4 ms·

> To his credit Sejnowski correctly picked on a weak point in the oped: Chomsky’s ChatGPT examples falling apples and gravity were anecdotal and insufficiently

> To his credit Sejnowski correctly picked on a weak point in the oped: Chomsky’s ChatGPT examples falling apples and gravity were anecdotal and insufficiently nuanced. This, I fail to understand. Clearly, Chomsky started with an exposition of the concept of negativity. How can this be missed by anyone with cursory familiarity with western philosophy? The rest of the article (including the passage regarding the apple) is pretty much a series of excursions on this theme, how LLMs systematically miss negativity.

masswerk 4y ago
It may be observed that Chomskys piece is in this sense much a modern take on Hegel's Phenomenology of Spirit, who also starts in his exposition with the core theme of negativity, as the very base of "sense certainty" (better: sensual certainty, sinnliche Gewissheit), what Hegel called the "teachings of Ceres", the very origin of conceptuality, in order to expand on this in his grand scheme. (I find it quite remarkable that Chomsky would fall back on this classic line of thought on this occasion.)
epups 4y ago
It's pretty bad because Chomsky provided only two examples of things ChatGPT cannot do - even implying its underlying architecture could not ever do that -, and it's obvious that neither he, his coauthors or the editors at the NYT bothered to fact check these assertions. I would bet that Chomsky himself has never played around with ChatGPT itself.
- masswerk 4y ago
  If we subscribe for a moment to the tradition that conceptuality, creativity, even certainty, are born from negativity, that crucial "what if not" assumption, the ponderability of existential alternatives, we may indeed defer that this is not something LLMs (in their current configuration) can do. (In a more statistical tradition, proximity to the semantic center is pretty much opposed to this.) Which poses yet another question, namely, how essential and significant are these things for us? Is the alleged promise aligned to our basic assumptions?
  - epups 4y ago
    I'm not sure why we should hold on to such tradition on creativity and certainty emerging from negativity - the reference that comes to mind when you talk about this is the "negation" concept from post-modern continental philosophers like Lacan. If so, I would say I hold zero weight to any argument appealing to such concepts as they have no empirical grounds. I would appreciate if you could correct me here if you mean something else.
    - masswerk 4y ago
      Lacan was much influenced by Hegel in his philosophical foundations, or rather, by the French interpretation as by Alexandre Kojève (who was in turn inspired by some, who started rather in a Neo-Kantian tradition, like Heidegger and Alexandre Koyré.) Moreover, Lacan had quite extensive empirical ground (having worked in psychiatric institutions for quite some time) and was informed by the state of art in Cybernetics and nascent AI research. (E.g., his prime example for fascination was derived from interlocking automata.) What all these have in common, is that reasoning about conceptuality, language, consciousness, expression and ethics, even hard along the edges of what is called the "language barrier", isn't exactly new. (Also, I'm personally part of a European tradition and can't help it.) Last but not least, there was already this post-structuralist notion of "text", which emphasized how text was essentially writing itself, without suggesting that this would imply us letting go of our agency. Anyways, what I tried to suggest was that there is some base for reasoning that what we're encountering here is more like a cargo cult of conceptuality than conceptuality itself at work. And that it is still on us to decide, if want to be persuaded by this to land our planes of reason next to it, or not.
      - epups 4y ago
        Thank you for your reply. I do understand where you are coming from (although my reading is more circumscribed to psychoanalysis itself) but I think our views diverge. I don't think the conceptual framework of psychoanalysis is based in true empiricism (ie the scientific method), so what we are left with is essentially a lot of self referential frameworks that ultimately did not lead us anywhere, and it was built upon fraud. The psychoanalytic conception of language did not help us treat any mental disorder and likely never will. So throwing this purely conceptual work away is thus very easy for someone like me - in fact, we largely already have. I cannot imagine that anyone will read Lacan in even 100 years from now, but I do think we will increasingly develop a computational understanding of language that will stand the test of time.
mannykannot 4y ago
I think Chomsky has a point here, but somewhat ironically, it applies much more broadly, raising doubts about linguistics being the route to understanding the mind. One thing grammar does not do is constrain grammatically-correct language to produce only the truth. It is as easy for humans as it is for LLMs to say false things, to construct false reasons for believing those falsehoods, and to be inconsistent about it. The argument Chomsky uses to claim LLMs will never lead to and understanding of the mind can equally be deployed against the prospect of understanding the mind through studying grammar. If anything, the success of LLMs, in producing grammatically-correct language about as well as humans do, is evidence for this position. They also present a significant challenge to the Poverty of Stimulus claim[1], which is often used in support of the universal grammar hypothesis. In the early 20th. century, western metaphysics took what has been called the 'linguistic turn'[2]. Personally, I am deeply skeptical of the proposition that studying human language will - or could, in principle - lead to knowledge about how the world must be, and the ability of people to hold inconsistent beliefs and to argue for falsehoods is one basis for that skepticism. [1] https://en.wikipedia.org/wiki/Poverty_of_the_stimulus https://en.wikipedia.org/wiki/Poverty_of_the_stimulus [2] https://en.wikipedia.org/wiki/Linguistic_turn https://en.wikipedia.org/wiki/Linguistic_turn
- masswerk 4y ago
  > One thing grammar does not do is constrain grammatically-correct language to produce only the truth. In all fairness, this belongs only to context-free grammars. Also, I'm somewhat concerned that we may see a resurgence of behaviorism promoted along the lines. (It may be worth remembering what kind of regime this had been.)
  - mannykannot 4y ago
    The statement you quote is true of all natural human languages, which is what matters here (if you do not agree, you should be able to refute it on purely grammatical grounds!) As for behaviorism, it can stand or fall on its own merits (or lack of them), just as should the alternatives.

masswerk 4y ago

It may be observed that Chomskys piece is in this sense much a modern take on Hegel's Phenomenology of Spirit, who also starts in his exposition with the core theme of negativity, as the very base of "sense certainty" (better: sensual certainty, sinnliche Gewissheit), what Hegel called the "teachings of Ceres", the very origin of conceptuality, in order to expand on this in his grand scheme. (I find it quite remarkable that Chomsky would fall back on this classic line of thought on this occasion.)

epups 4y ago

It's pretty bad because Chomsky provided only two examples of things ChatGPT cannot do - even implying its underlying architecture could not ever do that -, and it's obvious that neither he, his coauthors or the editors at the NYT bothered to fact check these assertions. I would bet that Chomsky himself has never played around with ChatGPT itself.

masswerk 4y ago
If we subscribe for a moment to the tradition that conceptuality, creativity, even certainty, are born from negativity, that crucial "what if not" assumption, the ponderability of existential alternatives, we may indeed defer that this is not something LLMs (in their current configuration) can do. (In a more statistical tradition, proximity to the semantic center is pretty much opposed to this.) Which poses yet another question, namely, how essential and significant are these things for us? Is the alleged promise aligned to our basic assumptions?
- epups 4y ago
  I'm not sure why we should hold on to such tradition on creativity and certainty emerging from negativity - the reference that comes to mind when you talk about this is the "negation" concept from post-modern continental philosophers like Lacan. If so, I would say I hold zero weight to any argument appealing to such concepts as they have no empirical grounds. I would appreciate if you could correct me here if you mean something else.
  - masswerk 4y ago
    Lacan was much influenced by Hegel in his philosophical foundations, or rather, by the French interpretation as by Alexandre Kojève (who was in turn inspired by some, who started rather in a Neo-Kantian tradition, like Heidegger and Alexandre Koyré.) Moreover, Lacan had quite extensive empirical ground (having worked in psychiatric institutions for quite some time) and was informed by the state of art in Cybernetics and nascent AI research. (E.g., his prime example for fascination was derived from interlocking automata.) What all these have in common, is that reasoning about conceptuality, language, consciousness, expression and ethics, even hard along the edges of what is called the "language barrier", isn't exactly new. (Also, I'm personally part of a European tradition and can't help it.) Last but not least, there was already this post-structuralist notion of "text", which emphasized how text was essentially writing itself, without suggesting that this would imply us letting go of our agency. Anyways, what I tried to suggest was that there is some base for reasoning that what we're encountering here is more like a cargo cult of conceptuality than conceptuality itself at work. And that it is still on us to decide, if want to be persuaded by this to land our planes of reason next to it, or not.
    - epups 4y ago
      Thank you for your reply. I do understand where you are coming from (although my reading is more circumscribed to psychoanalysis itself) but I think our views diverge. I don't think the conceptual framework of psychoanalysis is based in true empiricism (ie the scientific method), so what we are left with is essentially a lot of self referential frameworks that ultimately did not lead us anywhere, and it was built upon fraud. The psychoanalytic conception of language did not help us treat any mental disorder and likely never will. So throwing this purely conceptual work away is thus very easy for someone like me - in fact, we largely already have. I cannot imagine that anyone will read Lacan in even 100 years from now, but I do think we will increasingly develop a computational understanding of language that will stand the test of time.

mannykannot 4y ago

I think Chomsky has a point here, but somewhat ironically, it applies much more broadly, raising doubts about linguistics being the route to understanding the mind. One thing grammar does not do is constrain grammatically-correct language to produce only the truth. It is as easy for humans as it is for LLMs to say false things, to construct false reasons for believing those falsehoods, and to be inconsistent about it. The argument Chomsky uses to claim LLMs will never lead to and understanding of the mind can equally be deployed against the prospect of understanding the mind through studying grammar. If anything, the success of LLMs, in producing grammatically-correct language about as well as humans do, is evidence for this position. They also present a significant challenge to the Poverty of Stimulus claim[1], which is often used in support of the universal grammar hypothesis. In the early 20th. century, western metaphysics took what has been called the 'linguistic turn'[2]. Personally, I am deeply skeptical of the proposition that studying human language will - or could, in principle - lead to knowledge about how the world must be, and the ability of people to hold inconsistent beliefs and to argue for falsehoods is one basis for that skepticism. [1] https://en.wikipedia.org/wiki/Poverty_of_the_stimulus https://en.wikipedia.org/wiki/Poverty_of_the_stimulus [2] https://en.wikipedia.org/wiki/Linguistic_turn https://en.wikipedia.org/wiki/Linguistic_turn

masswerk 4y ago
> One thing grammar does not do is constrain grammatically-correct language to produce only the truth. In all fairness, this belongs only to context-free grammars. Also, I'm somewhat concerned that we may see a resurgence of behaviorism promoted along the lines. (It may be worth remembering what kind of regime this had been.)
- mannykannot 4y ago
  The statement you quote is true of all natural human languages, which is what matters here (if you do not agree, you should be able to refute it on purely grammatical grounds!) As for behaviorism, it can stand or fall on its own merits (or lack of them), just as should the alternatives.