Evolution Never Designed You to See the Truth

Donald Hoffman is a cognitive scientist at UC Irvine who works on evolutionary psychology, visual perception, and consciousness. He is the author of more than 120 papers and of *The Case Against Reality*, whose subtitle states the thesis bluntly: evolution hid the truth from our eyes. On Lex Fridman's podcast he restates the claim as a theorem. The technical question his team asked is the probability that natural selection would shape sensory systems to see true properties of objective reality. Except for one kind of structure — probability measures, the sigma-algebras any science needs if the world and our perceptions are not wholly unrelated — the answer, he says, is precisely zero.

Most of us, Hoffman included, once assumed the opposite. Surely the way to make senses fitter is to make them tell more of the truth, or at least the truths we need in our niche. Evolution by natural selection is not a hand-waving story, though. John Maynard Smith made it a mathematically precise theory in the 1970s, and there is evolutionary graph theory and genetic algorithms besides. Two of Hoffman's graduate students, Justin Mark and Brian Marion, ran simulations that first tipped him off. He then took the problem to mathematician Chetan Prakash, with Manish Singh and Chris Fields among the other collaborators. Prakash proved several theorems that, Hoffman says, uniformly give the same result.

The counting argument is almost childishly direct. Suppose the world has some structure — a total order, a topology, a metric — and consider all maps from that world into fitness payoffs. Count how many of those payoff functions are homomorphisms that preserve the structure, divide by the total, and take the limit as the number of world states and payoff values grows large. Every time, the ratio goes to zero. Fitness functions are not required, in standard evolutionary theory, to preserve any particular structure of objective reality. What we see, Hoffman concludes, is whatever we need to stay alive long enough to reproduce. Adaptive behavior, full stop.

Fridman presses the bold version of the claim. The table, the atoms, a century of particle physics, DNA, embryos — is none of that real? Hoffman will not say the perceptions are fake. They are real as perceptions, and they are a real adaptive set. What evolution precludes is the tight match we assume: that when you look up and see the moon, something in spacetime matches what you perceive. The senses are there to guide adaptive behavior, not to show the truth. In fact, he thinks they are there to hide it, because the truth is too complicated.

The analogy he returns to is the laptop desktop. When you write an email you are toggling voltages, but "if you had to know all that truth, your friends wouldn't hear from you." The blue folder icon does not exist in the hardware. It exists because it is useful. Evolution, Hoffman says, gave us perceptions that work the same way: an interface, with no manual, whose design goal never included fidelity. Fridman objects that Chrome and Python are still tethered to the zeros and ones. Hoffman grants the backstory on a computer. Evolution does not tell us the backstory of our interface. It only says that whatever reality is, you do not see it.

That is why looking at neurons, once you take the theorem seriously, is like staring at desktop icons and hoping to recover the voltages. The scientific theories we already have are precise enough to announce their own limits. They are not precise enough to say what the next move is. That, Hoffman argues, is where scientific creativity begins — and it is not a move alien to physics.

"Spacetime is doomed" is not Hoffman's line. He attributes it to Nima Arkani-Hamed at the Institute for Advanced Study, to Ed Witten, and to Nobel laureate David Gross. Einstein's gravity together with quantum field theory entail that spacetime cannot be fundamental. Particles in spacetime are irreducible representations of spacetime's symmetries. Put the two theories together and spacetime loses even an operational definition at 10^-33 centimeters and 10^-43 seconds. The theories tell you where the wall is. They do not hand you a flashlight for what is behind it.

Arkani-Hamed and colleagues have been guessing anyway, and some of the guesses are paying off. The amplituhedron and the cosmological polytope are high-dimensional geometric objects that encode scattering amplitudes at colliders. Once you drop spacetime, computations that took billions of terms collapse to one or a handful. Feynman diagrams, Hoffman says, force locality and unitarity inside spacetime and sprout endless loops; outside spacetime those notions are not primitive, and a dual conformal symmetry that is actually in the data becomes visible. Locality and unitarity then reappear as properties of the facets of the polytopes.

The pointer into that deeper realm is older than the names. In 1986, Parke and Taylor took a routine two-gluon-in, four-gluon-out amplitude — hundreds of pages, billions of terms, too heavy even for the supercomputers experimentalists needed — and reduced it first to nine terms, then to a single expression, the Parke-Taylor formula. No one knew where it came from. Hoffman reads it as a sign that spacetime has had a good ride and reductionism a useful methodology, and that both are over as fundamental stories. Whatever comes next must still project back into spacetime and recover evolution by natural selection, or it is the wrong next theory.

When Hoffman says a neuron does not exist unperceived, he wants a concrete picture, not a slogan. Sketch a Necker cube on paper and a viewer sees a 3D cube that flips which face is in front. Which face is in front when no one looks? Neither: there is no cube, only paper. You create the cube on the fly and garbage-collect it when you look away. Spacetime, he says, is that cube — a data structure sensory systems construct. New physics beyond spacetime, such as the map from an amplituhedron down into four dimensions, is what tethering would actually look like. What the physicists have not yet found, and what Hoffman wants, is dynamics, not only static geometry.

That is why the spacetime assumption bothers him more than any other interface. Everything else we think we know is a thing in spacetime. Kill the ontology and you also kill the methodology of reductionism: the bet that smaller scales yield more fundamental laws. Scientific explanation is broader than that bet. Einstein's postulates about the speed of light are not a march to smaller objects; they are deep principles. Thermodynamics reducing heat to molecular motion was a genuine success of the reductionist style. Hoffman thinks that success is likely an artifact of how we use the interface, and that the program has a known stop sign at the Planck scale. He is not impressed that the stop sign is "only" 10^-33 centimeters.

The history of elements is his preferred illustration. Deep thinkers once treated earth, air, fire, and water as fundamental. The periodic table showed that "earth" was silicon, iron, and many other things projecting into one icon. Then came quarks, leptons, and gluons. Then the punchline: those particles are not fundamental entities either, only irreducible representations of the Poincaré group — symmetries of the spacetime data structure itself. Drill from neural networks to neurons to chemistry to quarks and you feel yourself approaching reality. Hoffman says you are only describing the desktop more precisely. You never left the interface.

Fridman brings up a thought experiment by neuroscientists Eric Jonas and Konrad Kording: apply the tools of neuroscience to a microprocessor — lesions, recordings, statistics — and watch how little you learn about what the chip is actually doing. Hoffman lights up because the same demand pulled him into the field. As an undergraduate at UCLA he read David Marr and thought, who is this guy? Marr had done rigorous work on the cerebellum, then left for MIT's AI lab because he still did not know what the cerebellum was *for*.

Marr's three levels became Hoffman's standard. Computational theory: what is being computed, and why? Algorithm: how, exactly? Hardware: what implements it? If you cannot build it, Marr argued, you do not understand it. Hoffman went to MIT to work with him and had a year and a half before Marr died. Most of current brain science, he thinks, still lives at implementation and a bit of algorithm. The computational theory — what the brain is computing and why — is nearly blank. Evolution will not fill in that "why." It only tells you that whatever reality is, what you see is an adaptive fiction.

If the turtles never stop, is science paralyzed? Hoffman leans on Gödel: any finite axiom system rich enough for arithmetic has truths it cannot prove; add them and new unprovable truths appear. Physical theories and evolutionary theory are mathematical, so each will have miracles — assumptions it does not explain — and Gödel limits besides. That is job security, not a license for "anything goes." Experiments still constrain what an axiom system can say. One speculative payoff of going around spacetime rather than through it: galaxies receding faster than light might not be unreachable if spacetime is only a data structure. The best scientists, he guesses, will keep oscillating between conceptual work and the non-conceptual place where Einstein said the real insights arrived.

Marr thought primitive senses were less realistic and that human vision, with 86 billion neurons, actually estimates true shapes and distances. Hoffman's simulations point the other way. Justin Mark ran genetic algorithms on foraging agents with random sensors and random actions. Truth-seeing organisms barely appeared; when they did, they were gone in a generation. Agents that tracked fitness payoffs were far fitter. Hoffman is not doctrinaire about the model. He does not believe any current theory as gospel. He studies evolutionary game theory because it is the best tool in the area, and he thinks the 1890 physicists who told bright students the field was finished were making a mistake we should never repeat.

A Yale team offered the strongest pushback. Hoffman's first simulations used one or two payoff functions; with 20,000, the Yale group found conditions in which truth-seeing organisms evolved. Hoffman grants the point, and grants that probability zero is not impossibility. Draw a curve in the unit square and its area-measure is precisely zero, yet it contains infinitely many points. Probability-zero events can happen infinitely often. They are still measure-zero in the parameter space. He then reran the work after relaxing a hidden Yale assumption: that organisms could not carve the world into objects.

Allow hierarchical clustering of similar payoff functions into data structures — a bottle, an apple, a rock, each bundling actions and payoffs — and the Yale results vanish. An object, on this view, is not a true thing that exists unperceived. It is a convenient encoding of fitness. Fridman worries that game theory oversimplifies, and Hoffman agrees it is a fair rejoinder: if someone has a better mathematical capture of Darwin, use it. Graph theory and quantum game theory already generalize the toolkit. He will not pretend the current tool is sacred. He will not ignore what it currently entails.

Where do fitness functions come from, and why is there a line between living and dead? Hoffman suspects the line is not fundamental. It is an artifact of an uneven interface that hides most of reality — dark matter and dark energy included — and makes some things look inanimate. He would like a deeper, perhaps timeless dynamics, possibly all cooperation and no scarce resources, that projects into spacetime as nature red in tooth and claw. That derivation does not yet exist. He flags it as a hope, not a paper.

When he looks at Fridman he sees skin, hair, and eyes, a symbol he creates and deletes when he looks away. He does not delete Fridman. The other person's experiences, thoughts, hopes, and dreams are still there. "I'm only deleting an interface symbol, but that interface symbol is a portal" — imperfect, but genuine — which is why a conversation can happen at all. Two consciousnesses really affect each other through icons generated on the fly.

The portal degrades along a gradient. A cat still counts as living and, for Hoffman, likely conscious. An ant is living, maybe not conscious. A virus is a maybe. A proton looks thoroughly dead. He is not saying the proton is conscious, any more than he is saying Fridman's face is conscious. The face is a data structure in the viewer, an experience, not an experiencer. He is more confident about why elementary particles make such bad portals: they are symmetries of spacetime, irreducible representations of its Poincaré group, almost entirely about the desktop itself. Only at higher levels do we start to build doors behind it. The door being shut is not evidence that nobody is on the other side.

Hoffman estimates that 99 percent of his colleagues still take spacetime, particles, and matter as fundamental, then ask what complicated matter could make consciousness arise. Integrated information theory says the right phi yields experience. Hameroff and Penrose point to orchestrated collapse of quantum states in microtubules. Global workspace theory, attention schema theory, and Philip Goff's panpsychism — which adds a consciousness property to particles — all stay inside spacetime. Hoffman thinks they are brilliant people stuck with earth, air, fire, and water: a useful vocabulary that will not get them to the ground floor.

He has asked Giulio Tononi twice, once in the 1990s and again a few years before this conversation, to pick one experience. What precise integrated-information structure is the taste of chocolate, and why could it not be vanilla? The answer was silence. He means it literally: none of these theories has accounted for a single specific conscious experience. That does not falsify them. It means that after decades, including the last twenty years of Francis Crick's life, no one is saying they are a year away. There is a gulf.

Steven Pinker has written that he likes global workspace theory but that the last bit — what it is like to feel something — may have to be stipulated as a brute fact. Pinker calls it "the last dollop." Hoffman's reply is the hinge of the interview. The whole promise of physicalism was that you would not have to stipulate consciousness; you would derive it from matter. If you are going to stipulate anyway, stipulate less: grant experience, and try to boot up the physical from that. He calls the stance conscious realism, not idealism, because idealism carries anti-science and anti-realist baggage he does not want.

A theory's assumptions are its miracles: the things it will not explain. Hoffman wants as few as possible. First, there are conscious experiences — headache, garlic, red — primitive, not derived from physics. Second, experiences have consequences: one can trigger another, described with probabilities. Mathematically that is a probability space of possible experiences plus a Markovian kernel. He calls the pair a conscious agent, and he is careful about the words. Annaka Harris warned him that same morning that "agent" smuggles in self and agency. He accepts the warning. The term is technical. There is no self, no learning, no memory, no intelligence, no planning in the definition. Those have to be built.

The framework has a property he likes: when two conscious agents interact, the pair itself satisfies the definition. There is one agent with an analytic decomposition into as many as you wish, scale-free from a unit that only ever experiences "red 35" up to one infinitely complicated agent. Networks of conscious agents are computationally universal — a trivial theorem — so they can do whatever neural nets and automata can do. The events in the probability space need not be computable, which leaves a door open for non-computable interactions. Chris Fields led a paper, *Conscious Agent Networks*, showing how to build memory and primitive learning in this language.

The standard hard problem asks how unconscious neurons produce the taste of chocolate. Hoffman changes the problem. He will not boot experience up from neurons. He wants the reverse: starting from experience, how do you build spacetime, the physical world, brains? "Consciousness is not something that brains do. Brains are something that consciousness makes up." Right now, he says, he has no neurons. If you looked you would create that data structure, use it, and garbage-collect it — the Necker cube again. He is not saying neurons are useless. He is saying they have lost the status of observer-independent objects. They are symbols on an interface. Why chocolate tastes like chocolate remains a miracle of the theory. He hopes a later version chips at it. His current imagination does not go that deep, and he would like a twenty-year-old to prove him wrong.

Fridman wants to know why time is doomed, not just space, and how dynamics survives. Hoffman offers a three-line proof. Imagine a Markovian dynamics of consciousness outside spacetime that is stationary: entropy is constant, so there is no entropic time. Any projection by conditional probability induces a chain whose entropy increases. Stationarity plus the Markov property are enough. The arrow of time can be an artifact of the view. Limited time is a scarce resource; scarce resources invite competition; evolution and selfish selves may be what a projection looks like, even if the base has no tooth and claw. He has not shown that networks of unconstrained conscious agents yield natural selection. He marks that as speculation his brighter colleagues might finish.

That speculation is also his scientific version of a spiritual claim. The finite self that competes for money and prestige has to die in some sense; a deeper, formless awareness that precedes any particular experience is the ground. Don and Lex would be projections through which one agent learns about itself. Mathematics, in this picture, is to consciousness as bones are to an organism: necessary, not the whole animal. Max Tegmark's level-four multiverse takes mathematical structure as fundamental; Hoffman takes consciousness as fundamental and mathematics as its skeleton. He expects an infinite variety of data structures, spacetime among a vanishingly small fraction, and forms of experience as unimaginable to us as a tetrachromat's extra color dimension is to ordinary trichromats.

Can we visualize any of that, or only write Markov kernels? Hoffman thinks science has done this before: Faraday's experiments, Maxwell's equations, Planck's desperate packets of energy, then quantum theory. Psychedelics look to him like another unexpected pointer, an intervention on the interface that seems to open new experiential portals. The one technology we already know for building new portals is having children: before the birth the portal is not there; after it, you meet a consciousness you had never met, with its own personality. That technology is slow. Reverse-engineering it, or even tweaking an existing complex portal with a chemical, is why he thinks psychedelics are worth studying. Tweaking is not creating, and he does not blur the two.

Fridman asks where Hoffman sits relative to Immanuel Kant. There is a lot of agreement: Kant's transcendental idealism says we do not see nature as it is; we impose structure. Hoffman says something similar, but he will not wear the idealist label. He wants realism in the name — conscious realism — without anti-science connotations. Kant's further claim that Euclidean spacetime is a priori is, Hoffman says, simply false. Berkeley has ingenious arguments in the same family. Leibniz, in the *Monadology*, posed the hard problem, judged it unsolvable, and tried to build from monads outside spacetime. If Leibniz walked in today, Hoffman jokes, he would grasp conscious-agent theory in five minutes and run past him.

The practical bite is how ephemeral the interface is. Spiritual traditions already say possessions vanish; Hoffman says they vanish on the order of hundreds of milliseconds. He creates the water bottle when he looks and garbage-collects it when he looks away. A mansion, a Porsche in a headset — each observer renders a private copy. Wars over property look, in this frame, like fights over icons in a game. You are not a small object inside a preexisting spacetime whose status is raised by the special car. You are the entity that authors the car on the fly. The traditions are a mixed bag, he adds immediately: wisdom plus plenty of nonsense. The move is to cash the ephemeral insight in mathematics rather than in slogans.

Nick Bostrom's simulation argument, as usually told, sits inside physicalism: a programmer in spacetime runs a computer that somehow hosts conscious entities, then nested simulations. That picture swallows the hard problem whole and explains nothing about how circuits yield experience. What Fridman finds more interesting is the engineering question of minimal rendering. VR researchers already study how little you need for immersion: a tight perception-action loop and perceptions close to what the user expects, after which you can move their reality around. Hoffman offers a homemade test. Close your eyes, attend forward, and you get a dynamic dark gray. Attend behind you: not darkness, not a shape of absence — no quale at all — and yet a sense of being. Fridman, who has been on that side of the room, starts replaying memory. Hoffman takes the point: we can live entire lives without noticing that it has always been like nothing behind us. The immersion is that complete.

Living with the theory has been difficult. The first time Hoffman thought we might not see the truth was in 1986; the math hit him like a ton of bricks. House, car, reputation — headset prizes. He meditates three or four hours a day and still says his default mode is physicalist. The way out of the game is not better stories. It is letting go of thought, after which name, history, and credentials go nearly irrelevant because only the present experience is, as far as he can tell, real. Next second he sinks back in. He is attached to the body and the interface. A voice still says he is crazy. As a scientist he takes it as established that local realism is false and non-contextual realism is false: quantum theory already shows that observation can be fact-creation, in Chris Fuchs's phrase, even when you can predict an outcome with probability one.

A year before the interview, COVID attacked his heart. His rate stayed at 190 beats per minute for 36 hours in the emergency room. He texted his wife goodbye: "I love you, goodbye." He was terrified, and also so sick that some part of him just wanted it to stop. The near-death did not turn him into a sage. It riveted his attention. He reads spiritual traditions more now, with science as a check, and he is frank that evolutionary psychology still describes him: in-groups, reputation, the impulse to lie and to anger. Grok it at an emotional level, not only on the whiteboard. Love, for him, splits in two. Sexual love has a clean evolutionary account. The deeper layer is what was at play in that text. "Love your neighbor as yourself" is, in his hearing, a description of fact-structure. If you and I are separate, fighting is available. If we are not, love is arithmetic. He gets glimpses of that, especially in the gap when thought stops.

God, raised a Protestant minister's son, is not a being you can take sides with. The closest wording he has is being itself — the ground of conscious being. Two beings named God can go to war; shared being cannot. That picture, he says, happens to fit the mathematics of one agent under many projections, though matching theology was never the goal. His advice to young people is that the universe is more interesting than advertised, and so are they. You are not a hundred-pound irrelevance in billions of light-years. You are, in this theory, the being that authors that space every time you look. The textbooks will look silly in fifty years; someone has to write the next ones, after first actually learning the current ones. The meaning of life he will not pretend to know. The "I don't know" may be part of the answer. His guess is that consciousness wakes up by learning what it is not — immerses itself in a headset, gets lost, and slowly climbs out, needing the negative examples. Why does formless being bother with tables and moons? He does not know. Without "not this," it might not know that it is. Fridman closes by saying he is glad Hoffman is still here. Einstein gets the last word: time and space are modes by which we think, not conditions in which we live.