◆  A MEASUREMENT INSTRUMENT, NOT A FORECAST

Three clocks, one field of compressing time.

A public-interest instrument tracking the arrival of AGI, the Singularity, and Superintelligence — derived from a versioned ensemble of capability, compute, economic, and alignment signals. Every movement sourced. Every methodology auditable.

AGI · APR 2027 AGI SINGULARITY · DEC 2029 SINGULARITY ASI · DEC 2031 ASI NOW
Nearest Clock · AGI
010M13D
Est. Apr 2027 · projection, not prediction
Signal Oscillogram · trailing 90 days · 3 channels · reconstructed from changelog Y = Δ months since edition · X = daily ticks
AGI CLOCK 5 signals
AGIArtificial General Intelligence
10 mo · Est. Apr 2027
Holding · flat since Jun
SINGULARITY CLOCK 2 signals
SingularityRecursive Self-Improvement Inflection
3yr 6 mo · Est. Dec 2029
Holding · flat since Jun
ASI CLOCK 3 signals
ASIArtificial Superintelligence
5yr 6 mo · Est. Dec 2031
Holding · flat since Jun
LIVE · 14 SIGNALS ·
METR · TASK HORIZON 11.4 hr ▲ +38% QOQ SWE-BENCH VERIFIED 88.2% ▲ +4.1pp QOQ HUMANITY'S LAST EXAM 38.9% ▲ +16pp QOQ ARC-AGI-2 62.4% ▲ +18pp QOQ FRONTIERMATH 41.7% ▲ +12pp QOQ GPQA DIAMOND 94.8% ▲ +2.3pp QOQ FRONTIER TRAINING COMPUTE 8.2e+26 ■ 2.1× YoY AI R&D CAPEX · TOP LABS $512B ▲ +64% YOY AI-AUTHORED RESEARCH SHARE 14.2% ▲ +5.3pp YOY INTERPRETABILITY INDEX 0.31 ▲ −0.02 QOQ SAFETY HEADCOUNT · LABS 12.4% ▼ −1.1pp YOY MANIFOLD · AGI SENTIMENT 25% ■ 8 mkts ARXIV · CAPABILITY FLUX 1.77× ▲ 3631 papers FRONTIER RELEASE FLUX 0.94× ■ 462 posts METR · TASK HORIZON 11.4 hr ▲ +38% QOQ SWE-BENCH VERIFIED 88.2% ▲ +4.1pp QOQ HUMANITY'S LAST EXAM 38.9% ▲ +16pp QOQ ARC-AGI-2 62.4% ▲ +18pp QOQ FRONTIERMATH 41.7% ▲ +12pp QOQ GPQA DIAMOND 94.8% ▲ +2.3pp QOQ FRONTIER TRAINING COMPUTE 8.2e+26 ■ 2.1× YoY AI R&D CAPEX · TOP LABS $512B ▲ +64% YOY AI-AUTHORED RESEARCH SHARE 14.2% ▲ +5.3pp YOY INTERPRETABILITY INDEX 0.31 ▲ −0.02 QOQ SAFETY HEADCOUNT · LABS 12.4% ▼ −1.1pp YOY MANIFOLD · AGI SENTIMENT 25% ■ 8 mkts ARXIV · CAPABILITY FLUX 1.77× ▲ 3631 papers FRONTIER RELEASE FLUX 0.94× ■ 462 posts
Alignment Deficit · Gauge, not prediction

Capability is pulling 1.73× faster than safety.

v1.4.0 · 180-day trailing · refreshed 2026-06-03
CAPABILITY · SAFETY VELOCITY RATIO TRAILING 180 DAYS
0.5×1.0×1.5×2.0×2.5× PARITY · 1.00× TODAY · 1.73× −180d−135d−90d−45dTODAY
STATE Moderate
1.73×
Keeping paceModerateElevatedCritical

Weighted blend of the capability-velocity proxy, arXiv flux ratio, and frontier-release flux ratio. Clamped to [0.25, 6.0]; flux inputs skipped when the classified sample is too small.

Signals from the Field

What moved, in the wild.

3631 arXiv papers · 462 lab/org posts · LLM-classified
2026-06-01
MODEL LAUNCH
Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action
Launch of NVIDIA Cosmos 3, a major omni-model for physical AI reasoning and action.
Lab · Hugging Face
Mag 4
2026-05-28
CAPABILITY LEAP
Reasoning with Sampling: Cutting at Decision Points
Novel sampling method improves frontier reasoning models on MATH, HumanEval, GPQA without additional training; measurable gains.
arXiv · arXiv
Mag 4
2026-05-28
CAPABILITY LEAP
Self-Trained Verification for Training- and Test-Time Self-Improvement
Self-trained verification dramatically improves reasoning on hard problems (14x on science tasks), unlocking both test-time and training-time improvements via verifier-in-the-loop approach.
arXiv · arXiv
Mag 4
2026-05-28
CAPABILITY LEAP
Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
Unified embodied foundation model achieving SOTA across manipulation, navigation, and trajectory prediction with strong out-of-distribution generalization across robot embodiments.
arXiv · arXiv
Mag 4
2026-05-28
CAPABILITY LEAP
Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation
Novel relational task extrapolation framework enabling systematic generalization beyond training distribution; demonstrates substantial gains across multiple extrapolation regimes including compositional depth.
arXiv · arXiv
Mag 4
2026-05-28
CAPABILITY LEAP
PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers
First LLM achieving competitive expert-level poker without training or solvers, solving imperfect-information game at scale.
arXiv · arXiv
Mag 4
2026-05-28
CAPABILITY LEAP
VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies
Novel visual reasoning framework for VLAs achieving 22.8× latency speedup while improving success rates—meaningful advance in embodied AI capability and efficiency at scale.
arXiv · arXiv
Mag 4
2026-05-28
MODEL LAUNCH
Introducing Claude Opus 4.8
New frontier model version (Claude Opus 4.8) from major AI lab; significant capability release.
Lab · Anthropic
Mag 4
2026-05-19
SAFETY
Frontier Risk Report (February to March 2026)
Substantive misalignment risk assessment of frontier AI agents across major labs with detailed evaluations and findings.
Lab · METR
Mag 4
2026-05-19
SAFETY
前沿 AI 风险报告（2026 年 2–3 月）
Third-party evaluation of AI agent misalignment risks across frontier labs; substantive safety research.
Lab · METR
Mag 4
Live signal readings · 14 of 14

Every clock is downstream of a public number.

Refreshed daily · sourced · changelogged
METR · task horizon
11.4 hr
+38% QOQ METR
SWE-Bench Verified
88.2%
+4.1pp QOQ SWE-bench
Humanity's Last Exam
38.9%
+16pp QOQ Center for AI Safety
ARC-AGI-2
62.4%
+18pp QOQ ARC Prize
FrontierMath
41.7%
+12pp QOQ Epoch AI
GPQA Diamond
94.8%
+2.3pp QOQ GPQA
Frontier training compute
8.2e+26
2.1× YoY Epoch AI
AI R&D capex · top labs
$512B
+64% YOY SemiAnalysis
AI-authored research share
14.2%
+5.3pp YOY arXiv attribution analysis
Interpretability index
0.31
−0.02 QOQ TSC composite
Safety headcount · labs
12.4%
−1.1pp YOY TSC survey
Manifold · AGI sentiment
25%
8 mkts Manifold
arXiv · capability flux
1.77×
3631 papers arXiv cs.{AI,LG,CL}
Frontier release flux
0.94×
462 posts Labs + safety orgs RSS
Recent Clock Movements · last 90 days

Every shift, logged and attributed.

Why two streams?
Apr 15, 2026
AGI
METHODOLOGY Methodology v1.4.0. AGI projection moved later (2026-12-15 → 2027-04-16) after three changes. (1) Benchmark signals (SWE-bench, HLE, ARC-AGI-2, FrontierMath) now use sigmoid (saturating-exponential) extrapolation instead of linear. Benchmarks are bounded; linear models overstate how fast the last 10pp closes and were pulling the AGI date ~4 months too early. (2) METR task-horizon weight raised 0.35 → 0.45: it is the only direct economic-capability measure and should dominate benchmark proxies. (3) Metaculus weight raised 0.15 → 0.20 as a reserved slot; currently absent because Metaculus gates Community Prediction data to a whitelist that excludes AGI questions (access request sent to api-requests@metaculus.com on 2026-04-15). Benchmark aggregate weight dropped 0.50 → 0.35. Source: lib/projection.py v1.4.0 project_agi · 2026-04-15
+122 days
Apr 15, 2026
SING
METHODOLOGY Methodology v1.4.0. Singularity projection moved earlier (2032-11-26 → 2029-12-16) after reweighting. METR task horizon (→1000hr, month-scale autonomous research) weight raised 0.20 → 0.35 as a leading indicator; AI-authored research share (→50%) weight lowered 0.50 → 0.35 because it is a lagging proxy (a linear fit to a pre-phase-transition variable is structurally conservative once capability crosses research-scale horizons). Metaculus+3yr heuristic unchanged at 0.30. Linear extrapolation retained for AI-authored share since it has no natural ceiling at low values. Net: more weight on the leading indicator, which produces the more compressed date. Source: lib/projection.py v1.4.0 project_singularity · 2026-04-15
-1076 days
Apr 15, 2026
ASI
METHODOLOGY Methodology v1.4.0. ASI projection moved earlier (2033-11-26 → 2031-12-16). Three changes. (1) HLE→95% and FrontierMath→95% now use sigmoid extrapolation with weights raised 0.30 → 0.375 each. (2) Singularity+2yr weight reduced 0.40 → 0.25 so ASI does not inherit Singularity's inertia. (3) Cascade minimum buffer from Singularity to ASI widened from 365d to 730d (MIN_BUFFER_DAYS_SING_TO_ASI), reflecting that ASI (qualitatively beyond best human, all domains) is not a one-year hop past recursive-improvement onset. The new date is pinned by cascade (Singularity 2029-12 + 730d = 2031-12); benchmark signals independently project earlier (2028-08 and 2029-03). Source: lib/projection.py v1.4.0 project_asi · 2026-04-15
-711 days
Apr 15, 2026
ALIGN
METHODOLOGY Methodology v1.4.0. Alignment Deficit formula unchanged in this release; gauge reading holds at 1.7× (Moderate). v1.4.0 is a clocks-only rework: sigmoid benchmark extrapolation, reweighted AGI/Singularity/ASI ensembles, widened ASI cascade buffer, plus editorial additions (first-visit disclaimer modal, Limitations methodology section, api.json and RSS/JSON-Feed disclaimer fields). The Alignment blend stays at 0.6 structured + 0.2 arXiv flux + 0.2 frontier flux. Next alignment methodology work will integrate a direct safety-capex proxy when the data source stabilizes. Source: v1.4.0 release notes · 2026-04-15
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Gauge reading moved Keeping pace → Moderate (1.19× → 1.61×) with the v1.3.0 methodology in effect. This is expected given the structured rework: the prior flow-vs-stock formula mathematically couldn't see that safety headcount share was shrinking and interpretability was declining. Under growth-vs-growth the structured ratio is 1.98×, which blends with arxiv flux (1.11×, post-symmetrization) and frontier flux (1.00×) at 60/20/20 to 1.61×. No real-world change today, the gauge is reflecting an existing reality that prior methodology obscured. Source: v1.3 post-release reading · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Methodology v1.3.0 also symmetrized the arXiv capability-vs-alignment discount. Prior v1.2 formula weighted capability_incremental at 0.25× while alignment_eval_result got 0.5×, an asymmetric thumb on the scale that compressed the ratio toward 1.0 and systematically understated the deficit. Both 'softer' sub-categories now receive the same 0.5× multiplier; landmark categories (capability_leap, alignment_advance) keep full weight. Observed effect: arXiv ratio moved ~1.11× → ~1.48× in subsequent runs (exact value depends on daily classifier output). Source: lib/ingest/arxiv_classifier.py::_aggregate v1.3 · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Methodology v1.3.0: structured composite reworked from flow-vs-stock to growth-vs-growth. Capability velocity is YoY capex growth (1.64×); safety velocity is the equal-weight average of safety-headcount growth (0.91×, share is shrinking 1.1pp/yr off 12.4% base) and interpretability-index growth (0.74×, index is declining 0.02/quarter off 0.31 base). Structured ratio moved 1.28× → 1.98×. The prior formula compared a flow (capex growth) to a stock (headcount level), which mathematically hid the trend. Floored at 0.25 to prevent near-zero growth blowing up the ratio. New safety_headcount_growth and interpretability_growth fields exposed in api.json for audit. Source: lib/projection.py::project_alignment v1.3 · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Methodology v1.3.0: arXiv classifier discount symmetrized. Previous formula weighted capability_incremental papers at 0.25× while alignment_eval_result got 0.5×, an asymmetric thumb on the scale that compressed the ratio toward 1.0. Both 'softer' categories now get 0.5×; landmark categories (capability_leap, alignment_advance) keep full weight. arXiv-flux ratio will re-aggregate at the next ingest cycle. Source: lib/ingest/arxiv_classifier.py::_aggregate v1.3 · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Alignment Deficit reading moved Keeping pace → Moderate (1.19× → 1.61×) following the v1.3 methodology rework. This is a measurement-bias correction, not a real-world worsening: the prior reading was understating the deficit because the structured formula compared a growth rate (capex) to a level (safety headcount). With both sides now expressed as growth rates, the gauge correctly reflects that capability investment is growing 64% YoY while safety headcount share and interpretability are both declining. The previous 'Keeping pace' reading should be considered retrospectively corrected. Source: Alignment gauge v1.3 transition note · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Alignment Deficit reading moved Elevated → Keeping pace (3.0× → 1.29×). This is a methodology effect from the v1.2 blend introduction, not a real-world improvement: the structured ratio (1.28×) is now diluted with arXiv flux (1.35×) and frontier-release flux (1.28×) at 20% weights each. Note three known biases all pulling the blend low: (a) structured formula compares capability flow to safety stock, hiding that safety headcount is shrinking 1.1pp/yr while capex grows 64% YoY; (b) frontier-flux undercounts the safety side because Anthropic, Apollo, METR, Epoch, Meta, Mistral, and xAI lack discoverable RSS feeds; (c) arXiv flux applies a 4× discount to incremental capability that is asymmetric with the alignment-side discount. A more honest reading is likely Moderate (~1.8–2.4×). Restoring the safety-feed coverage and symmetrizing the arXiv discount are tracked for v1.2.1; structured-formula rework is tracked for v1.3. Source: Methodology doc: alignment blend transparency · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Frontier-feed coverage update: dropped 7 lab feeds that 404 with no discoverable RSS (Anthropic, Apollo, Epoch, Meta, Mistral, xAI, plus broken METR URL). Added 3 working safety-side feeds: METR (correct URL: metr.org/feed.xml), MIRI, and LessWrong curated. Added Google Research blog. Net effect: safety-side feed coverage roughly doubles, partially offsetting the structural undercount flagged in the prior entry. Source: data/frontier_feeds.json · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Methodology v1.2.0: Alignment Deficit blend now includes LLM-classified frontier-lab and safety-org RSS feeds (trailing 90 days). Captures launches, capability claims, and safety/policy posts that don't appear on arXiv. Final blend: 60% structured + 20% arXiv flux + 20% frontier-release flux. New 'Signals from the Field' site section surfaces highlighted papers and lab posts. Source: Methodology doc: frontier release monitor · 2026-04-14
gauge change
Apr 14, 2026
ALIGN
METHODOLOGY Methodology v1.1.0: Alignment Deficit now blends LLM-classified arXiv capability/alignment flux (trailing 30 days, cs.AI/cs.LG/cs.CL) with existing structured inputs at 30% weight. Grounds the gauge in observed research output, not only lab-reported headcount and spend. Clock projections unchanged. Source: Methodology doc: arXiv classifier · 2026-04-14
gauge change
Apr 12, 2026
AGI
SIGNAL AGI Clock moved 11 days closer following METR task-horizon update: autonomous task completion length crossed 11 hours, passing the 10-hour inflection 4 months ahead of model expectations. Source: METR Task Horizon v14 release · 2026-04-12
-11 days
Mar 28, 2026
SING
SIGNAL Singularity Clock moved 6 days closer after Anthropic publication on automated red-teaming showed AI-driven interpretability contributions at frontier lab level. Source: Anthropic research post · 2026-03-28
-6 days
Mar 19, 2026
ALIGN
SIGNAL Alignment Deficit gauge moved from Moderate to Elevated: capability/safety velocity ratio crossed 3.0× for first time since index inception. Source: TSC composite index · 2026-03-19
gauge change
Methodology

Open source, versioned, auditable.

projection: v1.4.0 · since Apr 15 2026
WHAT WE COUNT TO

Operational, not rhetorical, definitions.

AGI, Singularity, and ASI each have a single operational definition. Every signal either contributes evidence toward it, or it does not.

HOW WE PROJECT

Ensemble over trajectories, not point forecasts.

Each clock is a weighted ensemble of trajectory models across capability, compute, economic, and safety signals. We publish the ensemble, the weights, and the residuals.

CASCADE ORDERING

One signal can move one clock.

A signal promoting AGI cannot skip to ASI. Cascades are ordered and each promotion is logged as a changelog entry with the triggering print.

WHY WE VERSION THE MODEL

Transparency as a feature, not a disclosure.

An opaque clock is an opinion in a lab coat. We publish every revision, every weight, every threshold, every formula on GitHub so you can argue with the data and the method, and reproduce any past reading from a git hash.

The changelog's Signal / Methodology split exists for the same reason: a methodology revision can shift a clock by years on the day it ships, while the world hasn't changed. Separating them makes the instrument's own motion legible.

LIMITATIONS

What the instrument cannot do.

These projections are a research instrument, not a forecast you should act on.

  • Benchmark saturation is a proxy for capability, not proof of AGI.
  • Log-linear and linear extrapolation breaks near benchmark ceilings and at phase transitions; real capability curves are sigmoidal.
  • Signal availability changes (e.g. Metaculus community predictions are currently gated).
  • Wide confidence intervals reflect real uncertainty, not rounding.
  • The model is opinion, not prediction. Methodology and code are public so you can disagree with both.

Do not use for investment, career, medical, legal, or life decisions.

OPEN SOURCE

Everything lives on GitHub.

Model code, signal ingest, changelog, RFCs, and this page are in a public repo. Every methodology change is a merged PR.

Data & feeds

Subscribe, embed, or fork.

CC-BY-4.0 · stewarded by Anapto
LATEST COUNTDOWN

Just the numbers.

One-item feed that updates only when a clock actually moves.
RSS · JSON Feed

MOVEMENTS & METHODOLOGY

Everything, annotated.

Changelog entries + daily synthesis; signal & methodology moves both included.
RSS · JSON Feed

LIVE API

JSON.

Current projections, confidence intervals, and per-signal rationale.
api.json