My median estimate sits just below 1.0 (0.96), reflecting an expectation of roughly constant or slightly accelerating efficiency doublings. Evaluated alongside related indicators, the projected 4.6x improvement in sample efficiency supports the view of consistently sustained, rather than radically accelerated, algorithmic progress. A major factor anchoring and slightly tightening the distribution near 1.0 is the likelihood of a methodological artifact: if the resolving series relies on a fitted constant-exponential trend line—as Epoch's current 7.6-month dashboard figure does—the ratio will mechanically be exactly 1.0. This supports a 25th–75th percentile range of 0.78 to 1.14. However, if the resolver reads doublings discretely from frontier data points or rolling windows, the natural noise of measuring adjacent short intervals (roughly 4–9 months each) combined with genuine structural shifts and broader uncertainty about AI R&D automation and scaling bottlenecks justifies wide 10th–90th percentile tails ranging from 0.55 to 1.45.

Drivers for acceleration (ratio < 1.0) The primary force that could drive the ratio below 1.0 is the successful integration of AI agents into AI R&D. Automation is rising: OpenAI reported logging 3.14 agent-workdays per human researcher workday in August 2026, alongside a 1.6x increase in experiments per experimenter. While OpenAI notes that compute growth confounds these figures and developers have not yet reported a 2x pace increase in research value, Epoch has observed faster capability trends for reasoning models Have AI capabilities accelerated? Epoch AI and publicly called its older ~3x/year efficiency figure "outdated" Epoch AI's Post. If automated AI research establishes a robust feedback loop by 2029–2030, algorithmic progress could compress significantly.

Drivers for a slowdown (ratio > 1.0) Conversely, measured algorithmic efficiency gains are highly dependent on compute scale. Research has shown that efficiency improvements are often "tied to compute scale" [2511.21622] On the Origin of Algorithmic Progress in AI and that much of the recent efficiency gain is directly attributable to scaling rather than pure algorithmic breakthroughs The least understood driver of AI progress. If frontier compute scaling encounters friction around 2029–2030 due to power constraints, chip manufacturing limits Can AI scaling continue through 2030?, or capital exhaustion, measured efficiency doublings would likely lengthen. Furthermore, the immense compute required for AI-driven research experiments may bottleneck the automated feedback loop The software intelligence explosion debate needs ..., neutralizing potential acceleration.

Measurement noise and uncertainty Measurement methodology introduces profound variance, which dominates the outer percentiles. Epoch's own analysts warn that estimates can vary drastically depending on benchmark and threshold choices https://epoch.ai/publications/revisiting-algorithmic-progress. Treating scale-dependent innovations and fragile statistical assumptions as purely algorithmic gains might yield "drastic overestimates" The least understood driver of AI progress. Because the resolution depends entirely on the specific published series used, the forecast must balance the possibility of a rigid, trend-fitted 1.0 against the sheer noise of discrete measurements of short, highly volatile intervals.