⌘K
Change language Switch ThemeSign In
Narrow Mode
From MCP to Skill: Critiquing the Evolution and Limitations of AI Infrastructure ================================================================================
From MCP to Skill: Critiquing the Evolution and Limitations of AI Infrastructure ================================================================================  ### Yangyi
@yangyi
当mcp来的时候，我们说mcp最终会死
因为mcp的模式是全量上下文，而不是渐进式披露结果
那时候大家聊这个东西，就觉得不够本质
明明只是运行一个脚本，给一个结果，为什么要在不需要过程数据的时候把上下文也扔回来？
果然，mcp死了，skill出来了
但skill也不够本质
因为skill当前没有一种自我迭代的infra
它是个一次性的
它缺少目标奖励函数的输入
如果skill无法自我持续迭代 改进 优化
那么它就仍然还是一个中间态
最终被吞噬 被封装
mcp死了,skill也快了，不过不重要，能用就好
这就是事物发展过程必然经历的阶段
就像transformer也得死
但必须要承认他们在历史发展过程中塑造的巨大伟业
没有他们，现在的AI还不知道是啥样呢 Show More
Mar 14, 2026, 9:07 AM View on X
1 Replies
1 Retweets
11 Likes
1,602 Views  Yangyi @yangyi
One Sentence Summary
This tweet critically analyzes the inherent flaws of the MCP and Skill paradigms, arguing that architectures lacking self-iterative capabilities will ultimately be absorbed or encapsulated.
Summary
The tweet delves into the evolution logic of AI interaction protocols. The author believes that MCP (Model Context Protocol) has an inherent flaw by adopting a full context approach rather than progressive disclosure of results. It also points out that the current Skill model, lacking input from a target reward function and a self-iterating infrastructure, remains an intermediate state technology. The article predicts the historical limitations of these technologies, including the Transformer architecture, emphasizing that while they are significant milestones in the long river of AI development, they will eventually be replaced by more fundamental solutions.
AI Score
84
Influence Score 3
Published At Today
Language
Chinese
Tags
MCP
Skill
AI Infra
Transformer
Architectural Evolution HomeArticlesPodcastsVideosTweets
From MCP to Skill: Critiquing the Evolution and Limitatio... ===============