📄 中文摘要
在早期使用GPT-2和GPT-3时，生成的文本看似流畅且自信，能够在表面上满足提问的需求，但实际上并不基于真实信息，而是通过组合可能的响应来构建的。这种现象被称为“完美的胡说八道”。尽管如今多代理系统的输出更为精准，依然存在类似的特性，只是变得不那么明显。作者提到，早在2006年，彼得·瓦茨就已对这一现象进行了诊断，但当时并未意识到其重要性。
大型语言模型并非缺陷心智
出处: An LLM Is Not a Deficient Mind
发布: 2026年3月13日
📄 English Summary
An LLM Is Not a Deficient Mind
Early experiences with GPT-2 and GPT-3 revealed that the generated text appeared fluent and confident, seemingly providing answers without being grounded in reality. This phenomenon, termed 'the perfect bullshitter,' highlighted the model's ability to produce probable responses by assembling tokens that fit expected text patterns. Although current multi-agent systems yield sharper outputs, the underlying issue persists, albeit less visibly. The author references a diagnosis by Peter Watts from 2006 that addressed this phenomenon, which went unrecognized at the time.