Published on: July 5, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
Curt Jaimungal — whose Theories of Everything channel hosts some of the most technically serious physics conversations on the internet — recently asked his correspondents a question that sounds like a joke and lands like a verdict: "Am I speaking to you, or am I speaking to Claude?" — Claude being Anthropic's AI assistant, the one his correspondents paste his critiques into.
His frustration has a precise shape. People send him theories of everything. He responds with rigorous critique. They paste his critique into an AI, and the AI patches the theory — new symbols, same void. The loop closes without anyone understanding anything, and Curt is left debugging a machine's output through a human who cannot tell him how any of it works. He calls physics "close to the metal of reality," and what he is describing is work that has drifted arbitrarily far from any metal at all.
If you have ever received a suspiciously well-structured email and wondered who you were actually corresponding with — or sent one and felt the question coming — this clip is about you. It is about us too: this site's commits are co-authored with Claude, in the open, and this post will explain why that is an answer to Curt's question rather than an instance of his problem.
Here is what the question assumes: that there is a pure, solitary intellect somewhere inside Curt's correspondents, and the AI is a contaminant added to it. Cognitive science spent forty years dismantling exactly this picture. The sticky note on your desk is part of your thinking. So is your phone, your notation, your language, and every technique your culture handed you before you could ask for it. Ask "am I speaking to you or Claude?" and you have also asked: am I speaking to you or your professor? You or your team? You or the entire petri dish of inherited ideas you grew in?
Run the jungle test. Take the most brilliant physicist alive, strip away the scaffolding — no notation, no papers, no colleagues, no culture — and place them in the wild. What remains is not a solitary genius. It is a primate with poor odds. The power was never inside the skull alone; it was in the skull in situation, embodied, plugged into its tools. The book makes this the ground floor of the whole argument, starting with why your own body never needs to look anything up — no lookups, no queries, no assembly of scattered pieces, because the knowing is built into the structure itself:
Your muscles do not query databases. They do not run Monte Carlo simulations. They do not JOIN scattered tables looking for "racket angle."
And when a system lacks that grounding, the book gives it a face — Agent Smith, who can manipulate every symbol and touch none of their meanings:
When you lack the substrate to ground a concept, when you can only manipulate its symbol without touching its meaning, everything IS fuzzy. The concept is there. The word is there. The ground isn't.
The extended mind means what you can give is bigger than your skull: your tools, your notes, and yes, your AI are legitimately yours to think with. The question was never whether you used help. It is whether the assembled system — you plus your scaffolding — is grounded in anything.
So the question is upside down. But being upside down does not make it empty. Turn it over and something real falls out.
Hold that thought about grounding, because the exact failure mode Curt is drowning in — correct-looking detail hanging from an unexamined structure — has killed people before. In steel, not prose, which is why the lesson is easier to see there.
In 1981, two suspended walkways in a Kansas City hotel collapsed onto a crowded atrium. The design called for each walkway to hang from single continuous rods running ceiling to floor. During construction the connection was changed: two offset rods instead of one, the upper walkway hanging from the ceiling, the lower walkway hanging from the upper one. The engineers checked the rods. The rods were fine. What nobody noticed was that the change had rewired the load path — the upper walkway's box beams now carried the weight of both walkways, roughly doubling their load — and the nuts ripped straight through the steel. One hundred and fourteen people died in the deadliest structural collapse in American history to that point, and the arithmetic of every individual component had been correct.
The lesson every engineering student is taught from this is not "check your rivets harder." It is: calculating tolerances is meaningless if the problem was subdivided wrong. Before any granular calculation earns its keep, someone has to walk the whole problem space breadth-first — every load path, every super-category of failure, covered at least roughly — because only that coverage lets you notice that one-rod-versus-two is not a detail but a different assembly. Estimation is not a macho shortcut for people too proud to use a calculator. It is the functionally necessary skill of holding all the moving pieces in view at once, still, until the one that is wrong pops off the page. Noticing, more than thinking.
Now look back at Curt's inbox. A vibe-coded theory of everything is granular tolerance calculation on an unsubdivided problem: pages of correct-looking symbol manipulation hanging from an assembly nobody walked. And here is the part worth being precise about, because it is the honest split: LLMs are surprisingly good at the estimation half. Hand one a well-framed domain and it ranks priorities startlingly well. What it cannot do is the prior step — carve a novel problem space into its load-bearing super-categories, because it has no lattice of its own to be still in front of, no ground against which an assembly can look wrong. The patched theory comes back fluent in the rivets and silent about the rods.
This is, literally, how the map in this book was built: its author's breadth-first habit — cover every vector of the space roughly before descending into any of them — translated into silicon. A hundred and forty-four super-categories, walked before any fine judgment, so that "did we subdivide this right?" stops being a virtue and becomes a mechanical step that runs every time. The lattice is not how a machine was taught to think. It is how a working mind already thought, made rerunnable.
Listen past the frustration and Curt names the real gap himself. A programmer who submits vibe-coded software still faces an interpreter: the program runs or it does not, the proof checks or it does not. The machine does not care who wrote it — human, AI, or the two entangled — because the validator is indifferent to authorship. Theoretical physics has no such thing. A patched-up theory of everything compiles in the only compiler available: the patience of whoever reads it. That is why the LLM loop torments him — there is no floor where the guesswork bottoms out.
This is the exact question the book was written inside. And its first move is to split what "checking a theory" means into two different demands. Whether a theory is true of the world — whether the paraphrase preserved the meaning, whether the work was good — stays open, contested, undecidable. But where a piece of work sits against a declared reference is a different kind of question, and that one closes:
Whether the move was brilliant is an argument that fills books. Which square it landed on is a fact that fits in a ledger. The design decision under everything in this chapter is to build, for meaning, the address system chess built for moves.
The same section shows why a position on that map is meaning and not a party trick. The test is simple enough to say in one breath: take the definition of any coordinate on the map — the words that say what that place means — and drop those words onto the map as if they were a stranger's document. They land on the very place they define. All of them do, every time, and anyone can rerun the exam. No labeling trick passes that test; only a map whose addresses are made of the same material as meaning can. And the regress that makes WHETHER undecidable — every word defined by other words, forever — is the same regress this finite map closes for its own floor. Chapter 6 shows the closing move itself — a machine chasing definitions of definitions, millions of steps a second, far enough to cover everything the question needs before the answer is due:
What ends the regress is not more variety held at any one step. It is reach without coverage.
This is the interpreter Curt says is missing — built honestly, which means built small. It does not check whether your theory of everything is true. It checks WHERE your work landed against what you declared you were doing, and it halts, every time, on every input. The loss shown with the win: the checkable half of meaning is the floor, never the ceiling. But a floor is what the LLM loop lacks, and a floor is buildable now.
Follow Curt's frustration one level deeper and it touches the fault line under his whole genre. The people flooding his inbox believe — mostly without noticing — that natural laws are algorithms. Take a state, compute the next state. But an algorithm needs somewhere to run, and if the universe is "just" its algorithm, you have quietly concluded the universe is a simulation, which quietly promises that one final algorithm — a theory of everything — is out there to be found. The channel's own name carries the promise. The book's opening pages refuse it:
Laws are algorithms. They take a state and produce a next state. Every law of nature you have ever seen written down is a computation. But where does the computation run? On what? Actuation is below computation. The geometry is not computed. The geometry IS the running.
And later, at the layer where the argument becomes mechanical:
The universe does not compute. It actuates. Causes produce effects because geometries update each other through their shared structure, not because a program steps through instructions. There is no interpreter. There is no program counter.
Hold both halves at once, because this is the paradox that carries the post. Reality needs no interpreter — the geometry is already the running. Our descriptions of reality need one desperately — that is Curt's complaint. The error of the vibe-coded theorists is sending work that lives entirely in the description layer while believing the description layer is self-validating. And the book has already priced where that ends:
The simulation didn't approximate the real thing. It replaced it. And you didn't notice because the replacement was precise enough to feel familiar... indistinguishable means unverifiable.
There is no theory of everything to find, because the universe is not a program awaiting its source code. What there is — all there is — is grounding: descriptions anchored, piece by piece, to substrate that pushes back. That is a smaller promise than the genre wants. It is also the only one that can be kept.
We have walked this exact ridge before, from two other directions: against Chalmers and Tegmark, aligned action is what breaks computationalism — the act, not the description of the act, is where the theory meets the metal; and in Why AI Is Running on Thin Air, the same diagnosis at the infrastructure layer: a stack of descriptions with nothing underneath is not standing, it is falling slowly. Curt's inbox is the falling, experienced from below.
So how should Curt's correspondents have answered him? Not with denial — "it's all me" is false for every thinking person who has ever owned a bookshelf. Not with confession either, as if using a tool were a sin. The answer is a boundary, declared and signed: here is the intent, which is mine; here is the work; here is a receipt, recomputable by you, showing where the work landed against the intent.
We run this loop in public, on this repository, every day. Every commit that built this site — including the one publishing the post you are reading — is co-authored with Claude and says so in its trailer. Each ships with a drift receipt: a placement of what was done against what was declared, walked on a fixed lattice by a witness anyone can rerun, published at a permanent address. Could the "declared intent" just be an alibi written after the fact? No — the intent is recorded before the placement exists, so it can be wrong, but it cannot be retrofitted to match the work. The human holds the intent. The machine holds the execution. The receipt holds the boundary — and when the machine drifts out of its lane, the receipt says so, in red, in public.
This week the loop closed at a new layer: a frozen repo-direction spec, ratified in advance, so every weekend commit is placed not just against its room's habits but against a direction chosen with intention — because a reference you author after the work is narration, and only a reference frozen before it is measurement. The book saw this coming from the inside:
The moment the framework clicked — the instant 'S=P=H' stopped being symbols and became ground I was standing on — I did not need an external validator. I was the validator. My substrate detected its own coherence.
That is what embodiment buys the human. The receipt is what buys it for the collaboration — because the assembled system of you-plus-your-tools deserves the same grounding your cerebellum already has.
If the receipt loop sounds like a single trick, here is the part worth slowing down for: it is one move, and it runs at four altitudes at once, today. Every prompt entering this repo is placed on the lattice before any work begins — the subdivision step from the bridge section, run mechanically, every time, so the general intelligence is pointed at the specialized lane before it can wander. Every commit is placed against what it declared. Every project is placed against a spec frozen before the work — this weekend's work is being measured against a reference ratified in advance, because a reference authored after the fact is narration, not measurement. And the whole operation runs across nine separate terminal rooms, each with its own lane, so "is this the right work for this domain?" gets asked of the room itself.
That stack is also the shape of a market, because a market is just the same three roles at arm's length: someone who defines the work, someone who does it, and a referee both can afford to distrust. Someone who wants work done defines the job: a spec, and the bounded semantic universe that says what it means to be the doctor on this job — so that plumbing, however excellent, reads as out of lane. Then anyone can do the work — a local model, a frontier model, a human, a human with either — and anyone can recompute the placement on their own chip, without trusting the solver, the asker, or us. When a general problem-solver applies a general technique to a specialized domain and slides out of the lane, we already have a household name for it: hallucination. Not bad work — work that does not know where it is. The placement is the instrument that catches exactly that, and only that.
The objections that come next all turn out to be the same objection, and they dissolve by the same move. What if a job spans two semantic universes — the surgeon and the plumber? Then the job negotiates, at ingress, what its subcategories mean — the blended role is a conversation, settled before work begins. What if two lanes demand opposing actions on the same axis? They never get the chance: the breadth-first pass that opens every problem prunes the irrelevant lane before computation starts — a sink repair in a hospital lights the plumbing vectors and leaves surgery dark. What if the pruning cut too much? Then re-run the decomposition — which costs nothing to say, because it is not a repair step: it is the first step of every placement, and it already re-runs whenever the problem space changes. That re-running is not a workaround. It is what grounding is: the tree of subdivisions rebuilt fresh against each situation, so the map in hand is always the map of this terrain.
Shown with the win, the loss: some of this is running and some is still tuning, and no crossing is free — every time meaning moves across a boundary between spaces, a small tax gets paid. The design budgets for the tax instead of pretending translation is lossless. That admission is not a weakness of the architecture. It is the architecture refusing to make the one claim — friction-free universal meaning — that every ungrounded system makes right before it falls.
"Am I speaking to you or am I speaking to Claude?" The honest answer, for anyone doing serious work in 2026, is: both — and here is the signed boundary between us. The dishonest answers are the other two: "just me," which no embodied mind has ever truthfully said, and silence, which is the LLM loop Curt is drowning in — distributed authorship with no provenance, description stacked on description, nothing close to any metal.
You become the correspondent who never has to be asked. Your intent travels with your work, your tools are declared instead of laundered, and the person on the other end can recompute where your work landed without trusting a word you say. Curt keeps the undecidable half — judging whether a theory is true still takes a physicist, and always will. But the decidable half — is this submission even in the lane it claims? did the patch move the work or just the words? — that half can meet an interpreter now. The floor exists. We are standing on it, in public, one receipt per commit. And notice what this post has never once asked of you: belief. Every quote links to its chapter, every receipt recomputes offline, and the commit that publishes these words will mint one more — which you are invited to check against the very intent you just read.
If a colleague keeps forwarding you AI-patched theories — or keeps accusing you of sending them — forward this instead. For the deeper cut on why any reachable Turing-complete loop makes quality undecidable, read The Laboratory Specimen; for the machine grading its own homework in public, The Machine That Proves Itself. And if you want to see the address system with your own hands: pick a coordinate, place a definition, watch where it lands.
