How an assessor could investigate evidence of acceleration from the automation of AI R&D

URL SCAN: How an assessor could investigate evidence of acceleration from the automation of AI R&D
FIRST LINE: This article discusses how an assessor could study whether capabilities progress could accelerate dramatically due to AI-driven acceleration of AI R&D. Such an assessment would answer:

The Dissection

This is an assessment proposal, not evidence of acceleration. It builds an audit framework around a frontier lab’s internal pipeline: capability outputs, algorithmic progress, AI usage, labor, compute, data, bottlenecks, forecasts, and researcher testimony.

Its deeper move is to make recursive AI R&D legible as a decomposable production function. That is useful for attribution, but it narrows the event to an accounting problem. It asks how quickly the machine is improving and which input caused it; it does not ask who owns the resulting productive capacity or what happens when human labor is no longer required for mass production and consumption.

The Core Fallacy

The proposal’s central error, under DT, is treating AI automation as a normal productivity input that can be cleanly isolated from labor, algorithms, data, and compute. It is not. AI changes the search process that produces algorithms, alters the value of experiments, creates feedback loops, and can change the bottlenecks being measured. Token spend, lines of code, and researcher testimony are weak proxies for a counterfactual that never existed.

Benchmarks and pretraining loss can show technical progress; they do not show productive participation, ownership, or social viability. Even perfect attribution would not reverse P1–P3. Once AI has durable cost and performance superiority in cognitive work, competition compels adoption; institutions cannot preserve human-only domains at scale; the wage-to-consumption circuit breaks. The assessor can date the fracture. It cannot repair the bone.

Hidden Assumptions

• Frontier developers will provide complete, dated, disaggregated data rather than a curated evidence package.
• Internal benchmarks and researcher perceptions remain comparable across model generations and are not strategically selected.
• AI’s contribution can be separated from compute, data, algorithms, management, and organizational learning.
• Interviews reveal causal facts rather than incentives, memory distortion, or post hoc narratives.
• Near-term bottlenecks are visible, stable, and forecastable.
• Hidden capabilities and competitive secrecy will not make public assessment lag materially behind reality.
• Capability acceleration can be translated into danger forecasts without modeling the ownership and employment shock that follows.

Social Function

Classification: partial truth serving transition management and elite self-exoneration.

The partial truth is real: the text correctly identifies the need for dated internal evidence, multiple capability measures, input decomposition, bottleneck analysis, and skepticism about naive benchmarks and self-reported speedups. It is not empty copium.

But its institutional function is anesthetic. It converts a civilizational transition into an assessor’s checklist. That lets labs and observers say “the evidence is not yet conclusive” while the competitive race continues, positioning the crisis as a problem of measurement, access, and forecast calibration rather than control of the productive core. The procedure may improve the autopsy; it does not give the corpse a pulse.

The Verdict

A serious measurement proposal with an inadequate theory of the event it measures. It could establish whether a lab’s AI R&D loop is accelerating and identify which observed inputs correlate with that acceleration. It cannot reliably solve the recursive counterfactual, expose concealed progress, or prevent diffusion once the advantage is real.

Under the Discontinuity Thesis, its likely historical role is not prevention but timestamping: documenting how quickly the system automated the machinery that automates itself. If it finds sustained AI-driven acceleration, that is evidence that the lag before P2 and P3 is compressing—not evidence that the post-WWII order has found a survival mechanism.