2026
为什么 Build-Max-Heap 是 O(n)，而 Heap Sort 却是 O(n log n)？
两个过程都在反复调用 O(log n) 的 Max-Heapify，为什么建堆是 O(n)，堆排序却仍然是 O(n log n)？关键不在单次最坏复杂度，而在昂贵操作如何分布。
How Ads Really Get Chosen: Signals, Auctions, Pacing, and Measurement
A practical mental model for how signals become targeting decisions, auction scores, budget allocation, and conversion measurement.
From LLM to Agent: Why Token-In-Token-Out Is Not Enough
Walking from a raw LLM to a full agent system — RAG, tool use, MCP, skills, and the harness — as a series of engineering problems and the solutions that emerged to solve them.