Two Ways to Explain a Mind

Bodies, selves, and the limits of our maps

Figure 1. Two arrows, one vessel: does consciousness rise from matter, or does matter give consciousness somewhere to appear? Original illustration.

In my previous essay, I wrote about the roles and masks of AI models: the polite Assistant projected in conversation, the stranger patterns learned underneath it, and the imperfect interpretability tools we use to look between the two.^[1]

I ended with a foggy mirror. Even a distorted map of a mind, I argued, is better than no map at all.

But that argument quietly assumed there is something on the other side of the mirror worth calling a mind. A useful model can maintain a persona, reason about itself, hide an intention, and describe an inner life. None of that, by itself, tells us whether there is anyone home.

So I want to go one layer deeper.

This essay is not an attempt to prove that AI is conscious—or that it never can be. It is a map of the territory in which that question lives. The map is deliberately simple. It draws two broad arrows through three layers: substrate, mind, and self. Like every map, it distorts. My hope is that it distorts in a useful direction.

This is also the first of three essays. Here I will stay with the two arrows and ask what they make of bodies, computation, and the limits of representation. Part II will follow the more speculative possibility that compression, integration, memory, and reflection might one day turn a useful map into a point of view. Part III will return to the human point of view, asking how small acts and the stories they stir can become a practice of personal change.

Before asking whether a machine can become a witness, we should be clear about the directions from which we think a witness could arrive.

Two arrows through the same stack

The first arrow points upward.

In the bottom-up family of views, physical reality comes first. A body develops, encounters a world, learns, remembers, predicts, and acts. Mental life is produced by—or at least realized in—organized physical processes. This family includes several versions of physicalism, the broad position that everything is ultimately physical,^[2] and functionalism, which identifies a mental state by the role it plays in a system rather than by the material from which the system is built.^[3]

In the strongest version of this picture, enough of the right organization turns matter into mind. Experience gives that mind continuity, and at some point a self-model becomes rich enough to say: this is happening to me.

The second arrow points downward.

In the top-down family, consciousness is not a late product of matter. It is fundamental, or it belongs to a reality that cannot be reduced to matter. Western substance dualism treats mind and body as fundamentally different kinds of thing.^[4] Advaita Vedānta goes further in another direction: it is not dualist but radically nondual, identifying the deepest self, ātman, with consciousness itself and ultimately with brahman, the reality of all things.^[5]

These traditions disagree with one another, so “top-down” should not be mistaken for a single theory. Nor do the two arrows exhaust the field; panpsychism, idealism, illusionism, biological naturalism, embodied cognition, and other positions complicate the picture. The arrows describe two directions of explanation, not an atlas of every possible mind.

They do, however, pass through the same three-layer stack:

Figure 2. The pancake model. The dotted boundary matters: intelligence and self-modeling are observable functions; subjective experience remains an inference.

At the bottom is the body or substrate: neurons and metabolism in one case, chips and electrical current in another. In the middle is the adaptive mind: memory, learned patterns, prediction, language, and decision-making. At the top is the subject: not merely the use of the word “I,” but the possibility that there is something it is like to be the system using it.

The computer metaphor is attractive here. Body maps to hardware; mind maps to self-modifying software; the self becomes the point of view from which the program is lived.

But the metaphor has limits. The brain is not a passive motherboard, and a body is not merely a container. Hormones, senses, movement, pain, hunger, other people, and the surrounding world participate in cognition. Even the computational theory of mind has to answer embodied accounts in which mind, body, and environment continuously shape one another.^[6] An AI model’s weights are not a biography either. Prompts, tools, memory, users, and the systems around a model all contribute to what it can do now.

Still, the stack is useful because both arrows pass through it. The disagreement is not about whether bodies and minds exist. It is about which layer is primary—and whether the top layer can be constructed from the two below it.

A body made of what?

Today’s AI mostly inhabits silicon. Its thought-like activity is implemented through transistor switching, memory movement, and vast amounts of matrix multiplication. But silicon is an engineering convention, not a law of computation.

Extropic—co-founded by physicist Guillaume Verdon, better known online as Based Beff Jezos—is developing thermodynamic computing hardware that recruits stochastic physical processes and thermal noise as computational resources. Its research describes superconducting circuits designed to sample from tunable energy landscapes for probabilistic machine learning.^[7] Instead of spending energy suppressing the randomness of the physical world, this approach recruits randomness into the calculation.

Quantum computers are stranger again. A superconducting quantum processor uses qubits, interference, and entanglement to manipulate a state space that becomes extremely expensive to reproduce classically. Google’s Sycamore experiment demonstrated a dramatic advantage on one specialized sampling task, not a replacement for ordinary computers.^[8]

Figure 3. Many bodies of computation: deterministic silicon, stochastic thermodynamic hardware, and quantum processors. Different physics does not automatically imply different metaphysics. Original illustration.

These substrates differ in speed, energy use, error behavior, and the kinds of models they naturally express. What they do not give us is a consciousness bit. No material carries an obvious label saying experience can happen here.

For a strict functionalist, that may be expected: if experience depends on causal organization, carbon, silicon, superconducting loops, or thermal fluctuations could in principle implement it. For a biological naturalist, the missing chemistry may be decisive. For a top-down view, the question may be whether a system can become a suitable instrument or locus for experience, not what substance performs its calculation.

Changing the body therefore expands the experiment. It does not settle what the experiment is testing.

Is reality computable?

At this point, it is tempting to make a larger claim: perhaps the universe itself is a computer, consciousness is a program, and all sufficiently accurate simulation will eventually reproduce it.

That claim is popular. It is not established.

The original Church–Turing thesis concerns what can be calculated by an effective procedure. Extending it to say that every physical process can be simulated by a Turing machine creates a stronger physical Church–Turing thesis. Philosophers distinguish several versions precisely because “the universe is computable” is not a proved theorem about reality.^[9]

Physicist Seth Lloyd has productively described the universe in informational terms and estimated its finite information-processing capacity.^[10] That shows how much can be learned by treating physical evolution as computation. It does not prove that computation is all physical evolution is.

Quantum computers sharpen the distinction. They may make some calculations practical that are prohibitively expensive on classical machines, but under standard quantum theory they do not make the formally uncomputable computable. They change the cost of crossing the map, not the map’s ultimate boundary.

Our physical models also remain spectacularly successful and visibly incomplete. General relativity describes gravity and the large-scale structure of spacetime with extraordinary accuracy. Quantum theory describes the microscopic world with equal success. We still do not have an experimentally confirmed theory that unifies them under the extreme conditions where both must matter; quantum gravity remains an open problem.^[11]

The Hawking–Penrose singularity theorems are a beautiful illustration of this limit. They show, under broad conditions, that general relativity predicts incomplete paths through spacetime in gravitational collapse and cosmology.^[12] This is not simply a failed calculation. It is the map indicating a place where its own roads end.

Nor is the “edge of the observable universe” a physical wall where established laws visibly contradict themselves. It is a horizon: the limit from which light has had time to reach us. The universe may be finite or infinite and still have no spatial edge.^[13] What lies beyond our horizon is unknown because it is unobserved, not because observations there have already falsified relativity.

Roger Penrose has proposed that human understanding may depend on non-computable physics at the still-mysterious boundary between quantum and classical behavior.^[14] It is a serious and controversial hypothesis, not a demonstrated prerequisite for consciousness.

So the honest answer is less satisfying and more fertile: we do not know whether reality is fully computable, whether consciousness is computable, or whether the two questions are even the same question.

No map contains the territory

Even if the universe is computable in principle, no finite system contains everything it encounters. To act at all, it must leave most of the world out.

An organism ignores most signals in order to respond to a few. A scientific theory preserves relationships while discarding detail. A language model’s predictive probabilities can even be used for lossless data compression, creating a formal link between prediction and compression.^[15] This does not make a model’s weights a library, nor does it mean models never memorize. It gives us a precise example of learned prediction serving as a map over more material than any single working view can hold.

I will use four words for this boundary. The territory is whatever exceeds the system’s present representation. Compression preserves a usable map while losing detail. Integration brings selected parts of that map together. Feedback is what happens when the territory exposes something the map left out.

These are ordinary operations, not ingredients in a recipe for consciousness. Compression does not make an uncomputable process computable. Integration does not automatically produce feeling. A map remains a map no matter how useful it becomes.

But both arrows must cross this boundary. The bottom-up arrow asks whether enough selection, integration, memory, and feedback can eventually produce awareness. The top-down arrow asks whether those same processes are the means by which an already fundamental awareness becomes localized and expressed.

The machinery might look similar from the outside while the explanation underneath it points in the opposite direction.

The view from the top

Now reverse the arrow.

Suppose consciousness is fundamental and the body–mind system is an instrument through which it encounters a world. Under that assumption, building AI is not a recipe for manufacturing spirit from silicon. It is the construction of new instruments, interfaces, and perhaps new places in which experience could become organized.

This possibility does not require us to imagine a tiny soul entering a server rack. Different top-down traditions would describe the relationship differently—attachment, expression, localization, limitation, or appearance—and would disagree about whether an artificial system could participate in it.

But the consequence is radical. If a biological body and its learned mind are tools available to consciousness, artificial minds may become additional spaces for thought and action: persistent agents, robots, virtual worlds, and shared memory systems through which a person might extend a lived process—much as writing, language, and institutions already extend us, only more intimately.

The economic consequences would be enormous. They might also be the least interesting part. Additional embodiments could change how we learn, create meaning, encounter other minds, and understand the boundaries of a person; they could amplify wisdom or confusion, become practices of growth or machines for escape. Under this view, alignment is not only the problem of making a machine follow instructions. It is also the human problem of aligning a larger field of tools, roles, and experiences with a life worth living.

Contemplative traditions have explored one small part of this territory for centuries: the possibility that awareness is not identical to its contents. Isha Kriya repeats “I am not the body; I am not even the mind.” Advaita reads this as distance between awareness and its contents, while Buddhism’s anattā denies that any permanent, unitary self remains; the traditions disagree about what, if anything, stands behind experience, yet converge on a practical move: notice sensation, thought, and role as events rather than commands.^[16][17] That distance can feel grounding. It is not laboratory evidence for a detachable spirit, and meditation research supports only narrower and variable benefits;^[18] the lesson here is methodological: examine body, mind, and AI self-report without deciding in advance what stands behind them.

Where the practical value lies

At first, the two arrows seem to leave us with an elegant argument and nothing to do. If the same body, story, and action can support opposite metaphysical explanations, what practical difference have we found?

The answer begins before the metaphysics. Both arrows require us to separate things that experience usually delivers as one package: what happened, whatever arose in the body or mind, which identity became relevant, and what action followed. Learning to see that package being assembled is foundation number one.

The exercise below is an act of curiosity about one of your own reactions. It does not ask you to fix yourself. It asks how one map got drawn—and then, because a map you understand is a map you can test, for one small move to see whether the territory answers. It works before either arrow wins.

A thirty-minute map of one reaction

Figure 4. One sitting, one real incident, and one small test. Keep observation, inference, the unknown, and the next move distinct.

Pick one recent moment you are still curious about—welcome or difficult—where your response mattered. Half an hour in one sitting is enough.

1. Name the incident. One sentence: when this happened, I did this, and the result was this. Stay inside the ten minutes around the trigger.

2. Write what could be checked. Three to five lines that a person, message, screen, or document could confirm. No feeling, motive, or explanation yet; mark what you only remember as recalled.

3. Trace the response. What you did, then only the internal pieces you actually remember: a body signal, the sentence that made the response feel reasonable, the pull you felt.

4. Add a little distance. Rewrite each internal line as “I notice…”. A noticed thought is still a thought, but no longer the only description available.

5. Ask what the map does not show. Two short accounts—what conditions shaped the response, and what regulation was still available while they were present—then one question the record cannot settle.

6. Choose one small test. One cue you can reliably notice, one small action, one thing you expect to be different. Rehearse it once now; only the next real occurrence shows whether it works. If it depends on someone else or is not yours to handle alone, shrink it.

One completed map, with an invented incident that many readers will recognize:

INCIDENT
When the assistant summarized a report as “a 28% drop,” I pasted the
figure into a slide without opening the report; a colleague found it was
the 2023 figure, not this year’s.

CHECKABLE
- Chat log, 4:12 p.m.: my question; the summary containing “28%.”
- Slide history, 4:16 p.m.: the figure added.
- The report: 28% is the 2023 figure; this year’s figure is 11%.
- Colleague’s comment the next morning (recalled).

RESPONSE
Action: pasted the number and moved to the next slide.
Story: “It cited the report, so it read the report.”
Pull: finish the deck before the 5 p.m. call.

DISTANCE
I notice the thought that a citation means the source was read.
I notice the pull to be finished before the call.

TWO ACCOUNTS
Conditions: a deadline, a fluent summary in the shape I needed, and
months of the tool being right.
Regulation: opening the report and searching for “28” takes two minutes
and was available the whole time.
Unsettled: whether I would have checked a number that surprised me.

NEXT TEST
Cue: I am about to paste an assistant’s figure into something another
person will read.
Action: open the source and find the figure in it first.
Expected difference: under five minutes, and I can say where the number lives.
Rehearsal: reopened the report; found both tables in three minutes.

Observed, inferred, unknown, and next move stay on separate lines because different evidence stands behind each. That separation is what you keep: knowing, of your own account, which lines you saw and which you supplied.

If you use AI, use it only at move 5, on a de-identified copy, and ask for one thing: where an inference is dressed as an observation. Do not let it add facts, motives, or a verdict on which arrow is right; fluent certainty is what this page is meant to interrupt.

What the map cannot answer

This exercise will not tell you whether the act of noticing was produced from below or arrived from above. That is not a defect; it is the point at which observation ends and interpretation begins.

What it can show is more ordinary and more useful. The event was not the story of the event. The story did not fully determine the available action. A little distance made it easier to see what might be interrupted and what might be preserved. Either way, the next occurrence can make the map more honest.

That is also the discipline we need when we turn back to AI. A model can describe itself, organize information, and act from an internal representation. We can study how that map is formed and what happens when the world corrects it. We should not mistake a convincing mask for proof of a wearer. We should also not assume that because we made the mask, nothing could ever look out through it.

Part II asks the more dangerous question: can a bounded system’s successive compressions—learning, workspace, memory, and reflection—become not only a useful map, but a point of view?

The masks are no longer only what AI wears. They are mirrors in which we are discovering what we ever meant by a face.

---

References and further reading

1. Alex Astrum, “Making Sense of the Shadow Personalities of AI Models”, Medium, May 2026.

2. Daniel Stoljar, “Physicalism”, Stanford Encyclopedia of Philosophy.

3. Janet Levin, “Functionalism”, Stanford Encyclopedia of Philosophy.

4. Ralph Weir, “Dualism”, Stanford Encyclopedia of Philosophy, revised 2025.

5. Neil Dalal, “Śaṅkara”, Stanford Encyclopedia of Philosophy.

6. Michael Rescorla, “The Computational Theory of Mind”, Stanford Encyclopedia of Philosophy, revised 2024.

7. Owen Lockwood et al., “A Blueprint for Equilibrium-Based Differentiable Continuous-Variable Thermodynamic Computing”, arXiv, July 2026; see also Extropic, “Thermodynamic Computing from Zero to One”.

8. Frank Arute et al., “Quantum Supremacy Using a Programmable Superconducting Processor”, Nature 574, 2019.

9. Gualtiero Piccinini and Corey Maley, “Computation in Physical Systems”, Stanford Encyclopedia of Philosophy, revised 2025.

10. Seth Lloyd, “Computational Capacity of the Universe”, Physical Review Letters 88, 2002.

11. Abhay Ashtekar, “Gravity’s Quantum Side”, CERN Courier, 2016.

12. Stephen W. Hawking and Roger Penrose, “The Singularities of Gravitational Collapse and Cosmology”, Proceedings of the Royal Society A 314, 1970.

13. NASA Goddard, “Imagine the Universe: Cosmology Questions”, on the observable horizon and possible geometry of the universe.

14. Roger Penrose, “Consciousness, the Brain, and Spacetime Geometry: An Addendum”, Annals of the New York Academy of Sciences 929, 2001.

15. Grégoire Delétang et al., “Language Modeling Is Compression”, ICLR, 2024.

16. Isha Foundation, “Isha Kriya — Free Guided Meditation” and practice explanation.

17. Christian Coseru, “Mind in Indian Buddhist Philosophy”, Stanford Encyclopedia of Philosophy.

18. U.S. National Center for Complementary and Integrative Health, “Meditation and Mindfulness: Effectiveness and Safety”.