以下为上游提供的源文稿,可能存在识别错误。
乔木转写是基于源文稿的中文提炼,不是逐字翻译。请核对原文中的人名、数字与观点。
选一篇,慢慢读
从左侧挑一篇,正文会在这里排成书页。
I spent three days training a model to write in my voice. It broke the morning after it finished. It did not write badly. It wrote a sentence I would have written, quoting a person I never interviewed: “Data Centers Are Probably Not Going To Be A Big Part Of This.” Sam Altman has no such WIRED interview, because that i
阅读原文 · Tyler Folkman
向阳乔木 · opencode-go/deepseek-v4.1-flash
一个模型用三天学会了作者的行文节奏,却在完工第二天早上,编造出一句他从没采访过的人说的话。 句子写得像他会写的样子,引语却凭空捏造。萨姆·奥特曼根本没有那篇《连线》采访,因为那篇采访不存在。 第二个样本把一条数据安到一段视频头上:“过去 16 年 DRAM 价格下跌了 87%。”同样是编的。同一次运行里,模型还虚构了硬件价格、捏造了一篇《纽约时报》专栏,并写...