作者驳斥"没人理解大语言模型"的普遍说法

一篇在Hacker News上发表的文章对"没有人理解大语言模型如何工作"的说法提出直接质疑1。作者声称"我理解大语言模型如何工作。数千名花费数年研究、设计、训练和实现LLM的人也都理解"1。这一立场直指MIT Technology Review、TIME和WIRED等主流媒体发表的"Nobody knows how AI works"类文章，认为这些报道将可解释性研究的局限性误读为人类知识的极限1。

作者通过对LLM运作机制的具体分析来支撑这一观点。大语言模型的完整工作流程包括将文本转换为token，然后进行矩阵乘法、注意力计算、归一化和非线性变换，最后逐层传递并输出token分布1。作者强调，这些操作中的每一个都可以被检验——通过访问模型和配备适当工具，研究人员可以记录每层的激活情况、检查注意头权重，并追踪数值计算的全过程1。

作者举例说明了当前正在进行的研究工作。Transformer Circuits研究项目专门调查神经网络内部的计算机制，并使用circuit tracing方法来识别特定模型行为背后的具体路径1。作者还以核技术作类比：核反应链并未被称为"没人理解"，尽管掌握其工程细节的人极少；AI同样是一个专门领域，但其基础科学实际上比大多数实验学科更容易获取1。

An author has challenged the widespread claim that nobody understands how large language models function, contending instead that LLMs are comprehensible mechanical systems whose operations can be fully inspected and traced 1. The author asserts that "thousands of people who have spent years researching, designing, training, and implementing LLMs" do in fact understand how they work 1. This position directly contradicts reporting from major publications including MIT Technology Review, TIME, and WIRED, which have published articles declaring that "nobody knows how AI works"—coverage the author criticizes as conflating the limitations of current interpretability research with the boundaries of human knowledge 1.

According to the author, the complete operational workflow of an LLM is transparent and verifiable 1. The process begins with text converted into tokens, followed by a sequence of mathematical operations—matrix multiplication, attention calculations, normalization, and nonlinear transformations—that propagate through successive layers before producing a distribution of output tokens 1. The author emphasizes that "every one of these operations can be inspected," noting that with access to a model and appropriate tools, researchers can log activations at each layer, examine attention head weights, and trace numerical computations 1. The Transformer Circuits research project exemplifies this capability, using circuit tracing methods to identify specific pathways responsible for particular model behaviors 1. Drawing an analogy to nuclear technology, the author suggests that while nuclear chain reactions are not dismissed as universally incomprehensible despite being understood by relatively few specialists, AI should be similarly recognized as a specialized domain where the underlying science is more accessible than in many experimental disciplines 1.