“Full interpretability of every neuron in a model like Gemini may not be necessary as long as the model's overall output and characteristics are good.”