tribes in equatorial West Africa, and of course different answers still from Christian, Muslim, Hindu etc. etc. (each of those having dozens to hundreds of sects…) There is no global answer.
To the extent that emotion is expressed in lexical tokens it can process, AI can process those tokens in context and “feel” to the fidelity that writers can express what emotion is and how it should / should not, does / does not impact decisions.
Knowing the physics doesn’t make self-control any easier. Losing weight is a behavioral thing and many of the factors in the behavioral choices aren’t under rational control - the decision to eat is strongly driven by all kinds of biological factors that don’t care about your reason and philosophies. I wonder what similar problems AI may develop in the future as an emergent behavior of other things we are building into it.
You’re right, in the sense that we wouldn’t know what is a path to sapience until it happens, but we can only work with what we know, and what we do know is that the transformer framework language models use is not what a brain is, and a brain is, as far as we know, the only path. The artificial neural network technology transformers are based on is rooted on a mathematical representation of an extremely simplified model of how neurons work. Human minds don’t work in tokens, and brains have chemical signals and things that also have an impact. The physical wiring is part of the story.
We’d need to objectively define what sentience/sapience is to begin with, anyway. We can’t even do that for people and animals, expanding it to machines seems an exercise in futility. It wasn’t that long ago that babies and animals were considered to be little more than automata as far as consciousness goes.
It would be amazing to be proven wrong, because the implications for philosophy and neurology would be staggering, but I digress.