I'm always kind of interested to hear what Chomsky has to say, even if I mostly disagree with him. As expected, he trots out Universal Grammar because what kind of Chomsky opinion piece would it...
I'm always kind of interested to hear what Chomsky has to say, even if I mostly disagree with him. As expected, he trots out Universal Grammar because what kind of Chomsky opinion piece would it be without at least a passing reference to it, even when it has just about zero relevance.
This caught my eye though:
The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.
Now, is that really true? It seems to me that humans are processing a pretty vast amount of information. At a gigabyte of information per hour based on typical video data rates, at 12 hours a day, a 10 year old human has processed 43 terabytes of information. That's not even counting all the other data streaming in from our various sensors -- heat, touch, balance, smell, taste, proprioception, hunger, and so on.
If I say "Hey, how's it going?" do you automatically scour every bit of information you're currently feeling or processing and pair that with every bit of data you've ever felt or sensed? Likely,...
If I say "Hey, how's it going?" do you automatically scour every bit of information you're currently feeling or processing and pair that with every bit of data you've ever felt or sensed? Likely, no. You quickly check your relation to me (am I going to listen to you brag or vent) and your current status and then decide to say "I'm just tired."
ChatGPT quickly checks Shakespeare and 5,000,000 pages of Tumblr before it tells you it's having a good morning only to browse every weather model available online to determine if it's unreasonably hot right now. I think "Month, January. Temp, t-shirt weather" and say "Boy, warm January, huh?"
Neither does ChatGPT. I see a lot of people talk about a "database", because that's how traditional software works, but there is no such thing. ChatGPT isn't checking anything. 5 billion pages of...
Neither does ChatGPT. I see a lot of people talk about a "database", because that's how traditional software works, but there is no such thing. ChatGPT isn't checking anything. 5 billion pages of Tumblr pages imprinted the statistical connection between their words in ChatGPT's weights. It does not have a copy to reference to.
Actually, I would say that's one of the main differences between ChatGPT and us. I can access my long term memory to look for things. ChatGPT doesn't; it's a straight shot inference. If I forget something, I can remember and dredge things back into short term memory. If a word connection is too far gone in ChatGPT's weights, it's outta there.
Well, kindasorta. That's the way a lot of people use GPT, but it's not the only option. Much in the same way that humans have a short term working memory, GPT can also use its context window to do...
it's a straight shot inference
Well, kindasorta. That's the way a lot of people use GPT, but it's not the only option. Much in the same way that humans have a short term working memory, GPT can also use its context window to do the same thing, and we can get GPT to actively use that working memory to do better reasoning. Generally, this is referred to as chain of thought prompting, and we're only really just scratching the surface of how this sort of technique might be used, especially when combining with external data sources that give language models essentially unbounded memory and computational abilities.
It is still "checking stuff", it's just more cleverly stored than "5 billion pages of tumblr pages". It ultimately still does come down to databases, though. that's why a lot of the bigger LLMs...
ChatGPT isn't checking anything. 5 billion pages of Tumblr pages imprinted the statistical connection between their words in ChatGPT's weights. It does not have a copy to reference to.
It is still "checking stuff", it's just more cleverly stored than "5 billion pages of tumblr pages". It ultimately still does come down to databases, though. that's why a lot of the bigger LLMs can't simply be installed on a consumer computer (outside of companies not wanting clients to reverse engineer their golden goose). It still needs massive stores of servers to properly resolve its "inferences" in an accurate (and slightly untimely, but ultimately responsive enough) matter.
. If I forget something, I can remember and dredge things back into short term memory. If a word connection is too far gone in ChatGPT's weights, it's outta there.
I'll leave the somewhat obvious disclaimer here that our recollection isn't perfect. In some ways, ChatGPT hallucinating may be the most human aspect of its operation. Sadly, market forces and culture mean that these companies would rather an LLM try to act like it knows something more often than not as opposed to admitting its folly and admitting it cannot follow.
Market forces and culture are not related to this. Hallucination is a result of how the data is stored and recalled — imperfectly, as something resembling a statistical model of language. The...
Sadly, market forces and culture mean that these companies would rather an LLM try to act like it knows something more often than not as opposed to admitting its folly and admitting it cannot follow.
Market forces and culture are not related to this. Hallucination is a result of how the data is stored and recalled — imperfectly, as something resembling a statistical model of language.
The recall process is something like predicting what the next word will be based on the previous word. For example, if the model has so far said “the Prime Minister of Canada is”, it must now estimate the most likely next word. That might be “Justin Trudeau” or “divorcing his wife”, but we have to keep in mind two things: training data and generalization.
First, training data. There is training data out there that is factually true but in a different context. For example, at one point in time the prime minister was Chrétien. There’s also training data that is factually untrue, for example fiction, or Reddit comments by idiots and liars.
Second, and more importantly, generalization. Much like humans, GPT does not encode its training data exactly, and that is a feature. By trying to find patterns that more efficiently pack the data into our and their limited memory, we perform the actual process of learning. Instead of simply storing a list of all the attributes and instances of prime ministers, we clump that information into rules — prime ministers are generally people, not rocks or planets or vegetables. They have typically western sounding names. They are mostly male. They are bilingual. And so on. If we didn’t do this, we would not be able to store the vast terabytes of training data in the much smaller model sizes, but at the expense of not being able to accurately regurgitate the exact training data. At the same time, storing the data as patterns lets us do interesting things with that data since the things we were told about the prime minister of Canada often apply to the president of the United States or the fictional ruler of Omicron Persei 8. This ability to generalize is what gives GPT its power and the inexact storage of a raw language model is inextricably bound to the ability to hallucinate — or as humans would put it, less derogatorily, to imagine.
I wouldn't say it's checking anything, nor is any cleverer (actual databases work very well for their job, and LLMs are not particularly better than them). No data is stored losslessly in a LLM's...
I wouldn't say it's checking anything, nor is any cleverer (actual databases work very well for their job, and LLMs are not particularly better than them). No data is stored losslessly in a LLM's weights; none of the connections or context in the data is stored at all, only posterior probability of likelihoods. You can't query anything.
that's why a lot of the bigger LLMs can't simply be installed on a consumer computer
That's not why. Although ChatGPT is proprietary, we know that the largest model from Meta, Llama, is roughly equivalent to chatgpt-3.5-turbo. The size of its weights is only 80gb, which is about what a modern video game is. From this you can also tell just how lossey the process from training data to weights is, when the training data numbers in the petabytes.
It's not 1:1 with a traditional database of course, but the fundamentals of it comes down to "store data, search for relevant data," and then on top of that attempt to interpret the relevant data....
I wouldn't say it's checking anything, nor is any cleverer
It's not 1:1 with a traditional database of course, but the fundamentals of it comes down to "store data, search for relevant data," and then on top of that attempt to interpret the relevant data. The data doesn't comes from navigating the natural world, so something needs to be provided, and then corrected when it inevitably gets misinformation. Or simply receive curated information to begin with to minimize corrections. That's part of why LLMs are in a legal battlefield as of now.
The size of its weights is only 80gb, which is about what a modern video game is.
modern video games are mostly textures and sounds taking up that space. And games aren't usually doing anything particulaly novel with the assets, just caching them where needed and replaying data that is cached. Add in some minor simulation for physical contact and you got a game (on a bytecode level, at least)
I imagine an LLM is storing a lot more textual data as well as compressed assets in certain products. It's doing a lot more with that data than streaming it so you need a mix of better algorithms to reduce traversal, and better hardware to speed up the traversal. That's why a game can be optimized aggressively to run at 60fps, but you can easilly take seconds to filter through a mere thousands of nodes of data naively. 80GB of such data means billions of such nodes to process.
I mean, I would really say it isn't. For one, again you're not storing data. There happens to be data that is retained, but that's a bug, not a feature - it indicates overfitting. Secondly, you...
but the fundamentals of it comes down to "store data, search for relevant data,"
I mean, I would really say it isn't. For one, again you're not storing data. There happens to be data that is retained, but that's a bug, not a feature - it indicates overfitting. Secondly, you can't search for data; given that LLMs are autoregressive probability models, you can provide tokens that you think are likely to co-correlate with pieces of information, but it's an insult to SQL to call that search.
I only mentioned video games to say that the size of the weights is not a computing bottleneck for LLMs. The amount of data in raw bytes in model weights is fairly trivial by modern multimedia standards, where blurays for movies can easily be 30 or 40gb.
Be it a database, an ancient tome, a brain, or a markov chain, we need to have some mechanism to store events, and a mechanism to recall the events. the structure and medium may be different, but...
For one, again you're not storing data.
Be it a database, an ancient tome, a brain, or a markov chain, we need to have some mechanism to store events, and a mechanism to recall the events. the structure and medium may be different, but the effect is the same. If it's not actively scraping information in real time and discarding it, it is storing some data, somehow.
I'm not really suggesting that it literally stores or scrapes data because I don't have enough knowledge of the architecture to make such a call. Plus, the underlying details of such storage isn't really of interest outside of ongoing legal battles (and even then, it's a footprint among the real issues of LLMs as of now).
I only mentioned video games to say that the size of the weights is not a computing bottleneck for LLMs.
If it's not a lookup table (since you emphasize how this is in fact not a table), it's bound to take a non-trivial amount of time to compute the result to respond with. Sure, 80GB is trvial in an age of big data that crawls through hundreds of terabytes a day in any given company. But we still have to operate within the realm of physics. a CPU only does so many operations, so if you can't brute force it, you either need to filter down and cache the data, or get more novel techniques to traverse the data with.
I'm not up to date on the research into how human memory works, but it does seem pretty clear that there are different kinds of memory. Episodic memory varies quite a lot, and it often takes quite...
I'm not up to date on the research into how human memory works, but it does seem pretty clear that there are different kinds of memory. Episodic memory varies quite a lot, and it often takes quite a lot of practice to memorize things.
Consider playing a complex song on a music instrument. It doesn't necessarily mean you have a copy of the song in your head, or at least not one you can easily access without playing it again. (Relying on "muscle memory" alone isn't reliable, but it often works fairly well.) It's not too much of a stretch to consider what LLM's do as being fairly similar.
And then, how about language itself? Children make lots of mistakes when learning to talk. It's not memorizing dictionary definitions and formal grammar.
My guess is that what an LLM does is somewhat similar to certain kinds of human memory, but something more is needed.
Whether or not the human mind is processing a lot of information, there's a deeper issue with this line of argument. Since human babies are born helpless, it's easy to think of them as a blank...
Whether or not the human mind is processing a lot of information, there's a deeper issue with this line of argument. Since human babies are born helpless, it's easy to think of them as a blank slate, but they're not. Humans have been "trained" by millions of years of evolution, the equivalent of probably exabytes worth of "data". It's difficult to come up with a perfect analogy, but I think it's more accurate to consider a baby the equivalent of a trained model, and to consider any learning that happens after birth like fine-tuning, or perhaps in-context learning (or a mixture of both).
Oh, babies are not helpless. Not entirely at least. They have access to a large arsenal of behaviors and bodily features that will, in all likelihood, trigger a deep seated program that will take...
Oh, babies are not helpless. Not entirely at least. They have access to a large arsenal of behaviors and bodily features that will, in all likelihood, trigger a deep seated program that will take over their parent's lives with an explosion of hormonal synchronicity. The program is simple and effective. Like a broken record it says, with total clarity: "keep this thing alive".
Babies are not helpless. They are the most successful parasite to ever exist.
(...)
Have I told you about MY 7-month-old son? HE'S SO CUTE I'M GOING TO EXPLODE.
Although untrained LLMs (and other networks, especially convolutional networks) also have an immense amount of structure built into them that predispose them to learning the tasks we give them....
Although untrained LLMs (and other networks, especially convolutional networks) also have an immense amount of structure built into them that predispose them to learning the tasks we give them. They aren't exactly blank slates either.
That's true. It's kindof an apples to oranges comparison. LLMs are an attempt to bootstrap intelligence off of humanity without needing to go through the whole billion-year evolution gauntlet. So...
That's true. It's kindof an apples to oranges comparison. LLMs are an attempt to bootstrap intelligence off of humanity without needing to go through the whole billion-year evolution gauntlet. So they definitely have more initial structure imposed on them than the first micro organisms (which might truly be called blank slates). Given the complexity of the human body though, I suspect that the amount of data imposed on an untrained LLM is much less than that of a baby.
Okay but have you considered MY child is the CUTEST thing in the world and NO OTHER CHILD out of 2 billion could POSSIBLY be cuter - Seriously though, yeah. I wasn't crazy about babies until mine...
Have I told you about MY 7-month-old son? HE'S SO CUTE I'M GOING TO EXPLODE.
Okay but have you considered MY child is the CUTEST thing in the world and NO OTHER CHILD out of 2 billion could POSSIBLY be cuter -
Seriously though, yeah. I wasn't crazy about babies until mine popped out, it's insane what instincts kick in.
I studied linguistics and web technology, and this was a running gag in some linguistics courses... I can't deny that he's had his moments of brilliance, though.
As expected, he trots out Universal Grammar because what kind of Chomsky opinion piece would it be without at least a passing reference to it, even when it has just about zero relevance.
I studied linguistics and web technology, and this was a running gag in some linguistics courses... I can't deny that he's had his moments of brilliance, though.
Human vision is a lot sharper around the fovea, so that might reduce the bandwidth a bit? Some VR systems take advantage of this. Also, higher frame-rate video is noticeable due to artifacts, but...
Human vision is a lot sharper around the fovea, so that might reduce the bandwidth a bit? Some VR systems take advantage of this.
Also, higher frame-rate video is noticeable due to artifacts, but we're certainly not taking it all in. Change blindness suggests there's a lot we see that we don't remember very well.
I wonder what the bandwidth of long-term memory changes in the hippocampus is?
If you want to hear a similar argument (that the human mind is much more efficient than what we have now), but from a more technical/accomplished point of view, Yan LeCun, one of the father's of...
It seems to me that humans are processing a pretty vast amount of information. At a gigabyte of information per hour based on typical video data rates, at 12 hours a day, a 10 year old human has processed 43 terabytes of information.
If you want to hear a similar argument (that the human mind is much more efficient than what we have now), but from a more technical/accomplished point of view, Yan LeCun, one of the father's of modern "AI" and Meta's chief machine learning scientist, makes a similar one in a lecture here
We are still far from uncovering the full power of the brain, but I can see this point being true in some aspects. We process a lot of information, but we don't have all that information stored in...
Now, is that really true? It seems to me that humans are processing a pretty vast amount of information. At a gigabyte of information per hour based on typical video data rates, at 12 hours a day, a 10 year old human has processed 43 terabytes of information
We are still far from uncovering the full power of the brain, but I can see this point being true in some aspects. We process a lot of information, but we don't have all that information stored in our brain, What we do have of it, most is in a sort of archived state and we'd need the right prompting to even bring it up (and like ChatGPT, we can indeed "hallucinate" with imperfect memories as we try to recall). So the brain's short term memory is pretty efficient in keeping us focused and only recalling what is necessary for that context (Or not. some store information better than others).
But in other aspects the brain is indeed doing a LOT of heavy work. Similar to a computer, most of the processor isn't usually taken up from active programs in the foreground. There's so many subconcious operations maintaining our body, and actions we optimize to take very little thought to execute (walking being the most common one, something most of us probably spent 2-3 years as toddlers striving to master). So you can make an argument that we are indeed supercomputers in some regards.
But that was all tangential. I think the main sentiment of that phrasing was simply that "we don't/can't memorize massive tables of information and recall it on the fly". Computers are extremely good at this, but are awful (relatively) at interpreting such information. Human minds are limited in processing raw information, but tend to be much better at aspects like pattern recognition, identifying subtle outliers, and connecting/correlating it to other types of data.
We give context to data, computers (even LLM's as of today) are simply attempting to imitate this with brute force of data (but slightly more elegant). In some ways it's simply because we use more than our eyes to recall and retain information; a computer can't "feel" heat on a smooth metal surface, associate it with a slight burning smell, hear the sizzle of smoke, and see a slowing red surface as a reference when trying to recreate the depiction of a stove. It can probably generate a nice looking stove regardless, but that lack of context can throw the rest of the scene off when it comes to contextualizing what else should be around a stove.
When it comes to language specifically, humans are incredibly good at acquiring incredibly sophisticated grammar from a shockingly small amount of linguistic data -- far less than any LLM is...
Now, is that really true? It seems to me that humans are processing a pretty vast amount of information. At a gigabyte of information per hour based on typical video data rates, at 12 hours a day, a 10 year old human has processed 43 terabytes of information
When it comes to language specifically, humans are incredibly good at acquiring incredibly sophisticated grammar from a shockingly small amount of linguistic data -- far less than any LLM is trained on and certainly less than anything that equates to "a gigabyte per hour". This is referred to as "poverty of stimulus" within linguistics and is a pretty old argument for the existence of a human language faculty, so I'm surprised you haven't encountered it before if you're familiar with the existence of Universal Grammar as a concept. Even linguists who hate UG don't necessarily reject the existence of a human language faculty and those that do have to address poverty of stimulus in their arguments against it.
I broadly disagree with the poverty of stimulus argument, simply because there's no real empirical evidence to support the theory. I would instead argue that language is broadly metaphorical, and...
I broadly disagree with the poverty of stimulus argument, simply because there's no real empirical evidence to support the theory. I would instead argue that language is broadly metaphorical, and the roots that underlie individual linguistic structures, especially in early language acquisition, are built on a framework that comes from a firehose of visual/tactile/sensorial data that we start receiving even before birth.
In other words,
language specifically
I don’t think this is something we can talk about meaningfully.
It would be impressive if humans learned language from the words they hear, if they only got those words in the form of disembodied text devoid of sensation, but they don’t.
However, my dismissal of UG in this context isn't about whether UG is valid or not, but rather that it is pretty useless in his argument against LLMs being unable to model intelligence because of their lack of natural grammar. Even if humans are constrained by UG, and LLMs have a broader capacity for learning rules, that isn't a bounding factor on LLMs ability to learn this subset. It's kind of a bizarrely bad argument really, akin to arguing that higher order logic can't model peano arithmetic because higher order logic is too expressive.
I broadly agree with you on the facts here. That said, I think poverty of stimulus still works with the other sources of information a human baby has to work with while learning, and many of its...
I broadly agree with you on the facts here. That said, I think poverty of stimulus still works with the other sources of information a human baby has to work with while learning, and many of its arguments rely on learning specifically grammatical structures that are more or less arbitrary, and thus the paucity of linguistic data specifically is still relevant. Whether you believe this actually requires a specific language faculty and whether if so you believe Chomsky's theories about it is of course another matter.
I do agree with you more generally about the irrelevance of UG to LLMs, though. While it's true that LLMs currently don't approach anything remotely close to general intelligence, that's largely due to failures outside of what UG would address. The fact that pure statistics can lead to a language model that is capable of so consistently producing grammatical utterances is well-established, fascinating, and completely irrelevant to UG.
I don't think that current LLMs are anywhere close to general intelligence. I also don't think they're obviously a dead end. Which is not to say that they couldn't turn out to be, merely that I...
I don't think that current LLMs are anywhere close to general intelligence. I also don't think they're obviously a dead end. Which is not to say that they couldn't turn out to be, merely that I don't think at present we can say for sure.
That preface was just to say that I'm neither an AI true believer nor an AI doomsayer.
This piece ironically suffers from the same lack of imagination that it accuses AIs of. It's deeply rooted in a particular theory of human language and human cognition, and seems to argue that this is the only pathway through which true intelligence could emerge. I find this quite absurd. We have "alien intelligences" on Earth today in the form of cephalopods whose brains are radically different from mammals, and yet they are some of the most intelligent species outside of the great apes. To posit that human-level intelligence can necessarily only emerge from the particular pathway that humanity took - and a particular theory about how that intelligence works to boot - is, to me, putting the cart before the horse. You haven't done the theoretical or practical legwork to fully understand human intelligence, and yet you want to make grand statements about the nature of all possible intelligences.
Some of the arguments made in this essay are also cheap shots that make no logical sense. The authors criticize AI for being able to learn falsehoods as easily as truths. As if humans don't "learn" untruths all the time! Sometimes humans will deny things that are readily apparent to their own senses because of the things they've been taught.
Chomsky's main contribution to his actual field of expertise, linguistics, is the theory of universal grammar. He argues that the input that babies receive is simply insufficient (the so-called poverty of stimulus argument) to determine the structure of grammar, and yet children learn to speak their native language fluently in a few years. So there must be some kind of overarching template grammar built into the brain, with the linguistic input infants are exposed to serving more as a way of fine-tuning it rather than building it from scratch. The alternative theory would be that language emerges from more general-purpose intellectual machinery in the brain, and is not, in a sense, "preinstalled" at birth.
Given this theory, we should realize that Large Language Models aren't just trying to learn as infants do. They're trying to bootstrap the entire structure built by billions of years of evolution and also learn the concrete facts that are hung onto that scaffolding. It shouldn't be surprising that we humans, who don't understand ourselves nearly well enough yet, haven't succeeded in making computers do in a few years what took nature billions. That we've even come this far is miraculous.
If we don't understand ourselves, we can't understand intelligence in general. And this piece reeks of an inflated view of our own understanding of how human intelligence works. With all due respect to Chomsky and his colleagues, universal grammar is far from a complete theory of human intelligence. And without such a complete theory, it's foolish to think we can constrain the space of all possible intelligences.
There's a thing called the "AI effect" where whenever a computer solves a problem that used to require human intellligence, and we see that the computer solution doesn't look like the human solution, the problem becomes re-classified as not being an intelligent task in the first place, and thus the successful artificial solving of it isn't artificial intelligence. I have a feeling even if, in some distant future, we have robots with fully developed human-level minds, a lot of people will still deny that they're even slightly intelligent, because their intelligence isn't an exact replica of human intelligence.
Brainless slime molds exhibit intelligent behavior! They literally have no brain! How can we possibly claim to constrain the shape of intelligence given the variety already present in nature? And that's just what currently exists, that we know of, on this one planet among billions in the universe. Never mind potential future technologies or hypothetical intelligent lifeforms on distant planets.
In short, I think this perspective is too narrow-minded and lacks creativity. It seeks to constrain intelligence to a small box with a precisely defined shape, when the truth is we don't even half understand intelligence, but from what little we do know, it doesn't fit neatly into a little cube. It could be a pyramid or a hypersphere.
That's not to say I think we're definitely on track for AGI with ChatGPT or Gemini. Merely that we shouldn't dismiss current approaches based on a preconceived notion of what intelligence must look like.
I mean... yeah, but that's kinda what you're gonna get by asking Chomsky. It's like asking Marx a question about economics -- if you didn't want it to be informed by the specific theories he...
It's deeply rooted in a particular theory of human language and human cognition
I mean... yeah, but that's kinda what you're gonna get by asking Chomsky. It's like asking Marx a question about economics -- if you didn't want it to be informed by the specific theories he pioneered, why would you even be asking him?
Well, there are some scientists and thinkers in general who welcome an interdisciplinary approach, especially when it comes to a subject as complex as artificial intelligence. Even if it's...
Well, there are some scientists and thinkers in general who welcome an interdisciplinary approach, especially when it comes to a subject as complex as artificial intelligence. Even if it's expected of him, it's still a little disappointing. At a minimum, I think properly tackling the topic of AI requires expertise from biology (to understand natural intelligence), computer science and statistics (to understand the technology and math) and linguistics (because a lot of intelligent behavior is mediated by language, although I personally don't believe all of it is).
This is an op-ed with three authors, and one of them is a tech guy. But it turns out the AI guy also has a background in Chomskyan linguistics.
I don't think that Chomsky and his theories have nothing useful to say about AI. I just think it would be a lot more insightful if he allied himself with people who had expertise in other relevant fields.
I absolutely agree that we need an interdisciplinary approach when it comes to AI. I'm not particularly convinced about your arguments for a biologist (at least when it comes to language models)...
I absolutely agree that we need an interdisciplinary approach when it comes to AI. I'm not particularly convinced about your arguments for a biologist (at least when it comes to language models) but I think insights from all fields that are related to the problem can be beneficial. I just think if you're asking Noam Chomsky about language models it's probably not going to be a shock that you get something heavily informed by his other big theoretical positions. Chomsky was hugely influential in both linguistics and computer science, so his perspective is likely to at least be interesting even if he's no expert in model architecture. Whether you agree with him is another matter, of course.
As for the other author, having a linguistics background is more common in the AI world than you might think! I made that exact pivot myself once I burnt out on academia. NLP has long been the only remotely lucrative career path for a linguist outside of academia that has anything to do with linguistics, and a lot of "data scientist" positions are essentially the same as ones that used to be for a "computational linguist". Whether that linguistics is "Chomskyan" or not isn't super relevant (and almost every major linguistics department in the US that does syntax goes for something that could be called "Chomskyan"). Which theory of syntax a linguist preferred is generally pretty irrelevant to NLP stuff, especially given the current trend is to not explicitly encode any linguistic structures and just rely on statistics and shittons of data.
While I don't disagree, I also don't think it's a very relevant response to my comment? It's like, I'm responding to his ideas, not how surprising or expected it is to hear them coming from him....
I just think if you're asking Noam Chomsky about language models it's probably not going to be a shock that you get something heavily informed by his other big theoretical positions.
While I don't disagree, I also don't think it's a very relevant response to my comment? It's like, I'm responding to his ideas, not how surprising or expected it is to hear them coming from him.
I'm not particularly convinced about your arguments for a biologist (at least when it comes to language models)
That was in the context of a general understanding of intelligence. The space of possible pathways that could produce intelligence must at least be as large as the space of naturally occurring intelligence, and that's where biology comes in.
As for the other author, having a linguistics background is more common in the AI world than you might think!
It's not that it's not a legitimate background to work in or speak about AI. It's just that I think it would have been useful to add more of an outside perspective to the mix and that's what I initially assumed when I read the credentials of the authors.
I think it's time we all realize that very few people understand the intersection of the technological side of Generative AI (or any other kind) and the social impacts of it and just stop...
I think it's time we all realize that very few people understand the intersection of the technological side of Generative AI (or any other kind) and the social impacts of it and just stop listening to anyone's opinions because they aren't coming from experts.
I'm not sure who they are. I really do think you need to be an expert on both sides to really discuss the societal impacts. If you can't predict where the technology is going and when, how are you...
I'm not sure who they are. I really do think you need to be an expert on both sides to really discuss the societal impacts. If you can't predict where the technology is going and when, how are you going to be able to predict how something as complex as all of society will react?
I don't think you need to be an expert in generative AI to discuss its social implications, if you understand it in some consistent way at some level of abstraction and are mindful of what that...
I don't think you need to be an expert in generative AI to discuss its social implications, if you understand it in some consistent way at some level of abstraction and are mindful of what that level of abstraction is.
That's the rub, though. Very few traditional academics can have a coherent model of LLMs, or autoencoding denoisers (in the case of stable diffusion). In this very piece, Chomsky makes quite a few...
if you understand it in some consistent way at some level of abstraction
That's the rub, though. Very few traditional academics can have a coherent model of LLMs, or autoencoding denoisers (in the case of stable diffusion). In this very piece, Chomsky makes quite a few important mistakes in his model of how LLMs works, and very basic ones, that even undergraduate students would know.
as a counterargument, artists have been making social commnentary about the impact of technology decades before any of the ambitions could be realized. Sci-fi authors of the mid 1900's would...
as a counterargument, artists have been making social commnentary about the impact of technology decades before any of the ambitions could be realized. Sci-fi authors of the mid 1900's would barely live to see the first personal computer, but were writing about the (usually cynical) implications of advanced tech that we ourselves in 2024 may barely live to see.
None of us know the future, but we can indeed influence the future through such conversation. I think "listening to non-experts" is worth at least that much compared to shutting down the conversation to a specific lens.
Out of curiosity, I asked ChatGPT, Write me a 3 paragraph essay critical of Noam Chomsky in the style of Noam Chomsky and title it "The False Promise of Noam Chomsky" I don't know much about...
Out of curiosity, I asked ChatGPT, Write me a 3 paragraph essay critical of Noam Chomsky in the style of Noam Chomsky and title it "The False Promise of Noam Chomsky"
I don't know much about Chomsky. Maybe those that do can tell me how it did.
The False Promise of Noam Chomsky
Noam Chomsky, a figure revered in intellectual circles for his linguistic theories and political activism, nevertheless presents a paradoxical narrative. While lauded for his critique of power structures and corporate influence, Chomsky's discourse often falls short of addressing deeper systemic issues. His emphasis on media manipulation and government propaganda neglects the underlying economic mechanisms that perpetuate inequality and exploitation. By focusing narrowly on surface-level phenomena, Chomsky inadvertently obscures the root causes of societal malaise.
Moreover, Chomsky's advocacy for libertarian socialism fails to provide a viable alternative to the prevailing capitalist framework. While his vision of decentralized, worker-controlled enterprises offers a theoretical departure from neoliberalism, it lacks pragmatic implementation strategies. Without a comprehensive blueprint for transition, Chomsky's ideas remain confined to academic discourse, failing to inspire tangible social change. Furthermore, his reluctance to engage with alternative perspectives stifles constructive dialogue, hindering the development of innovative solutions to contemporary challenges.
In conclusion, while Noam Chomsky's contributions to linguistics and political discourse are undeniable, his limitations as a social theorist warrant critical examination. His narrow focus on media critique and ideological hegemony overlooks deeper structural inequalities, while his advocacy for libertarian socialism lacks practical viability. To truly advance social justice and emancipatory politics, it is imperative to transcend the confines of Chomsky's intellectual framework and embrace a more holistic understanding of power dynamics and systemic change.
I'm always kind of interested to hear what Chomsky has to say, even if I mostly disagree with him. As expected, he trots out Universal Grammar because what kind of Chomsky opinion piece would it be without at least a passing reference to it, even when it has just about zero relevance.
This caught my eye though:
Now, is that really true? It seems to me that humans are processing a pretty vast amount of information. At a gigabyte of information per hour based on typical video data rates, at 12 hours a day, a 10 year old human has processed 43 terabytes of information. That's not even counting all the other data streaming in from our various sensors -- heat, touch, balance, smell, taste, proprioception, hunger, and so on.
If I say "Hey, how's it going?" do you automatically scour every bit of information you're currently feeling or processing and pair that with every bit of data you've ever felt or sensed? Likely, no. You quickly check your relation to me (am I going to listen to you brag or vent) and your current status and then decide to say "I'm just tired."
ChatGPT quickly checks Shakespeare and 5,000,000 pages of Tumblr before it tells you it's having a good morning only to browse every weather model available online to determine if it's unreasonably hot right now. I think "Month, January. Temp, t-shirt weather" and say "Boy, warm January, huh?"
Neither does ChatGPT. I see a lot of people talk about a "database", because that's how traditional software works, but there is no such thing. ChatGPT isn't checking anything. 5 billion pages of Tumblr pages imprinted the statistical connection between their words in ChatGPT's weights. It does not have a copy to reference to.
Actually, I would say that's one of the main differences between ChatGPT and us. I can access my long term memory to look for things. ChatGPT doesn't; it's a straight shot inference. If I forget something, I can remember and dredge things back into short term memory. If a word connection is too far gone in ChatGPT's weights, it's outta there.
Well, kindasorta. That's the way a lot of people use GPT, but it's not the only option. Much in the same way that humans have a short term working memory, GPT can also use its context window to do the same thing, and we can get GPT to actively use that working memory to do better reasoning. Generally, this is referred to as chain of thought prompting, and we're only really just scratching the surface of how this sort of technique might be used, especially when combining with external data sources that give language models essentially unbounded memory and computational abilities.
It is still "checking stuff", it's just more cleverly stored than "5 billion pages of tumblr pages". It ultimately still does come down to databases, though. that's why a lot of the bigger LLMs can't simply be installed on a consumer computer (outside of companies not wanting clients to reverse engineer their golden goose). It still needs massive stores of servers to properly resolve its "inferences" in an accurate (and slightly untimely, but ultimately responsive enough) matter.
I'll leave the somewhat obvious disclaimer here that our recollection isn't perfect. In some ways, ChatGPT hallucinating may be the most human aspect of its operation. Sadly, market forces and culture mean that these companies would rather an LLM try to act like it knows something more often than not as opposed to admitting its folly and admitting it cannot follow.
Market forces and culture are not related to this. Hallucination is a result of how the data is stored and recalled — imperfectly, as something resembling a statistical model of language.
The recall process is something like predicting what the next word will be based on the previous word. For example, if the model has so far said “the Prime Minister of Canada is”, it must now estimate the most likely next word. That might be “Justin Trudeau” or “divorcing his wife”, but we have to keep in mind two things: training data and generalization.
First, training data. There is training data out there that is factually true but in a different context. For example, at one point in time the prime minister was Chrétien. There’s also training data that is factually untrue, for example fiction, or Reddit comments by idiots and liars.
Second, and more importantly, generalization. Much like humans, GPT does not encode its training data exactly, and that is a feature. By trying to find patterns that more efficiently pack the data into our and their limited memory, we perform the actual process of learning. Instead of simply storing a list of all the attributes and instances of prime ministers, we clump that information into rules — prime ministers are generally people, not rocks or planets or vegetables. They have typically western sounding names. They are mostly male. They are bilingual. And so on. If we didn’t do this, we would not be able to store the vast terabytes of training data in the much smaller model sizes, but at the expense of not being able to accurately regurgitate the exact training data. At the same time, storing the data as patterns lets us do interesting things with that data since the things we were told about the prime minister of Canada often apply to the president of the United States or the fictional ruler of Omicron Persei 8. This ability to generalize is what gives GPT its power and the inexact storage of a raw language model is inextricably bound to the ability to hallucinate — or as humans would put it, less derogatorily, to imagine.
I wouldn't say it's checking anything, nor is any cleverer (actual databases work very well for their job, and LLMs are not particularly better than them). No data is stored losslessly in a LLM's weights; none of the connections or context in the data is stored at all, only posterior probability of likelihoods. You can't query anything.
That's not why. Although ChatGPT is proprietary, we know that the largest model from Meta, Llama, is roughly equivalent to chatgpt-3.5-turbo. The size of its weights is only 80gb, which is about what a modern video game is. From this you can also tell just how lossey the process from training data to weights is, when the training data numbers in the petabytes.
It's not 1:1 with a traditional database of course, but the fundamentals of it comes down to "store data, search for relevant data," and then on top of that attempt to interpret the relevant data. The data doesn't comes from navigating the natural world, so something needs to be provided, and then corrected when it inevitably gets misinformation. Or simply receive curated information to begin with to minimize corrections. That's part of why LLMs are in a legal battlefield as of now.
modern video games are mostly textures and sounds taking up that space. And games aren't usually doing anything particulaly novel with the assets, just caching them where needed and replaying data that is cached. Add in some minor simulation for physical contact and you got a game (on a bytecode level, at least)
I imagine an LLM is storing a lot more textual data as well as compressed assets in certain products. It's doing a lot more with that data than streaming it so you need a mix of better algorithms to reduce traversal, and better hardware to speed up the traversal. That's why a game can be optimized aggressively to run at 60fps, but you can easilly take seconds to filter through a mere thousands of nodes of data naively. 80GB of such data means billions of such nodes to process.
I mean, I would really say it isn't. For one, again you're not storing data. There happens to be data that is retained, but that's a bug, not a feature - it indicates overfitting. Secondly, you can't search for data; given that LLMs are autoregressive probability models, you can provide tokens that you think are likely to co-correlate with pieces of information, but it's an insult to SQL to call that search.
I only mentioned video games to say that the size of the weights is not a computing bottleneck for LLMs. The amount of data in raw bytes in model weights is fairly trivial by modern multimedia standards, where blurays for movies can easily be 30 or 40gb.
.
Be it a database, an ancient tome, a brain, or a markov chain, we need to have some mechanism to store events, and a mechanism to recall the events. the structure and medium may be different, but the effect is the same. If it's not actively scraping information in real time and discarding it, it is storing some data, somehow.
I'm not really suggesting that it literally stores or scrapes data because I don't have enough knowledge of the architecture to make such a call. Plus, the underlying details of such storage isn't really of interest outside of ongoing legal battles (and even then, it's a footprint among the real issues of LLMs as of now).
If it's not a lookup table (since you emphasize how this is in fact not a table), it's bound to take a non-trivial amount of time to compute the result to respond with. Sure, 80GB is trvial in an age of big data that crawls through hundreds of terabytes a day in any given company. But we still have to operate within the realm of physics. a CPU only does so many operations, so if you can't brute force it, you either need to filter down and cache the data, or get more novel techniques to traverse the data with.
I'm not up to date on the research into how human memory works, but it does seem pretty clear that there are different kinds of memory. Episodic memory varies quite a lot, and it often takes quite a lot of practice to memorize things.
Consider playing a complex song on a music instrument. It doesn't necessarily mean you have a copy of the song in your head, or at least not one you can easily access without playing it again. (Relying on "muscle memory" alone isn't reliable, but it often works fairly well.) It's not too much of a stretch to consider what LLM's do as being fairly similar.
And then, how about language itself? Children make lots of mistakes when learning to talk. It's not memorizing dictionary definitions and formal grammar.
My guess is that what an LLM does is somewhat similar to certain kinds of human memory, but something more is needed.
Whether or not the human mind is processing a lot of information, there's a deeper issue with this line of argument. Since human babies are born helpless, it's easy to think of them as a blank slate, but they're not. Humans have been "trained" by millions of years of evolution, the equivalent of probably exabytes worth of "data". It's difficult to come up with a perfect analogy, but I think it's more accurate to consider a baby the equivalent of a trained model, and to consider any learning that happens after birth like fine-tuning, or perhaps in-context learning (or a mixture of both).
Oh, babies are not helpless. Not entirely at least. They have access to a large arsenal of behaviors and bodily features that will, in all likelihood, trigger a deep seated program that will take over their parent's lives with an explosion of hormonal synchronicity. The program is simple and effective. Like a broken record it says, with total clarity: "keep this thing alive".
Babies are not helpless. They are the most successful parasite to ever exist.
(...)
Have I told you about MY 7-month-old son? HE'S SO CUTE I'M GOING TO EXPLODE.
Exactly, babies have so much behavior and knowledge from birth it's surprising people still keep comparing them to untrained neural nets.
Although untrained LLMs (and other networks, especially convolutional networks) also have an immense amount of structure built into them that predispose them to learning the tasks we give them. They aren't exactly blank slates either.
I fed this conversation to my KID-BIO-AI. He looked overwhelmed and licked the floor in desperation. Not enough training data, it seems.
That's true. It's kindof an apples to oranges comparison. LLMs are an attempt to bootstrap intelligence off of humanity without needing to go through the whole billion-year evolution gauntlet. So they definitely have more initial structure imposed on them than the first micro organisms (which might truly be called blank slates). Given the complexity of the human body though, I suspect that the amount of data imposed on an untrained LLM is much less than that of a baby.
It's too late - he's been infected! The baby mind virus has taken hold!
Okay but have you considered MY child is the CUTEST thing in the world and NO OTHER CHILD out of 2 billion could POSSIBLY be cuter -
Seriously though, yeah. I wasn't crazy about babies until mine popped out, it's insane what instincts kick in.
I studied linguistics and web technology, and this was a running gag in some linguistics courses... I can't deny that he's had his moments of brilliance, though.
Human vision is a lot sharper around the fovea, so that might reduce the bandwidth a bit? Some VR systems take advantage of this.
Also, higher frame-rate video is noticeable due to artifacts, but we're certainly not taking it all in. Change blindness suggests there's a lot we see that we don't remember very well.
I wonder what the bandwidth of long-term memory changes in the hippocampus is?
If you want to hear a similar argument (that the human mind is much more efficient than what we have now), but from a more technical/accomplished point of view, Yan LeCun, one of the father's of modern "AI" and Meta's chief machine learning scientist, makes a similar one in a lecture here
We are still far from uncovering the full power of the brain, but I can see this point being true in some aspects. We process a lot of information, but we don't have all that information stored in our brain, What we do have of it, most is in a sort of archived state and we'd need the right prompting to even bring it up (and like ChatGPT, we can indeed "hallucinate" with imperfect memories as we try to recall). So the brain's short term memory is pretty efficient in keeping us focused and only recalling what is necessary for that context (Or not. some store information better than others).
But in other aspects the brain is indeed doing a LOT of heavy work. Similar to a computer, most of the processor isn't usually taken up from active programs in the foreground. There's so many subconcious operations maintaining our body, and actions we optimize to take very little thought to execute (walking being the most common one, something most of us probably spent 2-3 years as toddlers striving to master). So you can make an argument that we are indeed supercomputers in some regards.
But that was all tangential. I think the main sentiment of that phrasing was simply that "we don't/can't memorize massive tables of information and recall it on the fly". Computers are extremely good at this, but are awful (relatively) at interpreting such information. Human minds are limited in processing raw information, but tend to be much better at aspects like pattern recognition, identifying subtle outliers, and connecting/correlating it to other types of data.
We give context to data, computers (even LLM's as of today) are simply attempting to imitate this with brute force of data (but slightly more elegant). In some ways it's simply because we use more than our eyes to recall and retain information; a computer can't "feel" heat on a smooth metal surface, associate it with a slight burning smell, hear the sizzle of smoke, and see a slowing red surface as a reference when trying to recreate the depiction of a stove. It can probably generate a nice looking stove regardless, but that lack of context can throw the rest of the scene off when it comes to contextualizing what else should be around a stove.
When it comes to language specifically, humans are incredibly good at acquiring incredibly sophisticated grammar from a shockingly small amount of linguistic data -- far less than any LLM is trained on and certainly less than anything that equates to "a gigabyte per hour". This is referred to as "poverty of stimulus" within linguistics and is a pretty old argument for the existence of a human language faculty, so I'm surprised you haven't encountered it before if you're familiar with the existence of Universal Grammar as a concept. Even linguists who hate UG don't necessarily reject the existence of a human language faculty and those that do have to address poverty of stimulus in their arguments against it.
I broadly disagree with the poverty of stimulus argument, simply because there's no real empirical evidence to support the theory. I would instead argue that language is broadly metaphorical, and the roots that underlie individual linguistic structures, especially in early language acquisition, are built on a framework that comes from a firehose of visual/tactile/sensorial data that we start receiving even before birth.
In other words,
I don’t think this is something we can talk about meaningfully.
It would be impressive if humans learned language from the words they hear, if they only got those words in the form of disembodied text devoid of sensation, but they don’t.
However, my dismissal of UG in this context isn't about whether UG is valid or not, but rather that it is pretty useless in his argument against LLMs being unable to model intelligence because of their lack of natural grammar. Even if humans are constrained by UG, and LLMs have a broader capacity for learning rules, that isn't a bounding factor on LLMs ability to learn this subset. It's kind of a bizarrely bad argument really, akin to arguing that higher order logic can't model peano arithmetic because higher order logic is too expressive.
I broadly agree with you on the facts here. That said, I think poverty of stimulus still works with the other sources of information a human baby has to work with while learning, and many of its arguments rely on learning specifically grammatical structures that are more or less arbitrary, and thus the paucity of linguistic data specifically is still relevant. Whether you believe this actually requires a specific language faculty and whether if so you believe Chomsky's theories about it is of course another matter.
I do agree with you more generally about the irrelevance of UG to LLMs, though. While it's true that LLMs currently don't approach anything remotely close to general intelligence, that's largely due to failures outside of what UG would address. The fact that pure statistics can lead to a language model that is capable of so consistently producing grammatical utterances is well-established, fascinating, and completely irrelevant to UG.
I don't think that current LLMs are anywhere close to general intelligence. I also don't think they're obviously a dead end. Which is not to say that they couldn't turn out to be, merely that I don't think at present we can say for sure.
That preface was just to say that I'm neither an AI true believer nor an AI doomsayer.
This piece ironically suffers from the same lack of imagination that it accuses AIs of. It's deeply rooted in a particular theory of human language and human cognition, and seems to argue that this is the only pathway through which true intelligence could emerge. I find this quite absurd. We have "alien intelligences" on Earth today in the form of cephalopods whose brains are radically different from mammals, and yet they are some of the most intelligent species outside of the great apes. To posit that human-level intelligence can necessarily only emerge from the particular pathway that humanity took - and a particular theory about how that intelligence works to boot - is, to me, putting the cart before the horse. You haven't done the theoretical or practical legwork to fully understand human intelligence, and yet you want to make grand statements about the nature of all possible intelligences.
Some of the arguments made in this essay are also cheap shots that make no logical sense. The authors criticize AI for being able to learn falsehoods as easily as truths. As if humans don't "learn" untruths all the time! Sometimes humans will deny things that are readily apparent to their own senses because of the things they've been taught.
Chomsky's main contribution to his actual field of expertise, linguistics, is the theory of universal grammar. He argues that the input that babies receive is simply insufficient (the so-called poverty of stimulus argument) to determine the structure of grammar, and yet children learn to speak their native language fluently in a few years. So there must be some kind of overarching template grammar built into the brain, with the linguistic input infants are exposed to serving more as a way of fine-tuning it rather than building it from scratch. The alternative theory would be that language emerges from more general-purpose intellectual machinery in the brain, and is not, in a sense, "preinstalled" at birth.
Given this theory, we should realize that Large Language Models aren't just trying to learn as infants do. They're trying to bootstrap the entire structure built by billions of years of evolution and also learn the concrete facts that are hung onto that scaffolding. It shouldn't be surprising that we humans, who don't understand ourselves nearly well enough yet, haven't succeeded in making computers do in a few years what took nature billions. That we've even come this far is miraculous.
If we don't understand ourselves, we can't understand intelligence in general. And this piece reeks of an inflated view of our own understanding of how human intelligence works. With all due respect to Chomsky and his colleagues, universal grammar is far from a complete theory of human intelligence. And without such a complete theory, it's foolish to think we can constrain the space of all possible intelligences.
There's a thing called the "AI effect" where whenever a computer solves a problem that used to require human intellligence, and we see that the computer solution doesn't look like the human solution, the problem becomes re-classified as not being an intelligent task in the first place, and thus the successful artificial solving of it isn't artificial intelligence. I have a feeling even if, in some distant future, we have robots with fully developed human-level minds, a lot of people will still deny that they're even slightly intelligent, because their intelligence isn't an exact replica of human intelligence.
Brainless slime molds exhibit intelligent behavior! They literally have no brain! How can we possibly claim to constrain the shape of intelligence given the variety already present in nature? And that's just what currently exists, that we know of, on this one planet among billions in the universe. Never mind potential future technologies or hypothetical intelligent lifeforms on distant planets.
In short, I think this perspective is too narrow-minded and lacks creativity. It seeks to constrain intelligence to a small box with a precisely defined shape, when the truth is we don't even half understand intelligence, but from what little we do know, it doesn't fit neatly into a little cube. It could be a pyramid or a hypersphere.
That's not to say I think we're definitely on track for AGI with ChatGPT or Gemini. Merely that we shouldn't dismiss current approaches based on a preconceived notion of what intelligence must look like.
I mean... yeah, but that's kinda what you're gonna get by asking Chomsky. It's like asking Marx a question about economics -- if you didn't want it to be informed by the specific theories he pioneered, why would you even be asking him?
Well, there are some scientists and thinkers in general who welcome an interdisciplinary approach, especially when it comes to a subject as complex as artificial intelligence. Even if it's expected of him, it's still a little disappointing. At a minimum, I think properly tackling the topic of AI requires expertise from biology (to understand natural intelligence), computer science and statistics (to understand the technology and math) and linguistics (because a lot of intelligent behavior is mediated by language, although I personally don't believe all of it is).
This is an op-ed with three authors, and one of them is a tech guy. But it turns out the AI guy also has a background in Chomskyan linguistics.
I don't think that Chomsky and his theories have nothing useful to say about AI. I just think it would be a lot more insightful if he allied himself with people who had expertise in other relevant fields.
I absolutely agree that we need an interdisciplinary approach when it comes to AI. I'm not particularly convinced about your arguments for a biologist (at least when it comes to language models) but I think insights from all fields that are related to the problem can be beneficial. I just think if you're asking Noam Chomsky about language models it's probably not going to be a shock that you get something heavily informed by his other big theoretical positions. Chomsky was hugely influential in both linguistics and computer science, so his perspective is likely to at least be interesting even if he's no expert in model architecture. Whether you agree with him is another matter, of course.
As for the other author, having a linguistics background is more common in the AI world than you might think! I made that exact pivot myself once I burnt out on academia. NLP has long been the only remotely lucrative career path for a linguist outside of academia that has anything to do with linguistics, and a lot of "data scientist" positions are essentially the same as ones that used to be for a "computational linguist". Whether that linguistics is "Chomskyan" or not isn't super relevant (and almost every major linguistics department in the US that does syntax goes for something that could be called "Chomskyan"). Which theory of syntax a linguist preferred is generally pretty irrelevant to NLP stuff, especially given the current trend is to not explicitly encode any linguistic structures and just rely on statistics and shittons of data.
While I don't disagree, I also don't think it's a very relevant response to my comment? It's like, I'm responding to his ideas, not how surprising or expected it is to hear them coming from him.
That was in the context of a general understanding of intelligence. The space of possible pathways that could produce intelligence must at least be as large as the space of naturally occurring intelligence, and that's where biology comes in.
It's not that it's not a legitimate background to work in or speak about AI. It's just that I think it would have been useful to add more of an outside perspective to the mix and that's what I initially assumed when I read the credentials of the authors.
I think it's time we all realize that very few people understand the intersection of the technological side of Generative AI (or any other kind) and the social impacts of it and just stop listening to anyone's opinions because they aren't coming from experts.
I'm not sure who they are. I really do think you need to be an expert on both sides to really discuss the societal impacts. If you can't predict where the technology is going and when, how are you going to be able to predict how something as complex as all of society will react?
That’s where the magic of collaboration comes in.
That's kind of my point.
I don't think you need to be an expert in generative AI to discuss its social implications, if you understand it in some consistent way at some level of abstraction and are mindful of what that level of abstraction is.
That's the rub, though. Very few traditional academics can have a coherent model of LLMs, or autoencoding denoisers (in the case of stable diffusion). In this very piece, Chomsky makes quite a few important mistakes in his model of how LLMs works, and very basic ones, that even undergraduate students would know.
as a counterargument, artists have been making social commnentary about the impact of technology decades before any of the ambitions could be realized. Sci-fi authors of the mid 1900's would barely live to see the first personal computer, but were writing about the (usually cynical) implications of advanced tech that we ourselves in 2024 may barely live to see.
None of us know the future, but we can indeed influence the future through such conversation. I think "listening to non-experts" is worth at least that much compared to shutting down the conversation to a specific lens.
Out of curiosity, I asked ChatGPT,
Write me a 3 paragraph essay critical of Noam Chomsky in the style of Noam Chomsky and title it "The False Promise of Noam Chomsky"I don't know much about Chomsky. Maybe those that do can tell me how it did.
The False Promise of Noam Chomsky
Noam Chomsky, a figure revered in intellectual circles for his linguistic theories and political activism, nevertheless presents a paradoxical narrative. While lauded for his critique of power structures and corporate influence, Chomsky's discourse often falls short of addressing deeper systemic issues. His emphasis on media manipulation and government propaganda neglects the underlying economic mechanisms that perpetuate inequality and exploitation. By focusing narrowly on surface-level phenomena, Chomsky inadvertently obscures the root causes of societal malaise.
Moreover, Chomsky's advocacy for libertarian socialism fails to provide a viable alternative to the prevailing capitalist framework. While his vision of decentralized, worker-controlled enterprises offers a theoretical departure from neoliberalism, it lacks pragmatic implementation strategies. Without a comprehensive blueprint for transition, Chomsky's ideas remain confined to academic discourse, failing to inspire tangible social change. Furthermore, his reluctance to engage with alternative perspectives stifles constructive dialogue, hindering the development of innovative solutions to contemporary challenges.
In conclusion, while Noam Chomsky's contributions to linguistics and political discourse are undeniable, his limitations as a social theorist warrant critical examination. His narrow focus on media critique and ideological hegemony overlooks deeper structural inequalities, while his advocacy for libertarian socialism lacks practical viability. To truly advance social justice and emancipatory politics, it is imperative to transcend the confines of Chomsky's intellectual framework and embrace a more holistic understanding of power dynamics and systemic change.