> LLMs are awful at the spatial stuff
And some kid is going to come in, make an agent to play this, and accidentally figure out some clever trick to getting an LLM to understand spacial stuff!
This is exactly why "toys" are so critical, especially now.
Sam Earle's fractal neural network approach (https://arxiv.org/pdf/2002.03896) is exactly this kind of thing.
His key trick: recursive weight-sharing in fractal convolutional blocks, so each column of the network acts as a continuous-valued cellular automaton ticking different numbers of times. The deepest column gets a 33x33 receptive field -- enough to connect power across an entire 32x32 map in one forward pass.
The agents discovered power-plant + residential pairing, road placement for density, zone clustering by type, and traffic-avoiding road layouts. When stuck at local optima, a human player could intervene (deleting power plants) to force re-exploration -- and the agent would improve its design.
The paper was 2019, before LLMs were doing this kind of thing. Different paradigm (RL on tile grids vs. LLMs on coordinate text), same hard problem.