llmri maps self CKA in 3D

This is the kind of interpretability demo that looks like art first and tooling second, which is exactly why it might end up useful. The visuals are compelling, but the current evidence is still correlational and hypothesis-driven.

• –Self-CKA is a sensible lens for comparing representational structure across layers instead of just staring at activations.
• –The Q-head subspace choice is the sharpest move here because it tries to inspect structure before FFN mixing muddies it.
• –Cross-model comparisons are the real payoff: the project already shows family resemblance across GPT-style, RWKV, Mamba, and Qwen models.
• –The “circuits within circuits” framing is intriguing, but the blob shapes alone do not prove causal mechanistic circuits.
• –A hosted interactive viewer would matter more than another static screenshot; the method needs rotation and exploration to be persuasive.