Leiter Reports: A Philosophy Blog

News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.

  1. J.P. Loo's avatar
  2. Sebastian Sunday Grève's avatar
  3. Giovanni Molteni Tagliabue's avatar
  4. Fabien Muller's avatar
  5. Saul Smilansky's avatar

“Computers plainly can’t think”

Philosopher Alva Noe comments; an excerpt:

[T]here is another origin of our impulse to concede mind to devices of our own invention, and this is what I focus on here: the tendency of some scientists to take for granted what can only be described as a wildly simplistic picture of human and animal cognitive life. They rely unchecked on one-sided, indeed, milquetoast conceptions of human activity, skill and cognitive accomplishment. The surreptitious substitution (to use a phrase of Edmund Husserl’s) of this thin gruel version of the mind at work – a substitution that I hope to convince you traces back to Alan Turing and the very origins of AI – is the decisive move in the conjuring trick.

What scientists seem to have forgotten is that the human animal is a creature of disturbance. Or as the mid-20th-century philosopher of biology Hans Jonas wrote: ‘Irritability is the germ, and as it were the atom, of having a world…’ With us there is always, so to speak, a pebble in the shoe. And this is what moves us, turns us, orients us to reorient ourselves, to do things differently, so that we might carry on. It is irritation and disorientation that is the source of our concern. In the absence of disturbance, there is nothing: no language, no games, no goals, no tasks, no world, no care, and so, yes, no consciousness….

[H]uman beings are not merely doers (eg, games players) whose actions, at least when successful, conform to rules or norms. We are doers whose activity is always (at least potentially) the site of conflict. Second-order acts of reflection and criticism belong to the first-order performance itself. These are entangled, and with the consequence that you can never factor out, from the pure exercise of the activity itself, all the ways in which the activity challenges, retards, impedes and confounds. To play piano, for example – that other keyboard technology – is to fight with the machine, to battle against it.

Let me explain: the piano is the construction and elaboration of a particular musical culture and its values. It installs a conception of what is musically legible, intelligible, permitted and possible. A contraption made of approximately 12,000 pieces of wood, steel, felt and wire, the piano is a quasi-digital system, in which tones are the work of keystrokes, and in which intervals, scales and harmonic possibilities are controlled by the machine’s design and manufacture.

The piano was invented, to be sure, but not by you or me. We encounter it. It pre-exists us and solicits our submission. To learn to play is to be altered, made to adapt one’s posture, hands, fingers, legs and feet to the piano’s mechanical requirements. Under the regime of the piano keyboard, it is demanded that we ourselves become player pianos, that is to say, extensions of the machine itself.

But we can’t. And we won’t. To learn to play, to take on the machine, for us, is to struggle. It is hard to master the instrument’s demands.

To master the piano is not just to conform to the machine’s demands. It is to push back, to say no

And this fact – the difficulty we encounter in the face of the keyboard’s insistence – is productive. We make art out of it. It stops us being player pianos, but it is exactly what is required if we are to become piano players.

Leave a Reply

Your email address will not be published. Required fields are marked *

10 responses to ““Computers plainly can’t think””

  1. I'm a non-professional who has engaged with philosophy of mind since I was a teen (debating with friends at the bus stop) and then had the pleasure of taking an intro class from Thomas Nagel as an undergrad who handed me mimeos of his essays Panpsychism and WIILTBAB when I told him at office hours that I thought everything had a mind. I've read papers by Strawson and Chalmers, Dennett and others. That said, I am mystified (not in the McGinn way) by what Noe is arguing about here. Is it just aesthetics? That we complicated flesh & blood human entities shouldn't so easily give up the word "thought" to objects manufactured of silicon, oxygen, phosphorus and boron? That social being is the most important thing in the equation? In which case I assume he thinks bees and ants (as Hofstadter amusingly argued) would qualify as thinking creatures par excellence. They are engaged in the environment they find, they breathe, they interact in complex fashion with each other, etc. I presume that's not his intent.

    Let's grant that the hype around current LLMs is overdone, that they are stochastic parrots, locked inside their data centers humming away on Nvidia chips, that they present merely a simulacrum of human thought. Is there anything more to it than this (not new) objection, made by many of the same people involved in building the things. There are obviously a lot of problems to be solved before we have anything resembling the interactive robots of science fiction. But clearly they are not impossible! The progress towards such creations has, seemingly, far exceeded what anyone might have expected as recently as 20 years ago. And 20 years from now, when one of them sits down to play Chopin, having learned autonomously through backpropagation and gradient descent, what will Noe say about what's going on inside its chips?

  2. Siddharth Muthukrishnan

    So what exactly is the argument here? The best I can reconstruct is this:

    1. Human cognition is entangled. (An entity's cognition is *entangled* iff its first-order actions are not separable from second-order reflections as to why it should take this action at all and as to what the best norms to follow in the context of action are.)
    2. LLM/computer cognition is not entangled.
    3. For an entity to think or to do anything, its cognition needs to be entangled.
    4. Therefore, LLMs/computers can't think and they don't do anything.

    The premise (3) seems far too strong. In particular, it renders too much of human activity as not thinking or doing anything. When I'm engrossed in a well-specified mathematical problem, I'm not reflecting on whether the problem is a good problem or whether the rules of mathematical operations are justified; I'm just trying to solve the problem. Or when I'm driving a car, I'm usually not reflecting on whether the rules of the road make sense or whether the car is functioning properly; I'm just trying to navigate safely from point A to B. Am I not thinking or doing anything during those periods? To say I'm not seems clearly false.

    Now perhaps the point is that I *could* reflect on those second-order concerns, but then so can LLMs. The meta-problems can also be inputs to the model.

    Perhaps the point is that the second-order concerns can't become salient in the same way to the LLMs as they do to humans *while* they execute the first-order actions. But this too seems too false. If you give a poorly worded or ill-specified mathematical problem to a latest-generation LLM (e.g., Claude 3.5 Sonnet), it will pick up on the ambiguities, and will ask for clarification. That seems a clear case of engaging in a second-order concern. (And if one finds this skill too "weak" in some sense, I suspect it's only a matter of time before they become better at it.)

    And in any case even if you deny LLMs have this capability, my broader concerns still stands: Why is the possibility of having second-order concerns becoming salient during first-order activity *essential* to that activity counting as thinking or doing?

  3. The article has several threads but the excerpt quoted above is mostly about the flow of action, reflective thought, conflict, struggle, and control. Like in the example of learning to play the piano: We don’t just emit a series of actions like an automaton. As we are playing, we concurrently reflectively scrutinize what we are doing, sometimes we criticize it and strive to do something better, and some of those times we succeed. All of this seems exactly correct, and Noe’s prose is lovely to boot.

    But then things go quickly off the rails. We are invited to think that while a machine can do the first-order actions, it can’t do the second-order reflective control part—the scrutiny, striving for something different, resistance part. Why not? I don’t see an argument. I see metaphors, assertions, and lots of lovely prose, but I see no arguments.

    Meanwhile in computational cognitive science, there is a large body of work making real progress on the algorithmic workings of reflective control, albeit using terms like dual process theory, executive processing, automaticity and control, cognitive control, and the like. So, the question again: Why can’t algorithmic entities do the reflective control part?

    Interestingly, standard LLMs likely serve as a decent approximation of so-called system 1 cognition, the automatic kind, which Noe seems quite content to grant. But the newer o1 architecture engages in serial step-by-step processing that has an uncanny resemblance to system 2 thought, the deeper reflective kind that is supposed to be impossible for a machine.

    I suspect if you stitch together a traditional LLM for fast processing with an o1 architecture for slow, reflective scrutiny in light of deeper considerations, you will get precisely the entanglement, i.e., the conflict, reflective scrutiny, and resistance, that Noe says is impossible for a machine.

  4. Shane M. Wilkins

    On Noe’s account, what are we to make of animal cognition? I take it that shellfish cognition isn’t “entangled”, but surely it is cognition—there has to be something in between shellfish perception and shellfish behavior, no?

  5. Is a rough summary as follows? All thinkers have emotions. Computers do not have emotions. So, every computer is not a thinker. Premise two seems a bit strong. Sometimes it's said that the less someone emotes the better he or she thinks or acts.

  6. The argument isn't that machines cannot engage in second-order reflection (though they plainly cannot). It's that doing things isn't *difficult* for machines, isn't a *struggle* for them, in the way that is constitutive of our engagement in the activities under discussion. As Alva writes:

    >>Somehow Turing and his successors tend to forget that games are also contests; they are proving grounds, and it is we who are tested and we whose limitations are exposed, or whose powers as well as frailties are put on display on the kickball field, or the four square court. A child who plays competitive chess might suffer from anxiety so extreme they are nauseated. This visceral expression is no accidental epiphenomenon, an external of no essential value to the game. No, games without vomit – or at least that live possibility – would not be recognisable as human games at all.<<

    I am sure he would say the same thing about playing the piano: that this isn't just a matter of producing the right notes any more than chess is just about making winning moves, and so the requisite struggle with the machine is available only to a creature subject to anxiety and other forms of emotional distress, to visceral symptoms like nausea, to rage at the piano — or herself — when the necessary coupling isn't achieved, to feelings of pride and shame at how she is seen by others. Further, in this case the activity that's being struggled with is an essentially bodily one, a matter of achieving the right posture while getting a pair of organs to move in synchrony while also holding the notes in mind and achieving the desired expression. All of this belongs to our embodiment, our history, our way of existing among and in the eyes of others. Could you program a machine to realize all of this? I don't know what's even being asked. But it would definitely need a lot more than second-order reflective control.

  7. You're right about Alva's point, John, but finding a task difficult is a second-order reflection on it, and so Chandra's interpretation of Alva is relevant. You say that machines "plainly" cannot engage in second-order reflection. I don't understand why. I agree that if a machine's behaviour is entirely the product of a program, and if that program does not engage in second-order reflection, then the machine won't. But what if the machine is "improvising"–writing its own learning-to-play program? Might it not be able to rate the difficulty of its own task? (Learning Bach is hard, but learning Busoni is diabolical.) It might not be the same thing as a human finding a task difficult–it may not experience the same feelings of frustration and frailty. But (even allowing that struggle is indeed constitutive of human tasks, and not just an invariable accompaniment) why could the machine not find a task difficult?

  8. A small thing first: while *finding* a task difficult requires engaging in second order reflection, a task's *being* difficult might not. Alva does use the language of "encountering" difficulty, so maybe he is committed to the first characterization. But it might not be essential to the phenomenon of "struggling" as such.

    Now the more substantial thing: I don't wish to speak for Alva, but I take him to be saying, in part, that struggling is borne from a *desire* to engage, successfully, in the activity that one struggles with, and that a machine can no more desire to win at chess, or master the piano, than it can desire to eat. That is, a machine couldn't have the last desire unless it took in food for the sake of nutrition, and its way of being was conditioned by this need. Likewise, the desire to win at chess, or master the piano, is a desire that arises from our social entanglement (since it's more than just the desire to make certain moves or produce certain sounds) — so that "feelings of frustration and frailty" are as much a part of the difficulty of these activities as feelings of hunger are part of the desire to eat.

  9. John, I don't think your "small point" is right. Alva writes that human beings are "doers," for whom "[s]econd-order acts of reflection and criticism belong to the first-order performance itself." This seems very clear: it's the "entanglement" of second-order awareness that is characteristic of us as humans. Now, one sentence later, he says: "To play piano, for example . . . is to fight with the machine, to battle against it." Taken on its own, this is consistent with your first-order interpretation, but the context seems clearly to force the higher-order reflection interpretation.

    Alva's thesis, you say, is that machines cannot have desires that arise out of bodily or social "entanglements." Or maybe that they cannot have these entanglements in the first place. Why can they not? You say that it isn't obvious what it would mean for them to have "all of this." Fair point: nobody has explained machine-desire adequately. Unfortunately, though, acknowledging this sad fact doesn't take us any closer to answering the question, "Can a machine play the piano?"

  10. It might be something closer to an Aristotelian model. Animals can perceive and desire but not reflect. That would vibe with other commitments about perception as a kind of activity (rather than passive reception). So I think the reply is that the shellfish doesn't cognize if by "cognize" you mean engaging in 1st order activities that also enable 2nd order reflection

    —–
    KEYWORDS:
    Primary Blog

Designed with WordPress