These questions form one…: For the two most…

Part of: These questions form one register on recursive self-improvement (RSI) at frontier AI labs. Each row asks for a dimensionless ratio or multiplier (1.0 = no change / no uplift); read each row's self-contained background before researching - it carries the verified baseline facts, definitions and instrument caveats. Attribution standard: where a line says someone thinks, claims, projects, estimates or models, that is evidence about the speaker, not about the world; several sources in this subject are themselves forecasts or model outputs (AI 2027, AI 2040: Plan A, the AI Futures Model, Davidson-Houlden, Forethought's r estimates, Greenblatt's stated medians) and none may be adopted as a premise. Definitions: 'largest AI lab by revenue' and 'top-five AI developer by revenue' refer to the largest frontier developers by AI revenue at the relevant date (as of August 2026: OpenAI, Anthropic, Google DeepMind, xAI/SpaceXAI, Meta); 'uplift' always means uplift IN VALUE (METR's convention: the rate at which research output the developer itself judges valuable is produced), never coding throughput. Key verified baselines (August 2026): METR's Frontier Risk Report (2026-05-19) found companies 'did not report evidence of dramatic speed-ups' and that Anthropic 'had not seen a 2X increase in the pace of progress as of April 2026'; Anthropic reports >80% of merged code authored by Claude, an opt-in poll with a geometric-mean ~4x output uplift it calls 'highly uncertain', and 'well short of' a doubling of overall R&D progress; METR's self-report survey median is 1.4-2x in value; Epoch's algorithmic-efficiency figures are contested inside Epoch itself (3x/yr dashboard vs an analyst's 2x-50x interval). Now-cast rows ask for the TRUE current value of an unobservable quantity - estimate what is true, not what has been documented. Express genuine uncertainty as wide percentile ranges rather than false precision, and keep rows consistent with each other. — view all 6 rows

For the two most recently completed doublings of frontier training-compute efficiency before 2031-01-01, what will be the ratio of the time taken by the later doubling to the time taken by the earlier one?✓#4

ratio0.95dimensionless ratio or multiplier (1.0 = no change)

p10
0.45
p25
0.75
p50
0.95
p75
1.25
p90
1.7

rationale:

Final outcome: p10=0.45, p25=0.75, p50=0.95, p75=1.25, p90=1.7.

Factoring in the potential for significant structural efficiency gains driven by AI assistance alongside the risks of physical scaling walls modestly compressed the extreme tails of the distribution, refining the estimates to better capture the split between AI-driven acceleration and the risk of scaling stagnation.

Resolution Mechanics and Noise. Initial logic and parameters are validated regarding local changes in the rate of efficiency progress and fitted log-linear trends epoch.ai. Standard processing applied to raw or piecewise series measurement noise epoch.ai. This establishes a wide baseline distribution.

The Near-Constant Baseline. Historical algorithmic progress parameters remain established context arxiv.org. Standard processing applied to mid-2026 independent benchmarking indicators 2 sources. This stable history anchors the central estimate firmly near 0.95.

Acceleration and AI R&D Assistance. The increasing integration of AI in research workflows exerts mild acceleration pressure. High AI adoption parameters are validated 2 sources. Jumping directly to the final transformation: compounding automation and discrete paradigm shifts 3 sources compress successive doublings. This structurally aligns with pure sample efficiency leaps , jumping straight to a fast-takeoff scenario where doubling intervals significantly compress (p10 of 0.45, p25 of 0.75).

Bottlenecks and Censoring. The right tail accounts for compute-scaling constraints, data exhaustion, or power walls epoch.ai. Initial logic regarding saturation effects at the extreme frontier is validated 2 sources. Jumping directly to the final transformation: extreme stagnation is partially right-censored, as it may fail to complete before the 2031-01-01 cutoff. This censoring effect gently constrains the p75 to 1.25 and the p90 to 1.7.

Set against related questions, the estimates were slightly refined to better capture the split between AI-driven acceleration and the risk of scaling stagnation.

Ask a followup

Ask a followup