MCP was 3x slower per call, and 9.4x slower on first call including initialization
【洞察】MCP 每次调用比直接 REST API 慢 3 倍，首次调用含初始化慢 9.4 倍——这不是特定服务器的问题，而是架构层面的必然代价：每个 MCP 服务器都在 LLM 和底层 API 之间增加了一个进程层。作者的结论是：CLI/API 对 AI 来说其实是更自然的接口（它已经有大量训练数据），而 MCP 是为了「看起来像 USB-C」而引入的不必要抽象层。这是目前对 MCP 协议最有数据支撑的批评。
this happens with getClient and setClient because it is a svelte context which is only available at component initialization (construction) and cannot be in an event handler.
Setting it to a lambda will lazily set the default value for that input.
The rationale is that it's actually clearer to eager initialize. You don't need to worry about timing/concurrency that way. Lazy init is inherently more complex than eager init, so there should be a reason to choose lazy over eager rather than the other way around.
setContext must be called synchronously during component initialization. That is, from the root of the <script> tag
It looks like the issue stems from having "svelte" as a dependency instead of a devDependencies in package.json within the sapper project. This causes import 'svelte' to load the rollup excluded npm package's set_current_component instead of from within the sapper generated server.js.
RandomOut: Using a convolutional gradient norm to rescue convolutional filters
或许导师这回可以相信初始化网络后的稳定性一直就是一个问题了吧~ 另外，此文还是在优秀的 MXNet 框架上跑的，赞一个~
Fixup Initialization: Residual Learning Without Normalization
关于拟合的表现，Regularization 和 BN 的设计总是很微妙，尤其是 learning rate 再掺和进来以后。此 paper 的作者也就相关问题结合自己的文章在 Reddit 上有所讨论。
Generalization in Deep Networks: The Role of Distance from Initialization
Goodfellow 转推了此文。
作者强调了模型的初始化参数对解释泛化能力的重要性！
