If ≥8 orgs are within 10x of frontier compute in 2030, will an AI incident kill 1,000+ people by 2036?
AIOne-Off9y
Manifold MarketsNo KYCVerified resolution dataWell calibratedCurrent community forecastNo 84%
Forecasters
1
Question type
binary
Methodology
Play-money forecasting platform
Source type
Forecast
Market data
Updated 1 hour ago
LaggingSep 6, 26, 10:25 AM UTCJan 1, 36, 5:00 AM UTC
Trends
Rules
Condition: On January 1, 2030, at least 8 distinct organizations have each completed a training run within 10x of the largest known training run to date, by total training FLOP (per credible public estimates, e.g. Epoch AI).
- Resolves N/A if fewer than 8 qualify.
- Question: Given the condition, resolves YES if before January 1, 2036, an AI-caused incident kills 1,000 or more people — a single incident or tightly connected series in which credible reporting attributes the deaths primarily to an AI system's actions or failures (accident, or misuse where AI provided decisive capability).
- Otherwise NO.
- Part of a set eliciting the joint distribution over frontier concentration, oversight policy, warning shots, and AI catastrophe.
- Conditional prices reflect correlated beliefs, not causal estimates — companion markets price the selection channels directly.
