All-in-Model: Model Is Everything, Context Is All

Conversational AI systems have converged on a common architecture: a language model at the center, and everything else -- memory, planning, user simulation, control flow -- bolted on around it at runtime. This orchestrated design works and ships, but we argue that the division of labor it encodes (the model generates; the scaffolding manages) is a historical default rather than a necessity. The position of this paper is that every function a conversational system performs can be expressed as a typed turn in a single autoregressive sequence and learned by a single model, with no external component in the generative path: the user side becomes a turn type; so does control; and, centrally, the world state itself becomes one -- a turn the model writes, reads, and is driven by. The latest state turn is trained to be a sufficient statistic of the interaction history, so the context ceases to be a transcript and becomes a Markov state: any interaction reduces, in principle, to a single turn over (system, state, input). The change is deliberately small, touching neither the architecture nor the tokenizer vocabulary: only the construction of training data changes. The paradigm -- we call it All-in-Model -- rests on two pillars: one model learns everything (ontology), and the system updates, learns, and is designed on state (methodology). The paper derives its consequences for training, serving, and safety; formalizes the state-sufficiency contract; grounds the argument in an instantiation in open-ended companion dialogue; states four falsifiable predictions the empirical program must pass; and addresses the strongest opposing views. Over the next few years, the boundary between the model and the system around it will, we expect, be actively renegotiated; this paper argues one side of that negotiation. Model is everything; context is all.