The human is the feedback loop

He's right. But the question that follows is more interesting than the one he asked.

Why he's right

Code has a compiler. A test suite. A stack trace. The signal arrives in milliseconds.

You write something, you run it, you find out. The loop is so tight it disappears.

Prose has only the reader. And the reader is far away. At publish, in another tab, next quarter. Whatever the writer is trying to do, only a person can confirm whether it worked. Someone sitting somewhere else, reading.

The gap is structural. No model improvement closes it from the inside. The thing being judged is whether the writing lands in a person, and that judgment lives in the person, not in the text.

But the loop doesn't have to live there

The conclusion most people are drawing from Naval's piece is that AI needs a better internal reward signal. Better taste classifiers. Smarter models that can grade their own output.

That's one path. The other is older and quieter: the loop doesn't have to live inside the model. It can live in the environment.

When the writer, human or AI, sees the actual reading experience as the writing happens, judgment arrives in real time. Not at publish, not in a different tab. Right there, where the writing is. The text is being judged, and the judge is in the room.

Music figured this out decades ago, when DAWs put composition and listening into the same window.

What writing in a closed loop looks like

I've been writing with AI directly inside our editor for the past few months. The AI drafts. The page renders as it drafts. Typography, images, spacing, the actual reading experience.

Same surface, same shape, no translation.

What changes is timing. When something is off, the sense arrives immediately.

A sentence that reads fine alone but stalls on the page. A paragraph that breaks the rhythm of the layout.

Mid-sentence, not after publish. I redirect. The AI adjusts. The page updates.

The AI isn't getting better. The loop is. Judgment can land while there's still time to act on it. The writer becomes a continuous test harness because the environment lets the harness run continuously.

The same words, drafted in a chat window and pasted into a CMS, would have read fine right up until the moment they didn't. The judgment that catches this can only happen against the real surface. Without the surface, it arrives too late to matter.

Preview is a lie

Most editors don't work this way. The preview is a lie. You write in one window, the page renders in another, and the gap between them closes only at publish, if it closes at all.

This was tolerable when humans wrote alone. Slow, but you could imagine the page from the words. You filled in the missing rendering with practice. Now that humans and AI write together, the output arrives faster than imagination can keep up. The gap matters more than it did, and there's less time to bridge it.

The room

Naval thinks creative AI needs better feedback. So do creative humans. We've been writing in different rooms than our readers for so long we stopped noticing.

The model isn't the bottleneck. The room is.

I don't think there's only one way to close the loop. But it's worth asking, every time we open an editor, where the reader actually lives. If the answer is "somewhere else," the loop is broken.