6 ideas
Perception is not a bottom-up build of the world from sensory data. The brain keeps generating top-down guesses about the causes of its inputs, and the senses mainly act to check and correct those guesses. What you experience is the brain's best current hypothesis, held in check by incoming signals.
In the predictive hierarchy, each level sends predictions down to the level below. Only the mismatch, the prediction error, is passed upward. So the forward flow of sensory information is mostly news about what was not expected, and well-predicted input is quietly explained away.
Attention is the brain's estimate of how reliable a given prediction error signal is. It works by turning up the gain on errors judged trustworthy and turning down the gain on noisy ones. This precision-weighting decides whether the brain revises its model or dismisses a surprise as noise.
Prediction error can be reduced in two ways: change the model to fit the world (perception), or change the world to fit the model (action). Motor commands are recast as proprioceptive predictions, and the body moves to make those predictions come true. Perception and action are therefore one error-minimizing process.
Many perceptual and psychiatric oddities can be read as errors in how much weight the brain gives its priors versus its sensory evidence. Overweighted priors can produce hallucination-like experiences or functional symptoms.
Predictive brains do not build rich, complete inner models when cheaper options work. They lean on bodily dynamics, environmental structure and tools, learning frugal predictions that offload work onto the world. This makes predictive processing compatible with embodied and extended cognition rather than a return to a detached, representation-heavy mind.