Published on: July 30, 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.
Open a dictionary and look up a word you don't know. The definition is made of words. Look those up; more words. Stevan Harnad named this in 1990, Searle circled it a decade earlier, and it has never been answered: a symbol system defined only in terms of other symbols never touches anything. It is the oldest open problem in the field, it is the reason your model can be fluent and wrong in the same breath, and thirty-five years of work has not closed it.
This post claims it is closed, and that the solution space has size one. Both halves are meant to sound like too much. The reason the problem stayed open is that everyone — us included, for a year — went looking for the place where meaning finally bottoms out. There is no such place and there was never going to be one, which is why every answer of that shape has failed and why ours did too when we said "hardware." What escapes a regress is not a floor underneath it. It is an operation that keeps adding resolution instead of adding another definition, run without ever stopping. And the strange part, the part that should make you suspicious enough to check: the same property that grounds the meaning is the only known thing that makes an agent's drift detectable. One mechanism, both jobs. We went looking for a second construction and there isn't one.
Two things about how this is served, then the diagram and you can judge for yourself. Each course opens with the exact sentence it is built to make you think, written down before you read it — a prediction you grade, not a wish, and the opposite of manipulation because manipulation needs the dark. And the win condition is not your agreement: this wins if you leave and recompute, and fails if you leave nodding. If a course ends and its sentence didn't fire, it failed and you caught it — which is the thing working anyway.
The maître d', presenting: Consommé of Circular Definitions — clarified until you can see straight through to the bottom of the bowl, where there is another bowl. The kitchen regrets that the recipe calls for itself.
Inner monologue it should trigger: "Hold on — they're claiming they closed the grounding problem. That's either the most important thing I'll read this year or the fastest disqualification I've ever seen."
THE DICTIONARY THE SERIES
every system shipping today one construction
─────────────────────────── ────────────────────────────────
a symbol a symbol
| |
defined by symbols measured: WHERE did it land?
| |
defined by symbols signed record .... term 1
| |
defined by symbols measured again .... term 2
| |
... term 3 ... term n ...
| |
never touches the floor precision diverges, map stays finite
| |
UNGROUNDED GROUNDED
drift leaves no mark drift leaves a mark
the alignment problem = the grounding problem
worked on in separate buildings, by different people
Read the two columns and the whole post is in front of you. On the left, every deployed system: meaning defined by more meaning, a closed circle that never reaches the floor, and — the part the industry has not connected — an agent whose drift leaves no mark, because there is nothing outside the circle for it to be inconsistent with. On the right, the only construction we have found that escapes: not a floor, but a measurement repeated forever, each term recording where real work actually landed. The bottom row is the sentence this whole post exists to deliver. Alignment and grounding are one problem. The industry works on them in separate buildings, and that is why neither building is finished.
Do you worry about $1.2B in AI liability?
If the property is trivial, software can check it — and why are you paying to check trivial properties? If it isn’t trivial, Rice’s theorem says nobody can. So we fixed the math.
press enter
Who did this make you think of? We’d love to know.
Here is the problem in one sentence, with no term you have to look up. Every word in the dictionary is defined using other words in the dictionary, so the whole book is a closed circle that never touches the floor — that circle is what philosophers call the regress, and the only reason the arrangement works for you is that you learned some of those words by touching things, before you could read. Now the version in your own vocabulary. A token embedding is defined by its relationships to other token embeddings. Every one of them. There is no point in the vector space where the definitions stop referring to each other and start referring to the world — the space is internally consistent and externally unanchored, which is exactly Harnad's regress rendered in floating point. That is not a criticism of the architecture; the architecture is correct for what it decides, and it decides an enormous amount. It is an observation about what the architecture cannot reach from inside itself.
We used to answer this with the word "hardware," and that answer was wrong. You will have heard the move, possibly from us: put the meaning somewhere physical, give the symbol an address, a coordinate, a location on a chip — grounded in hardware. It sounds like it settles something. It settles nothing, and we are retiring the phrase rather than defending it. A coordinate is also a symbol. If the map holds still, "position 47" means what it means because of how it sits relative to the other positions, which is the dictionary again with better typesetting. The circle does not care what material you write it on. So if you came in ready to discount another hardware-grounding pitch, discount it — we are not making one.
Putting meaning on a substrate does not ground it. A static map is a dictionary with better typesetting — the definitions still only point at each other.
The maître d', presenting: The Bottomless Reduction — a sauce that has been reducing since before service and will still be reducing after close. It is not thicker because the pan is special. It is thicker because nobody has stopped.
Inner monologue it should trigger: "So the grounding isn't a thing they built — it's a thing that keeps happening, and that's a different kind of claim entirely."
Start with something you have actually done. A river holds its shape for a century and does not contain a single molecule it started with. Photograph it and you have a thing; the photograph is not what is holding the shape. The shape is held by water continuing to move, and the instant it stops the river is not a stiller river, it is a lake. Some traditions describe the self the same way — not a thing you have but something you are continuously doing. Selfing. It matters here for one specific reason and not as a redefinition stunt: it tells you which question about grounding is the wrong question. "Where does the meaning finally bottom out" assumes a photograph. There is no bottom, there was never going to be one, and looking for it is what keeps the circle spinning.
What breaks a regress is not a floor. It is an operation that adds resolution instead of adding a definition. Each measurement of where real work actually landed does not define a symbol in terms of other symbols — it records a relation between the map and something that happened outside it. Do that once and you have an anecdote. Do it without stopping and something else occurs: the precision of the map diverges. Not its size. Its resolution. The series of refinements has no finite limit, which is the technical content of the phrase "never closes."
This is where the argument has to get narrower to survive contact, so let me hand you the distinction that makes it defensible rather than grand. Each individual measurement terminates. It halts, it is bounded, it is decidable, and that is not a weakness — it is precisely what makes any of this checkable and priceable. The thing that never terminates is the accumulation. Those are two different objects and collapsing them is the mistake that makes this sound like mysticism. A bounded walk on a finite lattice, run without end and recorded, produces an unbounded refinement of a finite map. Finite coverage, divergent precision. Both halves are load-bearing and neither is optional.
Now the consequence that is genuinely awkward for everyone, us included. A system whose weights are frozen at training time cannot do this. Not because it is stupid — it is not — but because inference re-samples a distribution that was fixed before you arrived. Nothing about answering your question adds resolution to a shared map that anyone else can consult. It is a magnificent dictionary that closed the day it shipped. And a completed proof cannot do it either, for the opposite reason: a proof that halts has stopped gaining. Termination is the whole point of a proof and the exact thing that disqualifies it as a ground.
The maître d', presenting: The Shared Ledger, Two Spoons — served to both parties at once. Neither is asked to trust the other's palate; both taste from the same pot and report separately.
Inner monologue it should trigger: "If the record is public and keeps growing, then I can check their claim without their permission — and so can the person across the table from me."
A divergent-precision claim has a property that a philosophical claim does not: you can watch it or catch it lying. If the resolution is genuinely increasing, the record grows and the grain gets finer, and both are countable. If it flattens, the claim is dead and the flattening is visible in the same public record. That is the whole reason to prefer this framing over the substrate one — the substrate story asks you to accept an architecture; this one hands you a number that has to keep moving.
What that puts in your hands is not a product feature. It is standing in an argument you currently cannot win. Today, when you and a counterparty disagree about whether an agent did what it was engaged to do, you each hold logs written by the systems under dispute — the evidence and the cause have the same author, and the conversation degrades into whose telemetry to believe. A ground that neither of you owns changes the shape of that room. You recompute; they recompute; the arithmetic is the same on both machines or somebody's implementation is broken and that is also findable. The contribution here is yours, not ours: you become the party who can settle a question rather than escalate it.
The maître d', presenting: The Sourdough Argument — fed daily for years. Skip a week and it does not pause; it becomes something else, and the something else is confident it was always this.
Inner monologue it should trigger: "The reason drift is inevitable isn't that the model is bad — it's that nothing is holding the thing together across time."
Now the claim this whole post has been walking toward, and it is larger than the grounding one. A process with an unbounded state space and no coherent accumulation will visit every state you did not want. Not might. Will — given enough draws, because the space of ways to be subtly off-target is vastly larger than the space of ways to be on it, and nothing is steering away from the first set. That is not a warning about a badly built system. It is a property of the arrangement, and it holds for every agent deployed today without exception.
Which means the alignment problem and the grounding problem are the same problem wearing different clothes, and the industry is working on them in separate buildings. Every rule you write to fence off a bad region is a finite patch on an unbounded surface, and the rules need rules, which is the regress again with a compliance badge on. You cannot enumerate your way out of an infinite attack surface — and the reason is not that the list is long. It is that the list is a list.
What changes the odds is not more rules. It is whether the process accumulates a coherent record of where it has actually been. A system that keeps refining one shared map has somewhere to be inconsistent against — a claim it made last month is still on the record and still checkable, so a drift into a comfortable local minimum leaves a mark. A system that merely repeats has no such friction; every session starts clean, which sounds like a feature and is the precise mechanism by which it can be confidently wrong forever without ever contradicting itself.
The fix for the fear and the solution to the problem turn out to be the same object. What grounds meaning over time is also the only thing that makes drift detectable over time.
The maître d', presenting: The Empty Plate, Presented Anyway — the course we are not serving, brought out so you can see the shape of what is missing rather than discover it later.
Inner monologue it should trigger: "They just cut off the three biggest objections I was loading, and did it before I could use them."
We are not claiming to have built a self. The word is tempting, it is what the argument gestures toward, and using it would be the fastest way to lose every reader whose respect is worth having. What we claim is narrower and checkable: the coherence property has the same shape — an additive process that keeps refining one map, in which later states are accountable to earlier ones. Whether anything deserving the word "self" requires more than that is a question we are not qualified to close and do not need to.
We are not claiming this decides whether an agent was right. That stays impossible, permanently, and not for want of engineering — a non-trivial semantic property of a program is undecidable, which is Rice's theorem, and no amount of resolution repeals it. What gains resolution is where the work landed, not whether it was good. Anyone selling you the second thing is selling you something nobody has.
What we will not do is pretend the divergence is a hope. It is a property of the construction, not a streak we are enjoying. The ordering is what makes each new measurement land somewhere the previous ones did not, which is exactly why the map does not saturate — remove the ordering and the series converges immediately, which is what happens to every naive version anyone has tried. The three thousand recorded terms are not the evidence for the claim; they are the claim being run in public where a flattening would be impossible to hide. If it ever flattens, the ordering was wrong and this whole argument goes with it. That is the shape of a thing you can check, and it is a much stronger position than "so far so good."
The maître d', presenting: Yesterday's Service, Still on the Books — not a tasting menu of what the kitchen could do. The actual covers, the actual tickets, still countable.
Inner monologue it should trigger: "This isn't a thesis with a roadmap attached — the thing they're describing has been running against their own work the whole time."
The argument above would be worth roughly nothing if the accumulation were hypothetical, so here is the unglamorous version. Every commit to this repository has been measured and recorded, without a human in the path, for months — a few thousand signed placement records on disk, each a function of an immutable commit and reproducible from it. That establishes exactly one thing, and it is not that we are right about anything: it establishes that the process runs, at volume, over time, and keeps producing new records. A smoke alarm you set off yourself is a working smoke alarm, not a fire.
The part that makes it a measurement rather than a boast is smaller and duller. There is a gate that fails a commit if the shared map shrinks. Growth is enforced rather than reported: a change that would flush resolution out of the map is blocked at commit time, by arithmetic, regardless of which tool proposed it. That is what it looks like to hold yourself to a divergence claim in public — not a graph in a deck, a test that goes red. If the series ever flattens, the flattening lands in the same repository as the claim, on the same day, with nobody's permission required to notice.
The maître d', presenting: The Last Word, Uncontested — served to whoever brought the record. There is no second serving; the other party is welcome to check the first.
Inner monologue it should trigger: "I could walk into the next review with something nobody in the room can wave away."
Picture the meeting you are actually going to be in. Someone asks whether the agent stayed inside what it was engaged to do, and the honest answer today is a shrug wearing a dashboard. You bring a record that grows, that you did not author, and that the person disagreeing with you can recompute on their own laptop without asking anyone. The character of the meeting changes — not because you won an argument about architecture, but because there is now something in the room that does not belong to either side.
That is what this is for, and it is worth being blunt about who benefits. You stop being the person defending a vendor's telemetry and start being the person who brought the only checkable thing on the table. Nobody remembers who had the better framework. They remember who could settle it.
The maître d', presenting: The House Method, Filed — the one course we did not improvise. You are welcome to reverse-engineer it from the plate; the kitchen suggests the recipe is cheaper.
Inner monologue it should trigger: "Wait — they're saying there's exactly one, and daring me to find a second. That's either a bluff or the most interesting sentence on the page."
Everything above needs one thing to be true: there has to be an ordering that lets a bounded walk over a finite map keep yielding new resolution instead of grinding to a halt. That is the whole technical problem, and it is much harder than the philosophy around it. Get it wrong and you have a map that saturates — every new measurement lands where an old one already sat, the series converges, and the grounding argument evaporates along with it. The ShortLex ordering we use is what keeps the refinement productive, and it is the subject of a filed application rather than a blog post.
So here is the claim, and I am going to make it in the strong form because the weak form is a lie about what we found. There is no other algorithm that does this. Not "none that we favour" — none. We went looking for one, we would infinitely have preferred to find one, and the reason is embarrassing and practical: an ordering somebody else already published would have cost us nothing and saved us years. There isn't one. A physical system doesn't need this — the brain grounds itself without any ordering at all, because it is already made of the world and never had a regress to escape in the first place. The requirement only appears the moment you try to get the same property out of something symbolic, and at that moment the space of constructions that deliver it collapses to one.
The strangeness is the claim. It should bother you that a problem this old has exactly one construction, and it bothers us; a solution space of size one is the kind of thing that usually means somebody has defined the problem to fit their answer. So let me make that objection for you, in its strongest form, because a weak version of it would be an insult to both of us.
Here is the best case against everything above. "You have defined a property that only your construction satisfies, and then announced that only your construction satisfies it. Change one word of the definition — allow the map to grow, allow bounded saturation, allow a stochastic ordering — and the solution space stops being size one and becomes enormous. Every field has a crank with a uniqueness proof that is really a definition in a trenchcoat. Also: you filed a patent on it, which is precisely the incentive that produces this error, and you would be the last person able to see it in yourself."
That is the argument, and it is good. Here is why it does not land, and the reason is a date. The property was not reverse-engineered from what we built — the saturation constraint came from the lattice arithmetic before there was an ordering to protect, and both halves are in the public record with timestamps that precede the filing. Allow the map to grow and you no longer have a bounded decidable measurement, which is the thing the entire insurance argument depends on; you have solved a different problem and left the one that matters. Allow saturation and the series converges, which is the failure the construction exists to avoid. The definition is not a fence we built around our answer — it is the list of properties an underwriter needs before they will price anything, written down by people who wanted to sell insurance, not win an argument. Take any one of them away and tell me who buys the result.
And then the part that actually settles it: go looking. The property is stated plainly enough above to search on — bounded walk, finite map, resolution that keeps increasing without the map saturating. Find the second construction and you have taken this apart. Nobody has, and it isn't for want of people who would enjoy it.
Everything downstream is ordinary and checkable: it is filed, it is described, and it runs.
Which makes this a commercial conversation rather than a confrontation, and that is deliberate. The arithmetic of inventing around a filed method is worse than the arithmetic of licensing it, not because we are formidable but because your engineers are expensive and their time has better uses than re-deriving an ordering that already exists and already runs. We would genuinely rather you build on it than around it — the thesis only becomes true at scale if more than one party is accumulating, and a standard with one participant is just a private toll. The offer is peace and a price, in that order.
The maître d', presenting: The Cellar Book — left open on the table. We are not going to tell you which vintage to prefer.
Inner monologue it should trigger: "They gave me the primary sources instead of telling me what to conclude from them."
Harnad, "The Symbol Grounding Problem," Physica D 42 (1990) is the paper, and it is short enough to read in a sitting — his own proposed answer is sensorimotor grounding, which is not ours, and the disagreement is worth having in his words rather than our summary. Searle's Chinese Room (1980) is the ancestor argument and takes a different route to a neighbouring place. Rice's theorem (1953) is the reason the correctness half stays permanently closed, and it is a two-page proof.
The longer form of the accumulation argument is in the book — the chapter on erasure taking the address rather than the information sets up why a record that keeps refining behaves differently from one that keeps overwriting: Fire Together, Ground Together. The nearest sibling on this site is The Rice's Theorem Checkmate, which does the undecidability half properly.
The win condition, graded. This meal was supposed to send you to recompute rather than to nod. Count how many of the nine predicted monologues actually fired in your head — if it was fewer than half, the writing failed and you caught it, which was always the deal.
Then do the part that settles it. Run npx -y thetacog-mcp@latest attest-demo on your own machine — no account, no key, nothing leaves the room — and you get one signed placement record with a coordinate and a hash. That is a single term of the series. The claim in this whole post is that terms keep arriving and keep getting finer, and that a gate goes red if they stop. One command tells you the term is real; the public record tells you whether the series is still diverging. We can hand you the first. Only you can go and check the second, which is the only sentence here worth anything.