通用人工智能（AGI）被宣传为斯皮尔曼的“g因素”，但其架构却更像吉尔福特的智力结构模型。
3 分•作者: jatinkk•6 个月前
我不是技术专家，也没有在科技行业工作，所以这只是一个外行人的视角。
围绕通用人工智能（AGI）的营销承诺了斯皮尔曼的 g 因素：一种通用的、流畅的智能，能够适应新的、未曾遇到的问题。
但其工程设计——特别是“专家混合”和不同的模块——看起来完全像是 J.P. 吉尔福德的智力结构。吉尔福德将智力视为大约 150 种特定、独立能力的集合。
问题不仅仅在于这些部分是如何组合在一起的。我看到的问题是：当模型面临一个不适合其预定义部分的问题时会发生什么？当架构依赖于在专业的“专家”之间切换，而不是使用统一的推理核心时，他们将如何确保输出不会显得支离破碎？
特定技能的集合（吉尔福德）与适应任何事物的能力（斯皮尔曼）是不同的。通过优化特定组件，我们正在构建一个擅长已知任务的系统，但可能从根本上缺乏真正通用智能所需的流畅推理能力。
我并不反对人工智能；我只是觉得我们可能需要重新审视我们的方法。我们不能指望在错误的道路上到达正确的目的地。
查看原文
I am not a tech expert and not working in the tech industry, so this is an outsider's perspective.
The marketing around AGI promises Spearman’s g: a general, fluid intelligence that can adapt to new, unseen problems.
But the engineering—specifically "Mixture of Experts" and distinct modules—looks exactly like J.P. Guilford’s Structure of Intellect. Guilford viewed intelligence as a collection of ~150 specific, independent abilities.
The issue isn't just about how these parts are stitched together. The issue I see is: what happens when the model faces a problem that doesn't fit into one of its pre-defined parts? How will they ensure that the output doesn't look fragmented when the architecture relies on switching between specialized "experts" rather than using a unified reasoning core?
A collection of specific skills (Guilford) is not the same as the ability to adapt to anything (Spearman). By optimizing for specific components, we are building a system that is great at known tasks but may fundamentally lack the fluid reasoning needed for true general intelligence.
I am not anti-AI; I simply feel we might need to relook at our approach.We can't expect the right destination with the wrong highway.