AI Safety · Part 2 of 7

Context: the desk, not a warehouse

Watch: context rot and the lost middle · 74 seconds · animated, sound on

Bigger context is not automatically better

Researchers tested 18 frontier models and every single one got worse as input grew, long before the window was full. Managing what goes on the desk matters more than the size of the desk.

The desk

The context window (explained in AI words, translated) is the model's working desk. Everything it can use has to be on that desk right now. It cannot reach into a drawer it does not have. Three things go wrong with it:

Context rot

The overstuffing problem has hard numbers behind it. Chroma Research tested eighteen frontier models and found the same pattern in every one: as input grows, accuracy falls, and it starts falling long before the context window is full. Piling more into the prompt is not a free upgrade. Past a point, it is how you make a model worse.

The lost middle

Stranger still: models recall the edges of a long prompt best. A fact buried in the middle is the one most likely to be missed, so the answer to your question can depend on where in the pile the answer happens to sit. That result has held since 2023.

How Cronus keeps the desk deliberate

Sources: Hong, Troynikov, Huber, Context Rot, Chroma Research, 2025 · Liu et al., Lost in the Middle, 2023.

← Part 1: Why AI makes things up Part 3: Grounding →