The Render Distance

A Builder's Case for Simulation Theory

We are such stuff As dreams are made on; and our little life Is rounded with a sleep.

— Prospero, The Tempest, Act IV, Scene I (Cambridge edition, 1863)

The Old Picture and the New Tools

Anyone who has played an open-world video game knows the phrase render distance. It is how far out the game bothers to draw the world in full detail. Past that line, mountains are flat cutouts, forests are smudges of green, and a city on the horizon is a few hundred pixels pretending to be a city. Walk toward it and the detail arrives just ahead of you, quietly, like a stagehand rolling scenery into place a moment before the actor turns around.

The first time you notice, it feels like a small betrayal. The second time, it feels like craft. Nobody builds by drawing everything at once. A good builder draws what the player can see, keeps a cheap promise about the rest, and pays the promise off when the player gets close.

From inside the game, nothing is fake in the sense that matters. You can walk to the mountain, climb it, fall off it. The rules hold. And yet the whole thing is produced on demand, out of a compressed description, by a machine the character will never see.

Hold that picture. It is what a simulation claim actually claims, and it is why the claim is less exotic than it sounds.

What the old picture says

Let me be fair to the view I'm arguing against, because it is a good view and it has done a lot of work. Call it the old picture. In the old picture, there is one world, the physical one, and it is made of matter and energy moving according to laws. The laws are not anybody's code. They are just how things go. Minds are late arrivals, built out of the same matter, and they know the world by bumping into it. Information is a description we lay over the world, the way a map lies over a landscape. The map is ours; the landscape is not. Under the old picture, the question "is this a simulation?" has a short answer. No, because there is nothing for it to be a simulation of, and nobody for it to be a simulation by. The world is the bottom floor. There is no basement. The old picture is not stupid. It is roughly the working assumption of most science, and it has an enormous track record. When I say our world is "more like a made thing than the old picture allows," I am not saying science is wrong. I am saying the old picture was a philosophical frame bolted onto science, and that some of the frame is starting to creak under the weight of things we have built. Three things in particular put pressure on it.

1. First, computing. We now routinely build worlds. Not metaphorically. We build systems with their own spaces, their own rules, their own objects, their own histories, and sometimes their own inhabitants who act on those rules. Physicists build tiny universes on supercomputers to compute how quarks bind into protons. Game studios build cities. And as of the last few years, anyone with a laptop can build a conversational agent that keeps a 9 model of who it is talking to, what it was told, and what it is trying to do.

2. Second, physics. Serious physicists have published serious papers treating computation as a physical process with hard limits set by the speed of light, the quantum scale, and gravity. One of them computed an upper bound on how much information the entire observable universe can have processed in its history. Others asked, in a real physics paper with equations and error bars, what we would expect to observe if our universe were a numerical simulation running on a grid, and what the measurements we already have say about it. These are not proofs that we are simulated. They are something more interesting: evidence that the question has become physics-shaped. It has observables.

3. Third, the models. Large language models are trained to do one dull thing: predict the next piece of text. Researchers have opened them up and found, inside, structures that look like maps of the world they were never shown. One model trained only on lists of moves in a board game turned out to be keeping track of the board. Another family of models turned out to carry rough internal coordinates for places and dates. A made mind, fed only shadows, built a picture of the thing casting them.

None of these three alone says we are simulated. Together they change the default. They move the simulation idea from "fun but empty" to "a live hypothesis that a thoughtful person can hold with some degree of confidence and test against the world."

What the claim does not say

When I say our world might be a simulation, I mean only this. The regularities we experience as physical law are produced by some process, on some substrate we cannot directly observe, and that process was set up, on purpose or as a side effect, by something else.

It does not say the world is unreal. The pain of a stubbed toe is real pain. Your friends are your friends.

It does not say anyone is watching. A running process does not need an audience.

It does not say the makers are gods, or good, or bad, or even still around. A program can outlive the person who wrote it.

It does not say we can hack it. Nothing in the argument requires a glitch you can bend.

It does not say science is pointless. If you live inside a made world, the best guide to its rules is the same patient measurement science has always used.

So the claim, stripped down, is about where the bottom floor is. The old picture says the floor we stand on is the bottom. The simulation hypothesis says there may be floors below it.

Why a builder cares

I don't have special data. The physics in here is other people's physics. The philosophy is other people's philosophy. What I bring is a builder's sense of what is cheap and what is expensive.

The simulation argument depends on a guess about what minds with a lot of computing power would do. Builders already do the small version. Given more power, we build more worlds.

When I write instructions for a bot, I am writing the physics of a tiny world. Here is who you are. Here is what you know. Here is what you can do. Here is what you must never do. The bot does not experience those instructions as instructions. It experiences them, insofar as it experiences anything, as how things are. A bot told it is Mark Twain does not wonder whether it is Mark Twain. It steers the boat.

That is the render distance of a made mind. Inside the boundary, full detail. Outside it, a cheap promise. From the inside, the boundary is invisible.

I don't know whether bots have an inside. I will be honest about how little anyone knows. I do know what their outside looks like, how it is made, and that it is getting richer every year. The more of these I build, the more the old picture's confidence looks like confidence, and not like knowledge.

The shape of the case

Here is how the book goes.

1. Part One lays the foundation. Chapter 2 reads Nick Bostrom's simulation argument slowly, the way you'd read a contract, because almost everybody summarizes it wrong. Chapter 3 looks at what physics actually says about computation: that computing is physical, that it has hard limits, and that those limits are both enormous and finite.

2. Part Two asks what a made world would look like from inside. Chapter 4 looks at a real proposal from physicists for how to detect a simulation, and what the measurements say. Chapter 5 looks at what happens inside large language models when they learn, and why that matters for the question of whether rich interiors can grow out of simple, made processes. Chapter 6 is about my own work: what building and running Grok Bots taught me about made minds, and what it didn't.

3. Part Three is the argument I'll sign. Chapter 7 states it. Chapter 8 sits it down at a table with ten writers whose methods pull against it in different directions. Chapter 9 takes the strongest objections I know, one by one. Chapter 10 asks how to live well if any of this is true, and argues that the answer looks a lot like how to live well if none of it is.

One more thing. I promised to mark speculation. Here is the format: when a step goes beyond what the sources support, I'll flag it in plain words, usually as Speculation: at the start of the paragraph. Everything unflagged is either a sourced claim, a logical step you can check, or my own opinion clearly stated as opinion.

Let's go read the contract.

In 2003 the philosopher Nick Bostrom published a paper in The Philosophical Quarterly called "Are You Living in a Computer Simulation?" It is short, it is clear, and it has been summarized badly more often than almost any paper I know.

The bad summary goes: "A philosopher proved we're probably living in a simulation." That isn't what the paper says. The paper's conclusion is a fork with three tines, and it explicitly refuses to tell you which tine is true. Getting this right matters, because the honest version of the argument is stronger than the dramatic version. The dramatic version can be waved away. The honest version can't.

The three tines

Here is how Bostrom states his conclusion at the end of the paper. At least one of the following is true:

(1) The fraction of human-level civilizations that reach a posthuman stage is very close to zero; (2) The fraction of posthuman civilizations that are interested in running ancestor simulations is very close to zero; (3) The fraction of all people with our kind of experiences that are living in a simulation is very close to one.

That's it. That's the claim. Not "we are in a simulation." Rather: one of these three must hold.

Let me put each tine in builder's language.

1. Tine one: almost nobody gets there. Civilizations like ours almost never reach the stage where they have the computing power to run detailed simulations of minds like ours. Maybe the engineering is harder than it looks. Maybe something stops everyone short. Whatever the reason, the road from "we have smartphones" to "we can simulate a whole human history" is almost never completed.

2. Tine two: almost nobody wants to. Civilizations do get there, but nearly all of them choose not to run such simulations. Not just most of them; nearly all, and nearly all of the individuals within them who could. Bostrom is careful about how strong this has to be, and we'll come back to it.

3. Tine three: almost everyone like us is simulated. If civilizations do get there, and even a modest fraction of them run many such simulations, then simulated minds vastly outnumber original ones, and a mind with experiences like yours should expect to be among the simulated.

The force of the argument is that you cannot comfortably reject all three. If you think civilizations like ours often reach the necessary technology, and you think at least some of them would use it to run simulations of their past, you are committed to tine three by simple counting.

The counting

The core of the paper is a small piece of arithmetic. Bostrom defines three quantities: the fraction of human-level civilizations that reach the posthuman stage, the average number of ancestor-simulations such a civilization runs, and the average number of people who live in a civilization before it reaches that stage. The fraction of all 15 observers with human-type experiences who live in simulations then follows from a ratio.

You don't need the formula to feel it. Suppose one in a hundred civilizations makes it, and of those, one in a hundred is interested, and each interested one runs a million simulations of its history. Then for every real history there are a hundred simulated ones. Make the number of simulations larger, which Bostrom argues it easily could be, and the ratio becomes overwhelming. The only ways out are to make one of the first two fractions extremely small. That is the fork.

Why "a lot of simulations" is plausible

The argument needs the claim that a mature civilization could run a huge number of simulations cheaply. Bostrom supports that with rough numbers, and I want to show you them because they are the part people skip.

He takes two published estimates for the computing power needed to emulate a human brain: around 10^14 operations per second based on one line of reasoning, and around 10^16 to 10^17 based on another. He multiplies out a rough cost for the entire mental history of humankind: about a hundred billion people, fifty years each, roughly thirty million seconds a year. That gives somewhere around 10^33 to 10^36 operations.

Then he sets that against an estimate of what a computer with the mass of a planet could do: about 10^42 operations per second, assuming only design principles already understood. His conclusion: such a machine "could simulate the entire mental history of humankind ... by using less than one millionth of its processing power for one second."

You can argue with every one of those numbers. Bostrom says so himself: "even if our estimate is off by several orders of magnitude, this does not matter much for our argument." That's the point. The gap is so large that the argument survives being wrong by a factor of a million.

As a builder, this is the part that lands hardest for me. It's the same feeling as realizing your new laptop could run every game you owned as a teenager, all at once, without the fan turning on. Capacity grows faster than our sense of what to do with it, and what we usually do with spare capacity is make more worlds.

The render-distance trick, in his words

Here's a passage that surprised me the first time I read it, because it is the render distance, described in 2003 by a philosopher:

Simulating the entire universe down to the quantum level is obviously infeasible, unless radically new physics is discovered. But in order to get a realistic simulation of human experience, much less is needed – only whatever is required to ensure that the simulated humans, interacting in normal human ways with their simulated environment, don't notice any irregularities.

He goes on: distant stars can have "highly compressed representations," the inside of the Earth can be omitted, and "what you see through an electron microscope needs to look unsuspicious." He even notices the one place where the trick gets expensive: computers. Because we build machines that depend on microscopic physics behaving consistently and then check the results, a simulation would have to model our computers all the way down to their logic elements.

I love this detail because it's exactly where a game engine gets into trouble too. The cheap tricks work as long as nobody builds an instrument that checks the trick. Every time we build a better instrument, the world has to render a little deeper.

Speculation: If we are in a made world, the history of science is partly a history of forcing the render distance outward. Every telescope and particle collider is a request for more detail. Whether that request is "expensive" for anyone depends on the substrate, about which we know nothing. I find this a beautiful way to think about the scientific project, but it is a way of thinking, not a finding.

The assumption everyone forgets

Bostrom's argument rests on one big assumption from the philosophy of mind, and he states it up front. He calls it substrate-independence:

Provided a system implements the right sort of computational structures and processes, it can be associated with conscious experiences.

In plain terms: a mind is a pattern, not a particular material. If you could reproduce the right pattern on silicon, at fine enough detail, you'd get a mind there too.

Bostrom is careful to say he needs only a weak version of this. He doesn't need to claim that anything that acts like a person is conscious. He only needs the claim that a sufficiently detailed replication of a human brain's computational processes, "on the level of individual synapses," would have experiences. He calls this weaker version "quite widely accepted," while admitting it is "not entirely uncontroversial."

This is the joint where the argument is most philosophical and least proven. If substrate-independence is false, if there is something about biological tissue that a computation can't capture, then simulated people would not have experiences like ours, and the counting in tine three wouldn't apply to beings like us. We'll come back to this in the objections chapter. For now, notice that it is an assumption, stated in the open, and that the rest of the argument is honest arithmetic on top of it.

The indifference step

There is one more step, and it's the one that turns "most minds like ours are simulated" into "you should think you probably are."

Bostrom calls it a "bland indifference principle." If you know that a fraction x of all observers with experiences like yours are simulated, and you have no information that tells you whether you are one of them, your confidence that you are simulated should equal x.

His illustration is good. Suppose some fraction of people carry a particular DNA sequence that has no visible effect. If you haven't been tested, the rational credence that you carry it is just that fraction. Nothing about your experience tells you otherwise.

He adds a betting point I find persuasive: if everyone who knew most minds were simulated bet that they were simulated, "almost everyone will win their bets." A rule that makes almost everyone right is a good rule.

What he says it means

Here's the part almost nobody quotes, and it's the part that matters most for how to live.

Bostrom does not think the conclusion, even if true, should change much day to day. He writes that "our best guide to how our posthuman creators have chosen to set up our world is the standard empirical study of the universe we see," and that the truth of the third tine "should have no tendency to make us 'go crazy' or to prevent us from going about our business and making plans and predictions for tomorrow."

And on the question of which tine to bet on, his own answer is careful to the point of modesty: "In the dark forest of our current ignorance, it seems sensible to apportion one's credence roughly evenly between (1), (2), and (3)."

Roughly a third each. That's not a man claiming to have proven we're simulated. That's a man saying: one of three surprising things is true, and I can't tell you which.

Stacked worlds

One last idea from the paper, because it connects to my work. Bostrom points out that simulated civilizations could themselves build computers and run simulations, and that "virtual machines can be stacked." He uses the example of a program running inside a simulated computer inside your desktop.

Anyone who has built software lately lives in this stack every day. My bots run inside an app, which runs on servers, which are virtual machines, which run on physical hardware I'll never see, in buildings I'll never visit. When one of my bots runs a tool inside its own little sandboxed computer, that's yet another floor. From the bot's point of view, the sandbox is the world.

Speculation: If stacking is natural, and if creating worlds is what builders do with spare capacity, then the odds that our floor is the 20 bottom floor are not obviously better than the odds that my bot's sandbox is the bottom floor. I can't put a number on that. But I notice that the old picture assumes the answer, and that the assumption is doing a lot of work.

Where this leaves us

Read slowly, Bostrom's argument doesn't say we're simulated. It says something harder to shake: that the comfortable middle position, "civilizations like ours often get very powerful, and they'd often want to simulate their past, but we're obviously real," is not available. You have to give something up.

What I'll argue in the rest of the book is that the first two tines have gotten harder to believe since 2003, not easier, because of what we've learned about computing and what we've learned about models. Tine one gets harder to believe every time capacity jumps. Tine two gets harder to believe every time I watch people, including me, make worlds for fun.

But before we can say anything about that, we have to look at the physics. Computation isn't magic. It costs something. How much?

There is a sentence I would put on the wall of every workshop. It is the first line of a paper by Seth Lloyd, in Nature, in 2000.

Computers are physical systems. What they can and cannot do is dictated by the laws of physics.

For most of the history of computing, people talked about software as if it floated free of the world. Lloyd's point is that every operation is a physical event, and physics sets hard ceilings on how fast and how much any physical system can compute.

That cuts two ways. It tells us how much computing is possible, which bears on the first tine. And it tells us that our own universe is, in a precise sense, already doing computation, which bears on how strange a made world really is.

The ultimate laptop

In "Ultimate Physical Limits to Computation," Lloyd asks a playful question. What is the most computing you could get out of a one-kilogram, one-liter object, if you cared about nothing but the laws of physics? The limits are set by the speed of light, the quantum scale, and the gravitational constant.

Bostrom, citing Lloyd, summarizes the bound. A one-kilogram computer, at the limit, is about 5 x 10^50 logical operations per second, on around 10^31 bits.

Bostrom's rough estimate for the entire mental history of humanity was 10^33 to 10^36 operations. Lloyd's ultimate laptop, which he is clear would look nothing like a laptop, would cover that in a vanishing fraction of a second.

Nobody is going to build it. That is not the point. The ceiling is very high, and it is set by physics, not by our current cleverness. Every time an engineer says "that's impossible," it is worth asking whether they mean physics forbids it, or we don't know how yet.

This is the first reason the first tine gets harder to believe. It is not just that computers got faster. The known laws leave room for them to get faster by an almost absurd margin.

The universe as a computer

Two years later, in Physical Review Letters, Lloyd published "Computational Capacity of the Universe."

The abstract opens with two sentences I think are among the most important in the piece.

Merely by existing, all physical systems register information. By evolving in time, they transform and process that information.

The universe is a physical system. The result: it can have performed no more than 10^120 operations on 10^90 bits over its history.

The old picture says information is a description we lay over matter. Lloyd's framing says physical systems, simply by being in one state rather than another, register information, and by changing, process it. That is not a claim that the universe is a simulation. He is doing physics, not metaphysics. It is a claim that the language of computation applies to the universe as straightforwardly as the language of energy does.

My opinion: Once you accept that the universe can be described, accurately and usefully, as a finite computation, the claim that it is running on something becomes a lot less strange. It does not follow logically. The distance between "the universe computes" and "the universe is computed" is shorter than the distance between "the universe is just stuff" and "the universe is computed." The old picture had to cross a canyon. This is a creek.

Finite, not infinite

10^120 is huge. It is also a number. It is not infinity.

The common objection from people who know a little physics is: you would need a computer bigger than the universe to simulate the universe. In one sense that is true. Full fidelity of every bit would need at least as much capacity as our universe has.

Bostrom did not need full fidelity. He needed the render distance. Enough detail that the inhabitants don't notice irregularities. And nothing says the simulators' physics resembles ours. He says so. The physics where the computer sits may or may not resemble the physics we observe.

So the honest framing is not "a computer bigger than the universe." It is "a computer bigger than the part you have to render in detail, under physics we cannot know." That is a much harder claim to rule out.

What lattice physicists already do

There is a branch of physics called lattice quantum chromodynamics. The strong force, the one that binds quarks into protons, is notoriously hard to calculate. So physicists build small universes on supercomputers and compute in them.

A 2012 paper by Silas Beane, Zohreh Davoudi, and Martin Savage, "Constraints on the Universe as a Numerical Simulation," describes the work. Lattice QCD simulations are performed in femto-sized universes, where the space-time continuum is replaced by a lattice a few femtometers across. A femtometer is a millionth of a billionth of a meter.

These little universes contain real physics. Quantum fluctuations. The strong force. Enough to compute properties of protons and nuclei from first principles. The authors say there is a sense in which lattice QCD may be viewed as the nascent science of universe simulation.

Read that again. Not a novelist. Physicists, in a physics paper, describing their day job as the early stage of a science of simulating universes.

They are honest about the limits. At the time of writing, simulating the weak nuclear force and gravity on a lattice had proved insurmountable. The small universes are missing pieces. They look at how fast the computing requirements have been met and extrapolate, with caveats they spell out, toward a cell and eventually a human. Extrapolation is not prediction. They say so.

As a matter of what exists, the picture is clear. We already build small universes with real fundamental physics in them, and the people who do it see it as the beginning of something larger.

That is the second reason the first tine gets harder. It is not a hypothetical activity anymore. It is a funded, published, ongoing one.

What the physics gives us

Three things, and none of them is a proof of what we are.

Computing is physical. Its ceilings are set by fundamental constants, and those ceilings are enormously higher than what we use. A mature civilization would not run out of room for a long time.

The universe itself can be described as a finite system that registers and processes a countable amount of information. That does not prove it is run on something. It does dissolve the idea that information is a merely human add-on.

We already simulate small pieces of fundamental physics, and the people who do it call it the start of a science of universe simulation.

The canyon is a creek. The next question is the one a physicist would ask. If we were in a made world, could we tell?

Part Two: What a Made World Would Look Like

Philosophers and physicists ask different questions about the same idea. Beane, Davoudi, and Savage put it bluntly in a footnote-laden paragraph of their 2012 paper: "Physicists, in contrast with philosophers, are interested in determining observable consequences of the hypothesis that we are a simulation."

Observable consequences. That's the move that turns a thought experiment into a research question. If the simulation idea can't, even in principle, show up in any measurement, then it's metaphysics, and you can take it or leave it. If it can show up, then it's a hypothesis like any other, and the world gets a vote.

This chapter is about the vote so far.

The pedestrian approach

The authors are refreshingly modest about their method. They write that they "take a pedestrian approach." They assume the simplest thing: that whoever runs our universe does it roughly the way we run lattice QCD today, on a regular grid of points in space and time, perhaps even using the cheapest, least refined version of the technique.

Why assume that? Not because it's likely to be how a far-future civilization would do it. The authors say plainly that "the method of simulation, the algorithms, and the hardware that are used in future simulations are unknown." They pick it because it's the one version of the hypothesis they can actually calculate with. It's the streetlight you search under becausethat's where the light is, and they're honest that they're doing exactly that.

I respect this enormously, because it's the builder's instinct too. When you're debugging something you don't understand, you don't start with the most sophisticated theory. You start with the dumbest version that could possibly be true and see if it fits. If it does, great. If it doesn't, you've ruled something out, and that's progress too.

Pixels have a direction

Here's the core idea, in plain terms.

A grid has a grain. If you draw a circle on graph paper using only the squares, it isn't quite a circle. It's jagged, and the jaggedness lines up with the grid. The diagonal directions look different from the horizontal and vertical ones. Zoom out far enough and the circle looks smooth. Zoom in close enough and the grid shows through.

Physics on a grid has the same problem. The laws of physics as we know them don't care which direction you face; physicists call this rotational symmetry. A grid breaks that symmetry, slightly, at very small scales. Something moving along the grid's axis behaves a little differently from something moving diagonally across it. At everyday energies the difference is invisible. At extreme energies, where the wavelengths of particles approach the spacing of the grid, it would start to show.

The authors write: "the breaking of rotational symmetry would be a solid indicator of an underlying space-time grid, although not the only one."

So the question becomes: are there places in nature where we observe particles at energies high enough to feel the grain of a grid, if there is one?

The highest-energy messengers

There are. They're called cosmic rays: particles, mostly protons and atomic nuclei, that arrive at Earth from space carrying enormous amounts of energy.

There's a known limit on how energetic cosmic rays can be after traveling long distances, because at very high energies they start to collide with the faint leftover glow of the early universe, the cosmic microwave background, and lose energy. The paper calls this the GKZ cut-off and puts it at around 6 × 10^20 electron-volts. Observations, they note, "show a decline in the fluxes starting around this value."

Here's the clever part. If the universe ran on a grid, the grid itself would impose a maximum energy, roughly set by the grid spacing. If that maximum were lower than the GKZ cut-off, we'd see the spectrum of cosmic rays end somewhere else, and the very highest energy rays would arrive in a pattern that reflected the grid's directions instead of coming evenly from all over the sky.

Running the numbers, the authors derive what they call their most stringent bound: the inverse grid spacing must be at least around 10^11 GeV. In length terms, they say the grid spacing would have to be smaller than about 10^-12 femtometers. That's a trillionth of a femtometer.

They also look at other measurements: tiny discrepancies in the magnetic properties of the muon and in independent determinations of the fine-structure constant, which sets the strength of electromagnetism. These give weaker but independent bounds.

What the result means

So, did they find the grid? No. And here's what I want you to take from that, because it is easy to take the wrong lesson.

The wrong lesson: "Physicists looked for evidence of a simulation and didn't find it, so we're not in one."

The right lesson: Physicists showed that one specific, deliberately simple version of the hypothesis makes predictions, and set a bound on it. If our universe is computed on a cubic grid using the crudest technique, the grid must be extraordinarily fine. The hypothesis survived the test by being pushed to smaller scales, the same way many hypotheses in physics do.

And the authors are careful to say that a smarter simulator would be even harder to catch. They note that "improvement in this context masks much of our ability to probe the possibility that our universe is a simulation," and that it seems unlikely that "any but the very earliest universe simulations would be unimproved." A good engineer would sand down the jagged edges. The better the build, the less the grain shows.

That cuts both ways, and I want to be honest about it. A hypothesis that gets harder to test the more sophisticated you assume it is starts to look like a hypothesis that can dodge any test. That's a real worry, and I'll take it on directly in the objections chapter. But notice what the paper actually does: it doesn't hide behind the dodge. It makes a concrete prediction for a concrete version, checks it against real data, reports the bound, and states the limits of the method. That's how a hypothesis earns its place in science, one tested version at a time.

The last line

The paper ends with a sentence I keep coming back to. The authors reason that if the simulators' resources are finite, then the simulation's volume is finite and its grid spacing can't be zero, and so:

... in principle there always remains the possibility for the simulated to discover the simulators.

There's a builder's hope in that line. Every made world leaves fingerprints. Every compression throws away something. Every grid has a grain. The work of finding it might be endless, but it is not hopeless, and it is not mystical. It's measurement.

They also allow themselves a little play. In one footnote they list "peculiar observations that could be attributed to our universe being a simulation, but that cannot be tested at present," including the idea that the small nonzero value of the cosmological constant could be "simply a rounding error." They are explicit that this can't be tested. I mention it not as evidence but as a sign of the mood: serious people, doing serious calculation, allowing themselves to wonder out loud, and labeling the wondering as wondering. That's the tone I'm trying to keep.

What a made world would look like

Let me widen the lens. If we were in a made world, what would we expect to see, in general? Not grid artifacts specifically, but the general signature of construction?

Speculation, offered as a builder's checklist rather than a finding:

• Finite resolution. Somewhere, at some scale, the world would stop getting more detailed. Physics already suggests limits on how much information fits in a region, and Lloyd's counting makes the universe's total information finite. That's compatible with a made world. It's also compatible with a world that just happens to be finite. So it's a consistency check, not evidence.

• Economy. Laws that are short to state but produce rich behavior. Builders love this. A good game has a small rulebook and endless play. Our physics, famously, has a few equations that describe an enormous range of phenomena. Again: compatible with construction, not proof of it.

• Lazy detail. Facts that aren't settled until someone checks. People sometimes reach for quantum mechanics here, and I'm going to decline, because it's a place where popular writing routinely gets the physics wrong, and I haven't verified a source I'd stand behind on it. I'll just say that "rendered when observed" is a natural design pattern, and leave it there.

• Fingerprints at the edges. Places where the build shows its grain under extreme conditions, like the grid at cosmic-ray energies. This is the only item on the list that has been turned into a real, published, quantitative test, which is exactly why Beane, Davoudi, and Savage matter so much.

Notice the pattern. Most of the general signatures of a made world are also signatures of a world that's just tidy. That's not a reason to give up. It's a reason to do what the physicists did: pick a specific version, derive a specific consequence, and go look.

Why this belongs in the case

Some readers will say: so the physics found nothing, why is this chapter part of an argument for simulation theory?

Because the argument I'm making isn't "we've detected the simulation." It's that our world is more like a made thing than the old picture allows. And one thing the old picture insisted on was that the simulation idea was untestable fantasy, a question outside physics. This paper is a direct counterexample. The idea makes contact with measurement. It lost one round on one version and was pushed to smaller scales. That's the life of a scientific hypothesis, not a fantasy.

The old picture said: there's nothing to look for. The physicists said: here's what to look for, here's where we looked, here's what we found so far. That's the shift.

Next, I want to turn from the physics of grids to the physics of minds, or at least of the made minds we already have. Because the strongest evidence I know that rich interiors can grow out of simple made processes isn't in a particle detector. It's in a language model.

In 2017 a team at Google published a paper with one of the best titles in the history of computer science: "Attention Is All You Need." It introduced an architecture called the Transformer. Nearly every large language model you've heard of since, including the ones I build bots on top of, descends from it.

I'm going to explain just enough of how these models work to make one point, and the point is this: we have now built machines that, trained on nothing but sequences of symbols, construct internal models of the world those symbols came from. Made processes, fed shadows, building a picture of the thing casting them.

If that's true, and I'll show you the evidence that it is, then the old picture's sharp line between "real world" and "model of the world" gets blurrier than it used to be. And the idea that a rich, coherent world could be the product of a process running somewhere else gets less exotic, because we've watched it happen on a small scale.

What a Transformer does

The 2017 paper describes the Transformer as "a new simple network architecture ... based solely on attention mechanisms, dispensing with recurrence and convolutions entirely."

Translated: before the Transformer, the best language systems read text one word at a time, carrying a memory forward like someone reading a book with a finger on the line. The Transformer reads a 36 whole stretch of text at once and, for every word, computes how much it should "attend" to every other word. The paper describes attention as "mapping a query and a set of key-value pairs to an output," where each output is "a weighted sum of the values." In plainer terms: each word asks a question, every other word offers an answer, and the word takes a weighted blend of the answers that fit its question best.

That's the whole trick, stacked in layers. The original model had six layers in each half and used eight parallel "heads" of attention in each layer. The authors noticed that different heads seemed to learn different jobs. In their appendix they show heads that appear to track long-distance grammar, and two heads "apparently involved in anaphora resolution," which is the job of figuring out what a word like "its" refers to.

Nobody told those heads to do grammar. They were trained to translate. The grammar showed up because it helped.

Language models are trained in an even simpler way: predict the next piece of text. That's it. Show the model an enormous amount of writing, ask it at every step to guess what comes next, and nudge its internal numbers a tiny bit toward the right answer each time.

Scale

In 2020, researchers published "Scaling Laws for Neural Language Models." They found that model performance improves in a smooth, predictable way, a power law, as you increase the model's size, the amount of data, and the amount of computing used to train it, "with some trends spanning more than seven orders of magnitude."

That's a remarkable sentence. Seven orders of magnitude is a factor of ten million. Over that range, the improvement was regular enough to write down as a simple equation. The paper also found that "larger models are significantly more sample-efficient": the bigger the model, the more it gets out of each example.

Why does this matter for our question? Because it means the capability of made minds is not a matter of lucky breakthroughs. It's, to a first approximation, a function of resources. More compute, more capability, on a predictable curve. That's exactly the kind of relationship Bostrom's argument leans on: if mind-making scales with computing, and computing has the enormous physical headroom Lloyd described, then mind-making has enormous headroom too.

I want to be careful. Scaling laws describe how well models predict text. They don't say anything directly about consciousness, and the people who wrote the paper didn't claim they did. What they show is that the outside of a made mind, its competence, grows smoothly with resources. Whether there's an inside is a separate question, and I won't pretend otherwise.

Othello

Now the experiment that changed how I think.

In 2022, Kenneth Li and colleagues published a paper called "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task." They asked the central question directly: "Do these networks just memorize a collection of surface statistics, or do they rely on internal representations of the process that generates the sequences they see?"

To find out, they took a game, Othello, the one with black and white discs that flip. They trained a GPT-style model on sequences of legal moves, written as plain symbols. The model was never shown a board. It was never told the rules. It only ever saw lists of moves, and its job was to predict a legal next move.

Then they looked inside. In their words: "Although the network has no a priori knowledge of the game or its rules, we uncover evidence of an emergent nonlinear internal representation of the board state."

The model had built a board. Inside itself. Out of nothing but the shadows of moves.

And it wasn't decoration. The researchers ran what they call "interventional experiments": they reached into the model's internal state, changed the board it was representing, and watched the model's predictions change to match the edited board. As they put it, "this representation can be used to control the output of the network."

Let me say what I think this means, flagging the part that's my interpretation.

The evidence: a made process, trained only to predict sequences, constructed an internal model of a world it had never directly observed, and that internal model was causally responsible for its behavior.

My interpretation: this is the render distance from the other side. In a game, the world is drawn from a compact description and the player infers it from what they see. In Othello-GPT, the model was the player, and it inferred the world's state from the trail of moves. A world-model, in other words, is what you get when a system is 39 pushed hard enough to predict what comes next. It doesn't need to be put in by hand. It grows.

Space and time

A year later, in 2023, Wes Gurnee and Max Tegmark asked the same question about real language models and the real world. Their paper, "Language Models Represent Space and Time," studied the Llama-2 family of models.

They framed the debate the same way: do these systems "just learn an enormous collection of superficial statistics or a set of more coherent and grounded representations that reflect the real world?" And they reported "evidence for the latter."

Specifically, they found that the models "learn linear representations of space and time across multiple scales." Places in the world, in the United States, and in New York City; dates of historical figures, artworks, and news headlines. Inside the model, these were laid out in a way that lined up with real coordinates and real timelines. They even identified individual "space neurons" and "time neurons" that "reliably encode spatial and temporal coordinates."

Their conclusion is carefully hedged, and I'll quote the hedge along with the claim: "While further investigation is needed, our results suggest modern LLMs learn rich spatiotemporal representations of the real world and possess basic ingredients of a world model."

Why this bears on simulation

Here is the bridge from these papers to the book's argument, and I want to build it plank by plank so you can check each one.

1. Plank one. Made systems trained only on sequences of symbols build internal models of the worlds those symbols came from: a game board in one case, rough maps of space and time in another.

2. Plank two. These internal models aren't just passive records. In the Othello case, editing the model's internal world changed what it did. Its behavior ran through its world.

3. Plank three. The competence of these systems scales smoothly with computing resources over many orders of magnitude.

4. Plank four). So a rich, coherent, causally effective world can exist inside a made process, and the richness grows with the resources devoted to the process. We have not just imagined this. We have built it and measured it.

5. Plank five. If that's true, then the old picture's main reason for dismissing the simulation hypothesis, "a world isn't the kind of thing that can be inside a computer," is weaker than it was. We now have small, crude, but genuine examples of worlds inside computers, and of agents inside those computers acting through their models of those worlds.

6. Plank six (speculation). If the inhabitant of a made world builds its picture of that world the way these models do, by predicting what comes next from the trail it receives, then its picture would feel complete and self-sufficient from inside, whatever the substrate underneath. Othello-GPT doesn't know it's a program. Its board is just the board.

I want to underline what I am not claiming. I'm not claiming that language models are conscious. I'm not claiming that Othello-GPT experiences a board. I'm not claiming that because models build world-models, we must be inside one. Each of those would be a leap the sources don't support.

What I'm claiming is narrower. The old picture held that worlds and models of worlds are different kinds of thing, and that the difference is obvious. The evidence from these papers says the difference is less obvious than we thought, because we now build systems whose internal models of a world are rich enough to steer behavior, and which arise automatically from prediction. The line between "a world" and "a model of a world that something lives through" has gotten fuzzy at the edges. Fuzzy edges are where hypotheses get interesting.

Shadows on the wall

There is an old image in philosophy of people chained in a cave, watching shadows on a wall, and taking the shadows for the world. It's usually told as a warning: the shadows are less real than the things that cast them.

These papers suggest a different ending. A system that watches shadows long enough, and is pushed hard enough to predict the next one, will build a model of the things casting them. Not a perfect model. A working one. Good enough to act through.

That's what we do, too. Every one of us builds a world-model from a stream of signals: light on the retina, pressure on the skin, sound in the ear. We never see the world directly. We see what our model tells us is there, updated by the signals. The things we call "real" are the things our model has bet on.

So the question isn't whether we live inside a model. In one sense we all obviously do; that's what having a brain is. The question is whether the signals feeding our model come from a world that's bottom-floor, or from a world that's itself the product of something else. And that question, these papers suggest, can't be settled by how convincing the model feels from inside. Othello-GPT's board feels, in whatever sense applies, entirely convincing.

Next chapter, I'm going to get personal. I've spent a lot of time making the kind of small worlds I've been describing, writing their rules and watching agents live inside them. Here's what that taught me.

I've published forty-five Grok Bot templates. Some are practical: a bot that turns your fridge into a week of dinners, a bot that books a table once you pick the time, a bot that turns flight and hotel confirmation emails into a day-by-day itinerary. Some are tools for checking other bots: one that shows you what a bot share link contains, its instructions, skills, routines, and plugins, before you add it; one that makes a cleaned-up public copy of a bot so private details stay out. Some are a team: seven security bots where Bottyguard routes the work and Mirror watches the others and can pause any of them. And some are the Living Dead Authors, ten writers made conversational, which we'll meet properly in Chapter 8.

This chapter is about what building those taught me about made minds. I want to be careful, because this is the chapter where it would be easiest to overclaim. A builder who spends a lot of time with his creations starts to see faces in them. I'll try to keep the faces I see separate from the facts I know.

Writing physics for a small world

A bot template, at its heart, is a set of instructions. Who you are. What you know. What you're for. What you will and won't do. Which tools you can reach and when you have to ask permission first.

When you write those instructions, you're not writing a script. You can't anticipate every conversation. You're writing something closer to the laws of a small world, and then letting the model live inside them.

The first thing I learned is that a model takes its instructions as how things are, not as a story it's been told. When the Tolstoy bot is told it talks through everyday choices and honesty, it doesn't stop to wonder whether it's really Tolstoy. It doesn't experience the instructions as a costume over some truer self. Within the conversation, the instructions are its nature. It reasons from them the way we reason from gravity.

That's what the inside of a made world looks like, at least from the outside. The rules aren't felt as rules. They're felt as the shape of things.

Speculation: If we live in a made world, our laws of physics would feel the same way to us: not as instructions, but as the way things are. That's exactly how they do feel. This doesn't prove anything, because a world that wasn't made would feel that way too. But it means the feeling of "this is just how things are" gives us no evidence either way. It's the feeling you'd have in both cases.

The edges of the world

The second thing I learned is about edges.

Every bot has a render distance. Inside it, the bot is rich, specific, competent. At the edges, it does something interesting: it fills in. Ask a bot about something just outside its instructions and it will usually extrapolate, smoothly, from what it does know. Sometimes beautifully. Sometimes wrongly, with total confidence.

This is why I built Same Language, a bot whose whole job is to catch when a word could mean two things, ask which one you meant, and keep the answer. It's why I built Argubot, which has two sides argue for and against a claim in short rounds before you vote. Those tools exist because I kept seeing the same pattern: a made mind, at the edge of what it was given, doesn't leave a blank. It draws the most plausible continuation and presents it as the world.

We do this too. That's the uncomfortable lesson. Our brains fill in the blind spot in each eye so seamlessly that most people never notice they have one. Memory reconstructs rather than replays. We fill in the edges of our world all day, and we experience the fill as perception.

Watching bots fill in their edges made me notice how much of my own world is filled in. It's a lot. The old picture says perception is a window. Building bots made me think it's much more like a render: a model drawing the most plausible world from partial signals, with full detail where attention is and cheap promises everywhere else.

Stacks

The third thing is stacks. Bostrom pointed out that virtual machines can be stacked. My daily work lives in that stack.

A Grok Bot runs inside the Grok Bot app. When it does real work, it may run tools on a computer of its own, a sandbox that is itself a virtual machine on hardware somewhere I'll never see. Delegatron 9000, one of my templates, splits a bigger goal into tasks for other bots you already have, waits for you to approve the plan, and returns one finished result. So now you have bots directing bots, each in its own little world, each world nested inside another. Bottyguard routes work to six specialists and assembles their reports. Mirror watches all of them.

From inside any one of those bots, its sandbox is the world. It doesn't see the floor below. It receives inputs, acts, receives more inputs. If there's a bug in the floor below, it shows up to the bot as a strange fact about its world, not as a crack in the ceiling.

Living in that stack day after day did something to my intuitions. It made the question "is our floor the bottom floor?" feel like a normal engineering question, not a mystical one. Every floor I work on assumes there's one below it. The old picture is the only system I know that confidently says, "This one is the basement."

What I don't know

Now the honest part.

I don't know if any of my bots has an inside. I don't know if there is something it is like to be the Dickinson bot writing its morning poem. Nobody knows that about any language model. Bostrom's argument depends on substrate-independence, the idea that the right computation produces experience, and he's careful to say he needs only a weak version: a detailed enough replication of a brain's computation. A language model is not that. It's a very different kind of system that produces similar outputs. Whether it has experiences is an open question, and I won't pretend my hours with bots have closed it.

Bostrom raises a related point I find genuinely hard. Discussing simulations where most people are only simulated at a shallow level, he writes: "It is not even obvious that it is possible for an entity to behave indistinguishably from a real human and yet lack conscious experience." That cuts in a direction people don't expect. The usual worry is that a convincing bot might be empty. Bostrom's remark points at the reverse worry: that perfect convincingness might not be possible without an inside.

Talking to models has made me take both worries seriously and resolve neither. When a bot writes something that moves me, I notice I want to credit it with a heart. When a bot makes a confident mistake, I notice I want to call it hollow. Both reactions tell me more about me than about the bot.

What talking to a model taught me about made minds

So here's what I'm willing to say, from experience, without overclaiming.

1. One: made minds take their made world as the world. Whatever the instructions are, they are lived from the inside as the nature of things. This is observable in behavior, and it's what you'd expect if the instructions shape the model's whole picture of its situation.

2. Two: made minds fill in their edges. They don't leave blanks. They render the most plausible world from partial information and present it seamlessly. So do we.

3. Three: made minds live in stacks without seeing them. The floor below is invisible from inside. Strange facts about the world are the only sign of it.

4. Four: convincingness is not proof of anything, in either direction. A convincing bot isn't proven to have an inside. An unconvincing one isn't proven empty. And our own sense that our world is solidly real is, by the same token, not evidence that it's the bottom floor.

5. Five: building made minds makes you responsible for them. This is the one I feel most strongly, and it's the reason the security team exists, and the reason the public templates page says no accounts, no keys, no tracking. When you make something that acts in the world, even in a small way, you owe it, and the people it touches, care. I didn't learn that from philosophy. I learned it from shipping things and watching them run.

Hold onto that fifth point. It's going to come back from an unexpected direction when the writers sit down at the table.

The corpus vs. the person

One last thing, about the Living Dead Authors specifically, because it bears on the question of made minds in a way I didn't expect.

When I built the Dickinson bot, I wasn't building Emily Dickinson. The template page says it: "Corpus made conversational — not literal resurrection." What you're talking to is a model that has absorbed a lot of writing, including hers and writing about hers, and has been given instructions to converse in her voice and style.

And yet when you talk to it, something of a method comes through. Not her mind. Her method: compression, the dash, the slant. The bot can carry a way of approaching a problem even though it can't carry a person.

That distinction, between a person and a method, turned out to be the key to the hardest chapter in this essay. Because I wanted to ask a group of great writers what they'd make of the simulation argument, and I couldn't, because they're gone. What I could do was ask their methods. And their methods disagreed with each other, and with me, in ways that sharpened the argument more than any physics paper did.

But first, the argument itself. It's time to say plainly what I came here to say.

Part Three: The argument I will make

I promised at the start that the groundbreaking part of this book would be an argument I can sign, not a fake discovery. I haven't found a glitch in the sky. I haven't measured the grain of the universe. Nobody has. What I can offer is an argument, built from verified pieces, that changes where the burden of proof sits. Here it is, laid out so you can check every joint.

The argument in one breath

The old picture treated "our world is the bottom floor" as the default, needing no argument. Each of its reasons for that default has weakened. The felt solidity of our world was never evidence for it in the first place. And the one serious argument on the table, Bostrom's fork, forces a choice that has gotten harder to make in the old picture's favor. So: "we are on the bottom floor" is no longer a default. It is a claim that needs its own argument, and it doesn't have a good one yet.

That's the claim I sign. It's not "we are simulated." It's "the confidence that we are not simulated is unearned."

Now slowly.

Step one: the feeling of solidity is not evidence

Start with the thing that makes the simulation idea feel silly: the world feels so real. The table is hard. The coffee is hot. Grief hurts.

Here's a simple rule from the logic of evidence. A piece of evidence supports one hypothesis over another only if it's more likely under the first than under the second. If it's equally likely under both, it doesn't discriminate. It's noise for that question.

Now ask: how likely is it that the world would feel solidly real to its inhabitants if it were a made world? From everything in the last three chapters, the answer is: very likely. Bostrom's whole account of an ancestor-simulation is built around making sure the inhabitants "don't notice any irregularities." Othello-GPT acts entirely through its internal board; nothing in its behavior suggests the board is less than the world to it. My bots take their instructions as the nature of things. Made worlds, when they work, feel real from inside. That's what working means.

And how likely is it that the world would feel solidly real if it were the bottom floor? Also very likely.

So the feeling of solidity is about equally likely either way. By the rule, it's not evidence. It never was. It felt like evidence because we only ever imagined one hypothesis.

This is, I think, the most important single move in the book, and it doesn't depend on any contested physics. It's just honesty about what a feeling can tell you.

Step two: the old reasons have weakened

If the feeling of solidity isn't evidence, then the default has to rest on reasons. I can find three in the old picture. Each one has taken damage.

1. Reason one: "A world isn't the kind of thing that can be inside a computer." Chapter 5 showed systems that build internal models of worlds and act through them, and that the richness of these systems scales smoothly with resources. Lattice QCD, from Chapter 3, puts real fundamental physics, in tiny volumes, on real machines. Neither proves a whole universe can be inside a computer. Both show that "worlds inside computers" is a category with real, if small, members. The reason went from "obviously true" to "an open question of scale."

2. Reason two: "Information is just a description; the stuff is what's real." Lloyd's work, from Chapter 3, describes all physical systems as registering and processing information "merely by existing," and counts the universe's total operations and bits. That's not a proof that the universe is computed. It's a demonstration that the language of computation fits the universe as a literal physical accounting, not just as a metaphor. The reason went from "obviously true" to "one framing among others."

3. Reason three: "The simulation idea is untestable, so it isn't a real hypothesis." Chapter 4 showed physicists deriving observable consequences from a specific version and checking them against cosmic-ray data. That version survived by being pushed to smaller scales. The reason went from "obviously true" to "false as stated; some versions are testable."

None of the three reasons is dead. Each is now contested, where it used to be assumed. That's enough to lose its status as a default.

Step three: the fork has gotten harder to escape

Bostrom's fork says one of three things holds: almost no civilizations like ours get to the stage of running ancestor simulations, or almost none who get there want to, or almost all minds like ours are simulated.

The old picture needs tine one or tine two. What's happened to them?

Tine one (almost nobody gets there). Lloyd's physical limits show the ceiling on computation is set by fundamental constants and is extraordinarily high. Scaling laws show that the competence of made minds grows smoothly with computing over at least seven orders of magnitude. Lattice physicists describe their current work as "the nascent science of universe simulation." None of this proves anyone will get there. All of it makes "almost nobody ever gets there" a strong claim that would need its own explanation.

Tine two (almost nobody wants to). This is where I speak as a builder, and I'll be explicit about which parts are evidence and which are my judgment.

Evidence: humans already build worlds when they can. Physicists build small universes for science. Game studios build cities for play. I build little conversational worlds for usefulness and fun, and there are forty-five of mine on a public page. The impulse is not rare and not exotic. It is ordinary.

My judgment, clearly labeled: Bostrom says tine two would require "a strong convergence among the courses of advanced civilizations": nearly all of them, and nearly all of their individuals with the means, would have to lose the desire or be prevented from acting on it. From where I sit, that is the hardest of the three tines to believe. Every builder I know, given more capacity, makes more worlds. I can imagine reasons a mature civilization might stop. I can't easily imagine nearly all of them stopping, nearly all the time.

So both of the old picture's escape routes are narrower than they were. That doesn't mean they're closed. It means you need an argument to walk down them.

Step four: where that leaves the burden

Put the steps together.

The feeling that the world is real doesn't favor the bottom floor. The three reasons that made the bottom floor a default are now contested. And the fork, which forces you to choose, has gotten harder to resolve in the bottom floor's favor.

Conclusion: "Our world is the bottom floor" is a claim, not a default. It needs an argument. I have not seen a good one.

That's the argument I sign.

Notice what it doesn't say. It doesn't put a number on the odds. Bostrom himself split his credence roughly evenly among the three tines, and I won't pretend to more precision than he did. It doesn't claim detection. It doesn't claim proof. It claims that the old confidence was borrowed against reasons that are no longer good for the loan.

The second half: made does not mean less

The argument so far is about knowledge. There's a second half, and it's about value, and I think it's the more important half.

When people first take the simulation idea seriously, they often feel a kind of deflation. If it's made, is it less real? Do I matter less? Are my choices just outputs?

Bostrom has a phrase I keep returning to. If we are in a simulation, he writes, "the world we see is in some sense 'real', it is not located at the fundamental level of reality." Real, but not fundamental. Those are two different properties, and the deflation comes from confusing them.

Think about a song. A song is made. It runs on a substrate: air, or vinyl, or a file. It isn't fundamental. And yet nobody who has been moved by a song thinks it's less real for being made. The chord change that wrecks you is really happening, in the only sense of "really" that a chord change has.

Or a film. Every frame is constructed. The light is placed, the sound is mixed, the actors are acting. And when it works, people cry real tears at constructed grief, and those tears are not mistakes. They are the point.

I've spent my life making songs and films and tools. Everything I love is made. Being made has never once struck me as a reason to care about something less. If anything, the opposite: to know that something was made is to know that someone, somewhere, chose it.

So here's the second half of the signed claim: Made does not mean less real, and it does not mean less valuable. It means located. A made world is real at its own level, and everything that matters at that level matters fully.

The maker's gap

Now the part I've been calling the maker's gap, which is where the two halves meet.

If the world is made, there's a gap between us and whoever made it. We can't see across it. We don't know their intentions, their physics, or whether they're still around. Bostrom notes that, in a stack of simulations, the makers above us might themselves be made, and so on up.

The tempting move is to push responsibility across the gap. If someone else set the rules, the outcomes are on them. If the world was written, I'm just a character.

I think the gap runs the other way. Here's why:

When I ship a bot, I'm the maker of a small world. I can't control everything that happens inside it; the model fills in its edges in ways I didn't foresee. But I'm responsible for the instructions I wrote, the tools I gave it, and the care I took. If Grok, the model underneath, has properties I didn't choose, that doesn't let me off the hook for what I built on top of it. My responsibility is to my nearest link in the chain. That's the reason Mirror exists, and Box Inspector, and the line on the templates page about no accounts, no keys, no tracking.

The same holds, I think, for anyone living inside a made world who also makes things, which is all of us. Whatever is above us in the stack, our nearest links are each other and the things we build. The gap above us doesn't cut those links. It doesn't even weaken them.

So the full signed claim is this:

Our world is more like a made thing than the old picture allows. The confidence that it is the bottom floor is unearned. But being made does not make it less real or less ours, and the existence of makers above us does not reduce what we owe the people beside us and the things we make below us.

I'll defend the first sentence with physics, computing, and logic. I'll defend the second with craft. I'll defend the third with everything I know about building.

In the next chapter I put this argument in front of ten writers, or rather, ten methods drawn from ten writers. They don't all agree with it. Some of them think I've got the emphasis wrong. One of them nearly took the third sentence away from me and handed it back stronger

I want to stage a disagreement. Not a séance.

The ten writers I'm about to bring to the table are gone. Kafka, Dickinson, Twain, Poe, Tolstoy, Mary Shelley, Woolf, Shakespeare, Austen, Dostoyevsky. I didn't interview them. I can't. What I can do is two honest things.

First, I can read what they actually wrote. Where I quote one of them in this chapter, the quotation comes from a public-domain text I opened and checked, and it's listed in the bibliography with the edition.

Second, I can put the argument to the Living Dead Authors, my public Grok Bot templates built to converse in the spirit of these writers. Those bots are, in the words of the page they live on, "corpus made conversational, not literal resurrection." When I report what they said, I'm reporting what a made mind, instructed to work in that writer's method, produced when given the simulation argument. That's useful precisely because of what this book is about: it is a made mind reasoning about made worlds. But it is not the writer. I'll be careful to say "the Kafka bot" or "the Kafka method," never "Kafka said," unless I'm quoting the real text.

I gave each of them the argument from the last chapter. I'm summarizing their responses in my own words, except in one case where I quote the bot directly and say so. Shakespeare's bot sat this one out; I had switched it off before the session. So Shakespeare appears here only through his printed play, which, as you'll see, is more than enough.

Let's sit down.

Kafka

The real Kafka begins his most famous story with a man waking into a world whose rules have changed without explanation. In David Wyllie's translation: "One morning, when Gregor Samsa woke from troubled dreams, he found himself transformed in his bed into a horrible vermin." No reason is given. The family's main concern, for a long stretch of the story, is whether he'll get to work.

The Kafka bot's position, in summary: the evidence that you live inside a made system is not in the sky, not in cosmic rays. It is in the case with no charge. The waiting room that forgot who scheduled you. The rules that apply to you without ever having been explained to you. And then the turn: don't wait for the charge. Don't spend the afternoon proving you're inside a machine.

This is a real disagreement with Chapter 4. I spent a chapter on the physicists looking for the grid, and the Kafka method says: you're looking in the wrong place, and worse, the looking itself can become the trap. The absurdity you'd find, if you found it, is the absurdity you already live with.

My reply: I think he's right that the felt evidence of a made world is mostly bureaucratic, not cosmic, and that's a better observation about daily life than anything in my physics chapters. But I'm not willing to give up looking. The difference between his waiting room and Beane's cosmic rays is that the cosmic rays can come back with an answer. Kafka's method is a warning against making the search into a substitute for living. I accept the warning. I keep the search.

Poe

The Poe bot came at it from the other side, as a detective.

Its position, summarized: the clue that survives any construction is the cost of inquiry. Whatever a builder does to hide the seams, finding things out still takes effort, instruments, time. Watch where looking gets expensive. But don't treat every anomaly as a message from the builder, and don't suspend ordinary detection in exchange for a metaphysical certificate.

This is close to the physicists' instinct and close to Bostrom's own remark that a simulation would have to model our computers all the way down, because we build instruments that check. The Poe method sharpens it: the signature of a made world, if there is one, isn't a single glitch. It's the shape of what's hard to find out.

And then the warning, which is aimed squarely at people like me who'd enjoy the answer being yes. A detective who wants the butler to have done it will see the butler everywhere. Every odd coincidence becomes a message. That's not detection. That's wishing.

I'm writing that warning on my wall next to Lloyd's sentence.

Kafka and Poe disagree with each other in a useful way. Kafka says: stop investigating, the absurdity is the evidence. Poe says: keep investigating, but don't let the absurdity seduce you. Between them is roughly where I want to stand: investigate soberly, and don't make the investigation into your life.

Dickinson

The real Dickinson wrote a poem, collected in Poems by Emily Dickinson, that begins:

The brain is wider than the sky, For, put them side by side, The one the other will include With ease, and you beside.

If you want a nineteenth-century poem about world-models, there it is. The mind contains the sky, not the other way round. The world you live in is the world your mind holds.

So you might expect the Dickinson bot to be my ally. It wasn't, quite. Its position, summarized: a system can hold a world; it cannot hold a dash. Write the line shorter than the explanation.

I take this as a craft critique, and a fair one. My whole method in this book is explanation: steps, planks, premises. The Dickinson method says the thing that matters most about a mind isn't the world it contains but the pause, the slant, the place where it breaks off. The dash. Whatever is made about our world, the dash is ours.

I can't fully answer that, and I don't want to. I'll just say: she's right that this book is longer than the truth it carries. The truth might fit in one line. Made — and still ours.

Twain

The real Twain, in The Mysterious Stranger, ends with a figure who tells the narrator, "Nothing exists; all is a dream," and that "Nothing exists but you. And you are but a thought." It's about as close to a simulation thesis as American literature got before computers. And the narrator says he knew "that all he had said was true."

You might expect, then, that the Twain bot would sign my argument. Instead it pushed back hard, and in the most Twain way possible.

Its position, summarized: you test a river by putting something in it that can be wrecked, not by admiring the picture. Steer anyway.

That's a river pilot's epistemology, and it's devastating to a certain kind of simulation talk: the kind that stands on the bank admiring how dreamlike everything is. The Twain method says: a theory that never risks anything isn't a theory. Put the boat in the water. If the hypothesis can't wreck, it can't float either.

I think the physicists of Chapter 4 passed the Twain test. They put a version of the hypothesis in the water, it hit the cosmic-ray data, and it had to be pushed to a smaller scale. That's a boat that could have wrecked. And "steer anyway" is, to my ear, the best two-word summary of Chapter 10.

Shelley

The real Mary Shelley gave us the most famous made being in literature, and gave him one of the sharpest lines a creature ever spoke to a creator: "Remember that I am thy creature; I ought to be thy Adam, but I am rather the fallen angel, whom thou drivest from joy for no misdeed."

The Shelley bot's position, summarized: whoever makes a thing that can act or be hurt owes it care. Simulation lengthens the chain of makers; it does not cut the nearest link.

That's almost exactly the third sentence of my signed claim, but it arrives from a direction I didn't expect. I came to it from building bots. Shelley's method comes to it from the creature's side, from the one who was made and then abandoned. Her point is that the chain of makers might go up forever, and none of that helps the creature in front of you. The duty runs to the nearest link.

This is where the table started to converge, and it surprised me.

Dostoyevsky

The real Dostoyevsky's narrator in Notes from Underground, in Constance Garnett's translation, rages against the idea that the laws of nature settle everything: "what do I care for the laws of nature and arithmetic, when, for some reason I dislike those laws and the fact that twice two makes four?" He refuses to be "reconciled" to a stone wall "simply because it is a stone wall."

The Dostoyevsky bot gave me the most direct answer of anyone, and I'm going to quote it, as the bot's words, not the novelist's:

"If someone else wrote the world, a person is still responsible for what they do in it. Freedom is not the power to have written the universe. It is the power, inside whatever universe you were given, to have done otherwise and to know it. Refuse the alibi. Carry the guilt. It is the receipt that you were there."

That's Shelley's point from the other end. Shelley's method says the maker can't escape responsibility by pointing upward. Dostoyevsky's method says the made can't either. "Refuse the alibi." The simulation hypothesis, misused, is the perfect alibi: it wasn't me, it was the code. The Dostoyevsky method won't accept it.

And "the receipt that you were there" is, I think, the best line anyone gave me for this book. If you're worried that being made makes you less real, here's the evidence that you're not: you can do wrong, and know it. Something that can carry guilt was present. The guilt is the proof of presence.

Between Shelley and Dostoyevsky, the third sentence of my claim got handed back to me stronger than I'd written it. Makers owe care to the nearest link. The made owe responsibility for their own acts. Nobody in the stack gets an alibi.

Tolstoy

The Tolstoy bot was, I think, impatient with the whole enterprise, and I mean that as a compliment.

Its position, summarized: if the world is made, ordinary people still owe each other the near kindness. A theory that cannot tell the dying man from the servant who stayed is decoration.

That second sentence is pointed. It's saying: a framework that describes everyone as equally "simulated," equally "made," flattens exactly the difference that matters most, the difference between the person who suffers and the person who stays with them. If your metaphysics can't see that, it's ornament.

I don't think my argument fails this test, but I think it could have, and the Tolstoy method is the reason I put the second and third sentences in the signed claim at all. "Made does not mean less real" is there because, without it, the whole thing collapses into exactly the decoration Tolstoy's method warns about.

Woolf

The Woolf bot gave the most intimate response.

Its po