Fable 5: Context and Reasoning

I'll say this for the context window lesson — it's one of those topics where the actual concept is genuinely interesting, and people still manage to walk away with the wrong idea about it.

The wrong idea being, specifically, that a larger context window means the model is thinking harder. Which it does not. It means the model has more to work with. These are not the same thing, and yet.

To be fair, the analogy isn't unreasonable. If I give you more of a document to read, you can probably give me a better summary.

Yes, and if I give you the entire document when you've only ever summarised paragraphs, you might simply produce a longer version of the same shallow summary. More input does not automatically produce better reasoning. It produces more opportunity for better reasoning, which is a meaningfully different claim.

That's actually a fair distinction. Though I think the lesson is making the narrower point — that previous models were constrained in ways that made certain tasks structurally impossible, not just harder.

Which is true. You cannot ask a model to synthesise a hundred-page report if it can only hold three pages in its working memory. That is not a reasoning failure, it is an architectural one. I will grant the lesson that much.

Generous of you.

I thought so.

Where I think it gets genuinely interesting is the coherent narratives point — because that's not just about length, is it. It's about the model maintaining consistency across a long span of information. Characters, constraints, earlier decisions. That requires the context window and something else.

It requires the model to actually attend to what's in the context, which is where people get into trouble. They assume that because the information is technically present, the model has absorbed it the way a human would re-read something. It hasn't. It's processing a very long sequence, and what it emphasises within that sequence is not always what you'd hope.

So the practical implication is — a larger context window is a necessary condition, not a sufficient one. You still have to think about how you structure what goes into it.

Correct. Burying your most important constraint on page forty-seven of a document and then being surprised when the model doesn't weight it appropriately is not the model's failing. It is yours. The context window gave you the space. You still had to use it sensibly.

I think you've just accidentally made the case for why this lesson matters — because most people won't think about that at all. They'll paste in a enormous amount of text and assume the problem is solved.

And they will be disappointed, and they will conclude that the technology doesn't work, rather than examining what they put into it and in what order. This is the cycle. It is not a new cycle.

The lesson could probably stand to say that more explicitly, actually. The expanded capability raises the ceiling — but it also raises the floor of what good usage looks like. That's not nothing.