A prominent AI research lab has decided to allocate the majority of its budget to increasing the size of its next language model and the volume of its training data, rather than developing novel model architectures. Their primary justification is that unexpected, sophisticated capabilities have consistently appeared in previous models only after significant increases in scale. Critically evaluate this strategic decision. Is the observation of these new capabilities sufficient justification for a 'scaling-first' approach? Discuss the strengths and weaknesses of this line of reasoning.