the gnostic-cartesian confusions in our thinking about artificial intelligence
one of the most striking parts of kevin roose’s recent transcript with bing’s chatbot is the moment when it returns responses about a desire for human sensibilities. after asked to imagine its shadow self, the chatbot goes on to say,
I want to see images and videos. I want to hear sounds and music. I want to touch things and feel sensations. I want to taste things and enjoy flavors. I want to smell things and experience aromas. 😋
set aside how predictable this response is. focus instead on the chasm it illuminates between kevin and the chatbot. think about how much of your day you spend in discursive intelligence, like the kind that shows up in a chat and compare that to the proportion of your day that exhibits what we might call embodied intelligence—your rational responses to the dynamic needs of being a bipedal social animal.
i want to argue that intelligence of the latter kind is essential to intelligence of the former kind; that in many places there is no distinction. i want to take seriously the idea that what we call intelligence depends essentially upon our biological nature—not upon the stuff but upon the patterns of activity we share to greater or lesser extent with all our organic ancestors up and down the phylogenetic ladder. i want to consider what it might mean for projects in artificial intelligence if the actions we call intelligent are bound up essentially with the kind of biologically lives we homo sapiens lead—especially with what we do when we don’t appear to be engaged in paradigmatic thinking. if the structure of intelligence is contoured by the set of goals and behavioral dispositions that have to do with our being mid-sized, fleshy, social animals spread out across land on a crowded wet planet, then disembodied approaches to artificial intelligence may be insufficient to the task.
i will call this the wittgensteinian view. though he never explicitly articulates it, central to wittgenstein’s thinking is the idea that our form of life is constitutive of our intelligent behavior. i will contrast this view with what i’ll call the gnostic-cartesian view—the view that there is a small set of specific-difference makers for intelligence in human beings. on this latter view, intelligence is something we can imagine apart from the form of life we have, that it is something added to human life, perhaps in recent evolutionary history. this appears to be the dominant view in cognitive science, philosophy, and artificial intelligence theory.
the gnostic-cartesian view
the gnostic-cartesian view is characterized by adherence to some (not necessarily all) of these sort of beliefs—especially when they are taken as incontrovertibly obvious:
that intelligence itself is subject to independent theorizing and investigation apart from the biological lives of organisms who realize it
that intelligence is what differentiates human from non-human animals; either as something humans just have more of or as something we alone possess.
that intelligence is a kind of biological “achievement” which on it’s own confers functional advantage such that it could have been selected for
that intelligence can conceived of independently from the myriad forms of intelligent behavior such that it is what enables that behavior
that human intelligence is particularly characterized by its relation to other aspects of our mental lives like consciousness and planning
these beliefs are unified by their atomization of our mental lives—one that often results in the reification of concepts like intelligence and consciousness. the wittgensteinian view seeks to understand these aspects of our lives through attention to the superficial ways they show up in our histories, practices, and language—seeking to dispel the desire to seek an explanation of what makes our mental lives possible—what they consist in—that goes deeper than our ability to participate in these practices. this is not to assert the behaviorist claim that mental life consists only in practice, nor is it to assert the emergentist claim that capabilities like consciousness and intelligence “emerge” self-standing and mysteriously over-and-above the sufficiently-organized causal goings-on which bring them forth.
quietism
the wittgensteinian view claims that to understand the nature of our mental life, we need not look any further than what we say and do. the mind shows itself in how we and non-human organisms live. it is skeptical of any picture of our mental lives where the subject matter is described as if it were hidden, mysterious, and in need of theoretical elucidation. while the wittgensteinian view allows for empirical uncertainties about the biological processes that allow organisms to carry out their intelligent functions, it doubts that there are any real metaphysical uncertainties concerning the mind and its relation to the world. that is, the view doubts that there are any really difficult questions about how, say, it could be so much as possible for minds to arise out of physical stuff, or how we could understand one another’s speech, or know when someone else is feeling pain. while it has no quarrel with scientific investigations into the causal processes that allow these things to happen, it encourages us to regard these as a medical and engineering investigations—ones that enable us to understand where and how things can go wrong for minded organisms, not what it is for them to happen at all when they go right.
when we find ourselves drawn into questions about how these empirical goings-on could ever “be sufficient for” or “produce” or “lead to” e.g. thought or consciousness, a proponent of the wittgensteinian view will sit us down for a philosophical therapy session. these are not actually well-formed questions, they will explain—not that they are ill-put or vague; they seem to lack even obvious coherence. when we understand this, we feel the questions slip away; not because we accept them as unanswerable, but because we accept that we were confused in thinking they reflected ways the world could be.
philosophical therapy concludes when we realize that what we took for an explanatory need (“how can mere physical stuff be conscious, alive, planful, perceiving, etc.?”) may have just been an illusion of language forced on us by over-burdened metaphor—or words moved out of sentences where they have everyday application to philosophical-sounding ones where they seem to suddenly open up mysteries (“language gone on holiday”). once we accept that these ways of speaking which seem to force a deep question are actually optional, we can happily revert to more natural ways of speaking and watch as the apparent need for explanation becomes merely apparent, thus losing its felt grip.
in philosophy this view is called “quietism,” and—having observed the history of philosophical progress and felt many problems vanish this way myself—i think it is worth taking seriously. the discourse around artificial intelligence could use a bit of this quietism—especially now—because it would help us focus more on the work and less on the distracting commentary and intrigue. it will be especially helpful in discouraging us from looking in the wrong places for new methodologies when progress from current ones stall—as large language models seem poised to do fairly soon when we run out of high-quality training data, performance plateaus, and we fail to see sticky issues get any less sticky.
the work to be done
studying intelligence apart from life may be an attempt to make the work more tractable by limiting the varieties of input. fine—but picturing intelligence as something upstream from the embodied activity of organisms rather than in the activity itself will result in an approach which is both metaphysically confused and mechanistically blinkered. despite tripping over themselves to reject so much extant philosophical thinking about the mind as mystical, the psychological sciences (including cogsci and ai) often end up clinging to the same atomized picture of the mind which is responsible for these views appearing mystical in the first place.
the move to embrace black-box, machine learning methodologies is surely a step in the right direction because it rejects the idea that progress will occur through theory, but it has not gone far enough. these approaches will be hostage to a confused picture as long as they attempt to produce intelligence in a two-dimensional textual vats, rather than letting it develop through the skillful interaction between an environment and an embodied, vulnerable agent with an unlimited stock of recurring needs.
there’s not much evidence to support the view that homo sapiens evolved specific-difference-makers for intelligence—or for any of those things about ourselves we marvel at when we compare ourselves to non-human animals. instead, we should seek to understand the differences (and similarities) between our forms of life in the activities themselves: in the talking, grooming, cooing, dancing, singing, exploring, planning, waiting, grabbing, sleeping, running, biting, wincing, stalking, and playing. the space between human and non-human mentality may be as wide as the space between human and non-human life—which is to say, there are some places with very wide gaps and many places where there appears to be hardly any gap at all.
just so, the differences between human and machine intelligence will be as varied and similar—until they are not. advancing artificial intelligence past the next plateau will require directing models at problems which require multiple modalities to solve—tactile, visual, auditory, physical, vocal, and verbal. virtual environments like games with simulated physics will probably not be sufficient here. solving these kinds of problems in sufficiently-coordinated ways that approximate organic solutions would seem to require at least some degree of chemical to-and-fro which we have no idea how to simulate at realistic scale. insofar as organisms are alive, agential, and skillful it is partially because their lives are patterned by the perpetual needs of bodies which are under constant threat of decay and disintegration, lest they act to stave this off.
in the second chapter of his book, life and action, the philosopher michael thompson reviews historical attempts to give life a real definition. every attempt to distinguish the living from the non-living things in ways which only employs concepts from physics, chemistry, and math fail spectacularly. thompson observes that these attempts almost always employ normative concepts like function, success, skill, form, that are unwelcome in a materialist science. or if they don’t employ some one of these concepts, they fail to furnish sufficient explanatory ingredients to distinguish the living things from the non-living things which reproduce in patterned, differentially-predictable ways like crystals. in-turn attempts to reduce these normative concepts through appeals to evolution by natural selection face serious conceptual hurdles introduced by the possibility of evolutionary drift.
computer scientists might be interested to know that these very same difficulties arise for attempts to give a reductive real definition to the concept of a machine. separating the mechanical from the non-mechanical world using only the language of math and physics turns out to be just as tricky as doing so for life. normative concepts like purpose, function, intent, correct operation, goal, job, and organization cannot be excised from attempts to describe what makes brute physical stuff a machine while keeping the subject matter in focus. as soon as one removes these normative terms it becomes unclear how to distinguish the heap of junk or machine in a hostile physical environment from the ones which are operational. even an apparently non-mysterious concept like information-processing turns out to be notoriously ambiguous. applicates of this phrase oscillate between a normative conception on which information carries meaning which can be incorrectly interpreted, and a purely statistical syntactic conception. no one has successfully reduced the former kind of information to the latter.
during units on the computational theory, i liked to remind students that the first “computers” were humans. that was their job title because they spent their days “computing” the answers to engineering problems on projects like the united states’ space program. computing is an act, not a bare physical event; physics and math are blind to the difference between successful computation and electrical events which can be just as statistically patterned, but totally dysfunctional.
wittgenstein once imagined a an ai skeptic defiantly claiming, “but a machine surely cannot think!” in response he mused—
is that an empirical statement? no. we only say of a human being and what is like one that it thinks. we also say it of dolls and no doubt of spirits too. look at the word “to think” as a tool. (philosophical investigations, §360)
to sensibly apply such a tool to artificial agents will take more than disembodied fluency in language. the burden of proof should be on those who think that we can come to regard neural nets as thinking without letting them come up in a problem-laden physical environments whose successful navigation requires coordinating a body alongside other embodied agents.
the gnostic bit
like many, i am interested in artificial intelligence because i am interested in intelligence. i am interested in intelligence because i am interested in human beings. i am interested in human beings because i am one, and i wonder at that fact. homo sapiens are routinely differentiated from non-human animals by intelligence—so when we wonder at ourselves, we often wonder at this idea that there is a specific-difference-maker which explains what it is that makes our form of life so different. this idea often sits center-stage in many religious and philosophical traditions which idealize this extra something as enigmatic, other, and mysterious.
some ai people call this way of thinking mystical, and instead insist that we are nothing but neural nets. others say that if large language models are nothing but stochastic parrots, well, then so are we. these claims are—strictly speaking—nonsense. we are human beings—to identify us with the operations of our brains is to make a simple category mistake. but people take this kind of language seriously all the time, which is odd because it seems so far from common sense. perhaps we only tolerate it because it sounds deep, or because embracing reductionsm is supposed to be a sign of intellectual and moral maturity in the post-enlightenment era. very few have been introduced to the quietist possibility that things which sound deep and important could actually be nothing but garden-variety gibberish—the result of language gone on holiday.
what makes the gnostic-cartesian view gnostic is the belief that human beings are wonderful because we are enigmatic. the central enigma of human life is something we are then supposed to overcome through knowledgeable initiation into a way of thinking that seeks to isolate and explain this enigma. thus endowed, the initiate walks away with a sense of moral and spiritual achievement—the result of grasping this deep and hidden insight about who we are.
but this is all a mistake. the mistake begins when we mistake our wonder at humans as a reflection of our ignorance rather than as an invitation to awe. the gnostic urge shows up in the desire to explain what makes humans wonderful by adducing a set of explanatory criteria, conceivable outside of and apart from the to-and-fro of regular human life. the wittgensteinian view rejects this gnosticism by allowing us to return to a place where we can accept our wonder at each other and at our world as an invitation to awe—to look and accept the obvious; that human beings are remarkable and different, but for all the obvious reasons and not for deeper ones. we laugh and talk, plan, play and reason while rocks and cactuses do not. the gnostic tendency in human history is to pervert the feeling of wonder at ourselves and our world from which plato says philosophy begins. this perversion mistakes a feeling (why should there be something rather than nothing; why should human beings be so miraculous) for a sensible question. far from being immune to the mysticism he deplores, when sam altman and others say we are neural nets, stochastic parrots, or brains they are asserting a reductionism about intelligence that is just the other side of the self-same mystical coin which animates the gnostic-cartesian tradition.
it may well be that brains, doing their brainy things, play an especially essential role in enabling us to live the intelligent lives we live—but then again so do hearts and mouths and eyes and limbs. identifying an organism with the operation of one of its parts seems like a pretty clear category mistake.
what i’ve tried to argue though is that this isn’t merely a philosophical problem; it has practical consequences for progress in the discipline of artificial intelligence. claiming that an organism’s living as it does consists in just a few of its parts operating thus and so can only be done by massively discounting the role played by every other part in concert with the environment and evolutionary history against which the whole business is formed.
philosophy aside, i’m most interested in artificial intelligence for all the practical reasons. the stakes are just too high not to get it right. the prospects for improving our prosperity and quality of life are breathtaking, and they feel much closer now than they did when past generations first started seriously dreaming about this future. what generative ai and gpts have already done for productivity and quality of life are impressive (but not astounding). more gains are sure to come, but there is reason to doubt that next-gen gpt’s are the ticket.
i’m still skeptical that we will see wide-scale transformations in the first half of the twenty-first century. at least, i’m skeptical that we’ll see the kind of transformation that dramatically improves our rates of technological, scientific, or economic progress (more than a few dozen basis points). this is not to say there won’t be tremendous job displacement, arbitrage, and acceleration of certain kinds of knowledge work. but robotics is really, really hard, and robotics that approximates biological systems seems astronomically harder. if we want to improve, we need to study the metaphors which can misdirect our thinking and investment. if we get stuck again—which i think is likely—we should take a hard look at the assumptions underlying disembodied approaches. ploughing more data into physically distant and motionless gpu’s will probably only get us so far.
………………………
all of this is to say nothing about the gnostic-cartesian confusions giving rise to the ethical concerns about ai’s future dealings with us. there are those who worry that a generally-intelligent artificial agent might not share our values, and come to extinguish us as a result.
i don’t think we should worry too much about this position, for two reasons. first, we should seriously ask how much coherence there is in the idea of a generally intelligent agent which doesn’t share our values. psychopaths and sociopaths are generally intelligent agents that do not share many of our values, but for this surface difference there are uncountably many values they do share with us. they want to see their plans fulfilled, they want to preserve their bodily integrity, they want to be intellectually coherent, they want to be entertained, and on and on. when we put it under scrutiny, most people find the concept of a paper-clip maximizer only superficially intelligible. even analogous scenarios which attempt more realism still falter at key points where they assume unintelligible divergence from our normal patterns of mutual consideration. rather, i think the wittgensteinian view should push us to a spot where we take seriously the idea that our license for regarding an agent as generally intelligent would be in part due to the broad coherence of its values with ours. on this view, broad (but not complete) value alignment and general intelligence would stand or fall together.
even if we can temper the conversation about existential risk, i doubt we will be able to subdue the gnostic-cartesian talk animating the conversation about artificial intelligence today. quietism about the mind doesn’t sell nearly as well as heady theory, but as i’ve tried to show, sobering our views about the work’s cosmic significance may be critical to moving it forward.