The conceptual foundation of OpenML dates all the way back to 2013 when a college boy in Physics major was reading and implementing the Matrix Numerov Method for Solving Schrödinger Equation

When I tried to package a standard agent skill with image assets into a MCP server and an agent can't see or fetch those images through MCP, because MCP protocol (modelcontextprotocol.io) does not contain any specification or protocol standard regarding how an MCP server should deliver or serve an assets folder from skill (agentskills.io), I sensed a fundamental mismatch between skill and MCP which triggered this blog post

Just as DevOps brought structure to software deployment and MLOps did the same for training models, AgentOps is the emerging field dedicated to the lifecycle of agentic AI. Shared evaluation frameworks like Harbor is an example of AgentOps. The core philosophy driving this shift is moving away from "walled-garden" testing (where companies build private tests that happen to make their own agents look good) toward Standardized Evaluation Environments. It is the realization that a benchmark is only trustworthy if everyone can run it, inspect it, and use the exact same ruler to measure success.

If the Universal Approximation Theorem expresses the unprecedented power of AI, the Halting Problem poses the opposite - its computational limit.

I study Social Psychology in order to get insights about myself. A book on Social Psychology sparked my interests on this field so I picked up a text book about it. The notion of "Fundamental Attribution Error" and its application on World War II German society ignited my examination on my past 30 years of life as a Chinese.