Summary
OpenAI o1 (2024) and successors marked a second scaling axis: instead of scaling training compute, scale inference compute by letting the model 'think' — generate long chains of internal reasoning before answering. The result: order-of-magnitude gains on math, code, and scientific reasoning.
Objectives
- 01Explain the difference between training-time and inference-time scaling.
- 02State what reasoning models actually do internally.
- 03Name the domains where reasoning models most clearly outperform.
The Lesson
The idea
Train the model with RL to produce long internal reasoning traces before its final answer. Trade inference latency for reasoning depth. Human oversight becomes harder — the trace can be long — but capability increases.
The shift
The paradigm moves from 'better model' to 'let the model think longer.' o1, o3, DeepSeek R1, Claude 3.7 Sonnet with extended thinking all follow this template.
Key Ideas
- Inference-time compute is a second scaling axis.
- Reasoning traces trade latency for capability.