Neurotypica

Neural pathways are the brain's learned mappings from perception to action. They're also the substrate of prediction: the more often a particular input has led to a particular outcome, the stronger the pathway connecting them, and the more confidently the brain predicts that outcome will happen again. This is why familiar contexts make predictions cheap and fast---the pathways are well-worn, and the model is confident. But it's also why changing behaviour is hard: the prediction is baked into the infrastructure.

To update predictions, you need to retrain the pathways in context. Practice the new prediction until it's the confident one.

Key takeaway: strong pathways make predictions fast and sticky; change requires building alternative routes.

Circuits are reused across many different mappings, which means predictions about one thing can interfere with predictions about another. When two routines share infrastructure, the prediction errors in one can spill over into the other. This is why practicing a new routine can jostle old patterns---you're updating shared circuitry, and the model has to reconcile competing predictions.

The practical move is to separate contexts as much as possible when retraining, so the old and new predictions don't fight for the same neural real estate.

Key takeaway: circuit reuse means prediction updates can have spillover effects; design contexts to minimise interference.

The more you run a prediction, the more stable it becomes. Repetition consolidates the pathway, building infrastructure around it that makes it faster but harder to change. This trade-off is central to how prediction engines work: you want confident, stable predictions for familiar contexts (speed and efficiency), but that stability makes it costly to update when the context changes.

Adolescents, with less white matter infrastructure, have more plastic prediction systems---they update faster, but they're also less efficient. Adults have more stable predictions---efficient but inertial.

Key takeaway: stability speeds up familiar predictions but makes updating costly; plan for the trade-off.

Task sets configure which predictions are active right now. When you shift into 'making tea' mode, the brain loads a set of expectations about what you'll see, what you'll do, and what outcomes matter. That set biases which predictions run and which sensory signals get attention.

If the wrong task set is loaded at the cue, the brain generates the wrong predictions, and you get prediction errors even when nothing surprising is actually happening. Preload the right set to align predictions with context.

Key takeaway: task sets gate which predictions run; preload the intended set at the cue.

Predictions are layered: high-level goals generate predictions about subgoals, which generate predictions about actions, which generate predictions about sensory outcomes. Errors at any level can propagate up or down. A violated high-level prediction (the café is closed) cascades down into action updates (go somewhere else). A violated low-level prediction (this coffee tastes wrong) can bubble up into goal revision (maybe I don't want coffee after all).

The hierarchy lets you resolve errors at the right level: sometimes you fix the action, sometimes you revise the subgoal, sometimes you change the whole plan.

Key takeaway: errors propagate through the hierarchy; adjust at the level that makes most sense.

Neuromodulators tune how the brain responds to prediction errors. Dopamine, for instance, seems to signal the magnitude of prediction error---how surprising the outcome was relative to expectation. Norepinephrine and acetylcholine tune which errors get attention and how much the model updates in response.

This is why context and state matter so much for learning and behaviour. High arousal amplifies prediction errors; calm states allow subtler updates.

Key takeaway: neuromodulation tunes how prediction errors drive learning and action; state shapes updating.

Predictions aren't just about the external world; they're also about the body. The brain predicts your heart rate, your hunger, your tension, and when those predictions fail, you feel it as affect. That affect colours how you interpret external prediction errors: the same surprise can feel threatening or exciting depending on your bodily state.

This is Mandler's insight: emotions are interrupted predictions about bodily states, and those interruptions shape how you respond to the world. Regulate the body, and you regulate how prediction errors feel and what actions they trigger.

Key takeaway: bodily prediction errors generate affect; state management is prediction management.

Predictions depend on context. The brain uses environmental cues to decide which model to load---which set of predictions is relevant right now. Walk into your kitchen, and the brain predicts you're about to make food. Walk into your bedroom, and it predicts you're about to sleep or change clothes.

If the cues are wrong or ambiguous, the brain loads the wrong predictions, and you get errors even when your behaviour is fine. Build clear, consistent cues into contexts where you want particular predictions to run.

Key takeaway: context cues which predictions load; design cues to align predictions with intended behaviour.

Memory isn't playback; it's reconstruction guided by prediction. When you recall something, the brain generates a prediction about what should have happened based on current schemas, and fills in gaps with plausible details. This is why memory feels vivid even when it's wrong---the prediction is confident, even if it's not accurate.

This also explains confabulation and false memories: the brain's prediction engine produces a coherent story, and you experience that story as memory. The lesson is that what you 'remember' is shaped by your current predictions about what makes sense.

Key takeaway: memory is reconstructed via prediction; current models shape what you recall.