https://geohot.github.io//blog/jekyll/update/2021/10/29/an-architecture-for-life.html
A representation function takes in the sensor data and transforms the manifold to something more learnable, referred to as the “hidden state”. It may discard non task relevant information. This can be a VAE, GAN, pretrained ImageNet, custom trained conv net, or in the case of GPT-3, handcoded as BPE.
A dynamics function deals with long term temporal correspondences. It functions as both a summarizer of the past and a predictor of the future, in the same way a Kalman filter does. This can be a couple dense layers, a transformer, or an RNN/GRU/LSTM. I guess it even could be a Kalman filter itself.
A prediction function tells you how to act in a state. It can do this by directly outputting an action or by outputting a value of this state allowing you to search.