A parks assistant still describes last year's trailhead kiosk after the county moved it. The new location is in a posted notice the model was not given. The lab GPUs are healthy. Which limitation is this?
Select an answer to reveal the explanation.
Short Explanation
Think of still describing last year's kiosk after the county moved it, because nobody handed the model the new notice. Stale parametric knowledge needs grounding. A healthy GPU is not the problem, and this is not random phrasing.
Full Explanation
Foundation models store parametric knowledge that can be stale or incomplete. When facts change in the world, the model keeps inventing from old parameters unless new material is grounded via retrieval or updated context. The parks assistant still describes last year's trailhead kiosk after the county moved it, and the new location exists in a posted notice the model was not given—classic stale/missing facts, not hardware failure.
A GPU shortage, because outdated facts are a hardware fault, confuses capacity with knowledge freshness; healthy lab GPUs already rule this out. Overfitting to the inference server's clock is nonsense: clock speed is not a training label set that would encode a kiosk move. Nondeterminism is varying phrasing across runs, not systematically reporting an obsolete location when current notice text was never supplied.
Exam caveat: parametric knowledge is not a live database. Relocation and policy changes need grounding; do not blame GPUs or randomness for missing facts.
Operational check: retrieve the latest posted trailhead notices into the parks assistant context, invalidate cached answers when a notice updates, and spot-check kiosk locations against the county feed after every facilities change.