A lot of the codes are best represented by graphs where some of the nodes on the graph are controllable and some are not.

Make sure

• Karas layers can take dictionary input
• TFPL convert to tensor is non-idempotent

To do

• Separate sensors and actuators from controllable and uncontrollable nouns. The relationship is possession versus subclass
• Remove probabilistic components from the prediction node
• Neighbor influence works on the current timestep while parent targeting works on the next timestep
• Add energy management network of organs and makes sensors and actuators kinds of organs. Since they only have but not are nodes, they can reuse code when subclassing different sensors for 3-D world versus direct unity
• Make the f_act optional

The neural networks should employ sparse activation. Maybe even the nodes are sparsely activated

Vision

zu clip embedding

zc eye movements

Combined self and other external dialogue, inner voice, Internet query responses, and text prompt in the text modality:

zc next token

zu T5, blender, gpt-2, longformer

I will have to pre-train the latent translation functions from text to audio.

Audio: