This essay frequently uses the word "insight", and its primary topic is whether an empirically fitted statistical model can provide that (with Norvig arguing for yes, in my opinion convincingly). How does that differ from your concept of a "cause"?
> I agree that it can be difficult to make sense of a model containing billions of parameters. Certainly a human can't understand such a model by inspecting the values of each parameter individually. But one can gain insight by examing (sic) the properties of the model—where it succeeds and fails, how well it learns as a function of data, etc.
Unfortunately, studying the behavior of a system doesn't necessarily provide insight into why it behaves that way; it may not even provide a good predictive model.
Norvig's textbook surely appears on the bookshelf of researchers including those building current top LLMs. So it's odd to say that such an approach "may not even provide a good predictive model". As of today, it is unquestionably the best known predictive model for natural language, by huge margin. I don't think that's for lack of trying, with billions of dollars or more at stake.
Whether that model provides "insight" (or a "cause"; I still don't know if that's supposed to mean something different) is a deeper question, and e.g. the topic of countless papers trying to make sense of LLM activations. I don't think the answer is obvious, but I found Norvig's discussion to be thoughtful. I'm surprised to see it viewed so negatively here, dismissed with no engagement with his specific arguments and examples.
In his view - most ML algos are at level 1 - they look at data and draw associations, and "agents" have started some steps in level 2 - doing.
The smartest of humans operate mostly in level (3) of abstractions - where they see things, gain experience, and later build up a "strong causal model" of the world and become capable of answering "what if" questions.
Chomsky's talking about predictive models in the context of cognitive science. LLMs aren't really a predictive model of any aspect of human cognitive function.
The generation of natural language is an aspect of human cognition, and I'm not aware of any better model for that than current statistical LLMs. The papers mapping between EEG/fMRI/etc. and LLM activations have been generally oversold so far, but it's active area of research for good reason.
I'm not saying LLMs are a particularly good model, just that everything else is currently worse. This includes Chomsky's formal grammars, which fail to capture the ways humans actually use language per Norvig's many examples. Do you disagree? If so, what model is better and why?
I’m not really sure what you’re getting at. Could you point to some papers exemplifying the kind of work that you’re thinking of? Of course there are lots of people training LLMs and other statistical models on EEG data, but that does not show that, say, GPT-5, is a good model of any aspect of human cognition.
Chomsky, of course, never attempted to model the generation of natural language and was interested in a different set of problems, so LLMs are not really a competitor in that sense anyway (even if you take the dubious step of accepting them as scientific models).
I certainly don’t agree with Norvig, but he doesn’t really understand the basics of what Chomsky is trying to do, so there is not much to respond to. To give three specific examples, he (i) is confused in thinking that Gold’s theorem has anything to do with Chomsky’s arguments, (ii) appears to think that Chomsky studied the “generation of language” (because he he’s read so little of Chomsky’s work that he doesn’t know what a “generative grammar” is), and (iii) believes that Chomsky thinks that natural languages are formal languages in which every possible sentence is either in the language or not (again because he’s barely read anything that Chomsky wrote since the 1950s). Then, just to make absolutely sure not to be taken seriously, he compares Chomsky to Bill O’Reilly!
This comment and GP comment are why the word "causal model" is needed. LLMs are predictive* models of human language, but they are not causal models of language.
If you believe that some of human cognition is linguistic (even if e.g. inner monologue and spoken language are just the surface of deeper more unconscious processes), then, yes, we might say LLMs can predictively model some aspects of human cognition, but, again, they are certainly not causal models, and they are not predictive models of human cognition generally (as cognition is clearly far, far more than linguistic).
* I avoid calling LLMs "statistical" because they really aren't even that. They are not calibrated, and including a softmax and log-loss in things doesn't magically make your model statistical (especially since ad-hoc regularization methods, other loss functions and simplex mappings, e.g. sparsemax, often work better and then violate the assumptions that are needed to prove these things are behaving statistically). LLMs really are more accurately just doing (very, very fancy and impressive) curve/manifold-fitting.
They are not predictive models in the domains Chomsky investigated. LLMs make no predictions about, say, when non-surface quantifier scope should or should not be possible, or what should or shouldn’t be a wh-island. They are predictive in a sense that’s largely irrelevant to cognitive science. (Trying to guess what words might come after some other words isn’t a problem in cognitive science.)
"What should or shouldn’t be a wh-island" is literally a statement of "what words might come after some other words"! An LLM encodes billions of such statements, just unfortunately in a quantity and form that makes them incomprehensible to an unaided human. That part is strictly worse; but the LLM's statements model language well enough to generate it, and that part is strictly better.
As I read Norvig's essay, it's about that tradeoff, of whether a simple and comprehensible but inaccurate model shows more promise than a model that's incomprehensible except in statistical terms with the aid of a computer, but far more accurate. I understand there's a large group of people who think Norvig is wrong or incoherent; but when those people have no accomplishments except within the framework they themselves have constructed, what am I supposed to think?
Beyond that, if I have a model that tells me whether a sentence is valid, then I can always try different words until I find one that makes it valid. Any sufficiently good model is thus capable of generation. Chomsky never proposed anything capable of that; but that just means his models were bad, not that he was working on a different task.
As to the relationship between signals from biological neurons and ANN activations, I mean something like the paper linked below, whose authors write:
> Thus, even though the goal of contemporary AI is to improve model performance and not necessarily to build models of brain processing, this endeavor appears to be rapidly converging on architectures that might capture key aspects of language processing in the human mind and brain.
I emphasize again that I believe these results have been oversold in the popular press, but the idea that an ANN trained on brain output (including written language) might provide insight into the physical, causal structure of the brain is pretty mainstream now.
> What should or shouldn’t be a wh-island" is literally a statement of "what words might come after some other words"!
This gets at the nub of the misunderstanding. Chomsky is interested in modeling the range of grammatical structures and associated interpretations possible in natural languages. The wh-island condition is a universal structural constraint that only indirectly (and only sometimes) has implications for which sequences of words are ‘valid’ in a particular language.
LLMs make no prediction at all as to whether or not natural languages should have wh-islands: they’ll happily learn languages with or without such constraints.
If you want a more concrete example of why wh-islands can’t be understood in terms of permissible or impermissible sequences of words, consider cases like
How often did you ask why John took out the trash?
The wh-island created by ‘why’ removes one of the in-principle possible interpretations (the embedded question reading where ‘how often’ associates with ‘took’), but the sequence of words is fine.
> Chomsky never proposed anything capable of that; but that just means his models were bad, not that he was working on a different task.
No, Chomsky really was working on a different task: a solution to the logical problem of language acquisition and a theory of the range of possible grammatical variation across human languages. There is no reason to think that a perfect theory in this domain would be of any particular help in generating plausible-looking text. From a cognitive point of view, text generation rather obviously involves the contribution of many non-linguistic cognitive systems which are not modeled (nor intended to be modeled) by a generative grammar.
>the paper linked below
This paper doesn’t make any claims that are obviously incompatible with anything that Chomsky has said. The fundamental finding is unsurprising: brains are sensitive to surprisal. The better your language model is at modeling whether or not a sequence of words is likely, the better you can predict the brain’s surprisal reactions. There are no implications for cognitive architecture. This ought to be clear from that fact that a number of different neural net architectures are able to achieve a good degree of success, according to the paper’s own lights.
> LLMs make no prediction at all as to whether or not natural languages should have wh-islands: they’ll happily learn languages with or without such constraints.
The human-designed architecture of an LLM makes no such prediction; but after training, the overall system including the learned weights absolutely does, or else it couldn't generate valid language. If you'd prefer to run in the opposite direction, then you can feed in sentences with correct and incorrect wh-movement, and you'll find the incorrect ones are much less probable.
That prediction is commingled with billions of other predictions, which collectively model natural language better than any machine ever constructed before. It seems like you're discounting it because it wasn't made by and can't be understood by an unaided human; but it's not like the physicists at the LHC are analyzing with paper and pencil, right?
> There is no reason to think that a perfect theory in this domain would be of any particular help in generating plausible-looking text.
Imagine that claim in human form--I'm an expert in the structure of the Japanese language, but I'm unable to hold a basic conversation. Would you not feel some doubt? So why aren't you doubting the model here? Of course it would have been outlandish to expect that of a model five years ago, but it isn't today.
I see your statement that Chomsky isn't attempting to model the "many non-linguistic cognitive systems", but those don't seem to cause the LLM any trouble. The statistical modelers have solved problem after problem that was previously considered impossible, and the practical applications of that are (for better or mostly worse) reshaping major aspects of society. Meanwhile, every conversation I've had with a Chomsky supporter seems to reduce to "he is deliberately choosing not to produce any result evaluable by a person who hasn't spent years studying his theories". I guess that's true, but that mostly just makes me regret what time I've already spent.
> The human-designed architecture of an LLM makes no such prediction; but after training, the overall system including the learned weights absolutely does, or else it couldn't generate valid language.
It makes a prediction about whatever language(s) are in the training data, but it doesn’t make any (substantial) predictions about general constraints on human languages. It really seems that you’re missing the absolutely fundamental goal of Chomsky’s research program here. Remember that whole “universal grammar” thingy?
> -I'm an expert in the structure of the Japanese language, but I'm unable to hold a basic conversation. Would you not feel some doubt?
I expect anyone learning Japanese as a second language will get a chuckle out of this one. It’s in fact a common scenario. You can learn a lot about the grammar of a language, but conversation requires the ability to use that knowledge immediately and fluidly in a wide variety of situations. It is like the difference between “knowing how to solve a differential equation” and being able to answer 50 questions within an hour in a physics exam.
> I see your statement that Chomsky isn't attempting to model the "many non-linguistic cognitive systems", but those don't seem to cause the LLM any trouble.
Of course they don’t, because researchers creating LLMs are (in the vast majority of cases) not attempting to model any particular cognitive system; they have engineering goals, not scientific ones. You seem to be stuck in the view that Chomsky is somehow trying and completely failing to do the thing that LLMs do successfully. This certainly makes for a good straw man (if Chomsky had the same goals, then yeah, he never got anywhere), but it’s a misunderstanding of his research program.
> "he is deliberately choosing not to produce any result evaluable by a person who hasn't spent years studying his theories"
You could say this of many perfectly respectable fields. Andrew Wiles has not produced any result evaluable by me or by almost anyone else. It would certainly take me a lot more than “a few years” of study to evaluate his work.
I’m afraid there are no intellectual shortcuts. If you want to evaluate Chomsky’s work, you will have to at least read it, and maybe even think about it a bit too! It seems a bit churlish to whine about that. All you are being deprived of by opting out of this time investment is the opportunity to make informed criticisms of his work on the internet.
(The good news is that generative linguistics is actually pretty accessible, and one year of part time study would probably be enough to get the lay of the land.)
> Andrew Wiles has not produced any result evaluable by me or by almost anyone else.
Fermat wrote the theorem in the margin long before Wiles was born. There is no question that many people tried and failed to prove it. There is no question that Wiles succeeded, because the skill required to verify a proof is much less than the skill required to generate it. I haven't done so myself; but lots of other people have, and there is no dispute by any skilled person that his proof is correct. So I believe that Wiles has accomplished something significant.
I don't think Chomsky has any similar accomplishment. I roughly understand the grandiose final goal; I just see no evidence that he has made any progress towards it. Everything that I'd see as an interesting intermediate goal is dismissed as out of scope, especially when others achieve it. On the rare occasion that Chomsky has made externally intelligible predictions on the range of human language, they've been falsified anthropologically. I assume you followed the dispute on Pirahã, which I believe clarified that features like recursion were in fact optional, rendering the theory safely non-falsifiable again.
So what's his progress? Everything that I see turns inward, valuable only within the framework that he himself constructed. Anyone can build such a framework, so that's not an accomplishment. Convincing others to spend years of their lives on that framework is a sort of an achievement, but it's not a scientific one--homeopathy has many practitioners.
> I expect anyone learning Japanese as a second language will get a chuckle out of this one. It’s in fact a common scenario.
I think this view is just as wrong applied to a human as to a model. A beginning language student probably knows a lot more grammar rules than a native speaker, but their inability to converse doesn't come from their inability to quickly apply them. It comes from the fact that those rules capture only a small amount of the structure of natural language. You seem to acknowledge this yourself--if nothing Chomsky is working on would help a machine generate language, then it wouldn't help a human either. This also explains my teachers' usual advice to stop studying and converse as best I could, watch movies, etc.
Humans clearly learn language in a more structured way than LLMs do (since they don't need trillions of tokens), but they learn primarily from exposure, with partial structure but many exceptions. I don't think that's surprising, since most other things "designed" in an evolutionary manner have that same messy form. LLMs have succeeded spectacularly in modeling that, taking the usual definition in ML or other math for "modeling".
It's thus strange to me to see them dismissed as a source of insight into natural language. I guess most experts in LLMs are busy becoming billionaires right now; but if anything resembling Chomsky's universal grammar ever does get found to exist, then I'd guess it will be extracted computationally from models trained on corpora of different languages and not any human insight, in the same way that the Big Five personality traits fall out of a PCA.
> So what's his progress? Everything that I see turns inward, valuable only within the framework that he himself constructed.
It's really not true that the whole of generative linguistics is just some kind of self-referential parlor game. A lot of what we take for granted today as legitimate avenues of research in cognitive science were opened up as a direct consequence of Chomsky's critique of behaviorism and his insight that the mind is best understood as a computational system. Ironically, any respectable LLM will be perfectly happy to cover this in more detail if you probe it with some key terms like "behaviorism", "cognitive revolution" or "computational theory of mind".
> Pirahã
It's very unlikely that Everett's key claims about Pirahã are true (see e.g. https://dspace.mit.edu/bitstream/handle/1721.1/94631/Nevins-...). But anyway, the universality of recursive clausal embedding has never been a central issue in generative linguistics. Chomsky co-authored one speculative paper late in his career suggesting that recursion in some (vague) sense might be the core computational innovation responsible for the human language faculty. Everett latched on to that claim and the dispute went public, which has given a false impression of its overall centrality to the field.
> So what's his progress?
I don't see how we can discuss this question without getting into specifics, so let me try to push things in that direction. Here is a famous syntax paper by Chomsky: https://babel.ucsc.edu/~hank/On_WH-Movement.pdf It claims to achieve various things. Do you disagree, and if so, why?
> Japanese
A generative linguist studying Japanese wouldn't claim to be an expert on the structure of Japanese in your broad sense of the term. One thing to bear in mind is that generative linguistics is entirely opportunistic in its approach to individual languages. Generative linguists don't don't study Japanese because they give a fuck about Japanese as such (any more than physicists study balls rolling down inclined planes because balls and inclined planes are intrinsically fascinating). The aim is just to find data to distinguish competing hypotheses about the human language faculty, not to come to some kind of total understanding of Japanese (or whatever language).
> I guess most experts in LLMs are busy becoming billionaires right now; but if anything resembling Chomsky's universal grammar ever does get found to exist, then I'd guess it will be extracted computationally from models trained on corpora of different languages and not any human insight, in the same way that the Big Five personality traits fall out of a PCA.
This is a common pattern of argumentation. First, Chomsky's work is critically examined according to the highest possible scientific standards (every hypothesis must be strictly falsifiable, etc. etc.) Then when we finally get to see the concrete alternative proposal, it turns out to be nothing more than a promissory note.
> It's very unlikely that Everett's key claims about Pirahã are true
Everett achieved something unequivocally difficult--after twenty years of failed attempts by other missionaries, he was the first Westerner to learn Pirahã, living among the people and conversing with them in their language. In my view, that gives him significantly greater credibility than academics with no practical exposure to the language (and I assume you're aware of his response to the paper you linked).
I understand that to Chomsky's followers, Everett's achievement is meaningless, in the same way that LLMs saturating almost every prior benchmark in NLP is meaningless. But what achievements outside the "self-referential parlor game" are meaningful then? You must need something to ground yourself in outside reality, right?
> Then when we finally get to see the concrete alternative proposal, it turns out to be nothing more than a promissory note.
I'm certainly not claiming that statistical modeling has already achieved any significant insight into how physical structures in the brain map to an ability to generate language, and I don't think anyone else is either. We're just speculating that it might in future.
That seems a lot less grandiose to me than anything Chomsky has promised. In the present, that statistical modeling has delivered some pretty significant, strictly falsifiable, different but related achievements. Again, what does Chomsky's side have?
> I don't see how we can discuss this question without getting into specifics, so let me try to push things in that direction. Here is a famous syntax paper by Chomsky: https://babel.ucsc.edu/~hank/On_WH-Movement.pdf
And when I asked that before, you linked a sixty-page paper, with no further indication ("various things"?) of what you want to talk about. If you're trying to argue that Chomsky's theories are anything but a tarpit for a certain kind of intellectual curiosity, then I don't think that's helping.
Believe Everett if you want to, but it doesn’t make much difference to anything. Not every language has to exploit the option of recursive clausal embedding. The implications for generative linguistics are pretty minor. Yes, Everett responded to the paper I linked, and then there were further papers in the chain of responses (e.g. http://lingphil.mit.edu/papers/pesetsk/Nevins_Pesetsky_Rodri...).
> And when I asked that before, you linked a sixty-page paper, with no further indication ("various things"?) of what you want to talk about.
I was suggesting that we talk about the central claim of the paper (i.e. that the answer to question (50) is ‘yes’).
I don’t see how it’s reasonable to ask for something smaller than a paper if you want evidence that Chomsky’s research program has achieved some insight. That’s the space required to argue for a particular viewpoint rather than just state it.
In other words, if I concisely summarize Chomsky’s findings you’ll just dismiss them as bogus, and if I link to a paper arguing for a particular result, you’ll say it’s too long to read. So, essentially, you have decided not to engage with Chomsky’s work. That is a perfectly legitimate thing to do, but it does mean that you cannot make informed criticisms of it.
> So, essentially, you have decided not to engage with Chomsky’s work. That is a perfectly legitimate thing to do, but it does mean that you cannot make informed criticisms of it.
Any criticism that I'd make of homeopathy would be uninformed by the standards of a homeopath--I don't know which poison to use, or how many times to strike the bottle while I'm diluting it, or whatever else they think is important. But to their credit they're often willing to put their ideas to the external test (like with an RCT), and I know that evidence in aggregate shows no benefit. I'm therefore comfortable criticizing homeopathy despite my unfamiliarity with its internals.
I don't claim any qualifications to criticize the internals of Chomsky's linguistics, but I do feel qualified to observe the whole thing appears to be externally useless. It seems to reject the idea of falsifiable predictions entirely, and if one does get made and then falsified then "the implications for generative linguistics are pretty minor". After dominating academic linguistics for fifty years, it has never accomplished anything considered difficult outside the newly-created field. So why is this a place where society should expend more of its finite resources?
Hardy wrote his "Mathematician's Apology" to answer the corresponding question for his more ancient field, explicitly acknowledging the uselessness of many subfields but still defending them. He did that with a certain unease though, and his promises of uselessness also turned out to be mistaken--he repeatedly took number theory as his example, not knowing that in thirty years it would underly modern cryptography. Chomsky's linguists seem to me like the opposite of that, shouting down anyone who questions them (he called Everett a "charlatan") while proudly delivering nothing to the society funding their work. So why would I want to join them?
>but I do feel qualified to observe the whole thing [Chomskyian linguistics] appears to be externally useless
Sure, Chomsky's work doesn't have practical applications. Most scientific work doesn't. It's just that, for obvious reasons, you tend to hear more about the work that does. You mention number theory. Number theory had existed for a lot longer than Chomskyan linguistics has now when Hardy chose it as an example of a field with no practical applications.
> seems to reject the idea of falsifiable predictions entirely,
As a former syntactician who's constructed lots of theories that turned out to be false, I can't really relate to this one. If you look through the generative linguistics literature you can find innumerable instances of promising ideas rejected on empirical grounds. Chomsky himself has revised or rejected his earlier work many times. A concrete example would be the theory of parasitic gaps presented in Concepts and Consequences (quickly falsified by the observation that parasitic gap dependencies are subject to island constraints).
The irony here is that generative syntax is actually a field with a brutal peer review culture and extremely high standards of publication. Actual syntax papers are full of detailed empirical argumentation. Here is one relatively short and accessible example chosen at random: http://www.skase.sk/Volumes/JTL03/04.pdf
>After dominating academic linguistics for fifty years, it has never accomplished anything considered difficult outside the newly-created field
What does this even mean? Has geology accomplished something considered difficult outside of geology? I don't really understand what standard you are trying to apply here.
> Sure, Chomsky's work doesn't have practical applications. Most scientific work doesn't.
> Has geology accomplished something considered difficult outside of geology?
Ask an oilfield services company? A structural engineer who needs a foundation? If that work were easy, then their geologists wouldn't get paid.
I could have just said "economically important", but that seemed too limiting to me. For example, computer-aided proofs were a controversial subfield of math, but I'd take their success on the four-color theorem (which came from outside their subfield and had resisted proof by other means) as evidence of their value, despite the lack of practical application for the result. I think that broader kind of success could justify further investment, but I also don't see that here.
> As a former syntactician who's constructed lots of theories that turned out to be false
I should clarify that I do see a concept of falsifiability at that level, of whether a grammar fits a set of examples of a language. That seems pretty close to math or CS to me. I don't see how that small number of examples is supposed to scale to an entire natural language or to anything about the human brain's capability for language, and I don't see any falsifiable attempt to make that connection. (I don't see much progress towards the loftiest goals from the statistical approach either, but their spectacular engineering results break that tie for me.)
Anyways, Merry Christmas if you're celebrating. I guess we're unlikely to be the ones to settle this dispute, but I appreciate the insight into the worldview.
I am not arguing that people should be paid public money to do Chomskyan linguistics. That is an entirely separate question from the question of whether or not Chomsky's key claims are true and whether his research program has made progress. Again, you will have to throw out the majority of science if you hold to the criterion that only work with practical applications has any value.
I also think that you continue to underestimate Chomsky's overall influence on cognitive science. If you think that post-cognitive-revolution cognitive science has achieved anything of note, then you ought to give Chomsky partial credit for that.
>I don't see how that small number of examples is supposed to scale to an entire natural language
Wide coverage generative grammars certainly exist, though they were never something that Chomsky himself was interested in. Here is one in a Chomskyan idiom: https://aclanthology.org/P19-1238.pdf
I'm still puzzled by your point about falsifiability. I haven't seen anything close to a falsifiable claim from people who are excited about the cognitive implications of LLMs. The argument is little more than "look at the cool stuff these things can do – surely brains must work a bit like this too!" Read almost anything by Chomsky and you'll find it's full of quite specific claims that can be empirically tested. I guess people get excited about the fact that the architecture of LLMs is superficially brain-like, but it's doubtful that this gets us any closer to an understanding of the relevant computations at the neural level.
Also, in case you missed the recent big thread, fMRI has taught us almost nothing due to its serious limitations and various measurement and design issues in the field. IMO it is way too slow and clunky to ever yield insights into something as fast as linguistic thought.
Thanks for the response, but (per the omitted portion of my sentence before the semicolon) I was not talking about the M in LLM. I was talking about a conceptual or analytic model that a human might develop to try to predict the behavior of an LLM, per Norvig's claim of insight derived from behavioral observation.
But now that I think a bit about it, the observation that an LLM seems to frequently produce obviously and/or subtly incorrect output, is not robust to prompt rewording, etc. is perhaps a useful Norvig-style insight.
> I'm surprised to see it viewed so negatively here, dismissed with no engagement with his specific arguments and examples.
I struggle to motivate engaging with it because it is unfortunately quite out of touch with (or just ignores) some core issues and the major advances in causal modeling and causal modeling theory, i.e. Judea Pearl and do-calculus, structural equation modeling, counterfactuals, etc [1].
It also, IMO, makes a (highly idiosyncratic) distinction between "statistical" (meaning, trained / fitted to data) and "probabilistic" models, that doesn't really hold up too well.
I.e. probabilistic models in quantum physics are "fit" too, in that the values of fundamental constants are determined by experimental data, but these "statistical" models are clearly causal models regardless. Even most quantum physical models can be argued to be causal, just the causality is probabilistic rather than absolute (i.e. A ==> B is fuzzy implication rather than absolute implication). It's only if you ask deliberately broad ontological questions (e.g. "Does the wave function cause X") that you actually run into the problem of quantum models being causal or not, but for most quantum physical experiments and phenomena generally, the models are still definitely causal at the level of the particles / waves / fields involved.
IMO I don't want to engage much with the arguments because it starts on the wrong foot and begins by making, in my opinion, an incoherent / unsound distinction, while also ignoring or just being out of date with the actual scientific and philosophical progress and issues already made here.
I would also say there is a whole literature on tradeoffs between explanation (descriptive models in the worst case, causal models in the best case) and prediction (models that accurately reproduce some phenomenon, regardless of if they are based on and true description or causal model). There are also loads of examples of things that are perfectly deterministic and modeled by perfect "causal" models but which are of course still defy human comprehension / intuition, in that the equations need to be run on computers for us to make sense of them (differential equation models, chaotic systems, etc). Or just more practically, we can learn to do all sorts of physical and mental skills, but of course we understand barely anything about the brain and how it works and co-ordinates with the body. But obviously such an understanding is mostly irrelevant for learning how to operate effectively in the world.
I.e. in practice, if the phenomenon is sufficiently complex, an accurate causal model that also accurately models the system is likely to be too complex for us to "understand" anyway (or you just have identifiability issues so you can't decide between multiple different models; or you don't have the time / resources / measurement capacity to do all the experiments needed to solve the identifiability problem anyway), so there is almost always a tradeoff between accuracy/understanding. Understanding is a nice luxury, but in many cases not important, and in complex cases, probably not achievable at all. If you are coming from this perspective, the whole "quandary" of the essay seems just odd.
Unless and until neurologists find evidence of a universal grammar unit (or a biological Transformer, or whatever else) in the human connectome, I don't see how any of these models can be argued to be "causal" in the sense that they map closely to what's physically happening in the brain. That question seems so far beyond current human knowledge that any attempt at it now has about as much value as the ancient Greek philosophers' ideas on the subatomic structure of matter.
So in the meantime, Norvig et al. have built statistical models that can do stuff like predicting whether a given sequence of words is a valid English sentence. I can invent hundreds of novel sentences and run their model, checking each time whether their prediction agrees with my human judgement. If it doesn't, then their prediction has been falsified; but these models turned out to be quite accurate. That seems to me like clear evidence of some kind of progress.
You seem unimpressed with that work. So what do you think is better, and what falsifiable predictions has it made? If it doesn't make falsifiable predictions, then what makes you think it has value?
I feel like there's a significant contingent of quasi-scientists that have somehow managed to excuse their work from any objective metric by which to evaluate it. I believe that both Chomsky and Judea Pearl are among them. I don't think every human endeavor needs to make falsifiable predictions; but without that feedback, it's much easier to become untethered from any useful concept of reality.
I would think it was quite clear from my last two paragraphs that I agree causal models are generally not as important as people like Chomsky think, and that in general are achievable only in incredibly narrow cases. Besides, all models are wrong: but some are useful.
> You seem unimpressed with that work
I didn't say anything about Norvig's work, I was saying the linked essay is bad. It is correct that Chomsky is wrong, but is a bad essay because it tries to argue against Chomsky with a poorly-developed distinction while ignoring much stronger arguments and concepts that more clearly get at the issues. IMO the essay is also weirdly focused on language and language models, when this is a general issue about causal modeling and scientific and technological progress, and so the narrow focus here also just weakens the whole argument.
Also, Judea Pearl is a philosopher, and do-calculus is just one way to think about and work with causality. Talking about falsifiability here is odd, and sounds almost to me like saying "logic is unfalsifiable" or "modeling the world mathematically is unfalsifiable". If you meant something like "the very concept of causality is incoherent", that would be the more appropriate criticism here, and more arguable.
I could iterate with an LLM and Lean, and generate an unlimited amount of logic (or any other kind of math). This math would be correct, but it would almost surely be useless. For this reason, neither computer programs nor grad students are rewarded simply for generating logically correct math. They're instead expected to prove a theorem that other people have tried and failed to prove, or perhaps to make a conjecture with a form not obvious to others. The former is clearly an achievement, and the latter is a falsifiable prediction.
I feel like Norvig is coming from that standpoint of solving problems well-known to be difficult. This has the benefit that it's relatively easy to reach consensus on what's difficult--you can't claim something's easy if you can't do it, and you can't claim it's hard if someone else can. This makes it harder to waste your life on an internally consistent but useless sidetrack, as you might even agree (?) Chomsky has.
You, Chomsky, and Pearl seem to reject that worldview, instead believing the path to an important truth lies entirely within your and your collaborators' own minds. I believe that's consistent with the ancient philosophers. Such beliefs seem to me halfway to religious faith, accepting external feedback on logical consistency, but rejecting external evidence on the utility of the path. That doesn't make them necessarily bad--lots of people have done things I consider good in service of religions I don't believe in--but it makes them pretty hard to argue with.
I'm not sure how you can square anything you said in your last paragraph with anything I said about all models being wrong, and causal modeling being extremely limited.
> But it must be recognized that the notion of "probability of a sentence" is an entirely useless one, under any known interpretation of this term.
He was impressively early to the concept, but I think even those skeptical of the ultimate value of LLMs must agree that his position has aged terribly. That seems to have been a fundamental theoretical failing rather than the computational limits of the time, if he couldn't imagine any framework in which a novel sentence had probability other than zero.
I guess that position hasn't aged worse than his judgment of the Khmer Rouge (or Hugo Chavez, or Epstein, or ...) though. There's a cult of personality around Chomsky that's in no way justified by any scientific, political, or other achievements that I can see.
If Chomsky were known only as a mathematician and computer scientist, then my view of him would be favorable for the reasons you note. His formal grammars are good models for languages that machines can easily use, and that many humans can use with modest effort (i.e., computer programming languages).
The problem is that they're weak models for the languages that humans prefer to use with each other (i.e., natural languages). He seems to have convinced enough academic linguists otherwise to doom most of that field to uselessness for his entire working life, while the useful approach moved to the CS department as NLP.
As to politics, I don't think it's hard to find critics of the West's atrocities with less history of denying or excusing the West's enemies' atrocities. He's certainly not always wrong, but he's a net unfortunate choice of figurehead.
I have the feeling we're focusing on different time periods.
Chomsky already was very active and well-known by 1960.
He pioneered areas in Computer Science, before Computer Science was a formal field, that we still use today.
His political views haven't changed much, but they were beneficial back when America was more naive. They are harmful now only because we suffer from an absurd excess of cynicism.*
How would you feel about Chomsky and his influence if we ignored everything past 1990 (two years after Manufacturing Consent)?
---
* Just imagine if Nixon had been president in today's environment... the public would say "the tapes are a forgery!" or "why would I believe establishment shills like Woodward and Bernstein?" Too much skepticism is as bad as too little.
I wrote "when America was more naive" but that isn't entirely correct. Americans are more naive today in certain areas. If my comment weren't locked, I would change that sentence to something like "when Americans believed most of what they read in the newspaper"
I agree that his contributions to proto-computer-science were real and significant, though I think they're also overstated. Note the link to the Wikipedia page for BNF elsewhere in these comments. There's no evidence that Backus or Naur were aware of Chomsky's ideas vs. simply reinventing them, and Knuth argues that an ancient Indian Sanskrit grammarian deserves priority anyways.
I think Chomsky's political views were pretty terrible, especially before 1990. He spoke favorably of the Khmer Rouge. He dismissed "Murder of a Gentle Land", one of the first Western reports of their mass killing, as a "third rate propaganda tract". As the killing became impossible to completely deny, he downplayed its scale. Concern for human rights in distant lands tends to be a left-leaning concept in the West, but Chomsky's influence neutralized that here. This contributed significantly to the West's indifference, and the killing continued. (The Vietnamese communists ultimately stopped it.)
Anyone who thinks Chomsky had good political ideas should read the opinions of Westerners in Cambodia during that time. I'm not saying he didn't have other good ideas; but how many good ideas does it take to offset 1.5-2M deaths?
> Just imagine if Nixon had been president in today's environment... the public would say "the tapes are a forgery!" or "why would I believe establishment shills like Woodward and Bernstein?" Too much skepticism is as bad as too little.
Today it would not matter in the least if the president were understood to have covered up a conspiracy to break into the DNC headquarters. Much worse things have been dismissed or excused. Most of his party would approve of it and the rest would support him anyway so as not to damage "their side".
I'm not sure what you mean? As the length of a sequence increases (from word to n-gram to sentence to paragraph to ...), the probability that it actually ever appeared (in any corpus, whether that's a training set on disk, or every word ever spoken by any human even if not recorded, or anything else) quickly goes to exactly zero. That makes it computationally useless.
If we define perplexity in the usual way in NLP, then that probability approaches zero as the length of the sequence increases, but it does so smoothly and never reaches exactly zero. This makes it useful for sequences of arbitrary length. This latter metric seems so obviously better that it seems ridiculous to me to reject all statistical approaches based on the former. That's with the benefit of hindsight for me; but enough of Chomsky's less famous contemporaries did judge correctly that I get that benefit, that LLMs exist, etc.
My point is, that even in the new paradigm where probabilistic sequences do offer a sensible approximation of language, would novelty become an emergent feature of said system, or would such a system remain bound to the learned joint probabilities to generate sequences that appear novel, but are in fact (complex) recombinations of existing system states?
And again the question being, whether there is a difference at all between the two? Novelty in the human sense is also often a process of chaining and combining existing tools and thought.
Shannon first proposed Markov processes to generate natural language in 1948. That's inadequate for the reasons discussed extensively in this essay, but it seems like a pretty significant hint that methods beyond simply counting n-grams in the corpus could output useful probabilities.
In any case, do you see evidence that Chomsky changed his view? The quote from 2011 ("some successes, but a lot of failures") is softer but still quite negative.
"BNF itself emerged when John Backus, a programming language designer at IBM, proposed a metalanguage of metalinguistic formulas ... Whether Backus was directly influenced by Chomsky's work is uncertain."
Oh, lots of stuff gets invented multiple times, when it's "in the air". Nothing special about Chomsky here. And I wouldn't see that distracting from this particular achievement.
Don't you think people would have figured it out by themselves the moment programmers started writing parsers? I'm not sure his contribution was particularly needed.
Lots of things get invented / discovered multiple times when it's in the air. But just because Newton (or Leibnitz) existed, doesn't mean Leibnitz (or Newton) were any less visionary.
For your very specific question: have a look at the sorry state of what's called 'regular expressions' many programming languages and libraries to see what programmers left loose can do. (Most of these 'regular expressions' add things like back-references etc that make matching their franken-'xpressions take exponential time in the worst case; but they neglect to put in stuff like intersection or complement of expressions, which are matchable in linear time.
Just checked after reading your comment. Surprisingly to me, AFAs (Alternating Finite Automatons) do let you introduce the Complement operation into Regex while preserving the O(mn) complexity of running NFAs.
That's really subtle, because deciding Regex universality (i.e. whether a regex accepts every input) is PSPACE-COMPLETE. And since NFAs make it efficient to decide whether a regex matches NO inputs, any attempts to combine NFAs with regex Complement would trip on a massive landmine.
The complement of a regular language is a regular language, and for any given regular language we can check whether a string is a member of that language in O(length of the string) time.
Yes, depending on how you represent your regular language, the complement operator might not work play nicely with that representation. But eg it's fairly trivial for finite state machines or when matching via Brzozowski derivatives. See https://en.wikipedia.org/wiki/Brzozowski_derivative
I'm a big Chomsky nerd, Chomsky fan, and card-carrying ex Chomskyan linguist. I hate to break it to you, but not even Chomsky thought that the Chomsky hierarchy had any very interesting application to natural languages. Amongst linguists who (unlike Chomsky) are still interested in formal language classes, the general consensus these days is that the relevant class is one of the so-called 'mildly context sensitive' ones (see e.g. https://www.kornai.com/MatLing/mcsfin.pdf for an overview).
(I suppose I have to state for the record that Chomsky's ties to Epstein are indefensible and that I'm not a fan of his on a personal level.)
In this context, "does not support" means "the evidence is of low quality", not "the evidence says it probably doesn't work". Per the quotations in my other comment here, the paper and its references conclude that the best available RCT evidence is favorable to cannabis for those conditions. They're just not impressed with the statistical power and methodological rigor of those studies.
It's unfortunately common to report that situation of favorable but low-quality evidence as "does not support", despite the confusion that invariably results. This confusion has been noted for literally decades, for example in
>I'm sad to see it repeated here, and I hope we can avoid propagating it further.
Science educators have been fighting the scientific theory vs vernacular theory fight for decades without much progress, so I wouldn't hold my breath.
I think at some point, the scientific community needs to accept that many of the formal and precise ways they are taught to write in order to avoid ambiguity, have the exact opposite effect on everybody else. Unless we adjust the terminology so that the scientific and casual definitions more closely align, we're just going up have to keep explaining.
The acute pain paper they cited (linked in other comment) said "low-quality evidence [...] for a small but significant reduction", which seems clear and correct to me. If these authors think that's too favorable, then the paper I linked above suggests "insufficient evidence to confirm or exclude an important difference".
Either of those distinguishes "strong evidence this doesn't work, and more studies are probably wasted effort" vs. "weak evidence, more studies required". I don't see any benefit to a single phrase covering both cases unless the goal is to deliberately mislead.
That's true, but I believe the authors' complaint here is efficacy rather than safety. (I also think they're using terms of art from evidence-based medicine to make a statement the general public is likely to misinterpret, per my other comment here.)
Safety is barely discussed in this paper, probably because the available RCT evidence is favorable to cannabis. I'm not sure that means it's actually safe, since RCTs of tobacco cigarettes over the same study periods probably wouldn't show signal either. This again shows the downside of ignoring all scientific knowledge except RCT outcomes, just in the other direction.
Acute pain isn't discussed in detail in this paper, but here's a paper they cited:
> Conclusions: There is low-quality evidence indicating that cannabinoids may be a safe alternative for a small but significant reduction in subjective pain score when treating acute pain, with intramuscular administration resulting in a greater reduction relative to oral.
> meta-analysis of 39 RCTs, 38 of which evaluated oral cannabinoids and 1 administered inhaled cannabis, that included 5100 adult participants with chronic pain reported that cannabis and cannabinoid use, compared with placebo, resulted in a small improvement in sleep quality [...]
It goes on to criticize those studies, but we again see low-quality evidence in favor.
In the context of evidence-based medicine, "does not support" can mean the RCTs establish with reasonable confidence that the treatment doesn't work. It can also mean the RCTs show an effect in the good direction but with insufficient statistical power, so that an identical study with more participants would probably--but not certainly--reach our significance threshold. The failure to distinguish between those two quite different situations seems willful and unfortunate here.
Perhaps any statement in that context should be assumed to be oversimplified; but I don't think I can fault someone for taking words to mean what they literally say. The COVID vaccines look great so far on balance, but they absolutely were oversold to the public. We'll pay the price in public confidence for at least a generation.
Could you give the whole paragraph, and not just the last sentence in it?
Ah, heck, I'll do the work of pasting it in.
> But again, one last thing. I — we don’t talk enough to you about this, I don’t think. One last thing that’s really important is: We’re not in a position where we think that any virus — including the Delta virus, which is much more transmissible and more deadly in terms of non — unvaccinated people — the vi- — the various shots that people are getting now cover that. They’re — you’re okay. You’re not going to — you’re not going to get COVID if you have these vaccinations. -Biden
I'm not sure why out of all that Trump-lite-contradictory rambling (and the massive amounts of other words and ink spilled by both the 2020[1] and the 2021 administrations on this subject), that sentence is the singular, unqualified, pinky-swear blood-pact promise that you think the medical community made to the public regarding the vaccine.
---
As for Walensky:
> Three days later, on April 1, a CDC spokesperson seemingly walked back the director’s comments, telling The New York Times “Dr. Walensky spoke broadly during this interview” adding that “It’s possible that some people who are fully vaccinated could get Covid-19. The evidence isn’t clear whether they can spread the virus to others. We are continuing to evaluate the evidence.”
If you're only going to listen to the first thing that's said on a subject, and ignore everything that follows, I don't think that sort of approach will serve you very well. For one thing, it'll probably mean that you'll think that people who correct themselves are idiots.
---
[1] Which, if I may remind you, developed, recommended, and rolled out the vaccine and had nothing to do with Biden.
I'm not sure what the rest of the paragraph adds here? Nothing in that qualifies or contradicts the absolute that I quoted. Are you just saying that the statement was so generally inarticulate that any reasonable person should have ignored it completely? That was true here, but that's not great for public confidence either.
I'm aware that the scientific literature told a more nuanced and accurate story, but only a tiny fraction of the population have the skills and time to study that. I don't think you can fault people for trusting their elected leaders; and if you do, then who are you expecting them to trust next time?
> a CDC spokesperson seemingly walked back the director’s comments
So after widespread criticism by actual scientists, she didn't even correct herself in her own voice, instead sending an unnamed spokesperson to smooth it over without explicitly acknowledging error. I can't believe you don't see how the damage is done.
It's obviously almost all confounders, since COVID mortality is low now and choosing to obtain the vaccine correlates strongly with other medical and non-medical factors affecting all-cause mortality. The point is that no adverse effect from the vaccine offsets that confounding beneficial effect. More covariates would shrink the confounding effect, but they did the best they could with their dataset.
That confounding effect turned out to be massive, which is bad news for anyone hoping to tightly bound the vaccine risk. It's good personal news for anyone in the vaccinated group, just more as to their general life choices than as to COVID.
This isn't an RCT and COVID mortality is no longer high, so the effect on all-cause mortality is almost entirely confounders. So that result just means people 18-29 who chose to get the COVID vaccine have other characteristics that result in the much lower mortality from non-COVID causes.
I'm not sure why. The top causes for that age group are usually non-medical, accident, suicide, or homicide. We might speculate those would anticorrelate more strongly with the prudence that leads people to get the vaccine than unavoidable medical causes, but looking at the V, W, X, and Y causes from Table 2 that doesn't seem to be true. I guess it could be true but only for the 18-29 group (and if it's not then what causes are responsible?), since they don't break that down by age.
The "problem" is that vaccine recipients are so much healthier overall than non-recipients that the vaccine would have to be spectacularly unsafe to offset that. So this analysis doesn't actually tell us much, but it's consistent with all other evidence that the vaccine is safe.
The Census Bureau asks if firms are using AI "to help produce goods or services". I guess that's intended to exclude not-yet-productive investigations, and maybe also indirect uses--does LLM-powered OCR for the expense reports for the travelling sales representatives for a widget factory count? That's all vague enough that I guess it works mostly as a sentiment check, where the absolute value isn't meaningful but the time trend might be.
The Ramp chart seems to use actual payment information from companies using their accounting platform. That should be more objective, though they don't disclose much about their methodology (and their customers aren't necessarily representative, the purpose and intensity of use aren't captured at all, etc.).
> The Census Bureau asks if firms are using AI "to help produce goods or services"
That's odd. I use AI tools at work occasionally, but since our business involves selling physical goods, I guess we would not count as an AI adopter in this survey.
After more investigation, I'm not sure what question was asked. I quoted that exact language from Apollo, and the Census Bureau uses very similar language in the spreadsheet with the aggregated responses, at
I'm unable to find a questionnaire with that language, though. I found a different questionnaire with many AI-related questions, some of which I believe would usefully capture both your situation and my hypothetical above. None closely match that language, though. The closest might be question 23, but that asks about use "in any of its business functions".
After yet more investigation, I think I found the questionnaire. Question 7 asks:
> Between MMM DD – MMM DD, did this business use Artificial Intelligence (AI)
in producing goods or services? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)
Executive Order 13769 (the so-called "Muslim Ban") actually was effective immediately, resulting in a similar race to the airport during its chaotic implementation. Hundreds of people with previously lawful immigration status were denied entry or detained, including dozens with green cards despite those being excluded by the EO.
In any case this new order was not ambiguous. It plainly said the restriction was on "entry", not the issuance of new visas. Nobody reading the text of the order claimed otherwise. The Trump administration just changed their position, in informal guidance rather than a formal executive order but with similar practical and legal effect.
Anyone who returned early was effectively hedging, paying the airline change fee and some inconvenience to avoid a potential $100k fee or worse, with downsides including job loss--potentially leading to forced repatriation--if your employer doesn't pay, and indefinite detention under harsh conditions. That hedge currently seems likely to expire worthless; but with such an imbalanced payoff, are you really saying you wouldn't have paid it yourself?
reply