Tweet

The human nervous system holds context via micro-patterns of tension, which stabilize select circuits for the duration of a task. Context can’t drift so long as the tension is held, and our bodies are really good at holding tension (maybe too good sometimes)

Modern LLMs don’t have this capacity: there’s no clean mechanism for defining a task by holding the state of select feature detectors invariant, so over time representations drift and all sorts of weird artifacts leak out. AIs struggle to hold context because they can’t look inside themselves and choose what to hold constant — things just flow, and given enough time things will drift

I suspect vasocomputation could lead to novel AI architectures which could hold context better. It’s also possible these architectures would be easier to align, since you could use this capacity to define goals for a system across various levels of representation — either in post-training or at runtime

I also suspect this could lead to a redefinition of how AI architectures define tasks, which might help on the specifics mentioned below