Definition and Baseline. Meeting this threshold requires a documented tenfold reduction in the training compute needed for a fixed capability level within a single 12-month window. The current baseline is far below this: Epoch's pre-training efficiency trend is roughly 3x/year, and multiple independent estimates drawn from inference price-performance and shared algorithmic progress cluster firmly around 2.5–4x/year 3 sources. A true 10x jump would require roughly tripling the current rate of progress, marking a severe acceleration signature rather than a historical continuation.

Physical Evidence and Acceleration. Currently, there is no evidence of a dramatic near-term speed-up. METR's analysis of long-running optimization series found no appreciable change in slope metr.org, and frontier labs consistently report that they have not crossed internal AI-R&D automation thresholds 2 sources. While AI-assistance anecdotes exist—such as labs merging substantially more code—this labor productivity has not yet translated into the required compute-efficiency value uplift metr.org. Achieving a 10x single-year leap physically requires heavy-tailed algorithmic breakthroughs or a full transition to automated AI R&D that drastically compresses doubling intervals .

The Definitional Path. There is a tangible near-term risk that the milestone is reached through a measurement convention shift rather than a sudden physical discontinuity. Epoch analysts have noted that if post-training improvements are included, efficiency gains might already be near 10x/year epoch.ai. If a credible tracker formally shifts its headline series to a post-training-inclusive metric, a 10x single-year figure could be published soon. This definitional risk—along with the chance of an early, narrow algorithmic paradigm leap—accounts for the probability mass before 2030 (p10 in mid-2028, p25 in mid-2030).

Measurement Degradation and the Upper Tail. The primary reason the median sits in mid-2033 is the severe and worsening measurement bottleneck. Resolution requires a credible published estimate, meaning a genuine acceleration must be cleanly documented. However, instruments are degrading exactly where they are needed: no closed frontier model released after July 2025 carries a public training-compute estimate, and METR's time-horizon suite is unreliable above 16 hours with a long-static dashboard metr.org. Tightening corporate secrecy structurally suppresses the likelihood of credible public documentation for a single-year efficiency spike , a dynamic that equally suppresses expectations for related transparent efficiency milestones . Even if a genuine 10x capability leap is physically achieved in the early 2030s as AI-assisted research matures, the joint probability that it is simultaneously reached and credibly published is less than 50% by mid-2033. Furthermore, set against related questions, the upper percentiles reflect a cohesive view on the risk of scaling walls and algorithmic stagnation. This risk, combined with permanent opacity, justifies extending the upper tail into the late 2030s (p75 in early 2038) and capping the 90th percentile at 'never'.