Diffusion ASI // MOSAIC LABS

4.1 · Context-state direct editing

File pages in context — instead of reading a file, editing text, and writing it back, the model holds sliding windows over the file inside its context as directly editable pages. The edit and the file are the same object.
Reverse self-prompting — the model can freeze the tail of its context, unfreeze earlier memory pages, and prompt itself backwards, revising the setup in light of the conclusion. Autoregression only goes forward; diffusion editing does not have to.
2D intelligence ascension — mode collapse is easier to escape when the mutation space is 2D. There are simply more directions to move, so the model gets stuck less and explores more.
Infinite edit pathways — because any token can change at any step, the number of routes from a broken state to a working one grows combinatorially. Hard bugs have more ways to be fixed.

Internal paging policy — the model learns to manage its open pages the way a developer manages editor tabs: load what is relevant, close what is not, jump between them. Navigation becomes a learned skill, not a fixed loop.
State / memory management — it keeps virtual scratchpads for working notes and pulls recurring patterns out of the files it has loaded, building a working memory that outlives any single edit.
Dynamic window sizing — each page window can grow to pull in more surrounding code or shrink to force focus, adjusted per step as the task demands.
Contextual awareness — because the whole state is visible at once, the model decides what needs changing from global understanding rather than local pattern-matching on the current line.

Collapsed roundtrip — edit, apply and validate happen in one forward pass instead of three separate model calls with tool round-trips between them.
Direct state mutation — no diff or patch is generated as an intermediate artifact; the state is changed in place, so there is nothing to mis-apply.
Img2img of language — the operation is image-to-image translation applied to text: take a state, condition on an instruction, emit the mutated state, with LLM-level semantics as the guidance signal.
Vertical timeline impact — a model that can edit-apply-validate in one pass removes the slowest loop in software work. Applied at the frontier, that is the kind of step change that bends a capability curve upward.