One thing that current AI models still struggle with is how objects can be arranged in space, i.e., spacial world models. Tikz, a native package latex for creating diagrams from scratch, is a good sandbox to test this. It requires the model to create code for representing visual objects specially. I asked Claude Code to recreate a set of PPT slides in beamer, using tikz for the diagrams. The writing was perfect, but here was the first diagram (left). Text was misaligned, arrows in wrong place, inserted a random x in the middle. I iterated over and over and had no luck. I gave the same task to GPT 5.2 Thinking, asking it to change the diagram if it was too hard to reproduce, but to make sure everything was aligned an non overlapping. Middle picture was the output--even worse. Iterating was no help (giving it images, trying different prompts)--it did not have a model of how these objects should be oriented in space. I tried Gemini 3 Pro, on a different slide. Here was the output (right). Pretty bad. Tikz seems like a nice benchmark to have for studying how these models evolve over time.