Before it
Sudden capability jumps with scale were widely accepted and used in safety arguments.
The problem
The jumps were measured with all-or-nothing metrics like exact match.
What it introduced
Show that the same runs look smooth under continuous metrics, so the discontinuity may be in the ruler, not the model.
Why it matters
A rare, effective correction to a dominant narrative, and a caution about how capability claims are measured.
What modern models inherit
Scepticism about benchmark discontinuities as evidence.
What came next
More careful evaluation methodology, and continuing disagreement about what emergence means.