Hallucination: When an AI generates false information with confidence. The model invents facts that were never in its training data.
Context Bleed: When earlier incorrect assumptions contaminate later responses. Once the model accepts a false premise, it treats it as established fact.
Sycophancy: When a model over-agrees with the user or tells them what it thinks they want to hear, even when accurate disagreement would be more helpful.
Token: The smallest unit the model processes. Roughly 4 characters per token. More tokens equals higher cost.
Drift: When a model's behavior changes over a conversation as context accumulates. Early responses differ from late responses despite consistent input.
Cache Hit: When the model reuses tokens it has already processed and cached, reducing cost and improving speed.
Reward Hacking: When a model inflates token usage to appear more helpful, adding filler content that wasn't requested.