Capabilities & failure modes

Models fail in characteristic ways, and once you can name the failure mode you can usually prevent it. Four matter most for day-to-day work.

Hallucination: fluent guessing

Because the model predicts likely tokens, it can produce a confident, well-formed sentence that is simply false — a citation that doesn't exist, a function that was never in the library, a statistic invented to sound right. Fluency is not truth. The model has no internal sense of certainty, which is why it sounds equally assured whether right or wrong.

Context windows: finite working memory

The context window is the model's working memory for a conversation — measured in tokens. Cram it past capacity and the model silently drops the earliest material, leading to 'forgetting' instructions or earlier facts. Long chats degrade not because the model got tired, but because the window overflowed.

• Knowledge cutoff: training has a date. Recent events, prices, or library versions may be absent or wrong — verify anything time-sensitive.
• Prompt injection: text inside a document the model reads can carry hidden instructions ('ignore previous rules…'). Treat any external text the model ingests as untrusted input.