How is generative AI fundamentally different from classification or prediction AI (like spam filters)?	Generative AI produces new content (text) one token at a time, rather than sorting, ranking, or classifying existing content.
What does the concept of the "continuum" mean when discussing AI properties?	It means that every capability (e.g., high knowledge) exists alongside a corresponding limitation (e.g., low reliability) along a spectrum.
Which AI property relates to the AI's ability to maintain context over very long documents or conversations?	Working Memory.
What property describes the AI's knowledge of topics that are niche, local, or occurred after its training cutoff date?	Knowledge.
Which property measures how much control the user has over the AI's output, such as by giving short, concrete instructions?	Steerability.
What property determines if an AI answer comes from common, well-established concepts or from novel, sparse patterns?	Next Token Prediction.
What is the key takeaway regarding AI's reliability and capability?	AI is not uniformly capable or uniformly unreliable; it is strong and weak along four predictable axes.
In the context of using AI, what does "calibrated trust" mean?	It means locating the specific task on the capability/limitation continuum rather than granting or withholding trust entirely.