Contents
A few weeks ago I caught myself doing something strange. I was annoyed at a friend for being late, and before I’d even finished being annoyed, some other part of my head had already produced an explanation — traffic, probably, or he lost track of time, he always does — and handed it to me as if it were a fact rather than a guess. I hadn’t reasoned my way there. It just arrived, fully formed, the way a model’s output arrives: no visible working, just a completion.
That’s a strange thing to notice about your own mind. And once I noticed it, I couldn’t stop noticing it elsewhere. The way a song feels “already familiar” the first time you hear it, if it borrows the right chord progression. The way you can predict, almost to the word, what a certain relative is going to say about politics at dinner. The way learning your second language made the third one arrive almost pre-installed. None of these are conclusions I reasoned to. They’re completions — outputs of some process running underneath the reasoning, trained on a lifetime of examples I never consciously reviewed.
I spend most of my day thinking about that process in machines. Loss landscapes, representation spaces, what a model has actually learned versus what it’s memorized, why some architectures generalize and others just get very good at the training set. And at some point I stopped being able to keep that thinking confined to the lab. Once you’ve spent enough time asking “what did this network actually learn, and how do I know,” you start asking the same question about people — about yourself first, which is the uncomfortable part.
That’s what this series is going to be. Not machine learning content, exactly, and not personal essays either — something in between. I want to take some ordinary thing: an argument, a habit, a piece of advice everyone repeats without checking it, a feeling that seems to come from nowhere — and open it up using the tools I actually have on hand. Representation theory, to ask what space an idea or a memory or a skill is really living in, and whether the boxes we assume exist (this is “logic,” that is “emotion,” this is “the English part of my brain”) are as separate as they feel. Learning theory, to ask how something got trained into us in the first place, and whether the way we’re taught to learn it is anywhere near optimal. And machine learning more generally, because it’s the first field in history that’s had to be precise about what learning even is — precise enough to build it — and that precision turns out to be a useful flashlight to shine on our own much blurrier version of the same process.
I don’t think the metaphor is going to hold up everywhere. Brains are not neural networks, training is not a childhood, and I’d rather write something honest that breaks down in an interesting place than something tidy that’s quietly wrong. So expect some entries to end without a clean answer — a phenomenon I can describe well through this lens and one I can’t, sitting side by side, because the failure is often more informative than the fit.
What I can promise is that every entry starts from something real and small — a moment, a conversation, a thing I noticed about myself that I didn’t like noticing — rather than from a thesis I’m trying to prove. The machine learning comes in because it’s genuinely how I think now, not because it’s a device to dress up the observation. If it turns out to be a decent lens for looking at ordinary life, that’ll show up in the writing. If it doesn’t, that’ll show up too.
Let’s see what’s actually underneath the completions.