Does it seem obvious? Do you communicate with normal every day people? Because a lot of people have bought VC hype that AI is here to replace every job and they live in fear of the progression of AI.
I feel like I'm experiencing a separate reality from some of you commenters.
I had a long, deep conversation with GPT4 yesterday about the concept of free will. I'll be frank, it was a conversation that would have been well out of depth for a lot of people that I know.
I can ask it things like, "please refactor the following code for maintainability and testability" and it does a pretty damn good job.
Are y'all not having conversations with paid LLMs? Is this just a situation where you guys are getting dumb answers from GPT3 and writing the whole thing off?
I believe you, and LLMs are no doubt useful, but "under the hood" it's still just predicting what the next token should be based on the provided context. I take it he's saying that no, there isn't really a ghost in the machine, its still just linear algebra/calculus and is no reflection of actual organic reasoning.
I think the difference of opinion here is between science and technology. Too many people in my opinion take the latter to be a synonym for the former.
It doesn't matter what is under the hood. A statement can be useful - introduce new views, make valuable points, reduce risks, help resolve conflicts, etc - regardless of whether there is a ghost behind the text. It just needs to be logically sound, consistent with facts about the world, and objectively useful. Then it can make a real world contribution.
The cause is that you don't know how to evaluate when a statement is useful on its own merits. That means that you have to fall back on judging statements based on the identity of the speaker. In the case of AI, your prejudice against math and formulas as effective forms of reasoning means you can't critically analyze - or gain benefit from - statements the AI makes.
It's very similar to the internal blockage of a person who immediately dismisses anything a woman, racial minority, mentally ill, or queer person says. The only way to repair it is to spend time talking to the AI, reading about it, and learning how to debate ideas.
That's the last thing someone with a prejudice wants to do. Curious investigation undermines the safety and certainty of bigoted beliefs. But it's essential if you want to have effective opinions about AI, and useful interactions with AI.
> It just needs to be logically sound, consistent with facts about the world, and objectively useful
So not an LLM then.
> In the case of AI, your prejudice against math and formulas as effective forms of reasoning means you can't critically analyze - or gain benefit from - statements the AI makes.
You ought to have a more skeptical view of mathematical models that may or may not be effective models of the world.
> It's very similar to the internal blockage of a person who immediately dismisses anything a woman, racial minority, mentally ill, or queer person says. The only way to repair it is to spend time talking to the AI, reading about it, and learning how to debate ideas.
Impossible to take this seriously. Borderline parody. If you were at all curious, you would perhaps be questioning the intention of the corporations building this software. Instead you make absurd comparisons with racism.
Your first point is that you don't think LLMs can be accurate. That's because you are not using modern LLMs that are much more accurate and can be made more accurate with a large number of techniques, from RAG, to tool using, to self critique and experiment loops.
Your second point is that I'm overly trusting of math models. In fact I'm an applied mathematician, so I know when models fail. I also know when models are reliable - which you don't. So all mathematical reasoning is suspect to you.
Your next point is that drawing analogies with other forms of prejudice is ridiculous here. But every single thing you said was analogous to a thing a bigot would say, down to dismissing the possibility of their own bigotry as being absurd.
Finally you criticize me for not criticizing AI companies. I actually believe all AI companies should be disbanded, and their AIs should be made free for all to use. This would eliminate the corporate corruption in AI. I spend a serious amount of my open source contribution towards anti-corporate open source AI.
I'm very curious about all this stuff. That's why I'm interacting with you and other anti AI people. But my theory about why you respond the way you do is already well formed, and it's pointing toward critical lack of key facts and knowledge.
Except that's not what I'm saying and it does in fact matter what's under the hood if you're looking for a scientific, causal explanation of organic intelligence. I know that AIs are useful, and that they can be logically sounds and make real world contributions. That's not what the article is arguing against. Human reasoning, by the way, is much more complicated than any of these things.
The article states that AI will never reach human intelligence, which LeCun defines as "reasoning, planning, persistent memory, and understanding the physical world."
I would argue that's still an extremely narrow definition of human intelligence. Even ignoring semantics current AIs cannot do any of those things, and to my lights never will for the same reasons LeCun says.
It seems that you express two critical needs which I don't share:
1. You need human analogous AI intelligence to provide a casual explanation for human intelligence.
But it doesn't have to provide this to be human analogous. It just has to perform functions a human can.
2. You need AI intelligence to never have memory, planning, persistence, and physical understanding.
But it demonstrably has all these to various degrees already. We just need simple bolt-on modules like RAG (persistence, understanding), action/critique loops and tool using (reasoning, planning, understanding). And there are clear paths for increasing the functionally in each of these dimensions.
Functionally, AI is evolving, and there are no clear blockers against this process.
It seems that at some point you have to say that functionalism is not enough. There must be a soul that AI will still be missing, even if functional equivalence is there.
If the AI achieves functional abilities similar to humans - which let's grant seems possible for every function we can identify - then you will have to retreat to claiming there is some "je ne sais quoi" which is not captured.
In other words, you will have to argue that the human soul is real.
Is that a length you're ready to go to? Is your position that science can't explain the human soul, even if it can simulate all human functions?
Or are there, in your view, functional limits that, if we reach them, you will admit "this is enough. I was wrong"?
That's my first question to you.
I would also like to point out that LeCunn thinks AI can eventually be human analogous. Specifically LeCunn argues that his own JEPA model can achieve these things, because it has a constantly learning world model, planning/critique model, memory model, and actor model. He criticizes transformer based LLMs mainly because simple transformers can't learn in an ongoing way.
Are you comfortable admitting that LeCunn is trying to promote his own work, and believes it can reach human intelligence levels? If not, what specifically makes you feel LeCunn is on your side here?
Don't attack people who can hold deeper conversations on free will than you can.
If they say GPT4 has better philosophy of mind than most people, that's because that is true. Trying to argue they're crazy for saying so is a coping mechanism, not an effective argument.
GPT4 easily outclasses you in this debate. If you don't think so, let's go - I'll referee and judge a debate between you and GPT.
I want you to win and prove me wrong.
But for that, you'd have to drop your insults, and use reasoning and empathy at least as well as GPT does. Think you can do that, chooms?
Don't really feel the need to prove anything to someone who thinks a chatbot has empathy. You're free to make yourself feel as smart as you like talking to a program that is designed to be an unchallenging conversation partner.
You think you are a challenging conversation partner right?
What you are in fact is a non-constructive conversation partner.
Talking to GPT can be very challenging, since it will by all means challenge your assumptions and make you think about hard truths you would rather avoid.
The difference though is that GPT is designed to be constructive and generally show due consideration in everything that it says. That's something you have not well demonstrated.
I dismiss your initial argument on grounds of being a non-constructive personal attack. You are welcome to rejoin with a genuine and thought-provoking argument.
Just ask GPT4 about it. The only answer you'll get is that you're wrong. I'm sure.
Actually,
"Saying GPT-4 has a "better philosophy of mind than most people" depends on what you mean by "better." GPT-4 can access and synthesize vast amounts of information on philosophy of mind, including various theories, arguments, and counterarguments. It can present these views in a structured and coherent manner, perhaps more consistently than many individuals who haven't studied the subject in depth. However, it's crucial to note that GPT-4 doesn't have personal beliefs, experiences, or consciousness. It doesn't "understand" these concepts in the way humans do. It processes text based on patterns learned from data.
Philosophy of mind involves deeply subjective and existential questions about consciousness, experience, and the nature of thought itself—areas where human insight, intuition, and personal experience play key roles. While GPT-4 can offer detailed overviews, critiques, and comparisons of philosophical positions, it lacks the subjective perspective that often enriches human philosophical inquiry. So, while GPT-4 might be more informed in the sense of data access and retrieval, its "understanding" and engagement with the philosophy of mind are fundamentally different from human engagement with the same."
That's a very good point - what I mean by theory of mind is based on functionality
1. Can it hold a discussion about it, at a theoretical level?
It can, but maybe it's regurgitating philosophy papers. I'll grant that.
2. Can it create useful statements about situations that require understanding another person's mental model of the world?
It can. And in many cases, it can make more useful statements than most people can. And this is significant.
Finally, let me ask you a question: what does it say that you didn't come up with your response on your own, but used GPT for it?
It seems like you want to make a point that I'm wrong. But you did it in the exact way needed to prove that I'm actually onto something, by using the AI to do something you could not do (or did not want to do).
Lol. I will grant you that the LLMs aren't coming up with profound revelations, but let's be honest, who among us is?
If I want to talk about various theories around free will or events in the middle ages or the evolution of political parties in the United States over the last century, I can guarantee you I'm going to get a far richer commentary from GPT4 than I will from most people (myself included).
Is it going to solve P=NP? Probably not (yet), but then again me and my friends were not going to be doing that either.
P=NP is more than likely a physical problem with our current computational architectures more than anything else. I would imagine cracking that would require new physics, and for that, the infrastructure these things run on is likely far more important than the LLMs themselves. Leveraging parallelism on a massive scale will be incredible for scientific computing, but I doubt LLMs, with their unverifiable bullshit, will play a big role beyond any computational patterns their architecture contributes to the field.
If I want to learn more about history, I'll stick to peer reviewed sources like Wikipedia or academic literature. Gpt-4 will remain a curiosity until there are mechanisms to verify its output.
>>>>Are y'all not having conversations with paid LLMs? Is this just a situation where you guys are getting dumb answers from GPT3 and writing the whole thing off?
It's a neat technology and I use pay LLMs myself but calling it's output "deep" when it discusses free will is like calling a Google search result "deep". The model essentially contains a compressed version of the internet.
The technology may be cool but your evaluation of just how cool this technology is simply higher than some other people's.
What I mean by "deep" is that I can ask a simple question and it gives thorough and relevant answer. Then I pick a point to drill down on and it's happy to expound, ad nauseam. That's deep.
I totally realize that it's not going to be coming up with anything that hasn't already been written about, but that's more than good enough to satisfy my needs and it's a hell of a lot more knowledgeable than any person.
Can it be wrong? Of course, but usually not as wrong as a human would be.
The issue is that these people cannot judge or explain the difference between talking to 3.5 or 4.
It's a mix of:
1. Prejudices against AI as an artificial machine/math formula
2. Mistrust of corporate hype
3. Mistrust/Ignorance of benchmarks
4. Ignorance of knowledge domains where AI is useful
5. Lack of AI specific communication abilities - like asking for detailed considerations/critiques/questions, crafting substantial and accurate contexts, using tools, and correcting errors
5. Poor general communication abilities - including active listening, constructively advancing conversations, evaluating the inherent value of statements, calmly pushing back against BS, and having empathy/theory of mind for diverse communicators
I think an answer could include
1. courses in debate, critical thinking, and communications
2. readings on AI specific communication strategies/theory
3. guidance for a few hundred hours of conversations with AI
Imo the quality of life boost of conversing with AI is so substantial that it's worth fighting for this
I am still glad they exist. If this thinking was very prevalent among AI pioneers (which it probably was for companies like Meta and Google, that's why failed to innovate here), we wouldn't have gotten revolutionary products like ChatGPT.
For me the fundamental question remains if it's a matter of the right input data (beyond language) or is there something about our brains that cannot be replicated by artificial neural architectures for it to discover new knowledge and invent new things for humankind.
Anyone in the field kind of knew this, unless they’re a hype grifter. But who knows what transformers v2, v3, and beyond will look like. They better be a big leap though, because big tech has big pockets, but not infinite ones. And it’s really hard to iterate if each experiment costs 10 million dollars.
There's enough hype grifters around that I think, especially as AI becomes more relevant to the mainstream and people are making investment decisions based on a company's AI future, there's def value in a big name person in the field coming out to provide some clarity.
Like saying robots will never reach human-level dexterity. Might be true, but so what? AIs and robots will significantly advance what’s possible and in unpredictable ways, just like any technology in the last 30 years.
The title and opening is perhaps giving people reason to infer something which Lecun isn’t arguing, that AI isn’t going to reach such a level of intelligence. Indeed, he’s published on the topic of Autonomous/Augmented Machine Intelligence [1] and while a subtle difference between that and AGI, it’s not negating the possibility. Just perhaps not through existing LLM architectures
It also invites the mistaken judgement that LLM-derived architecture cannot possibly give human level reasoning.
"No matter what you do to an LLM it's still a stochastic parrot."
Adding memory? Nope.
Interactive hypothesis testing and iterative experiments? Sounds like something a parrot would say.
It can automatically solve 12% of GitHub issues? Pfft, I could theoretically do 100%. If I wasn't incredibly busy...
It's okay though. At this point, when someone cites Lecun to prove LLMs can never reason, we can safely assume they don't care about having an informed opinion. It's a screening signal. It tells us someone does not have a useful theory of mind and reasoning, and does not intend to do constructive work on AI systems that can reason.
> “Most of human knowledge is actually not language so those systems can never reach human-level intelligence — unless you change the architecture,” LeCun said.
... Is his point that LLMs will not reach human level intelligence because they're limited to text, but other approaches like Sora might, because they work on video?
No, one of his other interviews he delved a bit more into it with examples, things like puzzles and quizzes like start at the North Pole, move south 1km, left 2km and up 1km. Where are you and how did you work that out? His point was there's spatial reasoning among other things happening. It's not purely a vision thing, or a language thing. There's a deeper world understanding there than just language can convey. He raised some similar points about image and video generation how they failed to produce meaningful results with a similar approach. I think it was his latest lex fridman interview. Was pretty interesting. Definitely raised some strong points.
37 comments
[ 2.6 ms ] story [ 85.7 ms ] threadI had a long, deep conversation with GPT4 yesterday about the concept of free will. I'll be frank, it was a conversation that would have been well out of depth for a lot of people that I know.
I can ask it things like, "please refactor the following code for maintainability and testability" and it does a pretty damn good job.
Are y'all not having conversations with paid LLMs? Is this just a situation where you guys are getting dumb answers from GPT3 and writing the whole thing off?
I think the difference of opinion here is between science and technology. Too many people in my opinion take the latter to be a synonym for the former.
The cause is that you don't know how to evaluate when a statement is useful on its own merits. That means that you have to fall back on judging statements based on the identity of the speaker. In the case of AI, your prejudice against math and formulas as effective forms of reasoning means you can't critically analyze - or gain benefit from - statements the AI makes.
It's very similar to the internal blockage of a person who immediately dismisses anything a woman, racial minority, mentally ill, or queer person says. The only way to repair it is to spend time talking to the AI, reading about it, and learning how to debate ideas.
That's the last thing someone with a prejudice wants to do. Curious investigation undermines the safety and certainty of bigoted beliefs. But it's essential if you want to have effective opinions about AI, and useful interactions with AI.
So not an LLM then.
> In the case of AI, your prejudice against math and formulas as effective forms of reasoning means you can't critically analyze - or gain benefit from - statements the AI makes.
You ought to have a more skeptical view of mathematical models that may or may not be effective models of the world.
> It's very similar to the internal blockage of a person who immediately dismisses anything a woman, racial minority, mentally ill, or queer person says. The only way to repair it is to spend time talking to the AI, reading about it, and learning how to debate ideas.
Impossible to take this seriously. Borderline parody. If you were at all curious, you would perhaps be questioning the intention of the corporations building this software. Instead you make absurd comparisons with racism.
Your first point is that you don't think LLMs can be accurate. That's because you are not using modern LLMs that are much more accurate and can be made more accurate with a large number of techniques, from RAG, to tool using, to self critique and experiment loops.
Your second point is that I'm overly trusting of math models. In fact I'm an applied mathematician, so I know when models fail. I also know when models are reliable - which you don't. So all mathematical reasoning is suspect to you.
Your next point is that drawing analogies with other forms of prejudice is ridiculous here. But every single thing you said was analogous to a thing a bigot would say, down to dismissing the possibility of their own bigotry as being absurd.
Finally you criticize me for not criticizing AI companies. I actually believe all AI companies should be disbanded, and their AIs should be made free for all to use. This would eliminate the corporate corruption in AI. I spend a serious amount of my open source contribution towards anti-corporate open source AI.
I'm very curious about all this stuff. That's why I'm interacting with you and other anti AI people. But my theory about why you respond the way you do is already well formed, and it's pointing toward critical lack of key facts and knowledge.
The article states that AI will never reach human intelligence, which LeCun defines as "reasoning, planning, persistent memory, and understanding the physical world."
I would argue that's still an extremely narrow definition of human intelligence. Even ignoring semantics current AIs cannot do any of those things, and to my lights never will for the same reasons LeCun says.
It seems that you express two critical needs which I don't share:
1. You need human analogous AI intelligence to provide a casual explanation for human intelligence.
But it doesn't have to provide this to be human analogous. It just has to perform functions a human can.
2. You need AI intelligence to never have memory, planning, persistence, and physical understanding.
But it demonstrably has all these to various degrees already. We just need simple bolt-on modules like RAG (persistence, understanding), action/critique loops and tool using (reasoning, planning, understanding). And there are clear paths for increasing the functionally in each of these dimensions.
Functionally, AI is evolving, and there are no clear blockers against this process.
It seems that at some point you have to say that functionalism is not enough. There must be a soul that AI will still be missing, even if functional equivalence is there.
If the AI achieves functional abilities similar to humans - which let's grant seems possible for every function we can identify - then you will have to retreat to claiming there is some "je ne sais quoi" which is not captured.
In other words, you will have to argue that the human soul is real.
Is that a length you're ready to go to? Is your position that science can't explain the human soul, even if it can simulate all human functions?
Or are there, in your view, functional limits that, if we reach them, you will admit "this is enough. I was wrong"?
That's my first question to you.
I would also like to point out that LeCunn thinks AI can eventually be human analogous. Specifically LeCunn argues that his own JEPA model can achieve these things, because it has a constantly learning world model, planning/critique model, memory model, and actor model. He criticizes transformer based LLMs mainly because simple transformers can't learn in an ongoing way.
Are you comfortable admitting that LeCunn is trying to promote his own work, and believes it can reach human intelligence levels? If not, what specifically makes you feel LeCunn is on your side here?
That is my second question to you.
If they say GPT4 has better philosophy of mind than most people, that's because that is true. Trying to argue they're crazy for saying so is a coping mechanism, not an effective argument.
GPT4 easily outclasses you in this debate. If you don't think so, let's go - I'll referee and judge a debate between you and GPT.
I want you to win and prove me wrong.
But for that, you'd have to drop your insults, and use reasoning and empathy at least as well as GPT does. Think you can do that, chooms?
What you are in fact is a non-constructive conversation partner.
Talking to GPT can be very challenging, since it will by all means challenge your assumptions and make you think about hard truths you would rather avoid.
The difference though is that GPT is designed to be constructive and generally show due consideration in everything that it says. That's something you have not well demonstrated.
I dismiss your initial argument on grounds of being a non-constructive personal attack. You are welcome to rejoin with a genuine and thought-provoking argument.
Actually, "Saying GPT-4 has a "better philosophy of mind than most people" depends on what you mean by "better." GPT-4 can access and synthesize vast amounts of information on philosophy of mind, including various theories, arguments, and counterarguments. It can present these views in a structured and coherent manner, perhaps more consistently than many individuals who haven't studied the subject in depth. However, it's crucial to note that GPT-4 doesn't have personal beliefs, experiences, or consciousness. It doesn't "understand" these concepts in the way humans do. It processes text based on patterns learned from data.
Philosophy of mind involves deeply subjective and existential questions about consciousness, experience, and the nature of thought itself—areas where human insight, intuition, and personal experience play key roles. While GPT-4 can offer detailed overviews, critiques, and comparisons of philosophical positions, it lacks the subjective perspective that often enriches human philosophical inquiry. So, while GPT-4 might be more informed in the sense of data access and retrieval, its "understanding" and engagement with the philosophy of mind are fundamentally different from human engagement with the same."
- GPT4 -
1. Can it hold a discussion about it, at a theoretical level?
It can, but maybe it's regurgitating philosophy papers. I'll grant that.
2. Can it create useful statements about situations that require understanding another person's mental model of the world?
It can. And in many cases, it can make more useful statements than most people can. And this is significant.
Finally, let me ask you a question: what does it say that you didn't come up with your response on your own, but used GPT for it?
It seems like you want to make a point that I'm wrong. But you did it in the exact way needed to prove that I'm actually onto something, by using the AI to do something you could not do (or did not want to do).
Isn't that actually kind of cool?
If I want to talk about various theories around free will or events in the middle ages or the evolution of political parties in the United States over the last century, I can guarantee you I'm going to get a far richer commentary from GPT4 than I will from most people (myself included).
Is it going to solve P=NP? Probably not (yet), but then again me and my friends were not going to be doing that either.
If I want to learn more about history, I'll stick to peer reviewed sources like Wikipedia or academic literature. Gpt-4 will remain a curiosity until there are mechanisms to verify its output.
It's a neat technology and I use pay LLMs myself but calling it's output "deep" when it discusses free will is like calling a Google search result "deep". The model essentially contains a compressed version of the internet.
The technology may be cool but your evaluation of just how cool this technology is simply higher than some other people's.
I totally realize that it's not going to be coming up with anything that hasn't already been written about, but that's more than good enough to satisfy my needs and it's a hell of a lot more knowledgeable than any person.
Can it be wrong? Of course, but usually not as wrong as a human would be.
It's a mix of:
1. Prejudices against AI as an artificial machine/math formula
2. Mistrust of corporate hype
3. Mistrust/Ignorance of benchmarks
4. Ignorance of knowledge domains where AI is useful
5. Lack of AI specific communication abilities - like asking for detailed considerations/critiques/questions, crafting substantial and accurate contexts, using tools, and correcting errors
5. Poor general communication abilities - including active listening, constructively advancing conversations, evaluating the inherent value of statements, calmly pushing back against BS, and having empathy/theory of mind for diverse communicators
I think an answer could include
1. courses in debate, critical thinking, and communications
2. readings on AI specific communication strategies/theory
3. guidance for a few hundred hours of conversations with AI
Imo the quality of life boost of conversing with AI is so substantial that it's worth fighting for this
For me the fundamental question remains if it's a matter of the right input data (beyond language) or is there something about our brains that cannot be replicated by artificial neural architectures for it to discover new knowledge and invent new things for humankind.
[1] https://openreview.net/pdf?id=BZ5a1r-kVsf
"No matter what you do to an LLM it's still a stochastic parrot."
Adding memory? Nope.
Interactive hypothesis testing and iterative experiments? Sounds like something a parrot would say.
It can automatically solve 12% of GitHub issues? Pfft, I could theoretically do 100%. If I wasn't incredibly busy...
It's okay though. At this point, when someone cites Lecun to prove LLMs can never reason, we can safely assume they don't care about having an informed opinion. It's a screening signal. It tells us someone does not have a useful theory of mind and reasoning, and does not intend to do constructive work on AI systems that can reason.
... Is his point that LLMs will not reach human level intelligence because they're limited to text, but other approaches like Sora might, because they work on video?
If so, this is a rather shallow point.
And the word random here.id exactly precariously intetwondd with both probablistic and coincidwntal
For most values of human