Can lightweight adapters replace millions of personalized models?

The standard story about LoRA is economic: it is a cheaper substitute for full fine-tuning. This paper proposes a different role — PEFT as persistent local state layered on a strong shared base. The base supplies general competence; the adapter carries the learned consequences of repeated experience with one user: preferences, skills, tool habits, memory-like updates. The provocative phrasing "million personal models of trillion parameters" is explicitly not millions of owned checkpoints; it is a few trillion-scale bases plus millions of lightweight adapters serving as durable behavioral deltas. The thesis only holds if three scaling axes reinforce at once — Scale Up (a stronger base makes small updates more useful), Scale Down (how small the adaptive state can get while still learning reliably), and Scale Out (turning repeated updates into served populations). Remove any axis and it collapses.

The framing matters because the vault's personalization thread keeps hitting the wall this paper names: prompts, retrieval, and profiles "help but are not enough" because they do not persist and reshape future behavior. Why does chain-of-thought reasoning fail for personalization? is the cautionary detail — naive personalization fine-tuning destroys generalist capability, which is exactly why the adapter-as-bounded-state framing (not replacing the base, not storing the whole person) is the safer design. Can models dynamically activate expert skills at inference time? is the composability precedent the population vision needs, and it warns that LoRA adapters interfere when composed where orthogonal SVF vectors do not — a real obstacle to "compose at population scale."

The Scale Up axis quietly rests on How should finetuning scale with model and data size?: if fine-tuning gains track base-model scale, then bigger shared bases do make each tiny adapter more useful. The honest doubt is whether per-user adapters drift, stale, or leak across the population once served at scale — persistence is a liability as much as a feature, and the paper's bounded framing ("does not store the whole person, does not replace retrieval") reads partly as a hedge against that.