Why the argument about how capable AI is turns out to be a reading we already have, not a gap we have yet to close.
Ask ten people when machine intelligence becomes general and you get a spread of dates. Ask a hundred, and the natural expectation is that the spread tightens — the way a hundred thermometers in one room converge on a temperature, each one a noisy read of a single fact underneath. That is not what happens. The spread does not collapse toward a date. It structures: the short-timeline and long-timeline camps grow more distinct, their reasons more specific, the gap between them wider and more load-bearing than it was at ten. Almost everyone treats this as a sign that the field is early — that we are missing the information that would resolve it. The wager of this essay is the opposite. The widening is not the absence of a measurement. It is the measurement.
To see why, you have to notice that “capability” is doing two incompatible jobs, and that the case for a smooth runaway depends on letting one quietly stand for the other.
Two things called capability
There is the kind of capability that climbs a standard supplied from outside. A model improves against a benchmark someone curated, a loss someone chose, a test whose verdict the world delivers without a vote: the code runs or it doesn’t, the protein folds the way it folds, the game’s rules settle the score. This motion is real, often fast, and entirely bounded by the quality of the external yardstick. Where the yardstick is cheap and unfakeable, progress here is genuine and it compounds. Nothing in what follows denies this, and the denial would be foolish — this is where the actual money and the actual transformation live.
Then there is the kind of capability the word general is reaching for: the single magnitude you are supposed to be able to climb without limit, the thing “superintelligence, however defined” gestures at. And the “however defined” is the tell. A quantity with no fixed referent cannot be maximized without limit, because there is no one thing to maximize. The first kind of capability is a property the situation settles — measure it from enough angles and the readings converge, because there is a fact underneath and the perspectives are noisy access to it. The second kind is not a property the situation settles. It is a property you settle, the moment you choose which competences count as “general” and in what proportion. It is a standpoint wearing the costume of a discovered fact.
A benchmark, in other words, can be either of two things, and from the inside they feel identical. Where it tracks something the world adjudicates, it is a ruler laid against a real edge. Where it tracks a magnitude we assembled by choosing what to weigh, it is a seat that has forgotten it is a seat — a vantage we picked, reported back to us as the view from nowhere. The Capabilities Index, plotted as one line, presents the second as if it were the first: a single number, climbing a trajectory, as though “capability” named a fact about the model rather than a fact about the model under a standard we chose to demand.
The test that tells them apart
You do not have to resolve the philosophical question — is capability really a scalar? — to act. There is a behavioral test that sorts the two without ever settling the metaphysics, and it is the one instrument the whole debate has been missing.
Watch what more measurement does. Add observers, vantages, harder probes. If the disagreement shrinks as you aggregate, it was noise — error scattered around a fixed value — and the quantity is the first kind: situation-settled, witnessable, the zone where AI genuinely runs away. If the disagreement persists and structures as you aggregate — if more eyes resolve the disagreement into sharper, stable, position-correlated camps rather than dissolving it toward a center — then it was never error around a value. It was different settings of something the question left open, and no amount of further measurement converges it, because the measurements were never closing on one value. Be precise about what this establishes and what it does not. It establishes that the verdict is seated — a partition we chose — not that there is no fact anywhere beneath it. A real, observer-independent quantity can sit under a classification that still refuses to glue, with the disagreement living entirely in the verdict rather than the referent; the same protein can have one true structure and four observers who classify its function four ways. Structured non-convergence witnesses the chosenness of the partition. It is silent on whether something real is being partitioned.
This is the depth cue from binocular vision, repurposed as an audit. Two eyes give you distance precisely because they disagree, and the size of the disagreement is the signal; average them and you destroy the very quantity two views were for. Convergence is the signature of a hidden number. Structured divergence is the signature that no single number is being approached — which is necessary for a seated partition but not sufficient for it, and the gap is where the test can be fooled. Three things produce structured non-convergence, and only one is a seat. A genuine value-choice (which competences count as “general”) is the seated case. But a real, high-dimensional referent sampled by different subspaces produces the same signature with no seat at all: when one camp means proof-generation and another means competition math under time pressure, both are tracking something real, and their estimates can fork further as each builds a sharper instrument — structure without a seat, a measurement tradition splitting, not a value being chosen. And a young field with correlated incentives mimics both before the evidence is in. Climate sensitivity ran the second pattern for decades: structured disagreement across model families that was deep uncertainty, not a constructed quantity, and it narrowed slowly rather than revealing a chosen partition.
So the bare converge-or-structure reading is only the first cut, and it needs a second axis to earn its conclusion: the magnitude of the structured spread, and its behavior under shared instruments. A spread that is structured but narrow — where every defensible weighting lands in the same place — is practically a fact even if the partition was chosen, and the optimist who notes this is right that chosenness alone settles nothing; the deflation bites only when the spread is wide. A spread that collapses when camps adopt a common instrument was information asymmetry or model misspecification around a real referent, not a seat. A spread that forks further under sharper shared instruments is either the seated value-choice or the underdetermined-multivariate case — and those two are separated only by asking whether the fork is a disagreement about what to value or about which real projection to track. The test is cheap to state and not cheap to run honestly: you are not arguing about definitions, but you are watching three things at once — does the spread close, how wide is it, and does it survive a shared ruler.
There is a failure in the other direction too, and it is the one institutions cause. A seated partition can be hammered into behaving like a settled fact by coordination — a definition enforced, dissent priced out — at which point the spread collapses and the test reads “situation-settled” over a choice that never went away. IQ is the worked example: a chosen weighting institutionalized until it behaved like a discovered scalar. So manufactured convergence fools the test exactly as manufactured consensus does, and the only defense is the same one the test needs everywhere — tracking which vantages dropped out of the count, and why.
Run it on the AI dispute and the question comes apart in your hands — but not where the first draft of this essay said it did, and the correction is the whole point of running the test honestly. “Can the model pass this coding benchmark, predict this structure, win this game” — add vantages and the estimates converge, because the world holds the standard and cannot be talked out of its verdict. That half is the first kind of capability, and the optimist who says it compounds is simply right. On the “general” half, the lazy version of this essay would report that timeline estimates have only hardened into camps. The actual record is more interesting and partly cuts the other way. The aggregate median has converged, hard: the Metaculus community moved from roughly fifty years out in 2020 to a fifty-percent date around 2031–2033 by early 2026, one of the fastest revaluations a major forecast has ever undergone. A median walking toward a date that fast is exactly the situation-settled signature, and an honest test cannot route around it.
What saves the seated reading is not denying that convergence but locating it. The median converged conditional on a frozen definition that does not itself glue: the figure rides on a four-condition standard including general robotics, and toggling a single condition moves it years — strip robotics and the mass slides forward two to three; add long-horizon agency or novel-insight generation and it slides back; adopt the architecture-skeptic’s prior and it goes to “never.” So what converged is the probability of a specific benchmark bundle by a date, and the bundle is the chosen partition. Across that partition the spread stayed wide and prior-correlated — an interquartile interval spanning more than a decade, defensible forecasters ranged from “by year-end” to “current architectures cannot get there at all” — and the median did not close so much as overshoot and partially reverse, the community and several lab-aligned forecasters pushing timelines back out together in 2025–2026. Lockstep reversal is the signature of correlated repricing on a shared shock, not of a variance collapsing onto a hidden number.
Notice what actually carried the verdict there, because it was not the convergence test. The binocular apparatus gets you to “the median converged but the partition didn’t” — to not noise, and not one number. The judgment that this is a seated value-choice rather than a real-but-underdetermined referent came from reading the content of the disagreement: the camps are not arguing about which real projection of a fixed capability to measure, they are arguing about how much agency, novelty, robustness, and breadth should count as general — what to weight, not what is the case. That is the ordinary value/fact distinction, inspected directly, and it is doing the load-bearing work the flagship instrument was presented as doing. The test sorts noise from structure; it does not, by itself, sort seated from underdetermined, and the case turns on the second sort. Granting the reading the content inspection supports, the persistent disagreement is not the field failing to do its homework. It is the readout that there is no single date to read off except relative to a partition the world does not supply — a real date exists once you fix the bundle, and the bundle is ours to fix. Whatever real capabilities sit underneath, “general” is a line we draw across them, and the line is the thing in dispute.
Why the seat hides, and why it is cheap to mistake
Two forces keep the second kind of capability disguised as the first.
The first is phenomenological: a manufactured center does not feel manufactured. The fused vantage you reach by “considering the whole picture” feels like where you were standing all along — like nowhere in particular, which is exactly what makes it feel like everywhere. A benchmark we chose, invoked often enough, stops registering as a choice. It becomes the floor everyone shares, the neutral input we disagree from, when it was a seated selection the whole time.
The second is structural, and sharper. The phrase “capability, as measured” carries a silent clause in invisible ink — as measured by the standard we chose to demand — and that clause is not idle, because demand does not merely measure capability, it grows it. The competences a field develops are cultivated toward the standards that get funded and scored, the way a country’s stock of a particular engineer is grown by the orders that make becoming one worthwhile. Capability is endogenous: not drawn from a fixed urn of latent talent waiting to be counted, but integrated over the demand that called it into being — and it decays when the demand leaves, which is what a capability rust belt is. So the number on the index is manufactured twice over: once by which ruler we picked, and once by demand cultivating the measured thing toward that ruler. When the benchmark saturates, the curve flattens and the spread compresses — and both can be read, falsely, as a wall in the capability itself, when what actually happened is the ruler ran out of room and the demand moved on. The saturation is in the instrument and the incentive, not necessarily the world.
Orbit-cost sold as a learning curve; fusion sold as orbit-cost
This is why the timeline gets priced wrong in a specific, diagnosable way.
Some quantities are trajectory-bound: a continuous parameter, moving monotonically, driven by something that compounds, so the curve has a slope you can extrapolate and a date you can attach with a real interval on the date. The cost per kilogram to reach orbit is like this — reusability and scale push it down a learning curve, and “more of what is already happening” is a forecastable thing to bet on. Other quantities are threshold-bound: a gate at an unknown distance, where progress along the approach does not reveal how far the gate is, so there is no slope to extrapolate. Fusion is the standing example, and “always twenty years away” is not a forecast that keeps failing. It is a constant — a placeholder emitted whenever the true distance is unknown, encoding “past the planning horizon” and nothing more. The repetition of the identical prediction is itself the tell that the quantity is threshold-type and therefore temporally unforecastable.
The convergence test is what distinguishes the two from the outside, before you know the mechanism: a trajectory-bound quantity’s forecasts converge as you measure more, because the slope is being revealed; a threshold-bound quantity’s do not, because the distance to the gate is not. So when “AGI in N years” is sold with the apparatus of a learning curve — the compute trend, the benchmark slope, straight lines on log paper — the test asks the one question that matters, and the answer when you actually pull the spread is split, which is itself the finding. The conditional forecast behaves trajectory-bound: fix the definition and the median has marched toward a date, fast. The unconditional quantity behaves threshold-bound: the definitions do not converge, the cross-partition spread has not narrowed, and the median’s recent reversal is regime-update-shaped, not slope-revealing. So the thing sold as orbit-shaped is orbit-shaped only inside a frozen bundle, and the gate that makes it fusion-shaped has moved to a place the learning curve never plots — not “how many years to the date” but “which bundle counts as arrival.” This is the essay’s most exposed claim and its kill condition rides with it — but the kill has to be stated carefully, because the obvious version is circular. “The spread narrows as the field standardizes its definitions” does not kill the deflation, because standardizing definitions is exactly the institutional coordination that manufactures convergence over a choice that never went away; that event is the false-negative, not the falsifier, and reading it as vindication for either side is unfalsifiable. The genuine kill condition requires independence: if many groups each author their own “general” bundle without coordinating onto a common one, and the dates those independent partitions imply cluster anyway, that is robustness-across-partitions — the real-referent signature, the world adjudicating where no enforcement could — and the deflation loses. Convergence by everyone adopting one bundle proves nothing; convergence of bundles nobody coordinated proves the seat was illusory. I commit to the threshold-at-the-definition reading because the conditional convergence and the cross-partition persistence have moved in opposite directions over the same window, and because the present spread shows low enforcement upkeep — the skeptics are loud and funded, not suppressed, so today’s disagreement is genuine rather than manufactured. I hold it killable on the day independently-authored definitions start landing on the same date.
The objection that wins half the field
The strongest response to all of this is not a rebuttal but a reclassification, and it deserves its full weight because it is what most serious, scientifically-minded researchers actually believe, and they believe it for a good reason. You have described a measurement problem and dressed it as a structural one. The benchmarks are crude and saturate; build better, broader, harder-to-fool ones and the disagreement converges. Adding measurements does converge on a stable answer in domain after domain — pretending otherwise is its own dishonesty. The non-convergence you are calling a depth signal is just bad rulers, and better rulers fix it.
This wins, cleanly, for an entire class of cases — and the convergence test is exactly what marks the boundary of the class. Better benchmarks converge precisely where there was a fact to converge on. They extend the first kind of capability into territory that used to look intractable: a question that was wide-open because no one had built the instrument becomes settled once the instrument exists, and that conversion is real, repeated, and where most of the value is. Structure prediction was this. The objection is right that the frontier of buildable, witnessable capability keeps moving, and right that yesterday’s “impossible” is often just last year’s missing ruler.
What a better ruler cannot do is converge a quantity that was seated rather than hidden. Measuring a chosen partition more precisely makes the measurement sharper without making the partition less chosen. So the objection wins its cases and loses the one the essay is about, and the line between them is the test itself: where more measurement converges, the optimist is right and you should build; where more measurement structures, there was no single fact to find, and a sharper instrument resolves the disagreement into finer relief rather than dissolving it. The thesis survives by narrowing — not “AI capability is a mirage” but “the general magnitude is the seated one, and the convergence test is how you tell, case by case, which half of the claim you are holding.”
The half neither side can witness
Honesty requires naming where this whole apparatus stops, because it stops at the same place the deepest version of the question lives.
Everything above is about verification — the cost of the witness, what converges when you measure, which disagreements are noise and which are choice. None of it touches generation: where the hypothesis comes from, who authors the question, the move of representing a hole in your own model and setting, unprompted, the standard for an answer you do not yet have. The obvious objection is that machines already do this — self-play invents its own training signal, agents author their own subgoals, curiosity-driven learners generate their own targets. But each of those authors a proposal that a fixed external check then grades: the game’s rules, the reward, the held-out task. Self-supplied means authoring the standard, not the proposal — and that, no instance has shown. The distinction collapses only if you let “generates its own data” stand in for “validates a standard it set itself,” and those are not the same move.
This is not a separate, deeper worry bolted onto the end; it is the candidate mechanism for the structuring the whole essay has been reading. If the gate the optimist and the deflationist are actually fighting over is a generation gate, then verification progress — better benchmarks, more compute, the entire trajectory-bound half that genuinely compounds — cannot by construction close it, which is exactly why the timeline disagreement would persist and structure no matter how much of the measurable kind of capability accrues. The non-convergence and the generation gate are one phenomenon: a threshold the verification curve does not approach, producing a spread that more verification cannot collapse. And the convergence test cannot adjudicate the gate itself, for the plain reason that you cannot run it on a faculty no instance has exhibited. The map this essay draws is a map of the verification half; the generation half is the cell that stays empty, and it is exactly the cell where “is the gate real” would be decided. Whether generation is a distinct faculty or merely verification run over a stream of proposals is the load-bearing uncertainty, and I can neither settle it nor pretend the test reaches it.
Who the test serves, and who it doesn’t
A diagnostic is a tool, and a tool has a beneficiary structure, so it has to be turned on itself before it is prescribed.
Reading capability as situation-settled — a discovered scalar climbing a trajectory — is not a neutral error. It converts the timeline from a declared bet into a physics question, and that conversion is load-bearing for the valuations, the recruitment, and the case for being trusted to write the rules, all of which depend on the breakthrough being imminent and the path being inevitable rather than chosen. The reading that the number is a seat has a weaker and more diffuse constituency — but not none: incumbents, skeptics, and regulators all have uses for a wall, and the deflationary move shares a structure with the thing it critiques, since from the outside you cannot always tell exposing a seat from occupying one. The asymmetry is a skew, not a one-sided payoff, and it does not make the deflation true. It says which claim to audit hardest — the louder, better-funded one — while conceding that the audit’s own framing is interested too. The incentive is an amplifier, not the explanation: it reinforces the misreading of a seated quantity as a discovered one, but it does not generate the seatedness, which would be there with no one selling anything. The structural claim has to stand on its own, or the incentive story is just the genetic fallacy wearing a critique’s coat.
The test itself sorts by power, and pretending otherwise would be the exact coordination-washing it is meant to expose. Running it requires longitudinal access to many vantages and — harder — the standing to call non-convergence a signal rather than your own incompetence. A well-resourced lab can defer the expensive witness, watch the spread over time, and afford to say “this is structuring, not converging.” An actor under competitive pressure cannot: the race rewards throughput, the cheaper test gets called validation, and “slow is smooth” has to be taught precisely because it does not survive contact with a peloton. So the convergence test is most available to those who already have the slack to wait for it, and naming that is not a footnote — it is the condition under which the prescription is honest at all.
There is a deeper limit, and it is the one the test cannot see past. The test reads convergence off the observers who were admitted to the question — and a reading that never gets a seat leaves no disagreement to measure. This is the flaw that hides inside a clean result: zero spread has two causes that look identical from the inside, a genuinely settled question and an eye that was struck before the count. A manufactured consensus and a real one present the same flat number, and which one you are holding is decided upstream, by who got admitted, where the test has no purchase. The jury that convicts unanimously may have agreed because the case was clear or because the readings that would have split it were excluded in selection — and the verdict cannot tell you which. So convergence is evidence only over the admitted vantages; on the question of who was admitted, a flat result is not the absence of disparity but the absence of its measurement, and reading it as agreement is the cyclopean error relocated one level up, to the seat that decides which seats exist. The discipline of reading the disagreement is, in the end, a fortification of a niche: the institutional pockets — adversarial review, dated and falsifiable predictions, replication — that keep disparity visible by force, against an environment in which the confident single number always travels faster than the declared, structured one.
What the reading actually says
The disagreement we keep trying to resolve is not the thing standing between us and the answer. On the part of the question that has a fact underneath, it already converged, quietly, and we should build there without waiting for permission. On the part that has a choice underneath, the median may keep moving but the partition will not settle however long we wait, and the persistence of that disagreement is the finding: there is a date only once you fix a bundle of competences the world does not fix for you, so what is owed is not a forecast but a declaration — this basket, this threshold — and a bet priced honestly around it. The convergence test does not say nothing is knowable, and it does not, on its own, say the quantity is seated; it says where the spread closes and where it does not, and hands the rest to a harder look at what the camps are actually arguing about. It hands the buildable half to anyone, and takes away only the unconditional date — the one nobody could supply without first drawing the line and calling it the world’s.
Open Questions
Empirically resolvable — partly run, and it split. The discriminating measurement is no longer purely hypothetical: pulled against the public AGI-timeline record (Metaculus, superforecaster panels, markets, 2020–2026), the aggregate median converged hard (≈50 years out to a ≈2031–2033 date) while remaining conditional on a definition worth two-to-three years per toggled condition and “never” under a different architecture prior, and the cross-partition spread did not narrow. That supports the threshold-at-the-definition reading without clinching it. The decisive observation must be stated with care, because the naive version is circular: a spread that narrows because the field standardizes onto one definition is manufactured convergence, not a falsifier — it is the false-negative the body warns the test cannot distinguish from the real thing. The non-circular kill condition is convergence across independently-authored definitions: groups that did not coordinate onto a shared bundle nonetheless implying the same date. Secondary discriminators sharpen it — does the converged definition pay predictive rent as its sub-benchmarks resolve, and is dissent flourishing or being priced out (active enforcement is the manufactured signature). Run it that way and it can kill the thesis; run it as “did definitions standardize” and it can only confirm it, which is the tell that the wrong version was being run.
Where even this fails — the unfalsifiable regime, named not hidden. Independence is checkable only until the coordinating paradigm shapes which definitions are imaginable. When the institution authors not just the agreed bundle but the space of proposable alternatives, the “independent” definitions are correlated upstream, their clustering proves nothing, and genuine and manufactured convergence become observationally identical. In that regime the seated reading is genuinely unfalsifiable — and you cannot reliably tell from inside whether you are in it. This is the struck-juror limit one level up: independence cannot be audited from within a paradigm that wrote the alternatives. The test holds where definition-authoring is independent enough to check and goes silent where coordination is total; claiming otherwise would be the manufactured-consensus move performed by the argument against it.
Conceptually underspecified — separating the seated from the merely underdetermined. The false-positive the body now concedes — a real, high-dimensional referent sampled by different subspaces produces structuring with no seat — is bounded but not closed. The proposed separator is whether the fork is a disagreement about what to value (seated) or about which real projection to track (underdetermined), but that line is itself contestable, since a choice of projection can smuggle a value and a value can be defended as a projection. Whether “what to weight as general” is cleanly a value-choice rather than a disguised empirical claim about which capabilities co-vary is the residual conceptual question, and it is not one the convergence test can settle from outside.
Structurally — the false-negative the test cannot see in real time. Coordination can collapse the spread on a seated partition (the IQ pattern), and manufactured convergence is byte-identical, from inside, to the real thing. The only defense named is tracking which vantages dropped out and why — which is precisely what cannot be done in real time, because the excluded reading leaves no cell to diff against. This is not resolvable by the test; it is a standing monitoring burden on whoever uses it, and its decision authority sits with whoever controls admission to the question.
Structurally irresolvable — is the gate a generation problem? Whether the disputed threshold (“general” capability) is crossed by more of the verification-bound progress the curve measures, or requires the self-authored standard that sits upstream of verification entirely, is not a question the test can reach, because no instance of the second faculty exists to run the test on. Decision authority here is not analytical; it belongs to whoever is staking the bet, and the most the reading offers is to keep the stake declared and the date unpriced rather than laundered into a forecast.
cafebedouin.org 