Increasing AI capability by scaling up parameters is potentially less problematic than using architectures that obscure the model's reasoning process