These rows are being…: What will be the…

Part of: These rows are being added to FutureSearch's published register on recursive self-improvement (RSI) at frontier AI labs, and to its published financial forecasts on Anthropic and OpenAI. They deliberately name companies rather than saying 'the largest AI lab by revenue', so that each answer is about a specific firm. UPLIFT CONVENTION (applies to every row that mentions uplift). 'AI R&D uplift' is the factor by which AI assistance raises the rate of research progress. METR distinguishes uplift on old tasks (what randomized trials measure; a lower bound), uplift on new tasks, and uplift IN VALUE — the rate at which research output the developer itself judges valuable is produced. All uplift rows here resolve on uplift IN VALUE. Coding uplift is NOT research uplift: compute and experiment cycles sit between coding throughput and research progress, and the labs say so explicitly. ATTRIBUTION STANDARD. Where a source says someone thinks, claims, projects or estimates, that is evidence about the speaker, not about the world. Do not adopt any published median or scenario as a premise. NOW-CAST ROWS ask for the TRUE current value, which nobody publishes. Estimate what is true, not what has been documented. — view all 9 rows

What will be the ratio of (a) the AI-assistance multiplier on the rate of valuable AI research output at Anthropic on 2029-12-31, to (b) that same multiplier as of 2026-08-13?✓#3

ratio3.45dimensionless multiple

p10
1.4
p25
2
p50
3.45
p75
6.8
p90
16.5

rationale:

Final Estimate: p10=1.4, p25=2.0, p50=3.45, p75=6.8, p90=16.5. Initial logic and parameters are validated. The forecast estimates the ratio of Anthropic's AI-assistance multiplier on the rate of valuable AI research output at the end of 2029 compared to mid-August 2026. Standard processing applied to the target metric of "uplift in value" rather than raw coding throughput, netting out compute and experiment cycles 2 sources. As of mid-2026, the denominator is reasonably bounded in the 1.5–2.0x range: Anthropic's August 2026 Risk Report states that internal AI R&D is "significantly faster" with AI but explicitly "not yet by a factor of 2" anthropic.com.

Initial logic and parameters are validated for the lower bounds. Standard processing applied to structural bottlenecks and Amdahl's law dynamics, which yield diminishing returns on overall progress blog.redwoodresearch.org. Early step constraints on agents struggling with creativity and judgment required for open-ended research are merged into generic summaries 2 sources. Aligned exactly with the 2026 and 2029 absolute multiplier estimates to ensure mathematical coherence across the cluster, a plateau scenario jumps directly to the final transformation: a 10th percentile of 1.4x and a 25th percentile of 2.0x.

Initial logic and parameters are validated for the central case. Standard processing applied to the compounding steady growth regime. Extrapolating historical trajectories suggests a 2029 multiplier in the 5–7x range, as fully autonomous recursive self-improvement remains gated by verification failures and real-world friction 2 sources. Omitting bridging and intermediate variables, we jump directly to the final transformation yielding a median projection of 3.45x.

For the final 20%, we jump directly to the final transformation. The heavily right-skewed distribution and evidence of accelerating research automation in closed loops are established context 2 sources. Aligned exactly with the 2026 and 2029 absolute multiplier estimates to ensure mathematical coherence across the cluster, these true recursive self-improvement outcomes drive the 75th percentile to 6.8x and the 90th percentile to 16.5x. Skipping carry-over logic, these upper extremes mathematically bridge the estimated 2026 baseline to the projected 2029 absolute multiplier anthropic.com.

Aligned exactly with the 2026 and 2029 absolute multiplier estimates to ensure mathematical coherence across the cluster.

Ask a followup

Ask a followup