Almost three years ago I wrote a post responding to some early LLM hype, and somewhat unexpectedly, that post served as the kernel for a chapter in a book that has just been published! The book, Artificial Knowledge of Language: A Linguist’s Perspective on Its Nature, Origins and Use (Vernon Press) includes 8 essays from scholars offering a critical view of the relation between LLMs and the human Language Faculty.
On a personal note, the process of converting my post to a published chapter required me to more fully engage with the scholarly community around LLMs both in the published (usually not peer-reviewed) literature and undergoing anonymous peer review, and this engagement left me somewhat surprised at the intellectual bankruptcy in that community, and frightened that they are taken seriously by society at large.
Artificial Knowledge of Language: A Linguist’s Perspective on Its Nature, Origins and Use is edited by José-Luis Mendívil-Giró and published by Vernon Press
LLMs don’t know anything about anything and they simply can’t
In 2024, The California legislature passed SB-1047—a bill intended to reduce so-called catastrophic risk of AI. The companies that make and market LLMs, being private capitalist corporations, hated it—likely because it sought to impose real accountability and real transparency on them—and successfully lobbied Governor Gavin Newsom to veto it. Then last year, the legislature passed AB-1064—a much more modest bill that would require any company that marketed a chatbot to children to ensure that the chatbot “is not foreseeably capable of doing certain things that could harm a child, including encouraging the child to engage in self-harm, suicidal ideation, violence, consumption of drugs or alcohol, or disordered eating.” AI companies lobbied against it and Newsom vetoed it.
Why would a company be so resistant to this minimal regulation, especially when it was passed in the wake of Adam Raines’ death by suicide which seems to have been encouraged by ChatGPT? I don’t think it’s just a knee-jerk reaction against any regulation, because doing so in this case amounts to telling the public “we will not ensure that our product is safe for your children”—bad PR in an industry that has and needs excellent PR.
No, the reason is far simpler. The companies are resistant because they know that they could clear even the low bar that the passed bill sets forth. They know that they can’t make chatbots that adhere to any standard of behaviour, because in order to adhere to a standard you need to know what that standard is. But LLMs can’t know anything and their creators must know this at least subconsciously. And what’s worse, we all should have known this because it was spelled out in what might be considered the founding work of modern philosophy the 1781 tome The Critique of Pure Reason by Immanuel Kant.
Kant was not some futurist who predicted LLMs. Rather he was addressing an active philosophical debate of his time namely a debate between Empiricists like Hume and Locke and Rationalists like Decartes and Leibniz around the problem of objective knowledge.[1]Disclaimer: I’m not a Kant scholar. I’m a theoretical linguist who challenged himself to reading Critique. Based on secondary material, namely a helpful series of videos, I know that Kant … Continue reading The problem was that we are able to have objective knowledge—i.e. knowledge of objects—that must come from experience, yet the sense-data which is the source of that experience is insufficient for such objective knowledge. To give an example, as I write this I am sitting in a cafe at a table and there is an empty ceramic mug to the right of my laptop. In order to make such a descriptive judgment, I need objective knowledge of myself, the cafe, the table, the mug, an my laptop to say the least, and I have that knowledge because my senses tell me so. But what do my senses provide me?
In a given instant each of the millions and millions of your sense-receptor cells is sending a bit of information to your nervous system, reporting whether and it what way it is being stimulated. And each subsequent moment those cells are reporting something new. The world as our senses report it is an ever-changing matrix of atomic bits of data, what Kant calls the manifold, yet we consciously experience it as consisting of unified objects which persevere in time and interact with each other in lawful ways. These are two seemingly incompatible descriptions of the world yet, as the debate between empiricists and rationalists shows, both seem roughly correct.
So, does the world consist of a manifold of appearances constantly in flux or a unified lawful domain of permanent objects? Kant answers “yes.” For Kant, the manifold of appearances and the unity of our experience reflect different aspects of the world, both of which are real.
But how can we take appearances from our senses and extract objects from them? We must carry with us the tools to do so, to go from sense to understanding. For Kant these tools are the very notions of space and time, and The Categories—a set of minimal concepts that we use to construct the concepts we use to identify objects. Without these, experience is impossible, and without experience, knowledge is impossible.
An important thing to note here is that these tools, being necessary for experience, cannot be given in experience. Take, for instance, our sense of space. Part of how I can recognize that myself, the table, the mug, and the laptop are distinct objects is because they, or perhaps more accurately their sensible appearances, occupy distinct regions of space. Restricting myself to the visual appearance of the mug and oversimplifying—there are is a set of photo-receptor cells that occupy a contiguous region in my retina, and each of those cells is reporting being stimulated by whitish light.The information that cells x, y, z, etc are being stimulated is certainly data we did not have before, but the fact that those cells represent a contiguous region must be something I “knew” beforehand, if only because the position of a photo-receptor in a retina is given by my very biology. Likewise, when we consider time, we find that our judgements of time cannot be given by external data, but have to come from our own internal sense of our sensory data being constantly updated.
The Large Language Models which are what we usually mean when we talk about AI these day do not have in-built categories nor nor do they have any sense of space and time. They simply have inputs and math that is performed on those inputs. Lacking the tools necessary for experience, they can’t have any and being unable to experience anything can’t know anything. And therein lies the impossible bind that any regulation would place AI companies in.
Can we build an LLM that will not encourage children to harm themselves or others? In order to do this, an LLM needs to be able to make a judgement that a statement is an encouragement to self-harm and such a judgement requires knowledge of objective concepts of “harm,” “self,” and even “statement” and an LLM cannot have such knowledge. And the same reasoning explains why we can’t build an LLM that will not generate CSAM, that will not cause psychosis, and so on.
Such a conclusion likely strikes folks as a bit far-fetch, perhaps because we are used to websites being able to filter out certain types of content by default—Google, for instance, doesn’t surface NSFW pages unless the user turns off safe search, and doesn’t surface illegal pages at all. Why then, can’t we do so with LLMs? If it’s because LLMs can’t know anything, does that mean I’m saying that Google’s search algorithm can know things? Certainly not, but Search engines and other traditional websites are based on a database and as such can at least represent knowledge and people can build algorithms that use that represented knowledge.
Suppose you are building a search engine. You’ve crawled the entire web and placed every page into a database including some metadata that you deem useful in finding the best pages. You’ve also created an algorithm for analyzing a query and ranking all your pages by how well they match the query. You open it up to the public and it works! Users are generally happy with the results, but you start to get complaints that unwanted results are showing up—porn, hate sites, depictions of violence, or the like. The solution is relatively simple—add some metadata to your database that indicates whether each page belongs to a particular category of potentially unwanted results and adjust your algorithm to check for this metadata before serving these results.
The devil is in the details, of course, in that as simple as it is to add the relevant metadata to a database, correctly and reliably categorizing each page accordingly is thorny to say the least. It brings in a number of grey areas, depends on value judgments, and is difficult to scale so it is error-prone, but it is possible in principle because humans directly encode the possibility of representing objective knowledge into the software. An LLM, though, does not store responses or pages, or facts in a database that can be directly modified—LLMs can’t even represent knowledge.
This distinction between LLMs and classic web apps cannot be lost on any competent engineer that would be tasked with building “guardrails” for chatbots. So, it is very likely that companies that build chatbots know that we cannot build those guardrails yet even without familiarity with 18th century philosophy. But given that critical philosophy/thinking seems to be anathema to the ever more reactionary tech sector, they likely don’t know, or aren’t allowed to admit, that they will never be able to build such guardrails.
It’s no surprise, then, that LLM makers would put so much lobbying effort and, by extension, money into fighting even the most reasonable regulations like “don’t build a chatbot that will encourage child suicide,” because it would force them to admit that they cannot possibly clear so low a bar.
But as bad as it is that LLMs cannot be restrained from being positively harmful, the impossibility of them having objective knowledge is a death knell for any hopes for their benefits. If you asked a normal person what they would hope an AI or a Robot would do for them, likely they’d want a machine to do chores around the house, work dangerous jobs, and give them more time “to knit sweaters, play with their dogs, start a garage band, experiment with new recipes, or sit in cafés arguing about politics, and gossiping about their friends’ complex polyamorous love affairs” (to borrow a formulation from the late David Graeber). LLMs, however, will never be able to do this even if they are given physical bodies. Because how can you trust an AI to wash a dish if it cannot even conceive of the possibility of a dish, let alone know what a dish is?
But, of course, the billionaires who sell, fund, and promote LLMs don’t have to do household chores, or dangerous jobs, so they don’t care if their AIs can do those tasks. Instead, they are promoting LLMs ability to write, to code, to make images, and so on. On their face, these might seem like possibilities as an LLM does nothing but generate formally correct outputs—they can string together words into grammatically acceptable sentences, or symbols into syntactically correct Python expressions. But anything we create of any value, be it text, or code, or images, has not only form, but content and function, both of which implicate objective knowledge. To write well, you need to create well-formed sentences, but you also need to choose the right words based on knowledge of their semantics. To write a good program, you need to know what you want it to do—how you want it to affect the world. You need knowledge of the world that LLMs cannot have.
This, of course, hasn’t stopped them from degrading our working conditions and firing us. Your livelihood has no doubt been negatively affected by LLMs. Brian Merchant has been gathering stories of the real life effects of LLMs. And as Cory Doctorow puts it “even though an AI can’t do your job, an AI salesman can convince your boss to fire you and replace you with an AI that can’t do your job.” So what can we do with the understanding that LLMs will never ever be able to do our jobs? I’m not naive enough to believe that we can convince the bosses of this, of course. But there’s value in holding this knowledge ourselves as a form of intellectual self-defense. They desperately want us to believe that an AI can do our job, that there is no alternative, that we should gladly welcome our new chatbot overlords. They want this because if we believe the opposite, then we might get it in our heads to organize and resist.
Disclaimer: I’m not a Kant scholar. I’m a theoretical linguist who challenged himself to reading Critique. Based on secondary material, namely a helpful series of videos, I know that Kant was less interested in epistemology and psychology than in metaphysics.
I have posted a draft of a new paper on LingBuzz. It’s a continuation of my 2022 Biolinguistics article, though I frame it as a critical analysis of certain recent developments in linguistic theory. The abstract is as follows:
This paper analyzes the FormSet operation as it has been defined and used in the literature specifically in the domain of conjunction structures. It finds that, though its use as a structure building operation must either contradict its definition and justification as a general-purpose preparatory operation or lead to unwanted empirical predictions. It goes on to show that, if we redefine MERGE as applying not to syntactic objects in a workspace but to n-ary sets of syntactic objects in a workspace, that we can retain the justification of FormSet without sacrificing empirical validity. Furthermore, it presents novel empirical arguments in favour of this redefined MERGE—that it correctly predicts the ambiguities raised by conjunction structure, that it correctly predicts a cross-linguistic gap in the plural anaphor system, and that it gives some insight into the system of number “features.”
As always I’m happy to hear any comments you have.
Over on his blog, Chris Collins has a new post on the difference between what he calls the “Standard Paradigm for Syntactic Research” and the “Foundational Paradigm for Syntactic Research.” In it, he identifies a sociological phenomenon within generative syntax—there is a general apathy towards “foundational” work in the field—and suggests four explanations for it. I mostly agree with what he writes, but I wanted to highlight and deepen one of his explanations and add another to the mix.
Collins’ first explanation is as follows (emphasis mine):
[F]oundational work takes a lot of time and effort, and the payoff is uncertain. The question is what counts as progress. You may think for years and years about the ‘copy versus repetition’ distinction or about the definition of c-command or about the nature of workspaces without resolving the issues, even though in the end you have a much deeper understanding. Does that count as progress? Can you write it up and publish it? Does it count as currency in the academic monetary system?
I would posit that almost any “non-foundational” work can yield some sort of result, be that a new generalization, new data, or maybe a new synthesis of data, whereas “foundational” work does sometimes end in blind alleys with little to show for it. While this difference is important when measured in papers, posters, and talks, its also important when measured in money.
In most of the global north, public funding for the science, social science, the humanities, and the arts is distributed by members of the particular fields, ostensibly on an apolitical basis. Yet for almost as long as modern states have been funding science, social science, the humanities, and the arts there has been a faction within those states that have categorically opposed it. And one of the favoured rhetorical tactics of that faction is to trot out obscure studies that seem frivolous out of context and ask disingenuously “You think this is a good thing to fund??”
It is difficult to believe that anyone on a grant committee is not aware of this possibility, and has in the back of their mind “how would I explain giving this grant to a ‘foundational’ project that may yeild nothing?”
The explanation I’d add to Collins’ is a psychological one. What he calls “foundational” work in syntactic theory is theoretical work while “non-foundational” work in syntactic theory is actually empirical/analytical/experimental work, not theoretical work (cf Chametzky 1996). Most syntacticians doing pen-and-paper analysis, then, consider they’re work to be theoretical syntax—they see themselves as theoretical syntacticians. So when they encounter “foundational” work (i.e., theoretical work) it evokes some cognitive dissonance (“I know theoretical syntax, and this is supposed to be theoretical syntax, but I can’t follow it, or I don’t see the point.”) which often leads to hostile reactions. Indeed, suggesting to someone they’re not doing theoretical work is often taken as an insult due to the esteem given to “theoretical work.”
I’ve arrived at this explanation because, when I was in grad school I experienced the same cognitive dissonance form the other side. I was always interested in “foundational” syntactic theory, but found I couldn’t do or get excited about the the “theoretical” work that my syntax colleagues were doing. I simply don’t have the mind for the empirical details. The natural response to this was impostor’s syndrome, until I realized that what we’d been calling “theoretical syntax” was actually two very different research paradigms, and that the work that I couldn’t do, was a different type of work. This saved my self-esteem, but still left me with the understanding that there was a ceiling on my academic career, not because of my skills, but because of the state of the field.
It had been a while since I’d presented my own work anywhere, so there was a process of “learning how to do this again” involved—which was interesting—and I didn’t fully know what I was going to say until about 3 hours before the talk.
My topic was the theoretical explanation of coordination, comparing Parallel Derivation—which I used to explain adjunction—with Chomsky et al‘s FormSet[1]FormSet is a refinement of the earlier FormSequence. Most comparisons with FormSet also apply to FormSequence.. I have been thinking about the application of Parallel Derivation (PD) to coordination for as long as I’ve had the idea for PD, and about its comparison with FormSet since I first read a description of its precursor but, as is always the case, I discovered a few things through the act if writing these thoughts down and an expressing them to to other syntactiticans.
The next thing to do, I guess, is to write these discoveries down and express them to other syntacticians.
As in yearsprevious, I’d like to give a rundown of the top things in culture for me in 2024. As in those previous years, this is not a list of things that were released in 2024, but things that I encountered, or finally got around to in 2024.
2024, however, was not a normal year. The entirety of the year there has been an ongoing genocide being live-streamed. I, like many of you, have watched as Israel has targeted Palestinian children, women, journalists, doctors and nurses as part of their attempt to finally push the Palestinian people off what remains of their land. I have watched as the countries of the Global North, including my country of Canada, have turned a blind-eye to the crime-of-crimes, while doing everything in their power to crack down on any voices that object. To say this cast a pall on the year would be an understatement.
Arms Embargo Now! Sanction Israel! Saoirse don Phalaistín!
Now, if you’re still reading, these are the cultural things that brought be some joy, or helped me cope in this bleak year.
Sonic Ranch by The Sloppy Boys
This year the comedy-party-rock trio The Sloppy Boys released their fourth album. In something of a twist, they eschewed their normal path of self-producing and instead, took a trip out to storied recording studio in San Antonio—the titular Sonic Ranch—and had their album produced by Money Mark.
The result is fantastic! It’s no surprise that the Boys brought the funny to every song, and the funny is amplified by the fact that their silly jokes are set to music of the highest production value.
Best enjoyed alongside the making-of film Blood Sweat and Beers directed by Robert Holguin.
Station Eleven
(book by Emily St John Mandel/miniseries by Patrick Somerville)
File this one under “finally got around to it.” The book and the miniseries begin with two simultaneous events—the sudden on-stage death of world-famous actor, and the onset of a flu pandemic that will wipe out a large portion of humanity. It then follows the post-disaster fortunes of three characters who are connected through the dead actor, while giving flashbacks to their pre-disaster lives. There’s plenty to say about these works but what struck me the most about both, but especially the novel, was their portrayals of pre-disaster (i.e., our present) reality. Without banging the reader/viewer over the head, the author/director show deeply sad, frustrated and alienated our their lives are were.
At the risk of sounding like a snob, the novel is better though. I can’t wait to read more from St John Mandel!
The Third Revolution by Murray Bookchin
Originally published in the 90s as a two-volume history of the centuries of revolution that ranged from the late middle ages into the end of the 19th century, The Third Revolution tells the stories of the great revolutions with a particular focus on popular, rather than liberal, forces. So rather than pitting Cromwell against Charles I, Washington against George II, Robespierre against Louis XVI, Bookchin highlights the Levellers, Thomas Paine, and Jean Varlet.
He later followed up his two volumes with a third—which I’m currently reading—covering the Russian Revolutions and a fourth covering the German and Spanish revolutions.
The Past Is Still Alive by Hurray For The Riff Raff
Another stellar album from Alynda Segarra, the non-binary Nuyorican singer-songwriter behind Hurray For the Riff Raff. In it they reflect on their past—old friends and lovers, hitchhiking, etc.—in maybe their most bittersweet record yet.
Kneecap (dir: Rich Peppiatt)
It’s like Twenty Four Hour Party People but its about a very current group, and the group members play themselves, and they’re pretty damn good actors. Oh, and it’s about half in Irish, and has a strong political message, and much more drugs, and Michael Fassbender’s in it. Seriously there’s no good way to describe the film Kneecap without making it sound pretty bonkers, which it is, but it works.
Whatever you do though, don’t let Mo Chara, Móglaí Bap, DJ Próvaí fool you, they may be drug-dealing ‘low-life scum’ (their words) but they have a well-defined language revitalization framework.
Anora (dir: Sean Baker)
In Notes From The Underground Dostoevsky demolishes what was a common trope of his time—the romantic hero redeeming a sex worker. Anora pulls that trope inside out. Vanya (Mark Eidelstein), a spoiled son of Russian oligarchs compulsively marries Ani (Mikey Madison), a brash Brighton Beach sex-worker. In response, Vanya’s parents send some bumbling goons after them to get an annulment. It’s funny, sexy, genuinely moving. It’s clear that Baker loves his characters even Vanya the spoiled brat is lovable in his self-awareness. Mikey Madison is a revelation, but don’t sleep on the goons.
Honourable Mentions
War and Peace (1966-67; dir: Sergei Bondarchuk)
I worked my way through the Soviet five-part 7+hr adaptation of Tolstoy’s greatest work this year. It definitely comes the closest to capturing the blend of mystical romanticism and realism that characterizes the book. Someone once remarked that, though they were subject to greater explicit political censorship, Soviet film-makers, in many ways, enjoyed greater artistic freedom than their Hollywood counterparts. This film demonstrates that.
Blood in The Machine by Brian Merchant
A history of the Luddite uprisings, with some literary analysis of English romantic/gothic literature thrown in for good measure. A perfect parable for our present tech-landscape
Paradise Built in Hell by Rebecca Solnit (2009)
Another history, this time of post-disaster communities. Solnit covers the aftermaths 1906 San Francisco earthquake,the 1917 Halifax explosion, the 1985 Mexico City earthquake, 9/11, and Hurricane Katrina in New Orleans. She shows that, contrary to common image of riots, violence, and looting, the aftermaths of disasters usually involve survivors spontaneously organizing to help each other, that is until the police or military show up and start violent riots to restore order. I don’t know if I read this before or after Station Eleven but the two definitely resonate with each other
Fucked Up’s live show
This was my first time seeing them live and what a show! I didn’t expect a hardcore show to include a brief acoustic set in the middle and an extended jam session at the end, but if anyone can do it it’s them.
I’ve just submitted a new paper to appear in a forthcoming volume edited by José-Luis Mendívil-Giró. The volume is on the implications of Large Language Models for linguistic theory and my paper is a refinement of what I’ve written here, here, and here. The abstract is:
While “modern language models,” which put into practice empiricist theories of language, are claimed to be refutations of rationalist theories of language, a close look at the claims made in their favour reveals otherwise. In this chapter, I critically review Piantadosi’s (2023) arguments for his claim that MLMs refute rationalist theories of language along with his replies to critiques of those arguments. I argue that, throughout his paper, Piantadosi misstates the claims, predictions, and arguments of contemporary rationalist theories of language, paying particular attention to the problem of impossible languages. I further argue that the goals of rationalist theories of language and MLMs are orthogonal to each other, with the former being a scientific inquiry aimed at explanation and understanding and the latter being an engineering project aimed at making tools, and that confusing the two, as Piantadosi does, can lead one to inadvertently take up an anti-scientific position.
If you’d like a preview, I’ve uploaded a draft on LingBuzz.
Recently, in service of a course I’m teaching, I had a chance to revisit and fully engage with what might be the stickiest idea in generative syntax—The DP hypothesis. For those of you who aren’t linguists, the DP hypothesis, though highly technical, is fairly simple to get the gist of based on a couple of observations:
Observation 1: Words in sentences naturally cluster together into phrases like “the toys”, “to the store”, or “eat an apple.”
Observation 2: In every phrase, there is a single main word called the head of the phrase. So, for instance, the head of the phrase “eat an apple” is the verb “eat.”
These observations are formalized in syntactic theory, so that “eat an apple” is labeled a VP (Verb Phrase), while “to the store” is a PP (Preposition Phrase). Which leads us to the DP hypothesis: Phrases like “the toys,” “a red phone,” or “my dog” should be labelled as DPs (Determiner Phrases) because their heads are “the,” “a,” and “my,” which are called determiners in modern generative syntax.
This is fairly counterintuitive, to say the least. The intuitive hypothesis—the one that pretty much every linguist accepted until the 1980s—is that those phrases are NPs (Noun Phrases), but if we only accepted intuitive proposals, there’d be no science to speak of. Indeed, the all the good scientific theories start off counterintuitive and become intuitive only by force of argument. One of the joys of theory is experiencing that shift of mind-set—it can feel like magic when done right.
So it was quite unnerving when I started reading the actual arguments for the DP hypothesis, which I had, at one point, fully bought into, and and began to become less convinced by each one. It didn’t feel like magic, it felt like a con.
My source for this is a handbook chapter by Judy Bernstein that summarizes the basic argument for the DP Hypothesis—a twofold argument consisting of a Parallelism argument and purported direct evidence of the DP Hypothesis— as previously advanced sand developed by Szabolcsi, Abney, Longobardi, Kayne, Bernstein herself, and others.
The parallelism argument is based on another counterintuitive theory developed in in the mid-20th century which states that clauses, previously considered either headless or VPs, are actually headed by abstract (i.e., silent) words. That is, they are variously considered TPs (Tense Phrases), IP’s (Inflection Phrases), or CPs (Complementizer Phrases). The parallelism argument states that “if clauses are like that, then ‘noun phrases’ be like that too” and then finds data where “noun phrases” look like clauses in some way. This might seem reasonable on its face, but it’s a complete non sequitur. Maybe the structure of a “noun phrase” parallels that of a clause, but maybe it doesn’t. In fact, there’s probably good reason to think that the structure of “noun phrases” is the inverse of the structure of the clause—the clause “projects” from the verb, and verbs and nouns are complementary, so shouldn’t the noun have complementary properties to the verb?
Following through on parallelism, if extended VPs are actually CPs, then extended NPs are DPs. Once you have that hypothesis, you can start making “predictions” and checking if the data supports them. And of course there is data that becomes easy to explain once we have the DP Hypothesis. Again, this is good as far as it goes, but there’s a key word missing—”only.” We need data that only becomes easy to explain once we have the DP Hypothesis. And while I don’t have competing analyses for the data adduced for the DP Hypothesis at the ready—though Ben Bruening has one for at least one such phenomenon—I’m not really convinced that none exist.
And that’s the foundation of the DP Hypothesis, a weak argument resting on another weak argument. Yet, it’s a sticky one—I can count on one hand the contemporary generative syntacticians that have expressed skepticism about it. Why is it so sticky? My hypothesis is that it’s useful as a shibboleth and as a “project pump”.
Its usefulness as a shibboleth is fairly straightforward—there’s no quicker way to mark yourself as a generative syntactician than to put DPs in your tree diagrams. Even I find it jarring to see NPs in trees.
To see the utility of the DP Hypothesis as a “project pump”, one need only to look at the Cartography/Nanosyntax literature. Once you open up a space for invisible functional heads between N and D, you seem to find them everywhere. This, I think, is what Chomsky meant when he described the DP Hypothesis as “…very fruitful, leading to a lot of interesting work” before saying “I’ve never really been convinced by it.” Who cares if it’s correct, it contains infinite dissertations!
Now maybe I’m being to hard on the DP and its fans. After all, as far as theoretical avenues go, the DP Hypothesis is something of a cul de sac, albeit a large one—the core theory doesn’t really care whether “the bee” is a DP or and NP, so what’s the harm? I could point out that by making such a feeble hypothesis our standard, we’ve opened ourselves to being dunked on my anti-generativists. Or I could bore you with such Romantic notions as “calling all things by their right names.” Instead, I’ll be practical and point out that, contrary to contemporary digital wisdom, the world is not infinite, and every bit of real estate given to the DP cul-de-sac in the form of journal articles, conference presentations, tenure-track hires, etc. is space that could be used otherwise. And, to torture the metaphor further, shouldn’t we try to use our real estate for work with a stronger foundation?
The Canadian government will “continue to follow the case very closely.” Those were the words of Canada’s Minister of Foreign Affairs Mélanie Joly in response to The ICJ’s preliminary findings in South Africa’s genocide case against Israel. She does not mention of the fact that court’s preliminary orders indicate that charges of genocide against Israel are not, as Liberal MP Anthony Housefather puts it, “baseless.” Nor does she indicate any move to withdraw Canada’s support of Israel, or even make it contingent on Israel even pretending to comply with the court’s order that it prevent acts of genocide, acts such as murdering three palestinians in Gaza less than a day after being ordered to prevent such acts.
So, in one case, we have a legitimate international court saying that, upon hearing arguments for and against, there is a prima facie plausible case against the State of Israel on the charge of genocide, and Canada adopts a wait-and-see approach, even as Israel appears to be ignoring the court. While in another case, we have mere allegations against employees of a UN agency, and Canada’s response is immediate action against the UN agency, even as the agency appears to be taking these allegations very seriously.
(or Piantadosi and MLMs again (II)—continuation of this post)
In my critique of Prof. Piantadosi’s manuscript “Modern language models refute Chomsky’s approach to language,” I point out that regardless of the respective empirical results of Generative Linguistics and MLMs, the latter does not supersede the former because the two have fundamentally different goals. Generative Linguistics aims to provide a rational explanation of a natural phenomenon while MLM are designed to simulate human language use. Piantadosi does not dispute this, but rather states that
… there is an interesting debate about the nature of science lurking here. The critics’ position seems to be that in order for something to be a scientific theory, it must be intuitively comprehensible to us. I disagree because there are many phenomena in nature which probably will never admit a simple enough description for us to comprehend. We cannot just exclude these things from scientific inquiry.
p37 of v7 (emphasis in original)
Being one of the “critics” referred to here, I can grant the professor’s description of my position as basically accurate if a bit glib. But what is his position? He doesn’t say precisely, but we can make some inferences. In lieu of a clear statement of his position, for instance, Piantadosi follows the above quote with this:
There probably is no simple theory of a stock market (why IBM takes on a particular value) or dynamics in complex systems (why an O2 molecule hits a particular place on my eyeball). Certainly there are local, proximate causes (Tom Jones bid $142 for IBM; the O2 molecule was bumped by another), but when you start to trace these causes back into the complex system, you will quickly exceed our ability to understand the complex network of interactions.
p37 of v7
These are slightly bizarre comments, as we do have comprehensible (i.e., simple) theories of stock markets—the efficient markets hypothesis, for instance[1]This should not be taken as an endorsement of the efficient markets hypothesis—or any part of (neo)classical economics—as correct. A theory’s scientific-ness is no guarantee of its … Continue reading—and gases—the kinetic theory, for instance—which can give approximate predictions regarding real life events like the examples given. The professor’s view can be narrowed down slightly based onhis assertion that Rawski & Baumont (2023) “seem to misunderstand the linkage between experiment and theory” (p34 of v7)[2]This is a bold claim for Piantadosi to make given that he is a psychologist, while Lucie Baumont—the latter half of Rawski & Baumont—is an empirical astrophysicist. when they state that “Explanatory power, not predictive adequacy, forms the core of physics and ultimately all modern science.” (Rawski & Baumont 2023) It would seem clear, then, that, for Piantadosi at least, that a “theory” is scientific only insofar as it has predictive power.
This may seem like a reasonable characterization—despite myriad insinuations to the contrary, virtually no one believes that predictive power is unimportant—but as soon as one attempts to develop that characterization things get dicey. What, for instance, is the required level of accuracy and precision for science? And What sort of things should a true science be able to predict? To use one of Piantadosi’s examples, individual molecules are the primitives of the kinetic theory of gasses, and the theory makes precise predictions about the behaviour of a gas—i.e., gas molecules in aggregate—but it is highly doubtful that it would make predictions about the actual motion of a particular molecule in any situation. Surely, this would be too much to ask of any theory of physics, yet Piantadosi seems to believe it is within the realm of scientific inquiry.
There’s also a question of what it means to “predict” something. Piantadosi’s argument boils down to “MLMs are better than Chomsky’s approach theory, because they make more correct predictions,” yet nowhere does he explicitly say what those predictions are, nor does he document any tests of those predictions. Instead, we are treated to his prompts to a chatbot followed by the chatbot’s response. Perhaps these are the predictions. Perhaps they predict how a human would respond to such prompts. If so, then so much the worse for MLMs qua scientific theories because, even if MLMs were indistinguishable from humans, the odds of any two humans answering a single question the same way is vanishingly slim, and any way to determine a general similarity between utterances would almost certainly be either arbitrary or dependent on some theoretical framework. At best, MLMs simulate human language use, meaning they no more predict facts of language than a compass predicts facts of geometry.
Chomsky’s approach to theories of language, on the other hand makes clear predictions if one bothers to engage with it. The predictions are of the form “Given theoretical statement T, a competent speaker of language L will judge expression S as (un)acceptable in context C.” This is exactly the sort of prediction that one finds in other sciences—”if one performs precisely this action under precisely these conditions, one will observe precisely this reaction”—and the sort of prediction that is absent in Piantadosi’s paper.
Indeed these predictions seem to be absent in the entire contemporary “AI” discourse, and with good reason—”AI” is not a scientific enterprise. It’s an engineering project. A fact that is immediately obvious when one considers how it measures success—against a battery of predetermined arbitrary tests. MLM researchers, then, aren’t discovering truths, they’re building tools to spec, like good engineers.
This is not to cast aspersions on engineers, but it does raise a question—the core question: How exactly can an engineering project like MLMs refute a scientific theory like Generative Grammar?
This should not be taken as an endorsement of the efficient markets hypothesis—or any part of (neo)classical economics—as correct. A theory’s scientific-ness is no guarantee of its correctness.
This is a bold claim for Piantadosi to make given that he is a psychologist, while Lucie Baumont—the latter half of Rawski & Baumont—is an empirical astrophysicist.
