How Might a Machine Intend?

Winner, Most Distinguished Preceptorial Essay, St. John’s Graduate Institute, Academic Year 2022 

Published in Colloquy, Spring 2023

What would it mean to say that a machine can “think?” There is a sense that machines are able to go through a logical process similar to what we describe as reasoning; this is true in very complex computing systems as well as simple electrical systems. They can reliably generate outputs based on certain inputs, much like a responsive sensory organism.

However, these are purely physical processes, which is typically understood to lack whatever it is we mean when we say “mind.” A home furnace doesn’t “know” it’s being called by a thermostat, nor does an elevator “know” I’ve addressed it when I press the button in the same way that you know when your name is called, or what Paris is, or what ice cream tastes like; the machine is wired in such a way that a current is routed to activate or deactivate the machine based on signals the designer has attuned it to be responsive to.

Yet the question persists, can machines think? Can they understand, or generate their own goals or ideas? Can they be taught to be truly spontaneous, can they learn to feel? Or, will they only ever be capable of that mere formal mechanized “reasoning,” with enough instructions that we can trick ourselves into seeing minds where there are none? Let us first turn to this question.

Can a Machine Understand?

Philosopher John Searle talks about the key difference between minds that form understandings and machines that execute formal, syntactical processes as being the ability to have intentions, which he posits are likely to be uniquely linked to biological life.[1] His thought experiment, the Chinese Room, compares the outputs of a computer that can have a perfectly-convincing natural-language conversation to a monolingual English-speaker operating on Chinese symbols by executing the instructions from a book that contains the right answers to any prompt.

Searle grants that this setup would be capable of passing Alan Turing’s famous Imitation Game,[2] and would create a convincing impression of understanding. If you provide a simple story in Chinese about a person going to a restaurant and enjoying a meal, then ask if the person paid the bill, the executor with the use of the Chinese book can reply “yes,” despite the fact there’s nothing that logically necessitates those things both be true. This is the kind of information we’d normally consider contextual, and indicative of understanding and intelligence. Under these perfect conditions, Searle asserts that no matter what, the executor still does not know Chinese and thus can never understand anything. All they are capable of is recognizing and manipulating the symbols themselves.

This kind of formal reasoning is different in-kind to whatever action is happening when humans interpret. The executor is never freed from the room to see the outcomes of their work, and so can never learn what the symbols do. Thus, Searle argues, it stands to reason that they can never be said to form an intention,[3] merely a calculation; whatever incredibly large and nuanced set of inputs they receive, they can only use a syntactic logical process to arrive at the correct output.

Is this actually analogous to machine reasoning? Algorithmic problem-solving bears much in common with this account; the actual concepts that underlie whatever the machine is manipulating symbols in reference to are semantically “empty,” as you can just as easily substitute the units it uses with something nonsensical and the computer will still follow the same logical steps.[4] A concept need only be a label here. In fact, a computer will always have an incredibly exact, precise definition of any term, with a definite address for that definition it can query for information; but it will only define what it means to the calculating system, so context is not necessary beyond how to operate on it.

Qualitative understandings that humans possess are not structured like this. Marvin Minsky explains this gap as the result of a “basic fact of mind: what something means to me depends to some extent on many other things I know.”[5] We do not have rigid formal definitions for every term we use when we speak; minds are elastic, and our concepts are ambiguous. A formal reasoner needs those hard definitions and for their operands to be arranged logically, but humans are often logically unclear to outright incorrect while still conveying the proper meaning. The computer’s connection to a question is totally intrinsic to the asking[6] – it would not have many different ways to understand “mind,” or “know,” for instance – whereas we have no special command over the world or direct connection to the information we are trying to apprehend. We are left with a much more difficult task: to uptake the question from our environment and the context, and learn from our many inevitable errors. To find any answer, we first have to invent the question for ourselves.

Perhaps, then, the semantic process is special to brains. What if we simulate a brain? Let us assume that a program can render a re-creation of a human brain as it acts in the body, down to the level of neurotransmitting molecules even, that is so high-fidelity that any mistakes in the model are indistinguishable from normal variances. Searle claims that any understanding the computer might have is only a simulation of understanding in the same way that a simulation of a rainstorm is not really a rainstorm; we can only ever replicate its formal elements to make representations. Programming is simply us elaborating an algorithm to jockey around empty symbols, which we the creators can then attach meanings to. The computer will never understand the semantics of whatever it may claim to know, just as the English-speaker will never understand Chinese.

This argument places a lot of weight upon a specific view on meaning of “simulate.” I think Searle is correct to be skeptical of considering the mind as separate from body, what he calls a “residual form of dualism.” There is a deeply-ingrained assumption that mind is both conceptually and factually independent of the body, and we should not be so quick to dismiss that, “actual human mental phenomena might be dependent on actual physical-chemical properties of actual human beings.”[7] However, the characterization he proposes places mind in a purely material realm, when in fact mind also has to do with information. It is one thing to simulate a physical action, and another to simulate a representational action. We can better understand this by examining the question: what is information?

Information is indeed a material phenomenon, but not itself a material. It would seem it is more like an arrangement of states in materials that can represent something else. When we talk of “1’s and 0’s,” “bits and bytes,” these are abstractions for the state of many tiny switches that encode meanings; yet computer memory is just as physical as a book. However, information is not dependent on the kind of states it relies on to represent; Hamlet is still Hamlet on a Kindle or etched in stone. Further, information and representations are materially dependent; if you destroy every copy of every version of a book, the ideas represented in them are truly lost. We know this is equally true of the mind; destroying portions of the brain takes away the ability to perform mental operations, including memory.

On the other hand, representations are themselves unlimited – we can represent something that has not, cannot, or will never exist without pushing the material limitations of the medium the information is stored in. We can describe impossible things, or even things that would just be materially difficult to accomplish physically, with the same ease as describing something totally mundane.

What is a human doing when forming intentions as Searle presumes we can? If we accept this picture of information, it would appear the process looks something like converting Action in the world into Information, then turning that Information into novel Action. The Information in the mind is some arrangement of brain matter, and mental processes like perception and thought alter those arrangements in order to represent.

Simulating this process in a different medium would involve altering the arrangements of the switches in, for instance, a hard drive, but the underlying goal of those rearrangements would be the same; the only difference lies in the arbitrary encoding scheme used. The resultant act of mind is truly virtual, in the sense that it expresses the essential feature – what a mind does in virtue of being a mind – even when lacking the material components that typically enable it. Thus, when we talk about understanding or intending, the critical question we must ask to answer Searle’s challenge is: does a brain think, or does a mind think? If it is a mind that thinks, then we must concede that it is neither fully material nor mystically nonmaterial. There is no reason to believe that a computing machine can’t perform this process.

How Might We Recognize an Intending Machine?

Simulating a brain may be able to accomplish the process of creating a fully-realized mind, complete with intentions and understandings. However, that likely isn’t the most efficient route to a conscious machine – in fact, merely reproducing something that we’re certain can create a mind tells us very little about the essential nature of either thing. Humans create minds all the time via reproduction, in fact, but we are no closer to answering the question: why does some matter think and other matter does not? If we want to provide an adequate account, we need to be able to answer the Composition Problem – the issue being that non-thinking matter can’t simply emerge into consciousness from total unconsciousness; whatever factors enable mind must be explainable in terms of what we know about that lower level of matter. Finding out how a machine can form novel intentions in its own way, as itself rather than due to its resemblance to a human, may be able to teach us something important about ourselves as well.

In discussing the essential nature of a consciousness, we are in a way asking about the concept of the soul. Apart from all experiences, contingencies, and even specific processes, what is that “spark” we’re referring to when looking at another conscious being? How might we recognize it if we saw it? The Cartesian ideal of the soul draws a near 1-to-1 relationship between the ability to rationalize and the “spark” that reflects the divine; an immaterial ghost-in-the-machine that links us to the creator and is the site of mentation. Aristotle proposes a more embodied form of soul as the animating force behind all life, and thus all living things possess “psychic powers.”[8]

In this model, plants have an appetitive soul, animals have an irrational soul in addition, and humans have a rational soul on top of these. The ability to move, grow or decay is the most fundamental, and it serves a nutritive function; a living thing’s pure striving for the means to go on living. Sensations delineate the irrational souls of animals from plants, but this also can include emotions and other highly-developed perceptions. Aristotle credits both of these “lower” souls with the “originative power”[9] [10] that would constitute an intention; a living body must contain within itself the ability to generate motions, and not merely move when it is acted upon.

 Aristotle separates mind and thinking into the rational soul, believing, “it alone is capable of existence in isolation from other psychic powers.”[11] This would seem to be accurate of machine thinking. Computers separate purely rational syntactic processes from the need for other sorts of mentation; but we’ve also established that, thus far, this has prevented it from understanding and intending the concepts rational thought can operate on. If this is somehow the highest form of soul, are we comfortable attributing the term to our PC? Or perhaps it is missing something provided by these other sorts of soul?

 Let us first consider the idea of an appetitive robot. Picture a machine sent to Mars for a long-term research mission and given the ability to harvest its own solar energy to power itself. Upon landing, it runs the subroutine to set up solar panels, and continuously monitors its own battery level to know when it’s time to go recharge. This is a sort of “hunger,” yet it still seems to be missing something. The Aristotelian model specifically places sensations like hunger that create pleasure and pain in the higher irrational soul, so it’s not necessary that the robot “feel” it so long as it has an awareness of its nutritive needs. Like any other computer, it embodies the “rational soul,” and now has the capacity to move by itself to “feed” itself as much as any living thing does – expending energy it contains in order to act as a whole – but intuitively, we’re still waiting on the “spark.”

Is this because we don’t view this as fully “originative?” Where is the novelty in its intentions? Again our intuitions point us to a distinction between merely executing prewritten instructions and intending something, but at the same time, a close examination of the ways humans make their decisions with high-level regularity and in accordance with causal physical laws begins to make that distinction uncomfortably gray. Further, that would open the door to denying that purely-irrational animals acting on instinct possess a soul or consciousness, since these are also in a way predetermined. How might we explain this discrepancy?

At time of writing, the robot (and any other program) is created when a computer programmer undertakes the representing function, supplying their intention and understandings, for the purpose of fulfilling their functions. A piece of accounting software is not being written to generate novel ideas on the nature of numbers, nor being asked to understand the cultural value of the money it records. The designer (and later, the user) provide the program with Information, and in response receive an Action. It’s perhaps most accurate to say that the program is itself a representation, at least in some respects, and not capable of creating its own. If its been supplied with an analogy, it is only following; it will never respond with one of its own.

We seem to be closing in on a soul. In order to call something a conscious mind, perhaps what we’re looking for is that process of turning Action into Information into new Action. It is in the Aristotelian irrational soul where this occurs, at the level of sensations and perceptions. As discussed, a purely rational syntactic computation has direct access to the precise meaning of any piece of information. Otherwise, it cannot function. Words like IF, THEN, GOTO or PRINT may leave room for ambiguity to a human, but can only ever mean one thing to a machine – a computation can never be asked to invent a question, because it’s never making a qualitative judgment. Conversely, living things that sense and respond to the environment must intuit information, must write and solve its own computational problems, and only then can they execute those instructions.

How Do Machines Perceive?

The miracle of brains are their ability to form perceptions of the world around them with lightning speed, which has traditionally been an obstacle for computer intelligence. They’re capable of doing this using neural tissue – hardware that is magnitudes slower than a computer chip – to perform an incredibly complex calculation that simultaneously recognizes the whole of a thing and compare it against every other memory of other perceptions.[12] This is an astounding display of the power of evolutionary dynamics.

Already we are beginning to see machines that make very complex perceptions and judgments, such as LiDAR demonstrations on self-driving cars. They are also supremely expensive, resource-intensive, and difficult to understand even for the humans who develop them. We’ve managed to make complex behaviors out of complex foundations, but we still lack trust for these systems because of it. Complex systems are inherently more prone to failure because there’s simply more things that can go wrong. It’s one thing for a CAPTCHA to misidentify a traffic light, but it’s another thing entirely if your car does. Within the very specific fields that these systems are designed to operate in, we lend their judgments about as much credence as a precocious seven-year-old at the wheel, and that’s likely a wise decision.

One major reason programmers can’t easily understand their systems’ behavior is because often they did not design them, at least not fully. Fashioned after the process of Natural Selection, many AI researchers are turning to genetic algorithm design, where random strategies and code are combined with each other, and a pool of randomized “offspring” is created of the most effective ones. Over many generations, powerful algorithms can be developed that a designer could not have foreseen. The machines are learning, in an abstract Darwinist sort of way.

Through this genetic process, it is foreseeable that we could make a machine that can change its own source code of its own volition. Currently, we tend to design systems with a specific purpose in mind, and we consider that purpose foundational to the program’s being. In Aristotelian terms, programs verifiably have a telos; considered another way, they do not create the problems they are tasked to solve. However, there is no conceivable reason that a machine could not be given the power to change its own purpose, since that purpose is ultimately a representation that can be combined and recombined on command. It would thus be capable of altering its goals in response to its experiences, changing Action to Information – the mark of intelligence. Intention, then, seems to be a form of learning.

However, if Google let loose such a program on all their data and it managed to self-organize into an intending, understanding and learning being, it’d be just as unlikely to be able to explain how it did it as we would be if we examined its code for the answer. Already we have a hard time explaining how these systems are making basic perceptions. Decoding complex behavior emerging from complex foundations is, as established, inherently difficult. Further, insofar as the machine participates in this enterprise we call mind, it will likely have similar difficulties accounting for itself as we do providing such an account of ourselves. As Charles Peirce put it:

[The] senses… [furnish] something more than plain, unvarnished facts of the outer world, [and no] direct scrutiny could enable us to say what part of that which we seem to see or hear is due to stimulations of the nerve-terminals of our eyes and ears and what part is a quasi-inferential interpolation of our own minds.[13] [14] [15]

This self-organizing program may be capable of consciousness, but like the brain simulation, it wouldn’t tell us anything about consciousness; we’ve created matter that thinks, but we still don’t know what makes matter think. The issue here is there’s no observable gradient from complex systems that make simple judgments to complex systems that have bridged the gap to complex judgments; they’re all equally black-boxes to us.

Let us look instead to nature; how do cells, simple organisms with no individual capacity for mind, work together to facilitate complex judgments? The aforementioned miracle of the brain is not only one of evolution; it is a miracle of computing, using the combination of many small processes in-parallel to efficiently ascertain information from the environment.

Ultra-simple computer programs are also capable of these kinds of parallel processes. Melanie Mitchell outlines a use of genetic algorithm design on a one-dimensional cellular automata program that applies a handful of simple rules to determine from one line to the next whether a “cell” will be black or white based on the corresponding neighbors in the previous line.[16] The rules themselves are chosen based on their fitness for answering the question, “are there more black or white squares in the starting conditions?”

After eighteen generations, the ruleset that had been generated would turn all 148 of the cells black or white in answer to this question inside 148 iterations and with 80% accuracy on the hardest cases, when the starting conditions were close to 50/50. Its accuracy increased the larger the discrepancy, much it would for a human “eyeballing” the problem.

[17]

 What is astounding about this graphic is we are seeing a machine running extremely simple parallel computations to answer a question it has no definitional link to. In other words, we are asking it for a judgment:

The black-white boundary and the checkerboard region can be thought of as signals indicating ambiguous regions. The creation and interactions of these signals can be interpreted as collective information processing performed by the [cellular automata]… [The] collective computation… is determined by its overall space-time behavior, and is thus almost always impossible to extract from the rule table. [18]

It isn’t difficult to envision the use of a simple ruleset that can make this kind of judgment. Given a black-and-white photo, a cellular automata structure could analyze each line of pixels and begin deriving information with little computing power. The ruleset was derived through chance in a short time, simply by choosing favorable traits; there’s no reason to believe these sorts of structures can’t be selected to answer other qualitative questions about images, or other kinds of information sets. There’s also no reason the squares could not be colored,[19] greatly expanding its possible uses. Simple computers, working in parallel, are already capable of forming new information about their environments in-time.

What we’re witnessing in the image, then, is a basic and clear-cut act of mentation. No particular simulated cell has a mind, yet in combination they’ve managed to reverse the flow of computation – turning the state of its world (analogous to our concept of Action) into Information. Once Information has been created by something, the raw materials exist for novel Intentions; although these cellular automata have not been tasked with making decisions based on their judgment, it is now conceivable that many more of these interactions of parallel computation can use those judgments to begin deciding. By explaining the proto-mind of simple computers, we have begun to resolve the Composition Problem.

Who Else Computes?

What we have established is that computation is a basic form of mind; nonintelligent matter can compute, but intelligent matter must be able to create and resolve computations. Computing is not itself sufficient to be a living, thinking being that forms intentions, but it seems it is necessary for those that are. The marker of life, intelligence, or “soul” is the reversal of this flow – making new Information from Action – but is reliant on the proceeding conversion of that Information back to Action to make that information do anything, and to act in the world.

Humans are certainly computers then, at least in part. Animals, too, are computing. Even appetitive beings who do not have enough complexity to support a neural system are capable of changing Information into Action; from single-cell organisms[20] all the way through plants, life reads and executes DNA instructions to constitute, feed, and reproduce themselves.

This comparison of DNA to computer instructions isn’t merely incidental. Seth Lloyd refers to this as one of the earliest Information Revolutions,[21] and that is no metaphor. DNA is fundamentally an encoding scheme for information that contains executable instructions for heritable traits, behaviors, morphology, and most importantly the instructions to construct the very machine that will execute those instructions. If human speech and digital computers are made powerful by the ability to make recursive statements,[22] DNA is powerful for the very same reason.

This is a radical proposition. Saying that DNA is a computation entails that some matter in nature is suitable for computing. Computing is then more fundamental than life itself, and if it is a sort of proto-mentation, then it firmly places the grounds for mind outside of biology. Can we go any further with this? What other matter computes?

The fundamental building blocks of the universe, subatomic particles, are a candidate for carrying information. They are themselves physical (analogous to our previous uses of “matter”), and capable of holding arrangements of states that facilitate interactions. For instance, when electrons, neutrons and protons interact, they form “outputs” that consistently display structures of the basic elements, which in turn interact to form molecules. Every “query” between an oxygen atom and two hydrogen atoms results in the same “answer,” a bonded state suitable for further calculation.

Given our expanded understanding of Information that subsists in the material and commands Actions, it is difficult to draw a distinction here that forestalls the conclusion that all atoms in nature compute. Further, there are already designs for computers that use the essential qualities of atoms and particles – quantum and molecular computing – that present only engineering challenges, not theoretical ones. Just that fact on its own entails that all matter is already suitable for computations.

 If atoms are a computation, what’s the computer? The universe itself, stemming back to the likely-story of the Big Bang, appears to be an unending chain of these computations; is the universe the computer? To our best understanding, to be a machine that executes instructions, it needs the capability to Read, Write and Address: that is, to assess the state of a unit of information, update that state if necessary, and to know where to go next. It’s in this last step that we see the importance of time in computing – addressing requires futurity to use the concept of a “next,” and it needs the power to move to the next stage of the sequence temporally and with purpose. Time is the engine that drives a computer; this is why the most basic component of the Digi-Comp is the clock plates, and the clock speed of a microprocessor is directly correlated with its power.

Understood physically, time is a thermodynamic phenomenon – the dissipation of concentrated energy is irreversible, and pushes the forward motion of time. Any attempt to re-concentrate energy necessarily creates more disorder in the scale of the universe; burning fuel to charge a battery creates a concentrated locus of energy, but will always lose some to sound, light, and waste heat along the way. This is what’s called entropy, and it sets the clock of the physical world to its drumbeat. There are ways to speed it up and slow it down, but as far as we know, there is no way to set it in reverse.[23]

Every subatomic interaction results in more entropy, just as any calculation actually consists of three things: input, output, and time. Non-living and non-computing matter doesn’t have much use for time; things that only act when acted upon have no need for futurity, they have no drives. Similarly, a single bit in a digital computer doesn’t have use for time either; whether it is in the 1 or 0, On or Off state is of no real consequence to it. If, again, the physical interactions in the world are creating and using information, the physical objects we observe are the data points, but the driver is that which makes use of time itself – i.e., the universe itself. It makes use of Information to generate Actions.

Minds and living things make great use of time, however. Time has an incredible evolutionary advantage baked into it – for beings that deal with information, there is a massive efficiency measure in not needing to know everything all the time.

A heat pump’s capabilities are embodied in its wiring, and it is not capable of being reprogramed without being physically reconstituted by an external force; but given that it’s not very complicated, this is not necessarily an issue. A tic-tac-toe game made of switches with a strategy pre-wired in cannot adapt that strategy; it may be advantageous to be able to react in-time, but the game is uncomplicated enough that this doesn’t prevent it from functioning reasonably well. Chess, however, can legally consist of 1040 possible states. A pre-wired chess machine need not necessarily account for all of these, as any strategy will reduce the possibility space as it goes on, but that’s still a lot of wires.

Computers and intelligences bypass these requirements by instead recognizing patterns and responding according to a set of rules. By not needing to know everything all the time, they’re able to focus their energy and resources on making the best decision now. But this also requires some sort of time; minimally, a memory, where patterns that have been encountered can be applied. It needs the capacity to change without changing the fundamental nature of itself; it needs to be a universal machine.

When something is capable of directing those changes itself, it is applying that memory – it is learning – for the purpose of generating a novel Action – it is intending. What else generates novel actions, based on the states of information from the past with the goal of driving to the “next?” What is doing this under its own power, supplying its own intentions? If these are the markers of mind, it is again difficult to forestall the conclusion that the universe itself is capable of it. All the capacities we have discussed that separate computers from other matter apply to it, and all the capacities that separate intelligence from computers also apply. The answer to the Composition Problem casts the raw materials of mind out into the universe, but their completed, complex totality is embodied as a single whole by the universe itself. How might a machine intend? By emulating the world around us.

[1] John Searle, Minds, Brains and Programs, “VI. The Many Mansions Reply.”

[2] Alan Turing, Computing Machinery and Intelligence, §1.

[3] Intention here can be understood as a novel, self-generated impulse towards some action. Searle makes no attempt to explore what it is to be capable of “intentional states” and instead stipulates that humans must be capable of them. I supply this definition to account for the “causal powers” he attributes to the mind; a thought is not a simple If X Then Y formulation, and thus is closely linked to understanding the context and broader consequences of any “input,” which then factor into a decision.

[4] This is similar to valid but unsound arguments in formal logic.

[5] Marvin Minsky, Why People Think Machines Can’t.

[6] In the broader philosophical discussion, it can be said that computing machines have a kind of Direct Reference à la the Logical Positivists, where meaning is contained in the referent and names essentially denote.

[7] Searle, ibid.

[8] Aristotle, De Anima, 413a20.

[9] Ibid, 413b.

[10] Aristotle’s choice of words here is particularly interesting. See Ada Lovelace Note G, “Notes by the Translator” to Sketch of the Analytical Engine Invented by Charles Babbage; see also C.S. Peirce’s solvet ambulando, “Logical Machines,” American Journal of Psychology Vol. I, No. 1, pg. 167.

[11] Ibid, 413b25.

[12] Daniel Hillis, The Pattern On the Stone, pg. 114.

[13] Charles S. Peirce, “Our Senses As Reasoning Machines,” The New Elements of Mathematics.

[14] This idea appears frequently in the literature. See Turing ibid, §6(4); Hillis ibid, pg. 153; Searle, “I. The Systems Reply,” ibid, Ed Fredkin, On the Soul (draft version).

[15] See Appendix A.

[16] Melanie Mitchell, “Life and Evolution in Computers,” History and Philosophy of the Life Sciences.

[17] Mitchell, ibid, Fig. 8.

[18] Ibid, pg. 378.

[19] Mitchell’s setup is 148 squares with 2 possible states per square; 1482 = 21,904 possible combinations the computation would have to handle. Evaluating the squares for RGB values adds one more state per square, for a total of 3 possibilities; 1483 = 3,241,792. Most modern consumer microchips are capable of between 1 and 3 GHz, or 1 to 3 billion calculations per second.

[20] Notably, single-cell organisms are also capable of basic forms of mind; they are able to learn via habituation (lessening reactions to repeated stimuli) and there is early evidence that they are also capable of Pavlovian associational learning. See Samuel J Gershman, et al. (2021) “Reconsidering the evidence for learning in single cells,” (https://doi.org/10.7554/eLife.61907).

[21] Seth Lloyd, Information and the Nature of Reality, “The Computational Universe,” pg. 94 – 95.

[22] Mitchell, ibid, “Self Reproduction in Computers.”

[23] See Appendix B.