Here is the translation:
This article discusses a study on the expected difficulty and output supervision of large language models (LLM). The research found that LLMs have problems in some aspects, such as accuracy, reliability, and consistency. Researchers acknowledged external criticism and stated they are improving their research methods, planning to expand data sets, reduce variability, and improve model optimization.
The article highlights several key points:
1. The scale of a large language model's parameters determines its intelligence level, but even with increased parameters, low-level problems persist.
2. Researchers recognize that their models may not be able to accurately answer all questions and perhaps lack a major breakthrough to fill the "gap" for true intelligence.
3. The article mentions the concept of "Scaling Law", which means the model's parameter scale determines its intelligence level, but this study shows that even with increased parameters, low-level problems persist.
In summary, this article discusses the problems of large language models in expected difficulty and output supervision, calling for further improvement and optimization. It also emphasizes the need for a major breakthrough to fill the "gap" for true intelligence.
Translation
这篇文章讨论了一项研究关于大型语言模型（Large Language Model，LLM）的预期难度和输出监督。这项研究发现了LMM在某些方面存在问题，如准确率、可靠性和一致性等。研究人员接受了外界的质疑，并表示他们正在改进自己的研究方法，并计划进一步扩大数据集、减少变异性和提高模型优化。
文章提到了几个关键点：
- 大型语言模型的参数规模会决定其智能程度，但即使参数规模增加，低级问题仍然存在。
- 研究人员意识到他们的模型可能无法完全准确地回答所有问题，也许还缺乏一个巨大的发现来填补达到真正智能的“沟壑”。
- 文中提到了“Scaling Law”的概念，这意味着模型的参数规模会决定其智能程度，但这项研究表明即使参数规模增加，低级问题仍然存在。
总的来说，这篇文章讨论了大型语言模型在预期难度和输出监督方面存在的问题，并呼吁进一步改进和优化。这也强调了需要一个巨大的发现来填补达到真正智能的“沟壑”。
Reference:
https://www.nature.com/articles/s41586-024-07930-y