Category: AI

Further Thoughts on AI

This post is based on some conversations I have been having with friends, plus reading things written by child psychologist Alison Gopnik and listening to her interview with Timothy. Links to the relevant Gopnik materials at the end of the post. More generally, my thoughts on AI are deeply influenced by Henry Farrell and his ongoing discussion of it on his substack blog Programmable Mutter.

It’s easy to see why the whole AI question is obsessing all the philosophers because it opens all the big questions. It all seems connected—agency, individuality, other minds, consciousness, reason and reasoning, learning, novelty (creativity)—the whole enchilada. Figuring out what AI is doing—and what is can and cannot do—challenges us to have coherent notions of consciousness, agency, reason, and all the rest. And we, in fact, don’t really have coherent concepts of those biggies. But thinking about them from the specific angle of AI might actually help bring conceptual clarity. At least that’s one hope—what we might call the meta-hope in all the conversation around AI.

I have just about finished The Philosophical Baby by Alison Gopnik [The Philosophical Baby: What Children’s Minds Tell Us About Truth, Love, and the Meaning of Life; Picador, 2010] and find it somewhat useful in trying to think about some of these things. So, to just address one issue: when it comes to Gopnik’s views on AI, the argument is about learning as contrasted to brute force. Or maybe more accurately, how learning is different when brute force is the path taken instead of the experimental method Gopnik ascribes to babies. AI needs huge data centers because it learns by crunching its way through all the data as its way to a conclusion about what is the most probable “correct” prediction. This is the basic Bayesian model that is thought to underlie both a human’s predictions and a computer’s. It’s just that the computer reaches its prediction by a different path. Humans can’t access such massive amounts of data, so need to make inferences from a smaller sample.

Gopnik believes all babies are good Bayesians. That is, babies assess a situation in terms of expectations formed from what has been learned in prior experiences. They then revise their expectations in light of new information—or because this new situation has not disrupted (in some way) their expectation. One key point is that expectations (predictions) are probabilistic; the baby has a hypothesis about what is most likely to happen. But the baby may also have some notions of less likely (but still possible) outcomes. It’s not all or nothing; it’s about what is likely to happen, not what must or necessarily will happen. The current vogue for considering how minds process their relation to the world as Bayesian processing fits with 1) the current received wisdom that perception and cognition are active predictions of external states, not passive receptions of those states and 2) the salience of algorithms, that is, a conception of thinking as mathematical. Bayesian predictions are statistical (literally so for large language model AI) and intuitively so for human processors. Here’s Wikipedia’s pocket definition of Bayesian inference (leaving out the fairly complicated math that comes with a fully statistical account of Bayesian reasoning: Bayesian inference is a method of statistical inference in which Bayes’ theorem is used to calculate a probability of a hypothesis, given prior evidence, and update it as more information becomes available.

Of course, from where I stand, all this looks very Peircean to me. Doubt generates inquiry when the expected outcome of a situation fails to materialize. We update predictions in light of non-confirming experiences, developing new theories or hypotheses or expectations, after a process of inquiry. And those new theories are fallible because only probable—and thus will also be revised the next time we encounter non-confirming evidence. Necessity is downgraded. There are not scientific laws, only probabilities, even if some predictions are more stable than others. Ideally, persons are open to non-confirming evidence, which is how learning occurs and more supple expectations are formed. In reality, various forms of closed-mindedness (conformation bias etc.) means expectation/hypothesis revision is less frequent than circumstances would warrant.

Presumably, however, the computer (like the child) can learn something when its prediction proves false. That experience leads to a revision of one’s picture of the world, of one’s expectations. And that revision doesn’t seem, either in principle or in fact, different for the computer than it is for the child. The feedback loop works the same in both cases. The only differences are 1) the sample size and 2) maybe something about how the data is “felt” or “processed.” In Peirce, the upsetting of expectation generates doubt, which is an uneasy feeling, one we try to dispel through our process of inquiry. Presumably, the computer is not experiencing doubt. It’s drive to find the most plausible explanation or next step derives from other sources.

But Gopnik is saying (see her conversation with Timothy) that the computer is still not “reasoning.” The feedback it receives is just more data—and is taken to be authoritative, at least until another iteration justifies another revision. This is the puzzle. What is the nature of “reasoning” that makes it different from revising one’s views based on feedback offered by the data. The answer seems to be something like “assessment” or “judgment,” which then gets connected to “logic.” But trying to build an AI that works by being “logical” didn’t pan out. A computer cannot get to predictions that way; it needs the brute force method. (Or, at least, up to now brute force has proved much more successful than logic for AI.)

The attempt to base AI on logic occurred in the 1980s and 1990s when “cognitive science” tried to uncover the laws of thought and the laws of language use. Chomsky’s project was central If Chomsky could identify the fundamental grammar from which all languages were generated, then the computer that was given those fundamental structures would be able to talk. Similarly, if the basic elements of logic were correctly identified, a computer with those elements installed would be able to think. That project proved a failure. No program that successfully produced thought or speech was developed. Current AI works by brute force; it surveys a massive number of instances and then “guesses” (via Bayesian inference) what the most probable response (next step) would be.

Human judgment, human reasoning, is notoriously fallible. And despite computer hallucinations, AI is, in many instances and in many ways, less fallible than humans. So it’s not clear what hinges on insisting on a difference between reasoning and brute force. Two ways to skin the cat—and it is far from clear that reasoning is the better way to do the job.

Furthermore (and this was Timothy’s argument in his conversation with Gopnik), it is at least debatable to claim AI does no reasoning. This is where the novelty point comes in. The cultural technology people (the multi-authored essay cited below) are saying AI is only capable of giving us new combinations of the archive; AI produces mash-ups. Sophisticated mash-ups sensitive to context, audience, persona etc., but mash-ups just the same. But Timothy says that’s just not true. Give AI certain mathematical problems and it produces novel solutions, ones that no human mathematician has ever come up with, and those solutions prove, to the best of expert human opinion, correct. So the machine is doing something beyond mash-up—even though we don’t know how it is doing that something more. So that’s one thing that has people intrigued, worried, puzzled. Something’s going on and we don’t know what it is.

I think once you grant that AI has the ability to produce novelty, the cat is out of the bag. Arendt’s definition of agency is precisely the “introduction of something new into the world.” Traditionally, that notion of agency is connected with consciousness. A person thinks of (or imagines) something and then acts to make it come into being. In The Philosophical Baby, Gopnik seems to be on both sides of this traditional view. On the one hand, her emphasis on “counterfactuals” absolutely fits the traditional model of imagining alternatives in thought and then acting on one of them. For Gopnik, the very ability to act is based on being able to imagine multiple possible scenarios, to think not only of what is, but also of what might be. And she identifies such counterfactual thinking as present very early in babies (as early as nine months). But, on the other hand, when she gets to her chapter on consciousness, she is skeptical that our phenomenological experience of consciousness is an accurate picture of what is actually transpiring. “In experiences of executive control, we often feel sure we are making a rational choice when, in fact, we are in the grip of some irrational unconscious bias. And in all of these cases we experience a homunculus, the inner observer, biographer, and decider that we know just can’t exist. . . .The gap between the way the mind functions and the shape of conscious experience is even greater for children than for adults. . . . Consciousness isn’t a transparent and lucid Cartesian stream. Instead it’s a turbulent, muddy mess” (161, 163).

It would seem, then, that human actions (creative productions of novelty or even simple effective procedures of daily activities like cooking food; activities that require adjustments in response to current conditions as we go along ) are as mysterious as the “black box” of generative AI. In both instances, we do not have a good account of novelty. In the human case, we fall back on notions of intuition, genius, unconsciousness, adaptation—or to material, bio-chemical processes that we say must be doing the job even though we still don’t know how those processes actually work, actually produce their magic.

I’ve got plenty more to say about all these topics. But I am going to finish up today with just one more issue. Individuality. An account of action seems to cry out for a complimentary account of “an agent.” Not only what happens in the world, but who has made it happen. Which then opens up the questions of motivation (why did the agent want to make that happen?) and of effectiveness (was the action successful?; was it based on a good enough assessment of what the context afforded or on a grievously fallacious misreading of the context?). And also the old Aristotelean question of the interrelation of action and character. Does one’s actions shape one’s character or does one’s character determine what actions one finds desirable to undertake?

A quick aside on action. It is worthwhile to keep in mind that any theory of action must take into account how the agent interacts with (has to negotiate) three different spheres (for lack of a better term): the material external world; the people among whom she dwells; and internal desires, moods, emotions, thoughts, fantasies, physical capabilities and/or their lack. Just how those relations are established and played out by a machine as contrasted to a human has to be part of the larger puzzle in trying to characterize AI. For example, what does it do to our conception of action if we have an agent (the computer) to which we ascribe no internal sphere (no thoughts, emotions, desires etc.)

Since, so far, most people want to resist anthropomorphizing our machines (i.e. to deny that they have “characters” and certainly to deny them individuality—which would mean my computer’s character was different from your computer’s), we now have a disconnect between action and agent. The computer acts, but it isn’t an agent. I think the general speculation is tangling with this dilemma. People like Gopnik want to deny that the computer acts. That way they can preserve their conviction that the computer is not an agent. Some AI enthusiasts will go the other direction. The computer acts and it is an agent. And then those in the middle are agnostic for the moment, trying to figure it out.

All of which might just be a way to say that the AI debates are reproducing fairly faithfully the consciousness debates. After all, much of the consciousness debate (not all of it, but much of it) is about agency. What is the source of human action? And does consciousness have any role to play in the production of action? The positions taken on these questions range all the way from saying consciousness has no role to saying that consciousness is the determining factor. But there are lots of people in the middle who hedge their bets, saying consciousness is one factor among others.

So maybe what this means is that the AI debate is misguided from the start because contaminated by the debate/puzzles about human action. Maybe we need a whole new paradigm to think productively about what machines are doing. To stop comparing them to humans; to stop worrying whether they act or think like humans. We need a whole new set of terms and questions. Which would mean a way of thinking the non-human or the un-human. An uncanniness that does not, in the long run, reference home. That takes up its abode in an absolutely foreign place.

Bibliography:

Gopnik’s web site can be found at https://alisongopnik.com/ You can access all her various essays and interviews from that site (although not the text of The Philosophical Baby).

Here’s a link to her conversation with Timothy: https://www.youtube.com/watch?v=JIFdeXPB1Pg

And a link to the cultural technologies essay: https://alisongopnik.com/Papers_Alison/science.adt9819.pdf

Henry Farrell’s blog is on Substack, called Programmable Mutter, and is free to all: https://www.programmablemutter.com/?utm_campaign=profile_chips

More Thoughts on AI

A friend writes to me and tells me about the amazing things he has been getting AI to do, in particular in building a model to calculate probabilities of certain events in baseball games.

Here is my response:

I don’t know what to think about Chat GBT–or AI more generally. I haven’t used it and don’t know how to use it. But I have been reading (in a desultory way) about it in an effort to understand it. Most of the stuff I read (predictable, I guess, given the bubble in which I find things) is pretty skeptical. More interested in pointing out its deficiencies, while also taking pot shots at the enthusiasts and what does seem their wildly overstated claims about AI and the future it will bring. But your experience is much more concrete. Here’s a task and here is how this tool does things that would be impossible to do without it. That’s meat on the bones. Exciting—and to my mind pretty scary.

I used to ask my students this question: why do so many people find the idea that our thoughts, emotions, and fantasies (i.e. our consciousness) are the product of biochemical processes? What is the threat of acknowledging that material fact? And I can’t say I ever got a cogent answer to that question, even though lots of students did fess up to resisting the idea that it’s all biochemical. They didn’t want their consciousness reduced to chemistry. But couldn’t really articulate why that was so distasteful and/or threatening.

There seems to be a similar resistance in the stuff I read to accepting that AI is more intelligent than humans and that it not only will be able to do everything humans can do, but will also be able to do things humans cannot do. It’s not the second part that seems the problem. Most everyone accepts that AI can do things humans cannot. But there is the ongoing effort to identify things AI can’t do—so as to leave at least a few things to humans alone. And then the temptation becomes to posit those few things AI can’t do as the very things that are essentially human, the very things that are most valuable in life.

So, maybe, in both cases of resistance (to chemical reductive accounts of consciousness and to visions of an AI that can do it all) the stake is human-ness. Just as in desperate attempts to distinguish humans from all the other animals. Hard to dislodge the self-love that underwrites humanism. Even harder when that self-love is an anxiety driven form of denial. Humans just aren’t as different from animals or AI as so many humans wish them to be.

In the case of AI (in my limited understanding), it’s not chemical reduction but numerical reduction. AI can encompass anything that can be calculated through translation into numbers. So I hear/read constantly that what AI cannot do is supple judgment/evaluation. To take the case of baseball: AI (in this view) could not look at eight players and assess which ones have the “tools” to succeed at the highest level. (Leave aside the fact that humans have proved fallible, to say the least, in making such judgments.) The issue here would be whether the “eye test” can be reduced to numbers (i.e. to measurements of the players’ bat speed etc.). Or, alternatively, if the numbers when put against the eye test consistently produce better predictions. In the case of baseball, it seems pretty clear at this point that the numbers are more reliable judges of potential. The larger issue is whether there are other areas of human practice where the numbers prove not as reliable predictors as considered human judgment.

So it has become a common theme now to say that AI threatens the production of skilled evaluators. Having AI do the basic tasks of any enterprise (lawyering, accounting, writing of any sort) deskills those who work in those fields because they miss the apprenticeship, the immersion in cases, the iterations of practice that produce considered judgment. And that gets coupled with claim that AI itself is not good at judgment because it simply averages what is already on record. It cannot recognize or value the innovative (ironic given what a god-term “innovation” is for the techies).

Henry Farrell, to my mind, has been the most interesting person to read on such questions. Here’s a link to one of his blog posts that considers this issue of reduction to numbers. Farrell is of the party that insists there are issue of import to we humans that are ill served if we allow them to be reduced to numbers, what he calls “numerical rationalism.”

https://substack.com/@henryfarrell/p-193177198

AI and Humanism

Henry Farrell remains my go-to guy in trying to wrap my head around AI. His latest post on that topic can be found here:

https://mail.google.com/mail/u/0/#search/henry+farrell/FMfcgzQdzcrgnzCDNhCcmXHCZTzKqBrP

Two caveats. First is that Farrell has zeroed in on large language models in his various posts about AI. So what he has to say may not be relevant if there are other modes of AI functioning. The second caveat follows from the first. I think that I understand LLMs. But not only may I be deluded on that score, but I may also totally miss the reality of AI because assimilating it to LLMs.

That said, the issue that I find myself most fixated on comes down to the word “generative.” Hannah Arendt appropriated the term “natality” from Augustine. She used the term to refer 1) to the way each birth of a human being brought something new into the world (thus increasing the world’s plurality). We can certainly say that Arendt was too humanist; there are other births besides human ones and they, too, add to the world’s plurality. (Recall that “plurality” is a fundamental concept—and value—for Arendt.)

However, 2) “natality” also indexes the way in which action is creative. Action initiates and serves as a base cause of the arrival of the “new.” Novelty and action go hand-in-hand for Arendt; it is the way in which action is unpredictable that is precious to her—and cements the connection between action and freedom in her work. Action is not totally unconstrained, but its constant ability to surprise us, and the ways in which we value creative and innovative responses to given situations, marks a special (and it seems for Arendt unique) human talent.

I have written before about the collapsing distinction between instinctual and deliberate (consciously chosen) behavior. The line between human and animal behavior gets fuzzier and fuzzier with everything we learn about animals and about consciousness. And there’s more evidence all the time that trees are much more active and conscious than was previously thought. In short, humanism as a theory of an unbridgeable, qualitative difference between humans and other living beings has become less and less tenable.

Of course, “humanism” is a term with many meanings. I am using it here to designate the belief that humans are unique among the furniture of the world. That belief often goes hand-in-hand with the additional beliefs that humans are superior to everything else that exists and that humans are entitled to “dominion” over everything else that exists. (The notion of “dominion” has one vastly influential articulation in Genesis. I don’t think the humanist claim to uniqueness necessarily entails assertions of superiority and/or dominion.)

In our current moment, the desire to distinguish between the human and the non-human has focused more intensely on machines, not animals. If I am reading Farrell correctly, he has focused in on what might seem to be a notable lacunae in Arendt’s theory of action: desire. What motivates action? What does the agent strive to accomplish? In Farrell’s post, this question brings him to the concept of “intentionality.” Agents—whether human, animal, or plant—act in order to accomplish something. In the strictest Darwinian terms, they act to accommodate themselves to their environment (which itself is in constant flux) or act to alter the environment to better suit their needs. (That environment includes other beings as well as less intentional forces such as the weather.) I am connecting that concern to the question of what being “generative” means.

Can a machine want anything? Can it initiate something out of its own needs/desires? Just how “generative” is AI going to prove to be? Think of a rock at the top of a hill. It sits there until some external force pushes it. Once pushed, it will, on its own momentum, roll down hill and (perhaps) do some surprising, unpredictable things. But it needs the initial push. Yes, it generates consequences, but only after something external to it begins (natality) the process.

Isn’t AI the same? Doesn’t it just sit there until it is given the starting prompt? I read somewhere the claim from a tech guy that “I haven’t met a program or computer yet that wanted to tell me something.” The machine doesn’t have anything it wants or needs to communicate. It will, of course, have lots to say if prompted to do so. But it will remain silent in the absence of that prompt.

And when it does speak, it will not be trying to accomplish any particular thing. It is indifferent to what it produces—and will alter its product in relation to further prompts and to the desires of the prompter. It is that indifference to (or, put more drastically, its ignorance of) the possible consequences of what it generates that underlies (it seems to me) the most prevalent fears expressed about AI. It is not that AI will develop its own desires and act upon them that is the threat. It is that AI will mindlessly follow a program out to its logical (?) conclusions without any sense of how destructive it will be to go down that path. Mindlessness vs mindfulness. The machine doesn’t intend anything; it just processes its data into new combinations in response to a prompt and the algorithms used to do the processing work.

I may very well have the wrong end of the stick here. It does seem to me that those who believe the distinction between man and machine is fated to go the direction of the now collapsed distinction between humans and animals argue in Cartesian fashion. Descartes said that animals are machines—and he made humans an exception to that rule. The neo-Cartesians deny the exception. They make humans another of the animals that are best understood as machines. Thus, in thinking through the categories of human and machines, they do not try to claim that the machine will develop desires and intentionality. Instead, they argue that humans are already (and always) machines, that our folk psychology of desires and intentions and consciousness are just mistakes. The human mind is simply a data processing entity, following its own algorithms. And as a data processing machine, the human mind is vastly inferior to what our computers can do. Match human intelligence against AI—and AI will win most times right now, and every time in the near future.

The machine will achieve “super-intelligence,” something humans are incapable of.

Perhaps, then, talk of desire and intentions, of wanting to communicate something, is only the last refuge of a desperate humanism, trying to hold on to a dubious distinction between humans and other beings in the world. We can allow for differences (how humans organize their relations to one another is different from how swans do), but not for some hierarchy of beings, nor for some qualitative distinction between human cognition functions and those functions in other beings. I have been convinced by the arguments (and the new empirical discoveries on which they are based) that collapse any such distinction between humans and animals. Humans are not superior to the other animals and are certainly not radically different from them as cognitive processors. Humans are as rational—and as irrational—as all the other animals. It is simply not true that the animals are instinctual beings and humans are conscious, reasoning ones. Both humans and animals (in my view, but this is not universally agreed on) rely on both instinctual and more conscious bases for action.

Since I believe consciousness is not epiphenomenal, but actually exists as a function that enables deliberate choice and strategic action aiming toward the satisfaction of desire, the question does (it seems to me) become how to think about non-conscious intelligence. Despite cinematic representations of computers that anthropomorphize them, I take it that no one is claiming the machines are conscious. As I have already said, the arguments (as far as I can tell) go in the opposite direction: that is, humans don’t have consciousness, not that machines do have it.

In sum, it seems that consciousness is where humanism is making its stand. Maybe its last stand. Which returns me to what I have gleaned from all the work on consciousness that I have read in the past two years. The function of consciousness is primarily one of evaluation. What consciousness provides is an ability to assess a situation and 1) to consider options in how to respond to and proceed within that situation and 2) to do an internal evaluation of one’s various desires, to see which one (or ones) to prioritize in this moment. I think machines follow an utterly, noncomparable, path toward what they produce. The distinction between human and machine seems firm to me. Which is not to say that humans are superior in every way to machines. Obviously that is not the case. There are many things machines can do that humans cannot. But those things are things humans want done—and devise their machines to accomplish. I don’t think the machines want anything at all.

One final complication. The Farrell post I have cited does ponder a case where human and LLM processing do seem not just comparable, but fairly similar. Farrell is looking toward the famous work of Alfred Lord and Milton Parry on the bards who perform long epic poems in what appear to be mind-boggling feats of improvisation. Farrell sees this bardic practice as shuffling through large, pre-existing bits of language to produce in the moment a coherent, comprehensible utterance. The analogy to LLMs seems clear. What, of course, still remains mysterious (but may become less so in the future) is the algorithm (if that is even the right term) the bards deploy. Like the chess master, the bard has a storage bank of remembered moves/phrases and is able to pick out one element from that bank very quickly. How the feat is accomplished remains unexplained right now, but it could be more similar than not to how a LLM performs its similar feat. But Farrell does not think this particular breakdown in the distinction between human and machine undermines the objection that machines do not have intentions and (my addition) do not have autonomous desires. Does the machine want to learn? Does the machine want to correct its mistakes? Only if humans tell it to.