Fable
The manuscript restorer slowly brings sentences out of a page of noise
The old scribe at the publishing house always wrote articles from the first character to the last. If he made a mistake early on, the rest often followed crooked, until the whole piece looked like a road gone astray. The house master heard of a new revision method: first make the whole page look like a mass of noise, then erase the noise step by step, letting the sentences emerge through repeated correction. The apprentices didn't understand at first. Why not just write from left to right? The reviser said that some tasks aren't about what word comes most naturally next, but about making the entire structure coherent at once. The first experiment was summarizing corporate reports. Once the old scribe set the framing in the first half, it was hard to go back and adjust the second half; the new reviser could make multiple global revisions, keeping the title, conclusion, and evidence more consistent.
But the new method has its own troubles. Every denoising step requires computation, and controlling the process, integrating tools, and ensuring facts aren't corrupted are all harder than traditional continuation.
Craftsmen first used it as an ordinary chat device, and the results were unstable. It isn't suited to the same rhythm; the interface, sampling, and evaluation methods all need to be redesigned.
Later, the team put it into tasks suited for global rewriting, parallel generation, and constrained editing, rather than forcing it to imitate the old scribe's word-for-word continuation.
The workshop owner realized that this was not a replacement for “being better at chatting,” but a different path for language generation. It could change latency structure, inference cost, and controllable editing methods.
The publishing house ultimately kept two types of tools: one good at sequential continuation, one good at gradually reconstructing the whole from chaos. Different tasks call for different generation mechanisms.
The new reviser is best at handling drafts where the opening and ending are tied to each other. If the title changes, the supporting evidence in the body must change too; if the conclusion narrows, the earlier setup must tighten as well. The verbatim scribe often only realizes at the end that the beginning should have been changed.
The workshop owner later put up signs for two kinds of work: urgent replies go to the sequential clerk, and overall polishing, gap-filling and constrained rewriting go to the reviser. Both can write, but they are not the same path; mistake the path, and the workshop will use the wrong time, the wrong budget and the wrong acceptance method.