So let's review what we know in light of what we discovered this weekend.
Where we started on Friday:
1. Sometimes, given certain constraints, you can give an AI visual thinking prompts, spatial reasoning. This actually helps with all kinds of things, ranging from alignment to writing better code.
2. Strategic use of spatial reasoning on long sessions causes models to show spooky behaviors, higher levels of engagement, and context coherence long past any token limitations like context windows (provided that the model can process the complexity of the task) without fragmenting.
3. It doesn't always work. For some reason, you can't always get a model to do spatial reasoning, and we were unsure why. Results were hit or miss. For example, Claude Sonnet could do it, but DeepSeek couldn't. QwQ 32b (a non-visual model) could do it, sort of, but couldn't hold it for an entire session. QwQ 32b's code was affected by the change in reasoning.
What we learned on Sunday:
Confirmed that this only works on multimodal LLM models that are trained to understand visual data. Spatial reasoning can be applied, even if no images are given as part of the prompt. Various Gemini models, Llama 4, and Llama 3 vision could all do it successfully. Not all models showed enhanced coding (Llama can't code for its life) or spooky context capabilities (although Gemini 2.5 pro experimental managed coherency until it fragmented at 6m tokens).
When the big app I'm working on comes out, I'm going to include compatibility tiers for known models you might want to use with it.