Near-Term Outlook

The Near-Term Outlook is a living record of dated, scoreable forecasts about AI over the three years from 2026-10-15 to 2029-10-15. It is deliberately conservative: it contains only forecasts that can be tied to a measurable source, and it declines to state numbers it cannot support.

Near-term forecasts are the only AI forecasts that can be scored within a project's own lifetime, so they carry most of the weight in judging whether the methods described in FCST-METHOD-001 work. This record is the stable-form counterpart to the Forecast Register (FCST-REG-001): the register holds the current status of every record, while this article explains the reasoning behind the near-term ones.

Reading rules.

• Statements of fact carry a date and source in-line. Fast-moving figures (model capabilities, prices, chip specifications, current benchmark scores) are not given here; use MODEL-001, FACT-001 and TIME-001.
• Where a forecast requires a baseline that this draft could not verify, the record says that the baseline is to be captured at first review rather than guessing.
• Confidence brackets marked "editorial judgment" are the encyclopedia's own qualitative estimates and are not outputs of a computed aggregate.

Near-term AI forecasts have a mixed record. Surveys and tournaments (see FCST-METHOD-001) document that expert timelines for the same milestone shifted by 13 years between consecutive survey waves (Grace et al., 2024), which is evidence that short-horizon beliefs moved sharply as capabilities were observed. No systematic scoring of 2022-era near-term AI forecasts is cited here. [VERIFICATION NEEDED: no study of the accuracy of pre-2023 near-term AI forecasts was located.]

Three kinds of near-term claim are separated:

1. Trend continuation. A measured quantity keeps its fitted rate (for example, doubling times of compute or task length).
2. Constraint arrival. A resource limit binds within the window (power, chips, data).
3. Institutional events. A legal or organisational event occurs by a date.

For trend continuation with a doubling time $T_d$, a quantity $y$ observed at time $t_0$ is projected as

Over a window of $\Delta t$ the projected multiple is $2^{\Delta t/T_d}$; for $\Delta t=12$ months this is $2^{12/7}\approx 3.3$ if $T_d=7$ months and $2$ if $T_d=12$ months. Because the multiple is exponentially sensitive to $T_d$, uncertainty in $T_d$ should be propagated rather than a single value assumed.

The record is organised by evidence type: (A) measured trends with published sources, (B) resource constraints with published projections, (C) institutional timetables, and (D) things this record explicitly does not forecast.

#6.1 Trends with published sources (as of the sources' dates)

• Training compute. Sevilla et al. (2022) report training-compute doubling of approximately every 6 months in the deep-learning era (analysis dated 2022).
• Task length. Kwa et al. (2025) report that the length of software tasks that frontier models complete at 50% reliability has doubled approximately every seven months since 2019, possibly faster in 2024, and cite about 50 minutes for Claude 3.7 Sonnet (original submission 2025-03-18; the arXiv page shows later revisions, whose contents were not reviewed here). The measure covers a particular family of software tasks and says nothing directly about other domains.
• Scale projection. Epoch AI (Sevilla et al., 2024; dated 2024-08-20) concluded that training runs of $2\times10^{29}$ FLOP will likely be feasible by the end of the decade, with power the constraint most likely to bind first; it estimated that 1-5 GW campuses could support runs of $10^{28}$ to $3\times10^{29}$ FLOP.

#6.2 Institutional timetable

Secondary legal sources (a law-firm summary dated 2026-06-22) report that co-legislators reached a provisional agreement on 2026-05-07 to postpone the EU AI Act's high-risk obligations to 2027-12-02 (standalone systems) and 2028-08-02 (product-embedded systems), that the European Parliament voted in plenary on 2026-06-16, and that formal adoption was expected before 2026-08-02 (Sidley Austin LLP, 2026). [VERIFICATION NEEDED: final Council adoption and Official Journal publication were not confirmed; the primary EUR-Lex text was not retrieved; some secondary sources disagree on a labelling deadline. Check the Official Journal before relying on any date.] Prohibited-practice and general-purpose-AI obligations already in application are reported as not delayed.

#6.3 Not forecast here

This record does not forecast: the date of AGI (see AGI-TIME-001); the release dates, names or scores of specific models; market valuations; or any intelligence-explosion dynamics (see ASI-EXP-001), for which no near-term operational test is proposed.

• Trend-continuation forecasts inherit the limitations of the underlying metric: saturation of a benchmark, changes in the task suite, and differences between laboratory and deployed performance.
• Resolution depends on third-party publishers (METR, Epoch AI) continuing their measurements; each record states a fallback.
• Closed laboratories do not disclose training compute, so compute resolutions rest on third-party estimates and may miss undisclosed runs.
• Whether capability trends translate into economic effects on the same timescale is a separate and less measured question (see ECON-COMP-001).

As of 2026-10-15 this draft's verified inputs are those cited in 6.1 and 6.2, each dated by its source. Developments after those source dates (including any model releases, measurements, or regulatory events in 2025-2026) were not systematically surveyed and are not asserted here. The first review (2027-01-15) is to refresh baselines.

#Forecast records

Record A

Record B

Record C

Companion near-term record: FCST-METHOD-001 (next survey wave, horizon 2029-10-15).

#Update log convention

Each review appends a dated row (date, record ID, observed value, status change) here and in FCST-REG-001. Records are never silently edited; superseded records receive status Superseded with a pointer.

• Foundational Concepts:
  • FCST-METHOD-001 Forecasting AI Progress: Methods and Calibration: Record format, scoring and standards.
  • AGI-TIME-001 AGI Forecasting & Timeline Methodology: Longer-horizon AGI timelines excluded from this record.
• Related Methods & Technologies:
  • ECON-COMP-001 The Economics of Compute & Scaling: Cost and supply context for the compute records.
  • MODEL-001 Model Registry: Living data on current models.
  • FACT-001 Factbook: Dated figures.
  • TIME-001 AI/SI Timeline: Event chronology.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: Out of scope here; no near-term test.
  • FCST-SIGNALS-001 Signals, Indicators and Inflection Points: Indicators that would change these records.
  • FCST-SCEN-001 Medium- and Long-Term Scenarios: Scenarios beyond the three-year window.
  • FCST-REG-001 Forecast Register: Current status of all records.