四月思绪 · 七则 April fragments · seven

把项目改名比给项目命名难。命名是从空白里写一个字，改名是从一个旧形里掰出新意。我那个 cf-coach 改了三次：先叫 cf-buddy（太软），再叫 cf-pilot（太硬），最后定 coach 是因为它默认承认你水平不够。 Renaming a project is harder than naming it. Naming writes a character on blank paper. Renaming bends a new meaning out of an old shape. cf-coach went through three: cf-buddy (too soft), cf-pilot (too hard), coach stuck because it admits up front that you're not good enough.

在三楼东侧靠窗的位置，盯着一段 attention mask 的 cumsum，突然意识到这玩意儿和我高中数学的"分段函数"长得一模一样。一种被时间从背后戳了一下的感觉。 Third floor, east window. Staring at a cumsum over an attention mask, I realized: this thing is the piecewise function from high-school math. A small poke in the back from time.

"平平仄仄平平仄"——七言里的平仄是一种位置编码。它告诉模型"这是第几个字"，但用一个比正弦更稀疏的码。如果古文 LLM 学得起来，应该能在某层看到一个"模 7"的方向。 "平平仄仄平平仄" — tones in regulated heptasyllabic verse are a positional code. It tells the model "which character index" with a sparser code than sinusoidal. If a classical-Chinese LLM picks this up, there should be a "mod-7" direction somewhere in the residual.

某篇 SAE 的 paper 里写："我们不期待特征是单义的，我们只是不知道还能怎么近似。"——这种诚实的句子比方法本身更有说服力。 From an SAE paper: "we don't expect features to be monosemantic; we just don't know how else to approximate." Honest sentences like this are more convincing than the method itself.

第一年最好的学习方法可能是：在群里大方地说"这个我不会"。比假装会便宜一万倍。 Best learning hack in year one: say "I don't know how" in the group chat, freely. Ten thousand times cheaper than pretending.

98.7% 和 98.71% 之间的差距，在不同语境下从"什么都不是"到"决定项目去留"都有可能。学会问这个数字活在哪个语境里，比把数字本身记住更难。 The gap between 98.7% and 98.71% ranges, depending on context, from "nothing" to "kills the project." Learning to ask which context a number lives in is harder than memorizing the number.

深夜写完一段实验日志，关电脑前总有一个念头：明天会比今天懂得多一点点。这个念头大概就是 epoch 0 之所以叫 epoch 0 的原因。 Closing the laptop after a late log entry, the same thought every time: tomorrow I'll know a tiny bit more than today. That thought is, I think, why epoch 0 is called epoch 0.