Intelligence and context are multiplicative, not additive

There are two ways to make an AI system better. Give it more intelligence, meaning better reasoning, more parameters, longer thinking. Or give it more context, meaning more of the user’s data, more of the relevant history, more of the situation it’s working inside.

Most teams treat these as separate investments. The model team chases benchmarks. The product team builds RAG. Each side ships incremental wins on its own axis. Each side gets linear returns.

But intelligence and context compound. The two axes multiply each other.

A weak model in rich context can’t tell which 2,000 of the 10,000 available facts matter for the decision in front of it, so most of the context gets ignored or, worse, mistaken for noise. A strong model in starved context has nothing to reason over and ends up hallucinating with confidence. Pile more context onto a weak model and the signal-to-noise ratio gets worse, not better. Pile more reasoning onto a context-poor model and it just gets better at being wrong with conviction.

The compound is the whole story. A strong model can use more context productively, because it can tell what’s relevant. Better-organized context makes the reasoning sharper, because the model has the right inputs to reason over. Each axis raises the ceiling on the other. The team that moves both at once gets multiplicative returns. The teams moving them independently keep ramming into the same ceiling and assuming the ceiling is physics.

This is also why “we have a 1M-token context window” stops being impressive about thirty seconds after you actually try to use one. The window is the easy half. The model that can identify which 2,000 of those million tokens actually matter for the question in front of it, that’s the hard half. And the hard half only works when the model is strong enough to reason about its own retrieval.

Bet on the product of the two, not the sum.