Forecasts
Dated forecasts on AI, risk and markets
Each question carries a deadline, a probability and a confidence rating that reflects the strength of the evidence behind it. Probabilities change only on verified evidence, and every question is scored when it resolves.
- Open questions
- Due in 12 months
- Expected to happen
- Resolved
- Confidence
ForecastsRecursive AI R&D
Will METR publish a follow-up assessment estimating at least 3x acceleration of AI capabilities progress at any frontier lab by 2027-12-31?
Laurenceo estimate: 18% (Unlikely) by December 31, 2027. Last updated November 7, 2026.
How it resolves
Resolves yes if a METR publication by the deadline gives a point estimate of 3x or higher for acceleration of AI capability or R&D progress at any frontier lab, as shown on metr.org.
Reasoning
Set at 18% because the first estimate is 1.9x with stated uncertainty, and a jump to 3x in a follow-up needs strong compounding and a repeat evaluation.
Evidence
Revision log
- November 7, 2026: set at 18%. Set at 18% because the first estimate is 1.9x with stated uncertainty, and a jump to 3x in a follow-up needs strong compounding and a repeat evaluation.
Every question
AGI Observatory
When AGI arrives, and who is leading
AGI here means AI that can do the work of an AI researcher, end to end. We do not guess the date ourselves: we combine the leading published forecasts with a measured trend, by a fixed rule, every morning.
Our estimate
What the estimate is built on
Each bar runs from today to the date a source expects. Our estimate is the middle of the counted sources.
How the estimate has moved
What changed, and when
The forecasts behind it
Counted sources set the estimate. The others are shown for context.
Test the estimate yourself
Switch sources on or off, including the ones the rule leaves out, and see where the middle lands. Same rule as ours: the median of the sources you choose, with the middle half as the likely range when four or more are chosen.
Your estimate
The published estimate is this raw median passed through fixed guardrails that limit how fast it can move in a day, so the two can differ for a while after a big change.
Who holds the frontier records
The best published result on four tests of frontier AI, and who set it.
| Test | Best result | Held by |
|---|
Companies and countries
Records each holds now, and records set in the last 12 months.
United States and China, best result on each test
| Test | United States | China |
|---|
Timeline
Milestones, records, releases and forecast changes, newest first.
Steps to watch
Dated questions on the forecast board that would show AGI getting closer, scored when they resolve.
How the observatory works
Revisions
How expectations changed
Every change is a move from one probability to another, with the reason and the evidence. Nothing is overwritten.
Track record
Are the probabilities any good?
If Laurenceo says 70%, it should be right about 70% of the time, and sharper than a coin flip.
Scores
Reliability
Where the open questions sit
A board with everything near 50% would say little; one with nothing between 5% and 30% would be overconfident.
Method
How the board works
- Probability
- How likely the outcome is by its deadline, from 0.5% to 99.5%. In words: Very unlikely (10% or less), Unlikely (11 to 35%), Toss-up (36 to 65%), Likely (66 to 90%), Very likely (over 90%).
- Confidence
- How much to trust the probability, computed only from the evidence, never from the probability. Points for sourced facts (1 for one, 2 for two or three, 3 for four or more), 2 or more independent publishers (+1), a sourced fact from the last 180 days (+1), a measured trend with enough points (+1), that trend agreeing within 20 points (+1, or -1 when more than 35 apart), a deadline within a year (+1, or -1 beyond two years), more speculation than fact (-1). High at 6 or more, Medium at 4 or 5, Low at 3 or less. A probability of 10% or less, or over 90%, on Low confidence is flagged.
- Movement
- A change of 2 percentage points or more is an update and needs a verified fact and a written reason. Smaller changes are ignored. One day's move is capped.
- Evidence
- A fact is stated in a source the pipeline retrieved and checked, and links to it. Inference is Laurenceo reasoning from facts; speculation is flagged. Every question shows evidence against as well as for.
- Trend test
- Where a question depends on a measured series, four growth shapes are fitted and weighted by support. The fitted probability is a reference, not the forecast.
- Resolution and scoring
- An event resolves YES on a verified source, and NO after its deadline plus 14 days. Resolved questions are scored with Brier and log scores on a recency-weighted probability: every estimate the question held counts, the latest most, each earlier one at half the weight of the one after it. A daily ledger of every probability is stored and cannot be overwritten.