Input is pixel, and some units are action triggers, gave it minecraft as a world, Spoiler: It's coherent.

My discoveries:

It remembers what being in the water and being outside water is, inside water bodies it continues holding the space stroke, allowing it to float,

It knows in water bodies you can't collide with blocks to jump, thus it holds the "space" stroke.

It has on a level, goal seeking behaviour, it focuses on reaching specific positions , first it turns it's head, it tries to go to specific co-ordinate, if there are blocks, it backs off (likely to look at surrounding) , then it attempts other route.

It learns pretty quickly, after only-once seeing the effect of holding the space key in a body of water, it can immediately do it again,

This whole thing is built on top of predicting the next state of the entire module, learning happens through modelling an asynchronous states that reperesent portions of the experience it encounters,

The error signal is the pred error, uses a normal recurrent network for exec..

https://twitter.com/rami_mmo/status/1773475064630894672