OFF-CYCLE RESEARCH / FORECAST AUDIT · v1.0.0
10,000 forecasts do not give us one AGI date.
AIMultiple assembled an unusually useful map of surveys, named forecasts and prediction markets. Its headline date is not a single consensus estimate, however: the inputs ask different questions, use different definitions and update on different clocks. For capital, the disagreement is more valuable than the average.
OOP-AGI-FORECAST-CLOCKS-2026-09-15 · EVIDENCE CUTOFF 2026-09-15
Use the article as a source registry and a map of forecast dispersion—not as a portfolio base case. “AGI in the late 2020s or early 2030s” is a defensible description of today’s fast-moving market and individual forecasts, while the underlying researcher surveys are generally later. Averaging them into one calendar year creates false precision. We will instead track six separate clocks and ask which economic transmission path moved, who can capture it and what the market had already priced.
01 / DENOMINATOR
What “10,000” contains
research surveys
More than 6,000 participants in total; most cited surveys place a 50% AGI-like threshold in 2040–2061.
named individual forecasts
Leaders, entrepreneurs and researchers with very different incentives, definitions and information sets.
community and market forecasts
Contributors across platforms—not 3,900 independent answers to one frozen question.
02 / METHOD AUDIT
Where one consensus date breaks
The target changes between rows
AGI, weak AGI, HLMI, full automation of labor, a high-quality Turing test and a company announcement are not interchangeable endpoints.
The headline count is not one sample
Survey respondents, market participants and named forecasters have different sampling processes. A participant count cannot be treated as ten thousand equally weighted, independent probability distributions.
The clocks do not update together
The article was updated on 14 September, but several embedded market readings were August snapshots. On 15 September, the live pages already showed materially different values.
One weighted year destroys the contract
A calendar average hides distributions, confidence, liquidity, question wording and resolution rules. The method is useful for visualization, not for a capital decision rule.
Beliefs and incentives are entangled
Lab leaders can see private capability evidence but also raise capital, sell compute or shape policy. Markets add price discipline, but their contracts can resolve on rhetoric rather than capability.
03 / ONTOLOGY
The six clocks we will track
Capability
How broad and deep is performance on unfamiliar tasks?
Performance × generality, with contamination-resistant task distributions and disclosed failure modes.Reliable autonomy
For how long can the system complete useful work without rescue?
Accepted-task rate, intervention count, horizon, retries, cost and reproducibility.Declaration and contract
Has a named institution declared a precisely defined threshold?
Exact speaker, wording, definition, verifier, date and resolution source.Deployment and diffusion
Can organizations repeatedly deploy it in real workflows?
Paid usage, retention, cycle-time gain, switching cost, safety and regulatory acceptance.Cash-flow capture
Who converts the capability into durable owner economics?
Incremental revenue, gross profit per useful task, capex, dilution and free-cash conversion.Market pricing
What future is already embedded in securities and forecast contracts?
Contract probability and liquidity, reverse-DCF expectations, factor crowding and valuation dispersion.04 / POINT-IN-TIME
A frozen view—not live truth
50% aggregate HLMI forecast
1,714 answers; feasibility of unaided machines doing every task better and more cheaply than humans.
open source ↗50% aggregate full-labor-automation forecast
The same survey family moves 69 years when the endpoint changes—proof that wording is causal data.
open source ↗Metaculus: weak general AI
A devised, tested and publicly announced system; 1.8k forecasters.
open source ↗Metaculus: general AI
A separate, stricter public-announcement question; 1.9k forecasters.
open source ↗Kalshi: OpenAI AGI before 2030
Live contract probability with more than $1.0M reported volume; contract-specific, not universal AGI odds.
open source ↗Polymarket: OpenAI announces AGI before 2027
Resolves on an official announcement, not an independent capability test. The contract is therefore a rhetoric-and-governance clock.
open source ↗05 / ALPHA TRANSLATION
What this means for AGI capital
- 01
A shorter capability timeline can raise compute and power demand before it improves downstream software margins. “Sooner AGI” is not uniformly bullish across the stack.
- 02
Fast capability commoditization can expand usage while destroying model-layer pricing power. Volume growth and owner economics must be measured separately.
- 03
Prediction-market movement is an expectations receipt. It becomes investment-relevant only when paired with a capability, deployment or cash-flow receipt and compared with what the security price already assumes.
- 04
Our March 2027 forecast remains a pre-registered internal hypothesis. This review neither upgrades it to consensus nor changes its six gates.
- 05
The exploitable object is not a date. It is the revision chain: which clock moved first, which causal lane confirmed it, and which listed owner remained mispriced after the evidence changed.
06 / FORWARD RECORD
Tracking contract
Freeze contract snapshots
Probability, volume/liquidity, exact wording, resolution rules, source and retrieval timestamp. Never overwrite an older vintage.
Pair expectations with reality
Update capability, autonomous horizon, cost per accepted task, deployment and cash-capture receipts on their own clocks.
Audit contract comparability
Version definitions, eligible sources, settlement authority and platform changes. Split a series when the contract changes.
Refresh slow cohorts
Add survey vintages without merging HLMI, FAOL, singularity and public-announcement questions.
Record rhetoric separately
Store exact attributable claims and incentives, but require external evidence before they alter a capability clock.
07 / PRE-COMMITMENT
When the system may open a review
Market repricing
A move on the same contract, with a non-declining liquidity/volume receipt, opens EXPECTATIONS_REVIEW.
Timeline revision
A comparable median shift in the same frozen question opens FORECAST_REVIEW.
Capability confirmation
Two source-independent capability or autonomy receipts are required before a calendar move can alter a research prior.
Capital escalation
Capability/deployment evidence + owner-economics evidence + a point-in-time valuation gap may open issuer review. Human approval remains mandatory.
08 / SOURCE CONTROL
Evidence ledger
Useful source map and chronology; weighted-year aggregation is not treated as one comparable consensus distribution.
aimultiple.com ↗PRIMARY · RESEARCH FRAMEWORKGoogle DeepMind et al. · Levels of AGISeparates performance, generality and autonomy—the ontology behind our clock split.
arxiv.org ↗PRIMARY · SURVEY RECORDAI Impacts · 2023 Expert Survey on Progress in AIProvides definitions, response counts, aggregation method and the 2047 HLMI versus 2116 FAOL contrast.
wiki.aiimpacts.org ↗LIVE · COMMUNITY FORECASTMetaculus · Weakly General AIPoint-in-time reading and exact weak-AGI public-announcement question.
www.metaculus.com ↗LIVE · COMMUNITY FORECASTMetaculus · General AISeparate general-AI question and current community estimate.
www.metaculus.com ↗LIVE · REGULATED PREDICTION MARKETKalshi · When will OpenAI achieve AGI?Contract probabilities and reported volume frozen at the evidence cutoff.
kalshi.com ↗LIVE · PREDICTION MARKETPolymarket · OpenAI announces AGI before 2027Resolves on an official announcement; useful only as a declaration clock.
polymarket.com ↗LIVE · COMMUNITY MARKETManifold · AGI When? High-quality Turing testIts draft, updateable Turing-test criteria demonstrate why contract text must be versioned.
manifold.markets ↗