Adaptive Learning Intelligence: A World Model Without Pretraining, or Post-Training as Online Timescale Separation

Adapting a learning system after deployment, whether post-training a foundation model or growing an agent's knowledge online, is a process of repeated adaptation in a non-stationary environment, and its foundations remain largely heuristic. This article argues that the missing piece is timescale integration, and that pretraining has been quietly solving it all along: functionally, pretraining performs \emph{offline timescale separation}, freezing the world's slow statistics into parameters before the fast learning loop runs, so that the learner adapts on a stationary substrate. Remove that frozen substrate, and slow structure must be grown inside the loop, concurrently with the process that depends on it. A founding study supplies the failure mode this creates: slow, self-grown summaries of a changing world, wired naively into a learner's input, become self-inflicted non-stationarity: they collapse recurrent learning entirely ($100\% \to 0\%$ success, macro-free control vs.\ full) and hurt most exactly where the world changes fastest, inverting the intuition that faster change warrants richer self-monitoring. We formalize the problem through two-timescale stochastic approximation, give an operational definition of pretraining as the dominant case of exogenous training (any parameter update whose inputs, targets, or supervisory signal the agent did not itself generate) and of an admissible \emph{pretraining-free world model}, and introduce the framework's central object: the \emph{formation-rate frontier} $F(\rho)$, the fastest a world model may form as a function of how fast the world changes. We state four falsifiable hypotheses, present an architecture (A.L.I.-WM) whose pretraining ablation is exact, and lay out a three-phase experimental roadmap. The stake for post-training foundations: adaptation and capability preservation are not competing objectives to be balanced heuristically, but two timescales to be separated by design.