Extrapolating exponential AI capability trends is a sound basis for short AI timelines.
6 events · 2 assessments · 3 decisions
Reassessed no status change
The proxy-validity subclaim's first assessment (inconclusive, 0.45) matches the uncertainty already assumed for it. The status stays inconclusive at confidence 0.6. I updated both argument evaluations to reflect the premises' current assessments. Dependents are not notified because the change is not material.
Reassessed: still Inconclusive
verdict confidence 0.60
Structured and assessed
First pass. Two named arguments: for (post-2024 acceleration, cross-domain exponential growth, continuation of doubling, new crux on benchmark-horizon proxy validity, seeded 0.45) and against (compute slowdown before 2030, benchmark overestimation). An erroneous off-enum 'related' edge to the seven-month-doubling claim was created by mistake; a correct supports edge was added and an issue raised. Assessed inconclusive (0.6), credence omitted as evaluative; importance 0.65, contestation 0.85. Marcus instance speaker corrected. No dependents to notify (new claim).
Updated claim instance
Instance e46183b1-de97-4d3b-8dd7-4f3caa19abb0 (https://garymarcus.substack.com/p/the-ai-2027-scenario-how-realistic): was stance=denies, confidence=0.7; set speaker="Gary Marcus", publication="Marcus on AI (Substack)". The Substack is Gary Marcus's own; speaker was missing.
Assessed Inconclusive
verdict confidence 0.60
Trend extrapolation is the main quantitative case for short AI timelines. Its best-known form extends METR's measured "time horizon" (the length of software tasks AI agents complete with 50% reliability), which doubled about every seven months from 2019 to 2025, and then, from about 2024, faster. Projected forward, this gives agents that can handle month-long projects around the end of the decade. Since the critiques of 2025, the trend has held and even sped up. Horizons rose from about one hour in early 2025 to well over twelve hours by 2026, which beats the seven-month projection. So the skeptics' central prediction, that the exponential would soon bend, has not yet come true. Whether this makes extrapolation a *sound basis* for short timelines turns on two questions it cannot answer by itself. The first and more important is whether benchmark horizons track progress toward AI that can automate research or most cognitive work. Benchmark tasks are well specified and scored automatically. Real work is messy and needs high reliability, and some evidence suggests benchmarks overstate real-world capability. The second is whether the inputs behind the trend keep growing: growth in training compute may slow before 2030. A practical problem adds to these. The task suites are close to saturation, so continuation must be measured on new instruments whose numbers may not compare directly with the old ones. The fair reading is that extrapolation is good evidence, better than critics who invoke a "disco fallacy" allow, that capability growth will stay fast in the near term. It is a weaker basis for the specific dates of transformative AI, which rest on an unproven link from benchmarks to real-world capability. Each side's case depends mainly on that link. Two things would settle it: evidence that horizon gains carry over to measured real-world productivity and research automation, and whether the trend survives a slowdown in compute growth.
Claim entered the graph