AI Trajectory

Forecasts

Forecasts

Which claims are surviving contact with reality? Follow upcoming tests first, then compare later evidence, shared milestones, and changes in published expectations.

Watch next

The next claims reality can test.

These are the nearest comparable deadlines still in play. They are checkpoints to revisit—not probabilities that an event will happen.

Selected claims · not a scorecard

How are the tracker’s selected claims comparing with later evidence?

This is a coverage-limited status map, not an evaluation of whole bodies of work. Selection is not preregistered or representative; forecasts, scenarios, models, and intentions answer different questions, so counts are never grades.

Confirmed target met Ahead reality faster On-track consistent so far Behind reality lagging Pending not yet testable Indeterminate deadline passed; public evidence cannot resolve it Falsified target missed
Body of workCoverageCurrent evidence
Situational Awareness Leopold Aschenbrenner · 2024-06 5 comparable · 6 proxy/context 7 areas: capability, compute, algorithms, physical, capital, value, response
1 confirmed1 ahead1 on-track2 pending
Jump to individual claims
AI 2027 Kokotajlo/Lifland et al. · 2025-04 3 comparable · 4 proxy/context 4 areas: compute, automation, physical, value
1 on-track1 behind1 indeterminate
Jump to individual claims
Dec 2025 timelines update AI Futures Project · 2025-12 1 comparable · 0 proxy/context 1 area: automation
1 pending
Jump to individual claims
Forethought SIE series MacAskill/Davidson et al. · 2025 2 comparable · 0 proxy/context 2 areas: capability, algorithms
2 on-track
Jump to individual claims
Bio Anchors (2-yr update)Ajeya Cotra · 2022 Source registeredNo structured claim harvest yet

Not scored. An empty row describes tracker coverage, not the forecaster’s position.
various (vintaged) Epoch AI · varies 4 comparable · 3 proxy/context 4 areas: compute, physical, capital, value
3 on-track1 pending
Jump to individual claims
Compute-centric takeoff-speeds framework Tom Davidson · 2023 0 comparable · 3 proxy/context 3 areas: compute, algorithms, automation
Jump to individual claims
Case for multi-decade timelinesEge Erdil · 2025-04 Source registeredNo structured claim harvest yet

Not scored. An empty row describes tracker coverage, not the forecaster’s position.
AI Infrastructure Spending: Inelastic Demand and Rising Costs Morgan Stanley · 2026 0 comparable · 1 proxy/context 1 area: capital
Jump to individual claims
Internal-goals statement (livestream + X post) Sam Altman (OpenAI) · 2025-10 0 comparable · 1 proxy/context 1 area: automation
Jump to individual claims

Coverage varies by source and subject. Portfolio bars include only selected direct and formula-backed translated claims; proxy and context claims remain linked but excluded. Status and assessment confidence remain separate.

Forecasts versus reality

Different predictions, translated into the same milestones.

Authors use different definitions and dates. This ladder keeps their original wording, then shows where their claims overlap.

  1. Rung 1Agent completes 1-month human-equivalent research task

    No structured forecast is attached to this rung yet.

  2. Rung 2Superhuman Coder (AI 2027 definition)
    AI Futurespendingmedium confidence
    bull model direct AI Futures Project · 2025-12 ↗
    “the new model with median parameters predicts SC in Dec 2031”AI Futures Model Dec 2025 update (supersedes AI 2027's Mar 2027 scenario date; do NOT overwrite the original — record both)
    Prediction
    Superhuman Coder
    By
    2031-12

    Conditionality: firm model output, but authors retain 'highly uncertain... cannot confidently predict a specific year'

    Revision trail: ai2027-apr2025 SC date (2027-03)

    Statuspending
    Assessment confidencemedium

    Evidence · 2026-07Nothing close to the SC bar (any AGI-company coding task, 30x faster, 30x cheaper) exists: SWE-bench Verified 79.2% (Opus 4.6), Terminal-Bench 2.0 ~65%; the independent AI 2027 tracker grades SC 'Not Yet Testable' (Jun 2026). The authors' aggregate self-grading puts quantitative progress at ~65-75% of scenario pace.Canonical measurement source 1 ↗

    CounterargumentEven the revised median has wide CI (10th pct 2027.5), and individual forecaster medians (Daniel ~2029, Eli ~2032) still disagree by years — the Dec 2031 model output is not a settled team view.

    How this status was assessed
    Relationship to measurement
    The published target and assessment evidence refer to the same measured quantity or milestone.
    Test
    Compare the published target and deadline with the metric observation and recorded milestone evidence.
    Target / deadline
    Superhuman Coder · 2031-12
    1 recorded assessment
    1. 2026-07pendingmedium confidence
  3. Rung 3Automated AI researcher / ASARA (Leopold, Forethought)
    Leopoldpendinglow confidence
    bull forecast proxy · excluded from totals Leopold Aschenbrenner · 2024-06 ↗
    “it is strikingly plausible that by 2027, models will be able to do the work of an AI researcher/engineer.”Ch I, From GPT-4 to AGI
    Prediction
    AI researcher/engineer automation
    By
    2027

    Conditionality: conditional-scenario ('strikingly plausible')

    Statuspending
    Assessment confidencelow

    Evidence · 2026-07Strong coding-agent adoption and longer autonomy, but no autonomous AI-researcher demonstrated; RE-Bench-class evals still short of expert autonomy.Canonical measurement source 1 ↗

    CounterargumentThe AI Futures Project itself (Dec 2025) pushed its Superhuman Coder median from Jan 2027 to Dec 2031, evidence against the 2027 rung.

    How this status was assessed
    Relationship to measurement
    Task-horizon evidence is a proxy for drop-in researcher or engineer automation, not a direct labor-substitution measure.
    Test
    Use the named metric only as directional evidence; do not treat it as a like-for-like test.
    Target / deadline
    AI researcher/engineer automation · 2027
    1 recorded assessment
    1. 2026-07pendinglow confidence
    AltmanContext only · not status-assessed
    bull intention context · excluded from totals Sam Altman (OpenAI) · 2025-10 ↗
    “We have set internal goals of having an automated AI research intern by September of 2026 running on hundreds of thousands of GPUs, and a true automated AI researcher by March of 2028.”Livestream + X post, 2025-10-28 (x.com/sama/status/1983584366547829073)
    Prediction
    automated AI research intern, then true automated AI researcher
    By
    2026-09 / 2028-03

    Conditionality: explicit self-hedge in the same statement: 'We may totally fail at this goal.' Lab-leader statement = intention data (bottom of the evidence hierarchy), scored with the corresponding discount.

    TreatmentContext only · not status-assessed

    Evidence · 2026-07Six weeks from the intern deadline, no public demonstration exists. Strongest adjacent result is DeepMind's AlphaEvolve (novel matrix-multiplication algorithm — real but narrow-domain optimization, not open-ended research). METR's NanoGPT-speedrun analysis (Apr 2026): AI contributions real but 'none reached the deep or breakthrough end of the scale'.Canonical measurement source 1 ↗

    CounterargumentAn 'internal goal' can be declared met internally without public verification — this claim may resolve only on OpenAI's say-so, which the evidence hierarchy discounts to intention data either way.

    Why this is context only
    Relationship to measurement
    Preserved as dated context; excluded from status totals and charts.
    Test
    Do not include this statement in aggregate status results.
    Target / deadline
    automated AI research intern, then true automated AI researcher · 2026-09 / 2028-03
    1 recorded context note
    1. 2026-07Context note
  4. Rung 4AI R&D speedup multiplier crosses 2x / 10x
    AI 2027on-trackmedium confidence
    bull scenario proxy · excluded from totals Kokotajlo/Lifland et al. · 2025-04 ↗
    “Early 2026: AI R&D progress multiplier reaches 1.5x — algorithmic progress '50% faster' with AI assistants than without”AI 2027 scenario narrative + compute-forecast multiplier table. CANONICAL READING: the narrative ladder (1.5x early 2026 -> 3x Jan 2027) — the corpus has two conflicting multiplier tables (compute-forecast quarterly vs takeoff-forecast milestone-tied), recorded per schema rule 4.
    Prediction
    1.5x AI R&D progress multiplier
    By
    2026-06

    Conditionality: scenario narrative with the project-wide hedge ('not a prediction; 2027 was our modal year'). The authors' own end-2026 target was 1.9x, which they now grade 'behind pace'.

    Statuson-track
    Assessment confidencemedium
    Self-reported · discounted

    Evidence · 2026-07METR's self-report survey (May 2026) median value-multiplier is 1.4-2x; METR's production-function read of Anthropic's 8x code-volume figure puts researcher uplift 'plausibly >2x'; the third-party AI 2027 tracker marks the 1.5x rung On Track. Evidence is self-report-dominated, hence the discount.Canonical measurement source 1 ↗

    CounterargumentAnthropic's own system card says acceleration is 'well short of a sustained, AI-attributable doubling of the overall pace'; the only RCT-grade measurement found a 19% SLOWDOWN (since methodologically disowned by METR itself, but never replaced with a positive result); self-reports historically overstate measured savings by ~40 percentage points.

    How this status was assessed
    Relationship to measurement
    The R&D multiplier is assessed from a composite of lab self-reports and external evaluations rather than one standardized series.
    Test
    Use the named metric only as directional evidence; do not treat it as a like-for-like test.
    Target / deadline
    1.5x AI R&D progress multiplier · 2026-06
    1 recorded assessment
    1. 2026-07on-trackmedium confidence
  5. Rung 5AGI-class (each author's own definition recorded)

    No structured forecast is attached to this rung yet.

How expectations change

Are forecasters moving their timelines?

Each line follows one person’s published median over time. It shows revision, not consensus: these estimates are never averaged together.

PUBLISHED MEDIAN, BY VINTAGE203020402050predicts202120222023202420252026publishedCotra — transformative AI: median 2050 as of 2020-09 — Draft report on AI timelines: 'a median of ~2050 for transformative AI'Cotra — transformative AI: median 2040 as of 2022-08 — Two-year update: 'A median of ~2040 (a decrease of ~10 years from 2050)'Cotra — transformative AIKokotajlo — AGI: median 2028 as of 2025-04 — AI 2027 scenario vintageKokotajlo — AGI: median 2030.95 as of 2026-01 — AI Futures update: 'Jan 2026: Dec 2030 (2030.95)' — 'things seem to be going somewhat slower than the AI 2027 scenario'Kokotajlo — AGILifland — AGI: median 2031 as of 2025-04 — AI 2027 vintage (Automated Coder)Lifland — AGI: median 2035 as of 2025-12 — Dec 2025 update: ~Jan 2035, 'about 1.5 years later than the model's output'Lifland — AGI
View chart as data
ForecasterPublishedMedian yearMilestone
Cotra — transformative AI2020-092050Not recorded
Cotra — transformative AI2022-082040Not recorded
Kokotajlo — AGI2025-042028Not recorded
Kokotajlo — AGI2026-012030.95Not recorded
Lifland — AGI2025-042031Not recorded
Lifland — AGI2025-122035Not recorded