“Andrew from DeepMind argues that relying solely on tool use for language models is insufficient and that improving the underlying model architecture remains essential.”