Gary: Like you, I have a great deal of respect for the scientific method. I often describe myself as a data slave, and so appreciate the point of view that you put forward in this post. I am not an expert on language, but I am wondering what you make of this recent paper which seems to "suggest that autoregressive deep language models provide a new and biologically feasible computational framework for studying the neural basis of language". Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nat Neurosci 25, 369–380 (2022).
I disagree with Chomsky about many things linguistic, but his skepticism about LLMs is correct. It is a mistake to confuse abstract, disembodied text with language, which evolved in a big-brained primate that perceived and behaved in a complex social and physical environment. As Stevan Harnad has observed, symbols ultimately must be grounded, and the symbol sequences we call language are grounded in the active engagement of human organisms with the world. See also Per Linell's underrated book, "The Written Language Bias in Linguistics."
I was thinking along the same route. There is a certain irony here. Chomsky's universal grammar theory is in a way responsible for creating the ground for NLP like systems and you can extend that line thinking and say his abstract theory made LLM possible. But even he is skeptical about this. so that tells you something.
Hi Gary, interesting question (yours), and, Chomsky's response.
Here is how to tell that LLMs are all just a house of cards. If I create 50TB of text trash, by which I mean, it's full of absurdities about the world - the ocean is blue because it's in the process of cooling down after a giant fire, for example - and use it create an LLM, then it, or another LLM, will be unable to say "I call BS". That's because an LLM has derivative intelligence at best, it has zero (0) actual understanding of a single word, eg even 'a'.
Meaning doesn't reside in words, grammar, language.
*Meaning solely comes from directly and continuously interacting with the environment, physically.*
I know what passage of time, my right hand side, night time, death, large etc mean, on account of experience. Sure, I can read up on, hear about, watch, imagine etc, the rest (eg living in the ISS) but these are extrapolations atop fundamental, lived, experience which has zero computational substitute.
As they say, you can use a calculator as a hammer but you shouldn't - NLP is a processor in a computer paradigm - a hammer with one function to process anything - not its memory, human genetic code or common sense that is language - a set of right word combinations - NLU - with two languages reflected on each other as DNA.
Maybe I'll find a way to stop sounding like a broken record someday, but real AGI systems doing important work, not putting orange balls on blue squares, will need to be legally required to be transparent enough to understand why they said or acted so. To eliminate traceability in life-dependent or even important interdependent conditions is a long, mostly secret road to tragedy and inefficiency.
Quite right, including the 'too soft' comment from Chomsky. I played with GPT-3 myself recently, as I was curious to see what it would do with famous examples of ungrammatical sentences from the linguistics literature. The result was disappointing, but it was amusing to realise how much stuff the system could make up, as Marcus mentions here, in this case a (non-existent) paper GPT-3 was attributing to Chomsky, with the title of "Propagating Activation Through Syntax" (seems to be just a concatenation of buzz words, which is itself amusing...). A transcript can be found here: https://leiterreports.typepad.com/blog/2021/10/ai-chatbot-part-2.html
I agree with many of your criticisms of GPT-3, but think it rather unfair to see the limitations of these models as a strike against Locke or (moreso) Skinner, each of whom assumed causal agents in a world of cause and effect, and the latter of whom predicated his work on unconditioned responses/stimuli.
Thanks for sharing this article with me the other day, Gary. It's a very insightful article, and I agree with everything you say about the limitations of GPT-3 or large language models at large. At the same time, I find the criticism by you and Noam Chomsky a bit harsh. It's been a while since I read the GPT-related papers, but I don't think the authors were claiming to explain why "human language [is] the way that it is."
I see it more like this: A company developed a shoe that makes humans run faster. Some people want to know more about this shoe. So, the company shares all the materials the shoe is made from along with the exact manufacturing procedure. Still, people complain: "This shoe still sucks because it doesn't explain how humans run."
i don’t know exactly whole believes what but (a) I think Noam gets asked all the time whether LLMs refute his theories about human language works and (b) that’s the question he has spent 70 years thinking about and he is entitled to consider whether LLMs help him understand why human language as it is. and answer is no.
aside from that i do think some people think LLMs must cast light on language; certainly i get that kind of reaction on twitter sometimes. i agree with you that LLM is more like a shoe than an explanation of a foot. but see LeCun’s recent tweet re language and brains and LLMs and some of the discussion after
The reality is there will be no information that you will be able to trust going forward - there is no capacity to avoid misinformation because there is no clear, simple, objective way to tell if something is misinformation that can be fed to the AI.
The entirety of AI relies on statistical / empirical methods and not analytical methods naturally. There is no way to make the AI synthesize a new, artificial language that might actually make sense.
The main argument appears to be that of the premise of time and how context can change the meaning of language but what I find interesting would be working on the ability of robustness and adaptability which I think we''re going towards where you can't deny, AI is helping, such as healthcare. Language, I think is the exception to this rule because of how it is has evolved and how it is used, we're becoming more rational but the human mind is nuanced which I think is hard to capture so instead of mimicking based on conclusions, the processes of how a conclusion should be made would need to be highlighted.
Can you elaborate on your statement: “…the kinds of artificial languages we find in computer programming and mathematics.” as I would like to understand why they are artificial against English or Mandarin. What makes them artificial?
This article is a failure of someone to be concise ... I came for the clickbait on Noam's take and you went on endlessly before finally revealing what Noam responded to you! You wasted my time with your clickbait shit! So, in full glory of the spirit of the Internet, Fuck you very much
We actually are carving up living human brains to see how language works. See Eddie Chang’s work at UCSF.
Gary: Like you, I have a great deal of respect for the scientific method. I often describe myself as a data slave, and so appreciate the point of view that you put forward in this post. I am not an expert on language, but I am wondering what you make of this recent paper which seems to "suggest that autoregressive deep language models provide a new and biologically feasible computational framework for studying the neural basis of language". Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nat Neurosci 25, 369–380 (2022).
I disagree with Chomsky about many things linguistic, but his skepticism about LLMs is correct. It is a mistake to confuse abstract, disembodied text with language, which evolved in a big-brained primate that perceived and behaved in a complex social and physical environment. As Stevan Harnad has observed, symbols ultimately must be grounded, and the symbol sequences we call language are grounded in the active engagement of human organisms with the world. See also Per Linell's underrated book, "The Written Language Bias in Linguistics."
I was thinking along the same route. There is a certain irony here. Chomsky's universal grammar theory is in a way responsible for creating the ground for NLP like systems and you can extend that line thinking and say his abstract theory made LLM possible. But even he is skeptical about this. so that tells you something.
Hi Gary, interesting question (yours), and, Chomsky's response.
Here is how to tell that LLMs are all just a house of cards. If I create 50TB of text trash, by which I mean, it's full of absurdities about the world - the ocean is blue because it's in the process of cooling down after a giant fire, for example - and use it create an LLM, then it, or another LLM, will be unable to say "I call BS". That's because an LLM has derivative intelligence at best, it has zero (0) actual understanding of a single word, eg even 'a'.
Meaning doesn't reside in words, grammar, language.
*Meaning solely comes from directly and continuously interacting with the environment, physically.*
I know what passage of time, my right hand side, night time, death, large etc mean, on account of experience. Sure, I can read up on, hear about, watch, imagine etc, the rest (eg living in the ISS) but these are extrapolations atop fundamental, lived, experience which has zero computational substitute.
Noam has been "on a roll" for decades (https://blogdredd.blogspot.com/2017/06/the-machine-religion.html).
As they say, you can use a calculator as a hammer but you shouldn't - NLP is a processor in a computer paradigm - a hammer with one function to process anything - not its memory, human genetic code or common sense that is language - a set of right word combinations - NLU - with two languages reflected on each other as DNA.
Maybe I'll find a way to stop sounding like a broken record someday, but real AGI systems doing important work, not putting orange balls on blue squares, will need to be legally required to be transparent enough to understand why they said or acted so. To eliminate traceability in life-dependent or even important interdependent conditions is a long, mostly secret road to tragedy and inefficiency.
Quite right, including the 'too soft' comment from Chomsky. I played with GPT-3 myself recently, as I was curious to see what it would do with famous examples of ungrammatical sentences from the linguistics literature. The result was disappointing, but it was amusing to realise how much stuff the system could make up, as Marcus mentions here, in this case a (non-existent) paper GPT-3 was attributing to Chomsky, with the title of "Propagating Activation Through Syntax" (seems to be just a concatenation of buzz words, which is itself amusing...). A transcript can be found here: https://leiterreports.typepad.com/blog/2021/10/ai-chatbot-part-2.html
Just read this, wish I had earlier.
The capital of London is ?????
I agree with many of your criticisms of GPT-3, but think it rather unfair to see the limitations of these models as a strike against Locke or (moreso) Skinner, each of whom assumed causal agents in a world of cause and effect, and the latter of whom predicated his work on unconditioned responses/stimuli.
Thanks for sharing this article with me the other day, Gary. It's a very insightful article, and I agree with everything you say about the limitations of GPT-3 or large language models at large. At the same time, I find the criticism by you and Noam Chomsky a bit harsh. It's been a while since I read the GPT-related papers, but I don't think the authors were claiming to explain why "human language [is] the way that it is."
I see it more like this: A company developed a shoe that makes humans run faster. Some people want to know more about this shoe. So, the company shares all the materials the shoe is made from along with the exact manufacturing procedure. Still, people complain: "This shoe still sucks because it doesn't explain how humans run."
i don’t know exactly whole believes what but (a) I think Noam gets asked all the time whether LLMs refute his theories about human language works and (b) that’s the question he has spent 70 years thinking about and he is entitled to consider whether LLMs help him understand why human language as it is. and answer is no.
aside from that i do think some people think LLMs must cast light on language; certainly i get that kind of reaction on twitter sometimes. i agree with you that LLM is more like a shoe than an explanation of a foot. but see LeCun’s recent tweet re language and brains and LLMs and some of the discussion after
The reality is there will be no information that you will be able to trust going forward - there is no capacity to avoid misinformation because there is no clear, simple, objective way to tell if something is misinformation that can be fed to the AI.
The entirety of AI relies on statistical / empirical methods and not analytical methods naturally. There is no way to make the AI synthesize a new, artificial language that might actually make sense.
The main argument appears to be that of the premise of time and how context can change the meaning of language but what I find interesting would be working on the ability of robustness and adaptability which I think we''re going towards where you can't deny, AI is helping, such as healthcare. Language, I think is the exception to this rule because of how it is has evolved and how it is used, we're becoming more rational but the human mind is nuanced which I think is hard to capture so instead of mimicking based on conclusions, the processes of how a conclusion should be made would need to be highlighted.
Can you elaborate on your statement: “…the kinds of artificial languages we find in computer programming and mathematics.” as I would like to understand why they are artificial against English or Mandarin. What makes them artificial?
This article is a failure of someone to be concise ... I came for the clickbait on Noam's take and you went on endlessly before finally revealing what Noam responded to you! You wasted my time with your clickbait shit! So, in full glory of the spirit of the Internet, Fuck you very much