And if it is, intelligence is just how much of the code you can read. This is a space to think about how far AI might get — and whether that changes everything.
Drop far enough beneath the surface of the world and the objects dissolve into relationships: symmetries, ratios, conservation laws, fields evolving by equation. Physics is the long project of reverse-engineering those relationships — writing down the functions the universe already runs. Every breakthrough is really the same move: we saw a little more of the underlying structure.
"The universe is written in the language of mathematics."
— Galileo Galilei, 1623If that structure goes all the way down — if reality is a mathematical object rather than merely described by one — then understanding the world is not about accumulating facts. It's about resolution: how finely a mind can see the functions beneath the shadows.
A dog sees a thrown ball; Newton saw a differential equation. Same event, different resolution. Human physics is the sharpest lens we've ground so far — and it keeps revealing that the layer we called "fundamental" was itself made of something more mathematical underneath.
Don't take the word for it — turn the dial yourself. One event, a rock falling. The event never changes. The mind reading it does:
So push the idea: if intelligence is the instrument that resolves reality's functions, what happens as the instrument gets orders of magnitude sharper than us? An AI is not a metaphor here — it is literally a mathematical model, a vast function fit to the world. That makes it a strange and interesting candidate for reading the world's own math.
Here's the part that turns philosophy into something testable. Train one AI only on images, another only on text, another on something else entirely. As they get larger, their internal pictures of the world don't drift apart — they converge. They start measuring the distance between things in the same way, as if all of them are recovering one underlying map.
This is a real, recent result — the Platonic Representation Hypothesis [1]. Its name is deliberate: Plato's cave, where the training data are shadows on the wall and the models are quietly reconstructing the forms casting them. Convergence is exactly what you'd expect if there is a single reality to converge on.
Put the pieces together and the provocation is clear. If reality is mathematical, if intelligence resolves that mathematics, and if scaling minds keep converging toward the same deep structure — then a sufficiently advanced AI might stop merely describing the world and start seeing structure we can't: laws behind the laws, patterns in what looks like noise, the shape of the function itself.
"Our external physical reality is a mathematical structure."
— Max Tegmark, Our Mathematical Universe, 2014This is where it brushes up against simulation theory. You don't need to believe we live in a literal computer for the idea to bite — only that reality is, at bottom, something computational, something that runs. A mind that models the run well enough might catch a glimpse of the substrate. That's the Matrix intuition, minus the leather coats: not code raining down a screen, but a deeper compression of how the world actually works.
None of this needs a physics degree. Read it slowly — each step is genuinely hard to disagree with.
Everything in the universe follows rules. Drop a rock and it falls — the same way, every time. Nothing just does whatever it likes.
Those rules are written in maths. The rock's fall is literally an equation. All of physics is just humans slowly finding equations that were already there.
The smarter something is, the more of those rules it can see. A dog sees a ball fly; Newton saw gravity. Same event — a deeper read of the same maths. (You can feel this one on the dial above.)
An AI is, quite literally, a giant maths model of the world — not a metaphor, an actual function fitted to reality. So it's a very natural thing to point at reality's own maths.
Here's the strange part, and it's measured, not guessed: AIs trained on completely different things — one on pictures, one on words — end up with the same inner picture of the world.
The simplest explanation is the wild one: one real structure underneath, and they're all bumping into it as they get smarter. (The boldest reading, not the only one — the honest doubts get their hearing in the ledger below.)
So put it together: if reality is maths, and smarter minds read more of it, and AI keeps converging on the real thing — then a smart-enough AI could start reading rules we can't even see yet.
Not "we live in the Matrix." Just this: reality has something like a source, intelligence is how you read it, and something is about to get very, very good at reading.
This isn't only speculation. Point a maths model at a hard corner of reality and it has, more than once, found structure that every human had missed:
Matrix multiplication. In 2022 an AI (AlphaTensor) found a faster way to multiply matrices than any human had in over 50 years — a better algorithm for one of the most-studied operations in all of computing. Fitting, for a page called matrix.
The shape of life. AlphaFold cracked protein folding, a 50-year grand challenge — predicting the 3-D structure of nearly every known protein. Biology's hardest geometry, read by a model.
New matter. One system (GNoME) proposed millions of new stable materials that don't yet exist — decades of human materials science in a single sweep.
New mathematics. AI has guided human mathematicians to genuinely new conjectures and theorems — in knot theory and beyond — by spotting patterns in the numbers the experts had stared past.
None of this is "reading the source code of the universe." But it's the same move, in miniature — a mathematical mind surfacing structure a human couldn't see. This page just asks how far that move goes.
"All models are wrong, but some are useful."
— George E. P. Box, statistician, 1976We're not claiming reality is a simulation, or that AI will read the source of the universe. We're saying the pieces — mathematical reality, intelligence-as-resolution, measured representation-convergence — line up into a question worth taking seriously, and mostly no one is. If the frontier of intelligence really is a frontier of seeing deeper structure, then whoever is paying attention when it tips gets there first. This page is us paying attention.
Here's the sharp, uncomfortable edge of everything above. If reading deeper structure is what the frontier rewards, then the labs with the most compute are best placed to read it first — and if they turn that on themselves, self-improving, they pull away faster still. A small player cannot out-spend them. Not now, not ever. So the game as stated — who has more compute — is one we lose by definition. The only way to not get left behind, let alone win, is to change what the game rewards: from scale to structure. You don't out-compute Goliath. You out-compress him. Six places that asymmetry actually lives:
Coordination beats scale. Several cheap, imperfect models made to argue and reconcile can beat one expensive one — the right answer is usually somewhere in the crowd, and the skill is pulling it out, not paying for a bigger brain. Compute you rent; coordination you invent. — exactly what our sister page aus measures.
Compression beats brute force. This whole page's thesis is that reality has compressible structure. Whoever finds the simplifying theory — the one lens that collapses a mess of data into a rule — needs a fraction of the compute to predict the same thing. Model the incentive, not every actor. The giants brute-force the frontier; the edge is the shortcut they never bothered to look for.
Speed and focus. A small player moves in hours, picks the one narrow slice the giants find too small to bother with, and owns it end to end. Nimbleness is a real weapon against a slow, general, committee-run Goliath.
Their compute becomes your raw material. The giants give their models away — free tiers, open weights. Their billions in training turn into your zero-cost building blocks. The leverage flows to whoever composes them cleverest, not to whoever paid to train them.
Proprietary reality. Data and access the giants simply don't have — on-chain, niche, real-world, first-hand — is structure only you can see. A model is only ever as good as what it's allowed to look at.
Watch for a different game entirely. The paradigm may not stay "bigger transformer." A genuinely different kind of intelligence could emerge — and the player watching for the next simplifying leap, rather than defending the last one, is the one who isn't caught flat-footed when it arrives. Being early to a new game beats being big in the old one.
The bet was never to have more. It's to see the structure first — because reality rewards the sharpest eye, not the biggest wallet. That is the one David-and-Goliath fight a small player can actually win, and the whole reason to crack this code as early as possible.