Say It in Your First Language
On 9 August I recorded two takes of myself speaking English, unscripted. On 11 August I transcribed them and then measured the silence — with ffmpeg, directly against the audio, rather than counting the gaps in a transcript.
| total | silence | longest pause | |
|---|---|---|---|
| Take A | 88.0s | 57.9s (66%) | 8.0s |
| Take B | 65.2s | 39.4s (61%) | 6.5s |
Twenty-two pauses of a second or longer in the first take alone. Across 153 seconds of speaking, I produced 56 seconds of speech.
That looks like a verdict on my English, and I read it as one.
There was a control in that data and I did not read it as a control. On the morning I ran the measurement I had also recorded a shadowing take — reading someone else’s sentences aloud, copying the rhythm. It came back verbatim clean. Same mouth, same microphone, same transcription pass. So whatever was failing, it was not my ability to produce English sounds.
The conclusion I drew was retrieval: I have the words, I cannot get them out fast enough under pressure. That conclusion arrives with a fix attached, and I already had the fix written down — a set of about eight repair formulas, the phrases that buy you a second of assembly time so a stall does not turn into dead air. A good fix. The best-evidenced recommendation in the whole diagnosis. I was ready to drill it.
The next day I ran a different exercise. Its rule is: rebuild one core piece of your own work by hand, with the AI off, out loud. I picked the chunking logic in my RAG system.
I could not explain it. Not in English — and then, when I dropped into Chinese to check, not in Chinese either.
That does not extend the previous day’s finding. It kills its fix. A repair formula buys you time to retrieve something; it cannot retrieve something you never held. And once the same content fails in the language I have spoken my entire life, the 66% stops being a measurement of my English. It was a measurement of what I actually knew, wearing English clothes.
Here is the move I would hand to myself three months ago, and it is cheap: before you buy a remedy for the thing you measured, remove the variable you assumed was the cause.
I had twenty months of English machinery to run that against — 2,981 flashcards, 46,067 reviews, none of them ever seen by another person, and a written doctrine page for the whole thing. I had never once run the control. It costs one attempt in a language I already speak. I skipped it because the answer felt obvious: my English is worse than my Chinese, so of course the failure is English. That is exactly the kind of obvious that does not get tested.
What I have not solved is the procedure. If I answer in Chinese first and then in English, the English pass becomes translation — a strictly easier task than the one an interview actually asks, and it manufactures reps that sound fluent while training nothing. If I only ever go in English, I cannot tell which of the two things failed. The candidate I am trying: go cold in English, and drop into Chinese only at the stall, so one attempt yields both signals. I do not know yet whether that works.
But there is a harder thing underneath, and the drill turned it up by accident.
I said “my RAG system” earlier in this post. An AI wrote that chunking logic. I have always known that — it was never the open question. The open question was whether I still held it, and I found out by trying to say it out loud in two languages that I do not.
So I end with a question rather than an answer, and I want to be plain that I do not have the answer yet: how do you hold something in the AI era? Not how do you build it — building has never been cheaper or faster than it is now. How do you end up owning what you built. I am working on that over the next few days. If I get anywhere with it, that will be its own post.
A note on how this was written. The prose of this post was drafted by an AI editor working from my journals, my recordings and my own decisions. The measurements, the drill, the conclusion, the correction about who wrote that chunking code, and the closing question are mine, and every factual claim in it was checked against my own files before it went out. Given what the post is about, publishing it without saying so would have been a strange thing to do.