人工智能上下文的第一原理

The context windows of AI models are expanding dramatically, reaching millions of tokens, which is indeed impressive. However, the question of what content fills these windows is often overlooked. Currently, the approach is to dump everything into the context—documents, chat histories, and tool definitions—hoping the model will discern what is important. However, larger windows do not solve the context problem; they merely provide more room for error. A million tokens of unfocused, unstructured context is less effective than ten thousand tokens of the right context, as the model has to work harder to find the signal amidst the noise, leading to poorer performance.