Test-Time Compute Dominance
By 2027, test-time compute scaling (thinking longer) will contribute more to capability gains than pre-training scaling (larger models).
70% confidence
Shareable summaryInfinite Architects predicted test-time compute would overtake pre-training. The o1/o3 paradigm shift suggests this is happening.
Falsification criteria
- If by 2027, pre-training scaling still dominates capability improvements, this is falsified
- If test-time compute shows diminishing returns below pre-training scaling, this is falsified
- Measurement: Compare capability per dollar spent on training vs inference
Supporting evidence
-
OpenAI shifts focus to inference-time scaling with o1/o3
2024-09-12
Industry leader pivoting to test-time compute
Timeline
- Made public
- Expected resolution
Checkpoints
- 2025-12-01 Mid-term check on industry direction
- 2026-12-01 Pre-resolution assessment