Chomsky and the origins of AI research
« previous post | next post »
Melissa Heikkilä, "LeCun: 'Intelligence really is about learning'", Financial Times 1/2/2026:
(The AI pioneer on stepping down from Meta, the limits of large language models — and the launch of his new start-up)
LeCun’s lightbulb moment came as a student at the École Supérieure d’Ingénieurs en Électrotechnique et Électronique in Paris in the 1980s, when he read a book about a debate on nature versus nurture between the linguist Noam Chomsky and Jean Piaget, a psychologist. Chomsky argued that humans have an inbuilt capacity for language, while Piaget said there is some structure but most of it is learnt.
“I’m not gonna make friends saying this . . . ” he tells me, “but I was reading this and I thought everything that Chomsky . . . was saying could not possibly be true, [because] we learn everything. Intelligence really is about learning.”
AI research — or neural networks, as the technology was then called, which loosely mimic how the brain functions — was practically a dead field and considered taboo by the scientific community, after early iterations of the technology failed to impress. But LeCun sought out other researchers studying neural networks and found intellectual “soulmates” in the likes of Geoffrey Hinton, then a faculty member at Carnegie Mellon.
[You can't read the whole article at that link without a subscription, which I recommend despite its price. But as Kai von Fintel tells us in the comments, there's an open-access reprint at Ars Technica.]
For a sketch of Yann LeCun's opinions about current directions in AI research, see "AMI not AGI?", 8/2/2025.
And 1980's Yann seems to have fallen into the common error of seeing Noam as a proponent of epistemological nativism rather than rationalism, though Noam has often been misleading on this issue, including apparently in the debate with Piaget. See e.g.
"The Forever War", 2/20/2022
"Straw men and Bee Science", 6/4/2011
"JP versus FHC+CHF versus PU versus HCF", 8/25/2005
"Chomsky testifies in Kansas", 5/6/2005
The book that LeCun refers to is Massimo Piattelli-Palmarini, Ed., Language and Learning: The Debate Between Jean Piaget and Noam Chomsky — or presumably the French version Theories Du Language, Theories de L'apprentissage.
Philip Taylor said,
January 8, 2026 @ 6:59 am
A free, trial, subscription will also allow full access to the article — this is the route that I have just used in order to access it. But then if you're prepared to spend €434 on a lunch for two, the cost of a full subscription to the FT is barely going to cause a ripple in your bank balance …
Kai von Fintel said,
January 8, 2026 @ 7:24 am
The article is reprinted and freely readable at https://arstechnica.com/ai/2026/01/computer-scientist-yann-lecun-intelligence-really-is-about-learning/.
David Marjanović said,
January 8, 2026 @ 7:31 am
Are you sure he isn't "Le Cun" with a space? I've never seen CamelCase names (of any origin) outside the US and maybe Canada; the Breton names with Le all have a space in France in my experience.
Did he simply change his mind on this issue, as he has on so many others?
Philip Taylor said,
January 8, 2026 @ 9:03 am
He may well have been born Yann Le Cun, David, but as far as the author of the article is concerned, he is [Yann] LeCun consistently throughout the article. He also self-identifies as such — see http://yann.lecun.com/.
Mark Liberman said,
January 8, 2026 @ 9:11 am
@David Marjnanović: "Are you sure he isn't "Le Cun" with a space? I've never seen CamelCase names (of any origin) outside the US and maybe Canada; the Breton names with Le all have a space in France in my experience."
His home page puts no space between the 'e' and the 'C'.
"Did [Chomsky] simply change his mind on this issue, as he has on so many others?"
As the linked pages document, his basic epistemological position has been pretty consistent — namely that no fundamental aspects of language are acquired from experience, either in the neurome or in the genome.
His suggestions about how language development nevertheless occurs have changed many times, starting with the half-serious 1965 suggestion of reincarnation, and on to the 1988 quote about a mysterious mutation or mysteriously emergent "physical laws applying to a brain of a certain degree of complexity".
J.W. Brewer said,
January 8, 2026 @ 9:47 am
I found the elaborate discussion of what was eaten and drunk during the conversation a bit odd, but I did appreciate the backstory detail that LeCun played the crumhorn in his youth.
Jerry Packard said,
January 8, 2026 @ 9:51 am
The Piaget-Chomsky debate is most interesting, but LeCun’s lightbulb moment is critically and – in my view fallaciously – based on the his assumption that “…we learn everything. Intelligence really is about learning.”
That Chomsky is correct in saying that humans have an inbuilt capacity for human language (as opposed to Piaget who believed that most of it is learned) is supported by the simple observation that, if placed in the environment where human language is being spoken, humans are the only organisms on earth that will learn human language.
Gokul Madhavan said,
January 8, 2026 @ 9:59 am
@J. W. Brewer: this is a standard feature of this genre of Financial Times interviews.
Philip Taylor said,
January 8, 2026 @ 10:19 am
Jerry — "if placed in the environment where human language is being spoken, humans are the only organisms on earth that will learn human language" — on what evidence do you base this claim ? If by "learn" you mean "learn to speak", then I would suggest that that may well simply be a matter of anatomy; if, however, you mean "learn to understand", then I would suggest that there are many counter-examples, the most commonplace being the abilitity of dogs and horses to follow spoken instructions ("sit", "bleiben", "die for England" (!), "whoah", "walk on" and so on); whether primates can learn to respond to spoken language using sign language is still a matter of debate, but I believe that they can.
Jerry Packard said,
January 8, 2026 @ 10:33 am
Philip, I like your argumentation but am nowhere even near convinced, as I feel your counterexamples are weak to say the least. Other species only understand and can produce mere snippets of humans’ vast linguistic capacity.
Chris Button said,
January 8, 2026 @ 11:45 am
I think Philip's point about anatomy is a good one. We became bipedal, which allowed for major shifts in the pharynx and larynx to enable complex articulations and hence "speech".
Philip Taylor said,
January 8, 2026 @ 11:46 am
With "can produce" I have no problem whatsoever, Jerry, but I am by no means convinced that they cannot understand us. When, for example, I tell my cat (who is sitting outside the back door) "stay there, and I'll get you some food", she does just that. And I think that dogs and horses, quite apart from primates, are far more intelligent that cats …
Jerry Packard said,
January 8, 2026 @ 2:07 pm
Of course they understand us. But that understanding stops way short of even the most rudimentary subject-verb-object sentence.
Olaf Zimmermann said,
January 8, 2026 @ 3:23 pm
Funnily enough, I bought the Piattelli-Palmarini whilst studying in France; a few years later, I picked up the English version London, at Dillons, for five quid – 'coz it was remaindered!. Working my way through Chomky's Aspects, I'd long reached the conclusion that his might be a nice theory of second language acquisition, but not very convincing otherwise. And then there was Fo[n]dor on his Auntie trip … different days – but the pop music was still (mostly) hand-made.
Rod Johnson said,
January 8, 2026 @ 4:32 pm
Relatedly, this article came out today: Dogs with a large vocabulary of object labels learn new labels by overhearing like 1.5-year-old infants. (NYT summary)
Chester Draws said,
January 8, 2026 @ 5:00 pm
Of course they understand us. But that understanding stops way short of even the most rudimentary subject-verb-object sentence.
As most of us who have travelled widely know, we can understand what a person is telling us without a single word in the local language. That angry policeman pointing at me is not saying "have a nice day".
Once you add a couple of words, it often becomes pretty obvious from context what is going on.
That animals can understand what we are telling them does not really imply that they understand our language. It merely means that they are intelligent.
Chris Button said,
January 8, 2026 @ 5:10 pm
Can't the further development of brain capacity for speech be put down to the physical development of our speech organs on account of us becoming bipedal?
In short, we develop the ability to produce complex utterances, and our brains develop accordingly to support this ability.
Nat J said,
January 8, 2026 @ 5:16 pm
My dog has a much easier time following my gestures than my verbal commands. I think what some animals are doing is following cues. I don't think that language can be equated to a complex system of cues. That leaves out all grammatical structure. Is there any reason to think any animals can understand grammar? They just know or learn that certain sounds indicate certain actions. That's probably why gestures are easier to follow for animals: no grammar to confuse them. So, I'm much more sympathetic to Jerry Packard.
That said, I'm not sure how good the argument is that "humans are the only organisms on earth that will learn human language". There are many skills that only humans can learn from their environment.
JPL said,
January 8, 2026 @ 5:36 pm
@ Philip:
I once had a dog that, while lying calmly next to me, became instantly fearful, and got up and looked around toward the neighbour's house when in conversation I used (probably with normal stress) a word that happened to be the name of the neighbour's dog that was always bullying him. I took from that that a dog is capable of noting the significance of some human speech sounds; he knew not only his own name, but the name of another dog. (He probably heard the name spoken most often not by me, but from the mouth of the neighbour, who was always calling his dog off, or calling him to come.)
JPL said,
January 8, 2026 @ 8:59 pm
If the problem underlying those addressed in the book "Theories du langage, theories de l' apprentissage" is to understand what the phenomenon of human language is and how it's possible, why all the focus on psychology, learning, the brain, neurology, genetics, etc., and nothing on speech communities, the speech actions of speakers, and the construction of norms for language systems? To take scientific knowledge as a subcategory of knowledge, it's obvious that it is, in its most important part, knowledge of the world. So how do language systems, which we use, in the Kantian fashion, to make sense of the world make effective and adaptive contact with the world? It can't be explained by focusing only on either "logical form" or "forms" in the sense of Bloomfieldian forms. I would think that even Chomsky ("Language is a system for the expression of thought.") would say, as Frege would, that it's thought that makes contact with the world. Thoughts don't belong to one person, as Wittgenstein taught us. So how does it all work?
If I were to try to capture the notion of "intelligence" in a single phrase, I might offer that the term 'intelligence' refers to the capacity of an organism for solving problems, and understanding the world is probably the most pressing problem for most organisms. And btw, I'd rather focus on the phenomenon of understanding (which, if it's functioning properly, should be increasing in adaptive effectiveness) as more basic than what people think of when they're talking about "knowledge". Understanding the world involves more than mathematics and formal logic; it involves things like the use of the category of causality, which necessarily involves tools for understanding the dependency structure of the world. It's not a question of "either-or". Acts of natural language use, as well as the activity of interpretation of perceptions, always involve both logical forms and reference to the world. We grasp the world and bring it into our understanding. (I hope what I said makes sense, because I wrote it in a bit of necessary haste.)
J.W. Brewer said,
January 8, 2026 @ 11:42 pm
I would be interested in hearing from those with relevant knowledge (including but not limited to myl) whether the article's claim that AI was in fact in the mid-EIghties "practically a dead field and considered taboo by the scientific community." LeCun is a bit older than me, and finished his Ph.D. in France the year I got my B.A. in the U.S., but the field wasn't so dead/taboo that he couldn't get himself hired by Bell Labs (alas approaching the end of its own lifespan …) to work on it.
As I may have mentioned before I took an undergraduate class on Artificial Intelligence (offered in the Computer Science department but aimed at non-majors; I may have gotten credit toward the linguistics major for it?) circa '86, and came away with a somewhat cynical view that the field was at the time overpromising and underdelivering, but also the sense that funders were still shoveling money at it. (The wikibio of the professor who taught it, now emeritized, talks about her work in "machine learning" rather than AI, but I don't know that that means much.)
ktschwarz said,
January 9, 2026 @ 12:00 am
"Practically a dead field" is an exaggeration, but the phrase "AI winter" appeared in the 80s:
https://en.wikipedia.org/wiki/AI_winter
JPL said,
January 9, 2026 @ 1:16 am
@J W Brewer:
"The wikibio of the professor who taught it, now emeritized, talks about her work in "machine learning" rather than AI, but I don't know that that means much."
Makes me think of Margaret Boden, an interesting figure in these debates in those years, who has not yet been mentioned, but whose writings might give you some insight concerning the anti-Chomskyan positions in that era.
Philip Taylor said,
January 9, 2026 @ 5:17 am
Jerry — "Of course they understand us. But that understanding stops way short of even the most rudimentary subject-verb-object sentence" — I think that I agree (but see later for a counter-example). However, when someone is speaking to me, how often do I analyse what they are saying in terms of "subject-verb-object" ? I would suggest, "almost never". Just as when I read a word on the printer page, I almost never consider the letters of which it is composed (unless one of them is blatantly wrong). Rather, in both cases, I immediately infer a meaning. And I believe that animals (well, some at least) do just then on hearing the spoken word. Not inately, of course, but having learned to do so during the course of their lifetime.
Now, my counter-example. Many years ago I shared a German-shepherd cross "Bassie" with my mother, and I would regularly take Bassie out for walks. If we were going somewhere suitable, I would often take a frisbee with me, which Bassie would willing fetch, both from land and from water. On one occasion, out walking with Bassie near Chislehurst ponds (without frisbee), a boy came up to me and said "Mister, can your dog get my frisbee ? It's gone in the pond and I can't get it". I replied (with perhaps more confidence than deserved) "Of course he can", and then said to Bassie "Bassie, go fetch the frisbee". Now he, like I, had no idea where the frisbee was, yet he immediately plunged into the water, swam around until he located the frisbee, seized it in his jaws and swam back to land. Now you (Jerry) ask me to accept that he could not parse "Bassie, go fetch the frisbee" into its constituent parts, and you may well be right, yet his behaviour suggested that he had "understood every word". What do you think ?
Olaf Zimmermann said,
January 9, 2026 @ 6:32 am
@JPL FWIW, Maggie couldn't program her way out of a paper bag. And, yes, I read most of her stuff. But there were a lot of non-Chomskyats around who did have, and could make, a point. Polyglots, mostly. This being said, I still consider him to be mandatory reading (his linguistic work, that is).
@Philip Taylor Dogs go by vowels and intonation (i.e. supra-segmental phonology) only. And gestures. No consonants, and certainly no syntax – rythm, though.
Jerry Packard said,
January 9, 2026 @ 7:38 am
@ Nat J
“There are many skills that only humans can learn from their environment.”
Yes, thanks, I hadn’t thought of it that way before. But what it implies to me is that we are natively equipped to learn those skills. For example, reading and writing we are equipped to learn. But unlike human language, we won’t learn them unless we are explicitly taught. Baby ducks and turtles are natively equipped to do certain things – they don’t have to be taught.
Olaf Zimmermann said,
January 9, 2026 @ 7:47 am
For the fun of it: https://www.hup.harvard.edu/books/9780674921832
Made for interesting classroom discussions at the time …
Jerry Packard said,
January 9, 2026 @ 7:47 am
@Philip
That is quite a nice example. For Bessie to accomplish that task, she needed to identify the word frisbee and relate it to the thing floating out in the water. Now, if she saw it go in, I’m not too impressed. But if you arrived later on the scene or Bessie had less knowledge of the context for other reasons, I’d be impressed, because it would mean she (?) independently inferred there was a frisbee out there waiting to be retrieved, with less help from specific context.
Jerry Packard said,
January 9, 2026 @ 8:41 am
“…when someone is speaking to me, how often do I analyse what they are saying in terms of "subject-verb-object"…”
You *always* do. You can’t take ‘subject-verb-object’ literally. It is simply a heuristically useful way to make the point that humans parse the language input into constituents – entities whose thematic role (agent, patient, possessor/possessed) is understood by their relative order or grammatical marking. If you didn’t do that all the time (implicitly or not), you wouldn’t be clear about the thematic argument structure of the utterance you heard – that is, you wouldn’t be entirely clear about who was doing what to whom.
Philip Taylor said,
January 9, 2026 @ 5:02 pm
Fair enough. If the requirement is not to interpret "subject-verb-object" literally but simply to register "who is doing (or to do) what to (or with, or for) whom", then I think I agree. But equally I genuinely believe that Bassie (not Bessie — he was a he, and named after the Battersea Dogs' Home from which he came) understood that he (Bassie) was required to (locate and) fetch the frisbee. Which he did. Neither he nor I had seen the frisbee enter the water — it had been thrown before we arrived, and was visible to neither of us (nor to the boy who had lost it). Yet locate and retrieve it he did, in response to spoken English.
Jerry Packard said,
January 9, 2026 @ 8:58 pm
Yes, I believe Bassie was able to execute the command and therefore must have understood the command and understood that he was expected to execute the command. Would Bassie have understood the command in a different context, e.g., if the boy asked you to ask Bassie to fetch the frisbee out of a field or the woods?
Philip Taylor said,
January 10, 2026 @ 4:53 am
[Off-topic, only for those who love animals]
Sadly we will never know, Jerry, but since he was used to fetching frisbees on land as well as in water, I very much suspect so — Bassie has long since passed on, put to rest by a kindly vet who administered the fatal injection in the comfort of his own home, lying on his favourite armchair (Bassie had told me that very morning, without needing to use spoken language, that he had finally had enough, and that one more day of life would be one too many). A week after he died I went back to my vet's and said "You know, of course, that Bassie is no longer with us — do you have another dog that is looking for a good home ?" and the vet. replied "We're so glad you asked — we've been keeping this one for you, knowing that Bassie was dying and that you would almost certainly want another dog after he had gone. She was brought in after a road traffic accident but no-one came forward to claim her". They introduced me to Cleo, Cleo to me, she rolled over on her back, said (again, without any need for spoken words) "love me", I rubbed her belly and that was it. Cleo had entered my life just a week after Bassie had left it. Oh, and once I had said I would take her, the vets then operated on her at no expense to me to repair her road-traffic injury. Clearly having been involved in a road traffic accident she needed to learn road sense as a top priority, and within a few months she could be trusted never to cross a road without the word of command, even if she had up to that point been in full pursuit of a fox. The fox would cross the road and she would stay on the pavement, watching it disappear out of sight.
[/Off-topic]
Jerry Packard said,
January 10, 2026 @ 8:49 am
Reminds me of putting down our dog Suga 41 years ago.
Robot Therapist said,
January 11, 2026 @ 8:33 am
"I would be interested in hearing from those with relevant knowledge (including but not limited to myl) whether the article's claim that AI was in fact in the mid-EIghties "practically a dead field and considered taboo by the scientific community.""
Right at the start of the 80s I was involved in an AI group under Karen Sparck Jones, and then in the mid 80s the "Alvey project", so there was at least some life.
Mark Young said,
January 11, 2026 @ 9:10 am
"AI research — or neural networks, as the technology was then called"
It's important to note that at that time the term "AI research" was not restricted to neural networks. I don't think the author is right to so restrict it even today, since expert systems were clearly a product of AI research in spite of not using neural networks.
Nevertheless, in the 80s and 90s when I was a student studying AI, neural network research was growing in importance from a low base. Every neural network talk I attended spent quite some time at the beginning explaining what it was and how it worked. NNs were then a small part of the field, and I never did anything with them in either of my degree programs.