OpenAI4 min readSample
Deliberation budgets scale past the point of diminishing returns
A frontier laboratory reports that allocating substantially more inference compute to hard problems continues to yield gains well beyond where the curve was expected to flatten.
criticalAGI relevance: criticalreasoningmodels
Agitology analysis
Why it matters
The second scaling axis — compute at inference rather than at training — was widely expected to saturate. If it does not, the reasoning dimension has more headroom than the Index currently prices in, and the accelerated scenario gains weight.
Key developments
- Gains persist across two further orders of magnitude of inference compute
- The effect concentrates on problems requiring long dependent chains
- Cost per solved problem rises faster than accuracy, bounding practical deployment
Index impact
How this development moved — or failed to move — the dimensions it bears on.