木鸟杂记

在古诗文中，我们以月喻离合、以水比时光、以柳寄别恨。

从第一个我体感“有点不一样”的 Code Agent —— Claude Opus 4.5 发布（2025年11月24日）以来，竟然才过去半年。

Prologue When first entering the world of large models, due to gaps in foundational math knowledge such as linear algebra, probability theory, and information theory, it’s easy to get lost among numerous terms: logprob (log probability), likelihood, NLL (…

引子 刚入门大模型的时候，由于线性代数、概率论和信息论等数学知识的短板，很容易迷失在诸多术语中：logprob（对数概率）、likelihood（似然）、NLL（Negative Log Likelihood，负对数似然）、cross entropy（交叉熵）、perplexity（困惑度）。

I joined an LLM company in early 2024, having previously worked in the infra industry (databases, storage, etc.), so I have some very basic insights on switching careers.

我是 2024 年初到一家大模型公司工作，之前一直在数据库、存储等 infra 行业工作，因此有些很粗浅的转行认知。

Since becoming distinctly self-aware, never have I clashed so intensely with the world and with myself as I have this year—yet the result is strangely magical: I have become even more peaceful.

有明显的自我意识以来，从没有像今年这样和世界、和自己发生如此激烈的冲撞，但结果很神奇——反倒更加平和了。

Princeton COS 597R “Deep Dive into Large Language Models” is a graduate course at Princeton University that systematically explores the principles of large language models, their preparation and training, architectural evolution, and applications in cutti…

Princeton COS 597R “Deep Dive into Large Language Models” 是普林斯顿大学的一门研究生课程，系统探讨了大语言模型原理、准备和训练、架构演进及其在多模态、对齐、工具使用等前沿方向中的应用与一些问题。

As cloud infrastructure continues to mature, emerging companies typically move their infrastructure to the cloud in order to achieve business goals quickly.

随着云基础设施的不断成熟，新兴的公司为了快速实现业务目标，一般都会让基础设施上云。

streamlit is a Python library for quickly developing simple web apps.

streamlit 是一款可以快速进行简单网页开发的 Python 库，其 slogan 是： A faster way to build and share data apps 即“一种快速构建、分享数据应用的方法”。

Introduction In certain workloads, memory usage gradually grows over time until an OOM occurs.

引子 在某些工作负载中，随着时间的推移，内存的使用会逐渐增长，直到 OOM。

Snowflake was founded in 2012 by two former Oracle employees, targeting cloud-native data warehouses from the very beginning.

Snowflake 由甲骨文的两位员工在 2012 年出来创办，一开始就瞄准云原生数仓，因此架构设计（在当时看来）非常“激进”。

I first heard about it from a passing mention on a podcast, so I sought out the audiobook for my commute.

缘于某个播客提了一嘴，便找来书在通勤时听了。