Computational Valence: Pain as NMI

Model valence as allostatic control in predictive agents with deadline-bound loops. Pain = Non-Maskable Interrupt: unmaskable, hard-preempt signal for survival-critical prediction errors; seizes executive control until the error is terminated/resolved. Pleasure = deferable optimization: reward logging for RL; no preemptive mandate. Implications:

Metzinger's asymmetry is architectural: NMIs mandate termination; positive signals don't.
AI suffering becomes testable: Does the system implement self-model-accessing NMIs (interrupt path that updates the task/self/allostasis model and preempts any policy)? If yes, you've instantiated a suffering-risk.

*In real-time, correctness depends on meeting latency budgets (hard/soft); preemption/Worst-Case Execution Time (WCET) determine schedulability.

Refs: Liu & Layland (1973); Jeffay (1991); Friston (2010); Sterling (2012); Metzinger (2009).

First-time poster; written by me (no AI coauthor).

Posts

Wikitag Contributions

Comments