Steelman - the strongest arguments against this document#
Contents
Part VI lists the live uncertainties the document already concedes - parameters and scenarios. This page is different: these are the arguments that say the framework is wrong, not that a parameter in it is mis-set.
Each is stated in its strongest form, then answered honestly - including where the answer is weak.
Selection was not by loudness. Each of the three earns its place by passing three tests at once: it attacks a step the argument needs rather than a detail, it names a mechanism rather than a mood, and it makes at least one prediction that differs from this document's - so that evidence, not taste, can settle it. Most common objections fail the third test; "it's all hype" and "it's all inevitable" are both unfalsifiable in the forms usually offered, which is why neither appears here despite dominating the public argument.
1. "Verification cost stops being the ordering principle"#
The argument. The master asymmetry is a claim about today's training regime. It holds because RL needs a cheap reward signal. But reward modelling, process supervision, and learned critics are all attempts to manufacture verification where it was expensive - and they are working better each year. If a model can reliably judge the quality of a legal argument, a research direction, or a piece of management advice, then verification stops being scarce, and with it the entire ordering of Part III dissolves. The domains this document says lag do so only because of a temporary property of the current method.
Why it's strong. It attacks the central claim directly, it has a plausible mechanism, and it has empirical support: capability in "unverifiable" domains has improved considerably more than a strict reading of the asymmetry would predict. Long-form writing, strategic reasoning, and open-ended research assistance have all advanced without cheap ground truth.
The answer. A learned verifier is only as good as what trained it, so this recurses rather than resolves - it moves the ground-truth problem one level up rather than eliminating it, and where genuine ground truth never arrives, the loop can drift confidently. Self-verification without external grounding is exactly the regime where systematic error compounds invisibly.
Where the answer is weak. The same recursion objection applies to human expertise, which is also mostly peer-validated without external ground truth, and human expertise works well enough to run civilization. If learned verification is as good as peer review, that is sufficient for most economic purposes. This is the most serious objection in the document and it is not resolved. → Uncertainty 5 (promoted from this section in round 8); interacts with but is distinct from Uncertainty 1.
2. "The competitive-dissipation result is wrong because these markets aren't competitive"#
The argument. Game 3 assumes Bertrand-ish competition, which delivers the surplus to consumers. But most of the actual economy is oligopolistic, differentiated, and switching-cost-protected. In concentrated markets, a cost reduction available to all incumbents raises margins rather than cutting prices, because nobody has an incentive to start the price war. Measured industry concentration has been rising for two decades. So the prediction should be margin expansion at incumbents, and the "AI is a bubble because earnings aren't up" argument gets resolved in favor of the bulls, not by the mechanism this document gives.
Why it's strong. The empirical premise is correct - concentration is high and rising - and the model in Game 3 is the textbook case rather than the observed one. It also makes a cleanly opposite prediction, which is rare and valuable.
The answer. Both can be true at once, and the resolution is dispersion rather than the mean: consumer surplus in competitive segments, margin expansion in concentrated ones. That is why B3 specifies watching dispersion, not the average - the two hypotheses look identical in aggregate data and completely different in the cross-section.
Where the answer is weak. It is close to unfalsifiable as stated. If margins rise the document says "concentrated markets," and if they don't it says "Red Queen." The honest position: Game 3 is a claim about competitive markets specifically, and its scope should be narrowed to say so rather than presented as a general result. That is a correction, not a defense.
Scope note (round 27): Game 3 now states the scoring cross-section explicitly - adoption / margins / prices read together, with dispersion as the test - so objection 2 is no longer free to claim unfalsifiability against a single mean. The remaining weakness is measurement: "competitive sector" is itself a judgment call at the boundary.
3. "Institutional friction is a cope"#
The argument. Every one of this document's slow timelines rests on institutions being slow. But that inference comes from technologies that required physical retooling, capital replacement, or coordinated standard-setting. AI requires none of those. It is distributed instantly, self-serves, needs no procurement to try, and diffuses through individuals rather than organizations - a worker adopts it whether or not the employer has a policy. The diffusion analogy is simply the wrong reference class, and every date in Part IV is 3–5 years late.
Why it's strong. Consumer adoption of AI tools has been the fastest of any technology ever measured, by a wide margin. And the mechanism - bottom-up individual adoption rather than top-down organizational deployment - genuinely has no precedent in the reference class.
The answer. Individual adoption is not the same as organizational benefit realization, and the second is what productivity statistics measure. The METR result is the sharpest evidence: enthusiastic individual adoption coexisting with negative measured productivity. Adoption speed tells you about the tool's accessibility; it tells you nothing about the redesign of the workflows around it.
Where the answer is weak. METR is one study, on one population, at one capability level, and it is being asked to carry an enormous amount of weight throughout this document. If it fails to replicate at higher capability, the institutional-friction argument loses its best empirical support and this objection substantially wins. That replication is the single most valuable piece of research anyone could run against this document.
How each objection resolves#
None of these needs to remain a matter of judgment for long; each is wired to a different part of the dashboard, and the wiring was chosen so that both sides should accept the verdict in advance.
Objection 1 resolves through B8, the capability gap between verifiable and unverifiable domains. If that gap closes and stays closed under blinded evaluation - not benchmark scores, which the objection itself predicts will be gamed - then the asymmetry is dying and the document's domain ordering with it. If the gap persists while learned-verifier methods proliferate, the recursion answer is holding.
Objection 2 resolves through B3's dispersion reading. Margin expansion concentrated in high-concentration industries while competitive segments pass gains through is the split verdict both hypotheses should accept. Uniform margin expansion everywhere is objection 2 winning outright, and the scope-narrowing correction below becomes a euphemism for a miss.
Objection 3 resolves through replication. The METR result re-run at higher capability, on production work rather than study conditions, is the single experiment that most moves this document's confidence in either direction - and it is cheap relative to what rides on it. Until it exists, B2-style instrumented productivity measures are the proxy.
The pattern in all three answers#
Each objection says the same thing in a different register: this document's method - rate limits, reference classes, payoff structures - systematically predicts continuity, and the case for discontinuity cannot be made from inside it.
That is a fair criticism and it does not have a good answer. The mitigation is to state it plainly (Part VIII), give the discontinuity cases their own section (Part VI), and attach falsifiable triggers to the continuity claims (Part VII) so that being wrong is detectable early rather than in retrospect.
If this document is badly wrong, the most likely single reason is objection 1, and the most likely form is that it looks approximately right until roughly 2029 and then ages very poorly.
Objection 1's middle case is now the default watch#
Uncertainty 5 no longer treats dissolve-vs-hold as the only fork: process verification can improve while outcome verification stays expensive. That is a Part III re-score event, not full framework death. Objection 1 still wins if both tiers close; partial wins should not be scored as full spine failure.
Related: Part VI - Uncertainties · Base rates · Scoring · Game 3