I assume this is indeed talking of those "smarter models with fewer parameters". Whether that's really happening is another question.