When Do LLMs Suddenly Get Good at Things?

In 2022, researchers defined an ability as "emergent" if smaller models lack it and larger models have it — meaning you couldn't have predicted it from the smaller models you already had. Since then, a fierce debate has unfolded over whether these jumps are a real property of the models or an artifact of the metrics we choose. This episode works through the strongest arguments on both sides, from the "mirage" paper that manufactured emergence on demand in vision tasks to the loss-perspective rebuttal that says the jump is real but indexed to pre-training loss rather than parameter count. It then turns to the linguistics thread: peer-reviewed work now largely agrees LLMs master formal grammar they were never explicitly taught, while functional competence — reasoning, world knowledge, social understanding — remains spotty. The uncomfortable open question is whether the syntactic structure found inside these models actually causes their behavior, or merely correlates with it. Episode #869955 — open it directly at myweirdprompts.com/869955