Before it
Scaling meant training compute: more parameters, more data.
The problem
Nobody had systematically compared spending compute at training time against spending it at inference.
What it introduced
A compute-optimal analysis of inference strategies, showing extra thinking can beat a much larger model on some problems.
Why it matters
It gave the reasoning-model era its theoretical footing, weeks before o1 shipped.
What modern models inherit
Inference compute as a second scaling axis.
What came next
o1 and o3, then DeepSeek-R1 and the open reasoning models.