Oliver Sourbut unveils mechanistic theory behind exponential AI time horizons
The METR graph's exponential trend may be driven by compound subtask success probabilities.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Oliver Sourbut proposes a mechanistic explanation for the famous METR graph showing AI time horizons increasing over time. He argues that “time horizon” actually measures the number of subtasks a human expert would need, not AI’s clock time, and notes a model where the chance of failure compounds with the number of subtasks fits the data. The article draws an analogy to Moore’s Law being better explained by Wright’s Law, and aims to find a similar deeper mechanism for the METR graph.
- METR's 'time horizon' measures the number of distinct subtasks a human expert would need to complete, not AI execution time.
- Exponential growth in task length is driven by compounding improvements in per-subtask success probability.
- Mechanistic understanding enables better prediction of trend breaks and identification of key levers (compute, data, training methods).
Why It Matters
A mechanistic theory for AI progress helps investors and engineers anticipate capability jumps and plan for potential ceilings.