Hallucination.

Hallucination is when a model makes things up or starts forgetting, confidently telling you something that isn't true. The usual cause is an overloaded context. Picture a new hire you've buried in information on day one: past a certain point they drop the early stuff and fill the gaps with guesses.

Remy's rule of thumb is that once you're past about half your Context window, hallucinations start. Two things keep it low: clean context, and the bigger context windows newer models ship with.

The bigger tell is the model itself. The best ones admit when they don't know instead of bluffing. As Remy puts it, one of his favorites "has one of the lowest hallucination rates going, because it'll admit when it doesn't know instead of bluffing." When you're building real work off the output, that honesty matters more than raw smarts.