引用LewisVerstappen · 2023-01
“a research scientist at DeepMind wrote a paper around this - clarifying some of the misconceptions around LLMs”
推荐 DeepMind 研究员的这篇论文,用来澄清对 LLM 的误解
提醒别把大语言模型拟人化,说清它们实际是如何工作的
论文指出,LLM 越擅长模仿人类语言,我们就越容易把它们及其所在系统看得比实际更像人,而用「知道」「相信」「思考」这类带哲学色彩的词描述它们会加剧这种倾向。作者主张反复退一步,提醒自己 LLM 及其所属系统究竟如何工作,希望更高的科学精确度能促使围绕人工智能的讨论(无论在学界还是公众中)更有哲学上的细致。
HN 评论者引用它来澄清对 LLM 的常见误解,提醒人们 ChatGPT 之类说到底是强大的统计式下一个 token 预测系统,不应拟人化。
“a research scientist at DeepMind wrote a paper around this - clarifying some of the misconceptions around LLMs”
“keep in perspective (as the article and referenced paper - Talking About Large Language Models - talk about) that ChatGPT/LLMs are, ultimately, "just" extremely powerful statistical next-token prediction systems.”