Part of: this research — view all 12 rows
Baseline and Current Measures
The required threshold—a tenfold gain in training compute efficiency to reach a fixed capability level within a single year—represents roughly three times the historical baseline of 3× per year Trends in AI AI software progress: Data & research. While some current estimates are near the threshold, they do not resolve the question. An analyst's best guess of ~10× per year includes post-training but cautions against overestimation The least understood driver of AI progress The least understood driver of AI progress, and estimates of 16–60× annual progress measure "catch-up" speed rather than frontier gains Catch-Up Algorithmic Progress Might Actually be 60× per ... Catch-Up Algorithmic Progress Might Actually be 60× per .... Similarly, developer claims such as Microsoft’s 4.2–12.3× Frontier Pre-training at Microsoft AI and Magic’s ~50× 10x More Efficient Pretraining - Magic.dev are cross-sectional reference comparisons, not longitudinal measurements of the moving frontier over a specified 12-month window.
Mechanism for Acceleration
There are strong indirect signals that a 10× efficiency window may already be occurring inside frontier labs. METR’s independent assessment found capability growth at Anthropic running at roughly 1.9× its earlier rate, attributing ~80% of this acceleration to AI and expressing confidence it did not stem from faster physical compute scaling https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/ https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/. If capability progress doubles without a corresponding surge in raw training compute, effective software efficiency must increase dramatically. Extrapolating from this capability acceleration suggests that internal compute multipliers could be compounding at roughly 15–27× per year AI is dramatically accelerating AI research right now AI is dramatically accelerating AI research right now. Anthropic itself cites an "accelerating pace of producing compute multipliers" https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/ Anthropic, PBC - S-1/A-2, making near-term achievement technically plausible. Expectations about internal R&D compute allocations matter here: more compute devoted to experimentation can help produce efficiency gains, but capability acceleration alone does not isolate those gains or establish the qualifying training-compute comparison. Shrinking efficiency doubling times strengthen the case for a near-term 10× window without establishing when a credible public estimate will appear.
Documentation and Reporting Lag
The binding constraint for this forecast is not necessarily achieving the efficiency gain, but publishing a credible estimate of it. Closed labs heavily redact their algorithmic-efficiency data https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/, and third-party trackers like Epoch lack training-compute estimates for most closed frontier models released after mid-2025. Although an unnumbered IPO roadshow slide reportedly illustrates "superexponential" pre-training efficiency Coxon vs Anthropic IPO: Superexponential Compute Slide Anthropic can already retrain Mythos with "a fraction of the ..., it does not provide actionable figures Former OpenAI and Anthropic researcher Jacob Coxon .... Resolution will likely depend on retrospective analysis: Anthropic has promised "further reports" on acceleration trends, and METR’s embedded assessors could eventually publish sanitized findings. When a credible estimate is released, it will likely date the 12-month window retrospectively to 2026 or 2027. Establishing how internal R&D compute allocations contributed, and measuring gains across a full year of shrinking efficiency doubling times, could leave public documentation well behind internal achievement. I place the median publication date at December 15, 2028. Earlier disclosure remains plausible given the existing acceleration signals and prospects for further reports, with a 10% chance by April 10, 2027, and a 25% chance by December 5, 2027.
Constraints and Tail Risks
Despite the acceleration in coding throughput, evidence for fully autonomous algorithmic invention remains mixed. Epoch’s InnovationEval found that agents failed to rediscover a comparable ML innovation despite substantial experimental budgets Can AI automate AI R&D yet?, and an empirical study of small-scale recipes found only 1.57× annual efficiency gains Pretraining progress is mostly coming from data. Measured algorithmic gains are also highly dependent on scale and reference models On the Origin of Algorithmic Progress in AI. These technical bottlenecks, combined with potential safety pauses and chronic non-disclosure from labs facing IPO and regulatory scrutiny, stretch the upper percentiles into the 2030s. Retrospective reporting delays remain important even if internal R&D investment and shorter efficiency doubling times deliver the technical threshold much sooner: I put the 75th percentile at October 1, 2031, and the 90th percentile at December 31, 2036. This reflects a meaningful chance that no qualifying 10× estimate is ever credibly documented within the forecast horizon.
Integrating expectations about internal R&D compute allocations and shrinking efficiency doubling times pushed the median and upper percentiles slightly later to account for the lag in retrospective public reporting.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited