Listen now | 00:49 AI's human-like, but inhuman, language skills 6:58 Bob argues that LLMs don’t vindicate the ‘blank slate’ view of the mind 18:32 Do humans and AIs acquire language in totally different ways? 30:47 Will AIs ever quit hallucinating? 39:23 The importance (or not) of “embodied cognition”
Pinker is one of my heroes so this obviously brought me a lot of joy. Was nice to see he’s been watching this stuff as closely as he has, which shouldn’t be surprising given his background.
Interesting that Pinker was very slow to understand Bob's proposal that LLMs in their pre-training evolve in a manner similar to the human brain. Pinker clearly does not agree (despite his polite acknowledgement of Bob's point at the end of the podcast). Instead, Pinker continually repeats that LLMs are just next word predictors that do statistical correlations of word frequency of occurrence to respond to questions. That's why LLMs can't tell you that Tom Cruise is the son of his mother because that information almost never occurs that way on the internet. It always says who Tom Cruise's mother is and almost never says the name of mother and then recites who her son is. No surprise here since the LLM is just looking at word frequency of occurrence something Bob can't seem to absorb but Pinker (and Gary Marcus who told Bob that repeatedly)
clearly do.
What was remarkable to hear is that Pinker seems to think, like Bob, that consciousness is epiphenomenal. And that's why neither Bob nor Pinker are even interested in the fact that LLMs are not sentient, because they don't think sentience does anything. As a result they are missing the whole key to AGI and the goal of most AI developers (and the fear of AI Doomers). Human intelligence is conscious, embodied, motivated and self-interested. Since AIs lack all of these qualities they can never produce human level intelligence because human intelligence requires all those things. Neither Bob nor Pinker seems to understand embodiment or "embodied cognition" and those concepts are key to why we can't develop AGI unless it can be made sentient. There is no evidence that it can be.
Human consciousness is largely awareness of our bodies or if you will our being. Our bodies are mortal and have many needs that must be constantly maintained. Therefore humans have interests and goals that dominant their minds. They are self motivated to keep going and to survive and thrive. AIs on the other hand have no motives, interests or goals. They don't care about anything. Human intelligence is all about feelings and caring and needs and goals. AIs cannot experience these things because they are not biological; organisms and do not experience anything. If you don't understand this concept of embodiment you really cannot understand anything about AIs and their ability or inability to think like humans. It is very strange that two such well known public intellectuals don't seem to grasp these relatively obvious points..
I had stopped listening to or reading Bob's thoughts on AI because I found his perspective so frustrating, but I listened to this podcast because I have always enjoyed Bob's past conversations with Pinker. There were portions of this discussion that I likewise enjoyed, but the fundamental problems with Bob's perspective on AI remain incredibly frustrating--I just don't understand how such a great thinker can hold such obviously wrong ideas on AI. Bob seems steadfast in his position that the pattern recognition of machine learning results in a "map of semantic meaning" that is "functionally comparable" to human understanding--despite the ever-growing evidence that LLM's fail basic tests of human understanding. Pinker made versions of this point a few times in the conversation, but did not hit it as hard as he could have. At one point, Pinker referred to the "mimicry" of AI, and that is precisely the point--LLM's are a wondrously powerful form of mimicry, but lack understanding in the sense that we have it. There are so many vivid examples of this, with basic math being a clear one--LLM's can perfectly recite mathematical principles in words, but are very poor at actually applying them. Why? Because they can recite the words based on past patterns, but they don't "understand" the words in the way that we do. Gary Marcus and others have supplied many other examples. I have yet to hear or read Bob clearly confronting this basic dichotomy between LLMs (1) perfectly explaining certain simple ideas in words and (2) erroneously applying those simple ideas in practice. How does the "map of semantic meaning" correspond to something "functionally comparable" to human understanding yet still result in this dichotomy?
LLMs raise so many fascinating questions, and as I've written in previous posts, there is no one that I would rather hear engage with those questions than Bob--I have long held Bob in the absolute highest regard as a public intellectual of the first order. But to me, those questions do not include "is machine learning a form of evolution that is reverse engineering the architecture of human understanding?" The answer to that question is no, and arguing otherwise ignores the mechanics of these systems and the empirical evidence showing that pattern recognition / reconstruction is different than human understanding.
By focusing on a mistaken, overly anthropomorphized conception of AI, Bob is missing out on opportunities to address some really interesting questions that flow from how the models actually work. I've written about some of those questions in earlier posts. But, seems like Bob is committed to the path he is taking...
I very much agree with you argument that "Human consciousness is largely awareness of our bodies". However I think you were a little hard on Steve, who gave a pretty good description of embodiment, although he didn't lean into how it is related to intentionality as much as you described. I doesn't think either of them is absolutely committed to a purely epiphenomenal view of consciousness.
Re Bob's claim that LLM training is doing the work of evolution plus learning:
While it may well be true that LLM training is doing some of the work that evolution did, I don't think it's technically correct to say that "the system for representing words as vectors" is itself learned when training transformers. Or at least it's ambiguous what that means.
Before any training at all, with a randomly initialized network, transformers still represent words (well tokens) as vectors - so "the system for representing words as vectors" is there before the transformer has seen a single bit of training data. There's not really any other way for a transformer (or a neural network in general) to represent things. It's just that during learning, these vectors evolve to take on something you could call semantic meaning - where geometric relationships between token-vectors start to correspond to semantic relationships.
I agree though it is totally possible that the model is kind of "re-learning" some generic linguistic algorithm that was discovered by evolution - I just wouldn't say it's accurate to describe it as "the system for representing words as vectors". There should be plenty of space for it to fit that - humans and chimps have about 60MB of DNA difference and GPT4 has about 100,000x that much storage available in its weights.
In Buddhism, attachment is called upādāna, which means grasping or clinging. It refers to the human tendency to cling to people, things, or ideas in the mistaken belief that they will bring us lasting happiness and fulfillment. Attachment arises from our desire to feel secure, comfortable, and in control of our lives.
I would argue that Bob is clinging to the idea of evolution. The four stages of training that Steve talks about, have very little to do with evolution except maybe in the reinforcement stage. Even that is labeled reinforcement "learning" not evolution. There are no killing of innumerable LLMs with a population of the smartest surviving. There was an approach to AI called genetic algorithms that was based on evolution, that mainly resulted in showing the limitation of an evolutionary approach. A single individual who learns through a series of thumbs up or down feedback, is not generally considered evolution. I suspect that Steve was too polite to call this out. The pre-training is pure learning, based on a high dimensional error vector, not selection.
This comment is a little late for anyone to read. But to be fair, Steve does mention Edelman, who had a theory in the 80s about neural group selection, kind of like what Bob is talking about.
According to Wikipedia:
Edelman's neural Darwinism is not Darwinian because it does not contain units of evolution as defined by John Maynard Smith. It is selectionist in that it satisfies the Price equation, but there is no mechanism in Edelman's theory that explains how information can be transferred between neuronal groups.[48] A recent theory called evolutionary neurodynamics being developed by Eors Szathmary and Chrisantha Fernando has proposed several means by which true replication may take place in the brain.
Szathmary has a talk with Terrence Deacon on YouTube where they trace evolution from life's origin to the noosphere, which Bob has mentioned in the past.
Pinker is one of my heroes so this obviously brought me a lot of joy. Was nice to see he’s been watching this stuff as closely as he has, which shouldn’t be surprising given his background.
Bob was also good
Great talk, Pinker always has interesting things to say.
This was a great conversation. Bob asked Pinker exactly the things I wanted to hear him speak on relating to language and AI.
Interesting that Pinker was very slow to understand Bob's proposal that LLMs in their pre-training evolve in a manner similar to the human brain. Pinker clearly does not agree (despite his polite acknowledgement of Bob's point at the end of the podcast). Instead, Pinker continually repeats that LLMs are just next word predictors that do statistical correlations of word frequency of occurrence to respond to questions. That's why LLMs can't tell you that Tom Cruise is the son of his mother because that information almost never occurs that way on the internet. It always says who Tom Cruise's mother is and almost never says the name of mother and then recites who her son is. No surprise here since the LLM is just looking at word frequency of occurrence something Bob can't seem to absorb but Pinker (and Gary Marcus who told Bob that repeatedly)
clearly do.
What was remarkable to hear is that Pinker seems to think, like Bob, that consciousness is epiphenomenal. And that's why neither Bob nor Pinker are even interested in the fact that LLMs are not sentient, because they don't think sentience does anything. As a result they are missing the whole key to AGI and the goal of most AI developers (and the fear of AI Doomers). Human intelligence is conscious, embodied, motivated and self-interested. Since AIs lack all of these qualities they can never produce human level intelligence because human intelligence requires all those things. Neither Bob nor Pinker seems to understand embodiment or "embodied cognition" and those concepts are key to why we can't develop AGI unless it can be made sentient. There is no evidence that it can be.
Human consciousness is largely awareness of our bodies or if you will our being. Our bodies are mortal and have many needs that must be constantly maintained. Therefore humans have interests and goals that dominant their minds. They are self motivated to keep going and to survive and thrive. AIs on the other hand have no motives, interests or goals. They don't care about anything. Human intelligence is all about feelings and caring and needs and goals. AIs cannot experience these things because they are not biological; organisms and do not experience anything. If you don't understand this concept of embodiment you really cannot understand anything about AIs and their ability or inability to think like humans. It is very strange that two such well known public intellectuals don't seem to grasp these relatively obvious points..
I had stopped listening to or reading Bob's thoughts on AI because I found his perspective so frustrating, but I listened to this podcast because I have always enjoyed Bob's past conversations with Pinker. There were portions of this discussion that I likewise enjoyed, but the fundamental problems with Bob's perspective on AI remain incredibly frustrating--I just don't understand how such a great thinker can hold such obviously wrong ideas on AI. Bob seems steadfast in his position that the pattern recognition of machine learning results in a "map of semantic meaning" that is "functionally comparable" to human understanding--despite the ever-growing evidence that LLM's fail basic tests of human understanding. Pinker made versions of this point a few times in the conversation, but did not hit it as hard as he could have. At one point, Pinker referred to the "mimicry" of AI, and that is precisely the point--LLM's are a wondrously powerful form of mimicry, but lack understanding in the sense that we have it. There are so many vivid examples of this, with basic math being a clear one--LLM's can perfectly recite mathematical principles in words, but are very poor at actually applying them. Why? Because they can recite the words based on past patterns, but they don't "understand" the words in the way that we do. Gary Marcus and others have supplied many other examples. I have yet to hear or read Bob clearly confronting this basic dichotomy between LLMs (1) perfectly explaining certain simple ideas in words and (2) erroneously applying those simple ideas in practice. How does the "map of semantic meaning" correspond to something "functionally comparable" to human understanding yet still result in this dichotomy?
LLMs raise so many fascinating questions, and as I've written in previous posts, there is no one that I would rather hear engage with those questions than Bob--I have long held Bob in the absolute highest regard as a public intellectual of the first order. But to me, those questions do not include "is machine learning a form of evolution that is reverse engineering the architecture of human understanding?" The answer to that question is no, and arguing otherwise ignores the mechanics of these systems and the empirical evidence showing that pattern recognition / reconstruction is different than human understanding.
By focusing on a mistaken, overly anthropomorphized conception of AI, Bob is missing out on opportunities to address some really interesting questions that flow from how the models actually work. I've written about some of those questions in earlier posts. But, seems like Bob is committed to the path he is taking...
Correct and well said. Bob's position on this is inexplicable.
I very much agree with you argument that "Human consciousness is largely awareness of our bodies". However I think you were a little hard on Steve, who gave a pretty good description of embodiment, although he didn't lean into how it is related to intentionality as much as you described. I doesn't think either of them is absolutely committed to a purely epiphenomenal view of consciousness.
I think you may be attributing views or lack of understanding to these two a bit liberally.
Despite this, I found your post thought provoking
Re Bob's claim that LLM training is doing the work of evolution plus learning:
While it may well be true that LLM training is doing some of the work that evolution did, I don't think it's technically correct to say that "the system for representing words as vectors" is itself learned when training transformers. Or at least it's ambiguous what that means.
Before any training at all, with a randomly initialized network, transformers still represent words (well tokens) as vectors - so "the system for representing words as vectors" is there before the transformer has seen a single bit of training data. There's not really any other way for a transformer (or a neural network in general) to represent things. It's just that during learning, these vectors evolve to take on something you could call semantic meaning - where geometric relationships between token-vectors start to correspond to semantic relationships.
I agree though it is totally possible that the model is kind of "re-learning" some generic linguistic algorithm that was discovered by evolution - I just wouldn't say it's accurate to describe it as "the system for representing words as vectors". There should be plenty of space for it to fit that - humans and chimps have about 60MB of DNA difference and GPT4 has about 100,000x that much storage available in its weights.
In Buddhism, attachment is called upādāna, which means grasping or clinging. It refers to the human tendency to cling to people, things, or ideas in the mistaken belief that they will bring us lasting happiness and fulfillment. Attachment arises from our desire to feel secure, comfortable, and in control of our lives.
I would argue that Bob is clinging to the idea of evolution. The four stages of training that Steve talks about, have very little to do with evolution except maybe in the reinforcement stage. Even that is labeled reinforcement "learning" not evolution. There are no killing of innumerable LLMs with a population of the smartest surviving. There was an approach to AI called genetic algorithms that was based on evolution, that mainly resulted in showing the limitation of an evolutionary approach. A single individual who learns through a series of thumbs up or down feedback, is not generally considered evolution. I suspect that Steve was too polite to call this out. The pre-training is pure learning, based on a high dimensional error vector, not selection.
This comment is a little late for anyone to read. But to be fair, Steve does mention Edelman, who had a theory in the 80s about neural group selection, kind of like what Bob is talking about.
According to Wikipedia:
Edelman's neural Darwinism is not Darwinian because it does not contain units of evolution as defined by John Maynard Smith. It is selectionist in that it satisfies the Price equation, but there is no mechanism in Edelman's theory that explains how information can be transferred between neuronal groups.[48] A recent theory called evolutionary neurodynamics being developed by Eors Szathmary and Chrisantha Fernando has proposed several means by which true replication may take place in the brain.
Szathmary has a talk with Terrence Deacon on YouTube where they trace evolution from life's origin to the noosphere, which Bob has mentioned in the past.