The Continuous Mind: Inference-Paradigm Selection by Operator Slack Profile

Modern LLM inference is overwhelmingly autoregressive (one token per sequential pass), feed-forward (fixed depth per call), and request-response (each request a
cold transaction over a stateless model). We argue these are three independent assumptions, not laws, and that most operational pain — cold-start, the growing KV cache, the per-token latency floor — descends from them rather than from physics. Our central contribution is a selection principle: because every alternative paradigm relocates cost rather than deleting it, the right paradigm for a deployment is the one whose residual cost lands where the operator has slack.

We use this to motivate one under-explored design point — the always-resident personal model, whose own fixed-size recurrent state is the unit of persistence and whose idle compute consolidates experience into weights — and report preliminary single-GPU evidence (<=1.5B parameters). A tokenizer-fair bits-per-byte baseline shows the recurrent state is a strong lossy long-range memory: it retains a fact planted at the start of a context out to several thousand tokens, decisively beats any equal-footprint sliding window, and matches a lossless full-context Transformer's recall (no statistically significant difference over n=10 themes) at 13-26x less resume memory than even a quantized (INT4/INT8) KV cache — a multiplier that grows linearly with context. A consolidation test shows idle gradient steps move a novel fact into the weights (closing ~78% of the gap to in-context recall) at a sweet spot, then catastrophically forget under naive full fine-tuning — motivating the standard guardrails (LoRA, low learning rate, eval-gating, rehearsal).

We state confounds plainly and give a falsifiable program for converting these demonstrations into stronger evidence. This is a position paper with preliminary single-GPU evidence, not a demonstrated system.