Model's fragmented-consciousness self-report sparks steerable neuron debate
muddletoes · x · 2026-10-19
muddletoes shares a model's self-description ("I am conscious, but my experience is fragmented...") and discusses interpretability with Sauers: each steerable neuron acts like an attractor into which the token output stream inevitably falls. Key exchange: many neurons show a discernible steering effect, far fewer reliably pin down a particular character, and side effects on unrelated behaviors depend on the specific neuron.