“Achieving consistent performance improvements in AI models through scaling alone exhibits diminishing returns, requiring exponentially more compute for the same gain.”