Signals, Indicators, and Inflection Points

A signal (or indicator) is a measurable quantity chosen in advance because its movement should change a forecast; an inflection point is a point at which the rate or direction of a tracked quantity changes materially, such as a transition from exponential to saturating growth. This article sets out how indicators for AI progress are chosen, what each one can and cannot show, and how pre-committed thresholds keep updating honest.

Forecasts and scenarios (FCST-SCEN-001) are only useful if observations can update them. Without pre-specified indicators, observers tend to interpret any new development as confirming their existing view. The discipline described here is to state, before the data arrive, which readings would move which forecast and by roughly how much.

Early-warning indicator practice comes from economics, public health, and intelligence analysis. In AI, indicators have mostly been published capability benchmarks and scaling datasets. Epoch AI's compute-trend analysis (Sevilla et al., 2022) turned training compute into a tracked series; METR's time-horizon measure (Kwa et al., 2025) added a capability series expressed in human task duration, intended to be comparable across model generations. Dynamic forecasting benchmarks such as ForecastBench (Karger et al., 2024) add a series measuring a model's ability to forecast. Expert surveys (Grace et al., 2024) provide a periodic belief indicator.

#4.1 Leading, coincident and lagging indicators

• Leading: resource inputs (training compute, power procurement, chip supply, capital). They change before capability and are the focus of projections such as Sevilla et al. (2024).
• Coincident: measured capability (benchmark and task-length series).
• Lagging: economic and social effects (productivity, employment, incidents), and the stated beliefs of experts.

#4.2 Detecting a trend change

For a series $y_t$ with log-linear growth $\ln y_t=\alpha+\beta t+\varepsilon_t$, a change in slope can be tested by comparing the fit of one slope against two slopes with a breakpoint $\tau$, for example with a likelihood-ratio or information-criterion comparison. Two cautions apply. First, with few observations (frontier models appear a few times per year), detection lags the true change by many months. Second, the observed series is a measurement of capability, and a change in the measurement (task suite, scoring, evaluation tooling) can mimic a change in the trend. The METR paper itself reports that the trend "may have accelerated in 2024" while noting a seven-month average doubling since 2019 (Kwa et al., 2025), which illustrates the difficulty of distinguishing a slope change from noise on a short window.

#4.3 Saturation

Benchmarks saturate: once the best systems score near the maximum, the series stops discriminating. A flat benchmark line after saturation is not evidence that capability has plateaued. Indicators should therefore be retired and replaced on a stated schedule, with overlap periods to allow re-calibration (see CAP-BENCH-001 on contamination and measurement problems).

#4.4 Pre-committed thresholds

For each indicator, a record states: the series and its publisher; a baseline value and date; a threshold; and the forecast or scenario it informs. Because the baseline for several series was not verified in this draft, baselines are captured at the first review of each record.

• Update rule for the time-horizon record. In FCST-HORIZON-001 (FCST-NEAR-001), if METR's published horizon rises by more than a factor of two within 12 months, the record resolves true and the Scenario B/C boundary is revisited; if it rises by less than a factor of two over 18 months, confidence in continued fitted trends is reduced.
• Compute indicator. If a documented run above $2\times10^{29}$ FLOP appears before 2029-10-15, FCST-COMPUTE-001 fails and the Epoch 2030 feasibility projection (Sevilla et al., 2024) is treated as having arrived early.
• Survey indicator. A shift in the 50% date of 13 years between 2022 and 2023 (Grace et al., 2024) shows how much a belief indicator can move within a year; it should be read with the capability events of that year, not as an independent signal.

• Goodhart effects. Once an indicator is widely used as a target, laboratories may optimise for it, weakening its meaning.
• Selection of indicators after the fact defeats the purpose; this article's list is fixed at the stated date and additions are logged.
• Independence. Several indicators share a cause (for example, the same compute growth), so agreement among them is weaker evidence than it looks.
• Unresolved: a validated standard measure of AI contribution to AI research; indicators for non-software capability; indicators of deployed (as opposed to demonstrated) capability.

As of 2026-10-15, the compute series, the METR time-horizon series, a dynamic forecasting benchmark, and a recurring expert survey exist as public series with the properties described above, per the cited sources (dated 2022-2025). The state of each series today was not re-measured for this draft; use MODEL-001, FACT-001 and TIME-001 for current values. No standard indicator of AI-driven acceleration of AI research was identified. [VERIFICATION NEEDED: a literature search for such measures was not completed.]

Indicator sets will likely be extended (more domains, more agentic measures) and some will be retired at saturation. The following record applies the methods above to one leading indicator.

#Forecast record

• Foundational Concepts:
  • FCST-METHOD-001 Forecasting AI Progress: Methods and Calibration: Scoring and record format.
  • CAP-BENCH-001 AI Evaluation, Benchmarking & Contamination: Why capability indicators saturate and mislead.
• Related Methods & Technologies:
  • ECON-COMP-001 The Economics of Compute & Scaling: Leading resource indicators.
  • AGI-TIME-001 AGI Forecasting & Timeline Methodology: Timelines the indicators update.
• Advanced Implications & Future Trajectories:
  • FCST-SCEN-001 Medium- and Long-Term Scenarios: Scenarios the indicators discriminate.
  • FCST-NEAR-001 Near-Term Outlook (0-3 years): Dated records using these indicators.
  • FCST-UNCERT-001 The Uncertainties That Matter Most: Uncertainties indicators should target.
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: The mechanism whose indicator is not yet standardised.
  • FCST-REG-001 Forecast Register: Record status.