Context window

I closed the previous article with an odd promise: you can’t send everything, there’s a bottom. Here’s the bottom.

Context window.

At the beach the logic is straightforward: “fine, I’ll upload everything. My whole life story, the entire drive, every document, and AI will finally know everything.” No. You can put very little in. And that little has a name that lies.

“Window” suggests a pane of glass through which AI sees everything you put in front of it. It’s actually a container with a bottom. Call it by the right name: it’s a backpack. Finite. You pack it before you leave, and what you leave at home you don’t have on the road. It’s not blurry, it’s not “a bit compressed”: it’s out.

And that backpack isn’t even entirely yours. Before you open your mouth, the provider has already half-filled it: their hidden instructions, your file, the entire chat history. Your content and your question compete for whatever space is left.

“Fine, I’ll get the giant backpack — the one-million-token one — and I’ve solved it.” No. And here’s the thing nobody tells you: bigger isn’t better. An enormous backpack filled at random is worse than a small, well-curated one — for two concrete reasons.

First: AI reads the beginning and the end better, and skims the middle. What you bury halfway through the pile is as good as not written. It’s there, but buried. The more you pile in, the more you bury the thing that mattered.

Second: noise drowns the signal, and every extra piece costs and slows things down. Give the model the whole manual instead of the right page, and you get a worse answer. More is less.

And when the backpack is truly full, something has to come out. Either the oldest thing drops, which is why long chats forget how they started. Or it gets compressed into a summary that keeps what it decided was important and discards the rest. In both cases you lose something — and you don’t choose what. The detail you needed may be exactly the one that ended up at the bottom, and you only find out when an answer comes out wrong.

The takeaway: it’s not how much you put in that matters, it’s what, where, and with what weight. A small, well-curated backpack always beats a large, bloated one. It’s the same virtue as “send as little as possible”: choose the few pieces that shift the result, and put them at the top. More is less.

There’s also a financial reason not to stuff it. That backpack is measured in tokens, and every token inside gets re-read from scratch at every turn. But that’s the next article.

Scroll to Top