Period 2 · Capabilities, Limitations and Common Errors
[video] Lecture video — Period 2: Capabilities, Limitations and Common Errors — /api/lesson-media/videos/4d3653ff-a3ca-4f63-bf8e-bb93c553e599-1794840393120-796331ea660529ef.mp4
Flashcards · Period 2
Language model — A model that generates language from patterns it learned and from the information supplied in the conversation.
It continues text in a plausible way; it is not a database of verified facts.
Token — A unit of text that the model processes. A token is a piece of text, not a guaranteed fact.
A word, or part of a word.
Hallucination — Unsupported or fabricated content presented as if it were fact.
A citation to an article that does not exist.
Task drift — A response that moves away from the agreed task.
A four-row table request that turns into a strategy essay.
Causal claim — A claim that one factor produced another. It needs suitable evidence, not just a comparison.
"The discount caused the decline."
Overinterpretation — Reading more into the data than it shows. Fix it by labelling a hypothesis and asking for the missing evidence.
Low recorded sales described as weak demand.
Correct, qualify or remove — The three ways to handle a claim: fix a wrong fact, state the limits of a partly supported one, or delete an unsupported one.
Correct the region label; qualify the cause; remove the unnamed report.
Independent check — A check that does not depend on the AI's own explanation: open the source, recalculate, or review as a human.
Summing Net_Sales_Amount per region in a spreadsheet.
