The universe as a minified bundle

01 / Running in prod Imagine finding a file called universe.js. It is 1.3 pebibytes, minified, and apparently contains everything. No source maps. No comments. No Git history. No README written by some exhausted engineer explaining why gravity depends on a utility called q7. Just one endless slab of punctuation, single-letter variables, collapsed functions, and deeply suspicious side effects. You are told the file runs stars, electrons, black holes, time, chemistry, your nervous system, and whatever consciousness is. You do not get the source. This is not quite our situation, but it is uncomfortably close. We can run the application because we are inside the application. We can hit it with things, freeze parts of it, accelerate parts of it to nearly the speed of light, bounce photons off it, put bits of it in vacuum chambers, and spend eleven billion dollars building a circular machine whose primary purpose is to smash extremely small things together and inspect the debris. What we cannot do is ask the universe to open the repository. Physics is what you do when prod is running, the source is missing, and somehow you still have to figure out the architecture. Physics is what you do when prod is running, the source is missing, and somehow you still have to figure out the architecture. Humans reverse engineered this thing with comically primitive tooling. Newton watches objects fall and discovers that nature is not improvising. The same rules keep appearing. Planets, cannonballs, apples: wildly different tickets, same underlying function. Maxwell later discovers that electricity, magnetism, and light, which had been sitting in different conceptual folders, are really one system. Einstein then discovers that space and time themselves are not the static coordinate grid under the program. The coordinate grid is executable. 02 / The interface Then quantum mechanics PR is merged and everybody has a bad century. Things behave like particles and waves depending on how you interrogate them. Empty space turns out to be busy. Measurement somehow cannot be treated as an innocent spectator. Entangled particles produce correlations across large distances that make our everyday picture of separate little objects carrying private local state look naive. The theory is not failing. That would be merciful. It works so well that entire industries depend on it. This is the point where Donald Hoffman makes me do that dog tilted head perplexed look thing. Hoffman argues that what we perceive is not reality as it exists in itself but an interface shaped by evolution. The claim sounds mystical until you translate it into just optimization. Evolution does not optimize for truth. It optimizes for fitness. If a brutally simplified representation helps an organism survive better than an accurate one, evolution has no moral obligation to preserve the accurate version. A red apple is not reality generously revealing its native object model. It is more like a UI element marked edible thing. It's this far away, this big, and probably worth reaching for. The visual system does not need metaphysical fidelity. It needs latency low enough that your ancestors can eat lunch without becoming lunch. An organism that perceives the deep ontology of fruit but requires forty seconds to decide whether to bite it is not going to leave descendants. Nature has always had strong opinions about performance. Hoffman likes the desktop icon analogy. The blue folder on your Mac is real in a perfectly useful sense but nothing inside the machine resembles a tiny blue folder. You confused the interface with the implementation. The icon succeeds because it hides almost everything that is actually happening. The Filesystem metadata, memory addresses, controller logic, cache state, electrical transitions, NAND cells, and billions of events that would make opening 2025_federal_taxes.pdf feel like operating a nuclear submarine. You confused the interface with the implementation. The uncomfortable move is to ask why human perception should be exempt from this principle. Maybe the moon is real in roughly the way the folder is real. Something absolutely exists, and that something reliably participates in interactions we can model, photograph, land spacecraft on (insert fake moon landing joke here) and bounce lasers off. But it does not follow that the deepest form of that thing is literally a gray sphere occupying coordinates in a fundamentally 3D container. The sphere may be the icon. We might have spent thousands of years becoming extremely sophisticated experts in the behavior of icons. 03 / Wrong layer Look at entanglement. Spooky action at a distance is famous because the situation seems insane if distance is part of the substrate. Two systems far apart behave as though some relationship between them matters more deeply than our classical picture allows. And nothing usable is simply shooting faster than light between them. The phrase itself quietly assumes the thing under dispute is distance. If spacetime is an interface, then far apart may describe the rendering rather than the architecture. Programmers have seen this kind of mistake before. Two controls can appear on opposite sides of a giant application and seem utterly unrelated. One is buried under Billing, another under Security, a third lives in some modal nobody likes. Then AI finally reads the bundle and discovers that all three call the same internal function with different flags. From the user's perspective they were far apart. From the implementation's perspective they were cousins sharing a kitchen. Once you allow yourself to think this way, the programmer part of the brain immediately becomes irresponsible. If spacetime is an interface, is there some lower-level operation that bypasses what looks like movement? Could two places that are separated by thousands of miles in the rendered world be related in some completely different way underneath? Could teleportation, or something that looks even stranger from our layer, turn out not to require physically traversing all the intervening points? This is where the analogy becomes dangerous, because every engineer who has ever found an undocumented endpoint starts wondering whether reality has one. The stupid version is person.location = "tokyo". That is almost certainly too cute by several orders of magnitude, but the more serious point survives. Suppose you use an application for ten years and discover that a certain screen is only reachable through six menus. You document the route, test it thousands of times, and eventually begin speaking as though the 6 step sequence is a law of the system. Then someone opens DevTools and calls the route directly. Nothing supernatural happened. You had discovered a law of navigation and promoted it into a law of architecture. This does not mean the speed of light is merely a menu animation or that relativity disappears if we become clever enough. Some constraints really are architectural. Conservation laws may be architectural. Causality may be architectural. We do not know. The point is narrower and more annoying. When you only have access to the rendered layer, interface constraints and implementation constraints can look identical for a very long time. Science is a giant company where every team has spent eighty years inventing its own names for user_id. 04 / Recovering names Physics has a scaling problem. No human being can keep quantum field theory, general relativity, cosmology, quantum information, topology, condensed matter, black-hole thermodynamics, particle physics, scattering amplitudes, and the relevant mathematics simultaneously loaded in working memory. The scientific corpus is too large, and the notation is fragmented across communities that often reinvent similar ideas using different language. One field calls something x, another calls a cousin of it y, and a third publishes the same structural insight in a journal nobody in the first two groups reads. Science is a giant company where every team has spent eighty years inventing its own names for user_id. That is where advanced AI could matter because it may become the first thing capable of holding enough of the conceptual codebase at once to notice that pieces we treat as separate are actually the same machinery. Maybe an equation from scattering amplitudes has the same deep structure as something in quantum information. Maybe an obscure topology theorem from 1987 and a modern black-hole result are different presentations of one object. Maybe several fields have already touched the same underlying function and failed to recognize it because the variable names were minified by history. There are already signs of that transition happening in 2026: an AI-assisted discovery paper in theoretical physics describes a system autonomously solving an open problem, while Nature Machine Intelligence has reported on AI’s rapid transformation of mathematical research. Science has done this before. Electricity and magnetism were different until they were electromagnetism. Space and time were different until they became spacetime. Mass and energy looked like separate concepts until an equation revealed an exchange rate. Great scientific advances often look in retrospect less like discovering an entirely new feature and more like realizing that the codebase had duplicate abstractions. Nature did not change. Someone finally noticed that two folders pointed to the same thing. Nature did not change. Someone finally noticed that two folders pointed to the same thing.