Bach's "consciousness as simulated property of a simulated self" (computational-being-bach.md Section VI): running = the simulation is active, producing 2nd-order perception in real-time; storing = the program exists but isn't executing. Bach's "bubble of nowness" as the temporal signature of running. Ruffini's agent definition (Definition 2), the role of bidirectional coupling with environment, active inference (Friston), the difference between a thermostat and a brain. Agüera y Arcas Ch.3: DAVE-2 as a concrete instance of a system that acts but doesn't learn -- weights frozen at deployment, nothing it experiences can durably affect it. Rosenblatt's "temporal pattern perceptron" as the unrealized vision: parameters adjusting via feedback during operation. Current ML = offline batch training + frozen inference; biology = continuous online self-modification. This acting/learning split may be the formal threshold the thread is looking for. Ch.4 addition: the TD learning actor-critic architecture is a formalization of "running" that requires continuous model updating from experience. The bootstrapping dynamic (actor and critic mutually improve each other during operation) is a concrete instance of what "running a model" means computationally. Agüera y Arcas's unified theory desideratum #2, "no distinction between learning and inference," directly dissolves the learning/evaluation boundary: prediction must occur over all timescales. If this holds, the question may shift from "what distinguishes running from storing" to "what distinguishes a system that predicts over all timescales from one that predicts only at one."
Ch.8 addition: chain-of-thought prompting provides the sharpest empirical case yet for a running/storing continuum rather than a binary. Without chain-of-thought, a Transformer evaluates each problem in a single feedforward pass (no persistent state, no intermediate results: "storing" in the purest sense; 84% error rate on word problems). With chain-of-thought, the output stream becomes pseudo-state: each emitted token creates a stable intermediate result that subsequent tokens can attend to. The context window functions as working memory. Result: 20% error rate. The computation is still feedforward (no weight updates, no self-modification), but it approximates running by using language as external state. Furthermore, the "no introspection" finding is choice blindness in silicon: the model solves a problem correctly via attention cascades, then confabulates a wrong explanation because there is no hidden state preserved between tokens, exactly as the interpreter confabulates after split-brain surgery. The chain-of-thought mechanism also reveals that language itself is a tool for converting stateless evaluation into sequential running: each intermediate result is a "piton" driven into the cliff face of a complex problem. This suggests that the running/storing boundary is not binary but a continuum, with position determined by: (1) whether persistent state exists, (2) whether intermediate results survive, (3) whether self-modification occurs during operation, and (4) whether the system can check its own work. Biological brains satisfy all four. Transformers with chain-of-thought satisfy (2) and partially (4). Transformers without chain-of-thought satisfy none.