“The recent plateau in large language model performance, observed since GPT-4, is largely attributable to exhausting the available high-quality data from the internet.”