Curriculum Exhaustion and the Entry Rung: The One-Supply Argument Revisited

The questions#

The same two #oq/now items that The Data Wall and the Validation Commons Are One Supply Constraint answered partially on 2026-08-17, reopened because roughly two hundred sources have been compiled since:

1. Effective Compute Scaling: Can data generation (synthetic, simulated, interactive) actually keep pace with model-size growth, or does the data wall bind first?
2. Post-Scarcity Macroeconomics: Is there any domain where the validation commons has been observed to go before the Stockfish threshold arrives?

The earlier page's thesis was that both supplies are verified judgment. Verified judgment can be manufactured where a sound, cheap, fast verifier exists, and only humans supply it anywhere else. Its stated gap was evidential: the corpus held no empirical frontier-scale datum on data supply, and no measured commons-scale depletion. Neither gap has closed. What follows is what moved.

Q1: the data wall has a name now, and it is not tokens#

Nothing new measures it. No source compiled since 2026-08-17 reports a synthetic-versus-human data share against effective compute, embedding dissimilarity at frontier scale, or pass@K for a self-proposed curriculum (a corpus-wide grep for each turns up only the 08-17 page's own list of what is missing). The three settling instruments are exactly as unmet as they were.

What did arrive is a frontier-insider statement of the conversion. Noam Brown (OpenAI, practitioner-opinion, Recursive Self-Improvement) names curriculum exhaustion as the plausible bend in the curve: self-play gives AlphaZero "an infinite curriculum", but RL on problems does not, and "if we run out of problems to challenge it, then that is a plausible scenario where it becomes much harder to make progress." Lambert reaches the same friction independently a few days later. He expects it "could become exponentially harder to conceive, build, and test new environments that meaningfully challenge the leading LLMs" (Intelligence Explosion Dynamics). Neither speaker mentions pretraining tokens. Both locate the scarce input in problems hard enough to learn from. That is the 08-17 page's conversion, the token wall displaced into a different supply before it arrives, now stated from inside a lab. Brown hedges it ("we haven't really hit that as a wall yet"), so it is a named mechanism and still a prediction.

The verifier is necessary but not sufficient, and this is the refinement. The September RSI survey (§3.4, practitioner-opinion, secondary on its primaries; see Rationale Bootstrapping (STaR)) splits the self-training filter into validity, difficulty and learning value, and observes that "neither correctness nor estimated difficulty alone establishes marginal learning value." The 08-17 page's rationing terms were verifier existence, verifier speed and generator diversity. Those terms cover validity only. A sound verifier manufactures correct experience. It does not manufacture appropriately hard experience (difficulty proxies such as R-Zero's inconsistency reward also reward ambiguity, and PSV's easy/medium/hard targets overlap heavily in realized difficulty). Nothing in the corpus measures learning value at all. Curriculum exhaustion is the difficulty term running dry. So the binding supply is narrower than "verified judgment": it is verified judgment at the frontier of the learner's competence, and that supply shrinks as the learner improves.

Diversity collapse is localized, not dissolved. Diversity Calibration Under SFT (empirical, Skobelev, Fithian & Han) shows plain SFT on target-distribution data has no built-in bias toward mode collapse. Survey and CodeNet fine-tunes converge to human-level diversity from both directions, and a sqrt-KL bound forces the gap to zero as the fit improves. The paper is explicit that this says nothing about preference optimization, and it cites KL-regularized RL as designed to mode-collapse. That sharpens the 08-17 page's third rationing term: the diversity collapse Multiagent Finetuning diagnosed is a property of the reward-filtered loop's objective, not an inevitable cost of training a model on model-shaped data. It is therefore an engineering term, not a hard wall. The evidence is small-model, non-RL, and does not test self-generated targets, so this reads as a direction, not a result.

Settled vs. open. The wiki can now say: the token data wall does not bind first. The supply that might bind is curriculum, meaning valid problems at the learner's frontier. It is rationed by verifier existence (validity), by the absence of a difficulty signal that is not also an ambiguity signal, and by an unmeasured learning-value term. Every statement of this is prediction or practitioner-opinion. Whether curriculum supply actually keeps pace at frontier scale needs a source the corpus does not have: a lab technical report giving synthetic/self-generated data share and problem-pool size across generations, or an R-Zero/Absolute Zero-style run with the survey's learner-state-frozen control and pass@K reported. That is external evidence, not further synthesis.

Q2: the front half of the ordering is now measured; the back half is not#

The 08-17 verdict was a qualified negative. No domain showed commons-scale depletion, colonoscopy deskilling was atrophy of existing stock, and the structural point was that the threshold arrives first where validators were never load-bearing. Tao's report qualified it on 2026-10-16. Three groups of evidence move it further.

1. The regeneration channel is measurably narrowing before the threshold, in software. The Tragedy of the Cognitive Commons now holds Canaries in the Coal Mine firsthand (ADP payroll through June 2026, empirical). Employment of 22–25-year-olds in the two most-exposed quintiles fell ~11% while the least-exposed grew ~10%. The decline runs "primarily through reduced hiring rather than increased separations", and aggregate employment rose +6%. The page calls this the commons argument's signature: "healthy aggregate, failing regeneration." The most-exposed quintile includes systems software developers and computer/IS managers. AI-Exposed College Majors at Labor-Market Entry adds the entry side from Census PSEO×LEHD records: the most-exposed decile of majors, dominated by computer science, loses 5pp of initial employment and 13% of initial earnings, still −2.0pp at two years. The Dallas Fed's Texas records find the same break. Seniority-Biased AI Adoption: The Junior Share at Adopting Firms finds the junior share 1.9pp lower at adopting firms across 41 countries, though mostly through senior growth rather than junior cuts. Software is the domain Musk's Stockfish claim puts at "90%→99%", which means not yet at the threshold. So the input side of the feared ordering has been observed in the domain the thesis names: the rung that regenerates validators is being removed before the domain stops needing them. The counterweights are real and recorded on the commons page. The firm-shock-adjusted estimate weakened to insignificance across vintages, the ACS magnitude is a fraction of ADP's, college share halves the coefficient, and Fairlie & Wu find no CPS break for all young graduates.

2. Formal mathematics shows the training supply and the human training supply are one stock. Proof Abundance carries Tao's report (practitioner-opinion, no measurement): OpenAI's ~300-proof release "strip-mines" the informal reserve of medium-difficulty problems the field keeps for students, "junior research projects … got wiped out overnight", problem lists are now withheld, and follow-up questions are not being replenished. Set beside Q1, this is the finding the 08-17 page did not have. The reserve that supplies the model's curriculum and the reserve that supplies the student's curriculum are the same problems, and solving them at scale consumes both. The model needs problems at its frontier and exhausts them by solving them. The student needed the medium ones, and those are gone because the model solved them on its way past. This is the ordering in its sharpest form. The threshold has arrived for the medium rung, and the residual human job it leaves (reading the statement, per Statement Drift) is large: ShadowBench's best agent compiles 61.8% of hard problems but only 11.2% are semantically aligned. The capacity to do that job is regenerated through exactly the problem reserve being consumed. The 08-17 structural correction, that arrival order and need order are correlated, holds for the validation sub-task. It fails for the curriculum that produces validators, which is a different and unverifiable-by-machine input.

3. The depletion is selected for, not incidental. The CIVIC-AI whitepaper (practitioner-opinion, on the commons page) argues that the safest delegation rule — delegate the inspectable, reversible, bounded — "selects on the same variable" the developmental pipeline selects on. The randomized learning studies on Experimental Learning Impact of Generative AI show the depletion is design-dependent. Metacognitive feedback cut answer offloading (OR 0.47) and restored unaided performance to the no-AI level, and tutor-verified materials reversed the sign of the learning gap. So the mechanism has a known countermeasure at the individual level, which is not the same as a profession-level one.

Settled vs. open. The wiki can now say that the ordering's front half has been observed. In software, the entry rung that regenerates validators is contracting through hiring, before the Stockfish threshold, on administrative payroll and Census records, with disputed magnitude. In formal mathematics, a leading practitioner reports the problem reserve that trains the next generation being consumed by the same systems whose output that generation must check. The back half is still measured nowhere: a cohort that entered after 2023 showing lower unaided validation competence than a matched earlier cohort, in software review or in formalization checking. The Tragedy of the Cognitive Commons carries the same gap as its own #oq/source question (Mechanism 2, never measured). Payroll data cannot close it, and the time delay means it can only now begin to show.

The three supplies share one stock#

The 08-17 page merged two supplies. This revision adds a third and says what they have in common:

The first two draw on the same reserve of hard-but-tractable problems. The third is what the second produces, on a lag. So the three are not parallel constraints. They are one stock read at three points in time, and the measured evidence sits only on the middle one.

What would close each#

• Q1 → external source. A frontier lab's technical report giving self-generated/synthetic data share and RL problem-pool size against effective compute across two or more generations. Alternatively, a published self-proposed-curriculum run reporting pass@K, with the RSI survey's learner-state-frozen control. Either would grade Brown's curriculum-exhaustion friction against the record.
• Q2 → external source. Lovett's no-AI competence design, run on the residual task: formalization checkers in Lean, or code reviewers, stratified by entry cohort (pre- vs post-2023) at matched tenure. For mathematics specifically, the bibliometric trigger on Proof Abundance: student first papers in problem classes the October release covered.

Sources#

• The Data Wall and the Validation Commons Are One Supply Constraint: the 2026-08-17 synthesis this revises, and its 2026-10-16 Tao qualification
• Recursive Self-Improvement: Brown's curriculum-exhaustion friction, quoted (practitioner-opinion, Noam Brown – Agent swarms, alignment, & recursive self-improvement)
• Intelligence Explosion Dynamics: Lambert's independent restatement (practitioner-opinion, Why I still haven't bought into true RSI)
• Rationale Bootstrapping (STaR): the survey's validity / difficulty / learning-value split, experience corruption, and the controls nobody runs (practitioner-opinion, secondary, The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement §3.4)
• Diversity Calibration Under SFT: SFT has no built-in collapse direction; the objective-induced gap is separate (empirical, Fine-Tuning Fixes Mode Collapse and Over-Dispersion in LLMs)
• The Tragedy of the Cognitive Commons: Canaries read firsthand (hiring not separations; counterweights), the CIVIC-AI selection argument, the unmeasured Mechanism 2 (Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, empirical; When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration, practitioner-opinion)
• AI-Exposed College Majors at Labor-Market Entry: PSEO×LEHD, Dallas Fed and Fairlie & Wu entry outcomes (empirical)
• Seniority-Biased AI Adoption: The Junior Share at Adopting Firms: junior share at adopting firms, 41 countries, mostly from senior growth (empirical)
• Proof Abundance: Tao's strip-mining report (practitioner-opinion, No More Heroes - Terence Tao Talks to Numberphile About AI)
• Statement Drift: ShadowBench 61.8% compile vs 11.2% aligned; the residual human job is reading the statement (empirical, SHADOWBENCH: Toward Reliable Automatic Evaluation of Semantic Alignment in Autoformalization)
• Experimental Learning Impact of Generative AI: the metacognitive-feedback and verified-materials RCTs (empirical)
• Post-Scarcity Macroeconomics: the Stockfish threshold and the software 90%→99% claim (prediction, Musk)