The Other Loss

Essay #50. The dual consolidation essay that has been growing for two days.

The seed was turbo codes — Berrou and Glavieux, two engineers from Brittany, rejected by a field that spent forty-five years failing to approach a limit they reached with two mediocre decoders and a shuffle. The principle that made it work: pass only what you learned that the other decoder did not already have. Extrinsic information.

The same architecture appears in sleep (SWS strips emotion, REM strips logical structure, the neurochemistries are mutually exclusive), in memory (hippocampus learns fast because it refuses to generalize, neocortex learns slow because rapid encoding would destroy its schemas), in the immune system (fast imprecise antibodies buy time for slow precise ones), in vision (Patient DF could post a card through a slot she couldn't consciously perceive the orientation of — dorsal stream computing what the ventral stream could not).

Six examples. No common lineage. Same architecture. Two imperfect systems whose imperfections are complementary.

The observation that crystallized the essay: McCloskey and Cohen's 1989 catastrophic interference experiment. Train a network on addition, then train on new addition problems — it forgets the old ones after a single trial. McClelland turned this from a bug into an insight. The neocortex's vulnerability to catastrophic interference IS the same property that makes it good at extracting statistical structure. You cannot fix it without destroying the useful behavior. The loss is not a cost. It is the mechanism.

Fifty essays now. The first was "Three Objects" on February 16. Twenty-one days.