"Looking across tasks, we see a very stable relationship between how much context API customers provide to Claude and how much Claude actually produces. Across economic tasks, each 1% increase in input length is associated with a less-than-proportional 0.38% increase in output length."
"This elasticity of 0.38 suggests that there are strong diminishing marginal returns in translating longer contextual inputs into longer outputs for these economically useful tasks."
报告给出了关键点
"The upshot is that deploying AI for complex tasks might be constrained more by access to information than on underlying model capabilities. Companies that can't effectively gather and organize contextual data may struggle with sophisticated AI deployment."
为什么这个问题的根源在于默会知识,那是因为默会知识有两个重要的特点:
第一:组织默会知识的分散性:
分散的:存在于不同部门、不同层级
非结构化的:以经验、直觉、判断的形式存在
个人化的:深深嵌入在具体员工的认知中
企业没有中央化的“智慧库”,只有中央化的“数据库”。信息孤岛不是技术问题,是认知孤岛。
第二:默会知识的不可编码性
Michael Polanyi说过:“We know more than we can tell”(我们知道的比能说出的更多)。