ML//emergent behavior
- Capabilities that appear suddenly at scale — arithmetic, chain-of-thought reasoning, translation to unseen languages.
Capabilities that appear suddenly at scale — arithmetic, chain-of-thought reasoning, translation to unseen languages.
Not present in small models, then "switch on" past a compute threshold.
Controversial: Wei et al. (2022) catalogued them, but Schaeffer et al. (2023) argued it's a mirage of metric choice.
The subjective experience of watching a model suddenly "get" something is hard to dismiss.