AI Capabilities And Limitations
← Hoch {lessons}
{lesson} 07AI Capabilities And Limitations

Study Notes

wab

wab — yIlaD.

{study notes}
  • {String Matching} QaPla': Qapla'pu'chugh "car" yIjatlhbe'chugh "automobile" vay' yIjatlhbe'. {Semantic meaning} pagh.
  • {Embedding} = Qapla': Text coordinates-Daq jatlhlu'. Similar concepts neighborhood-Daq Qapla'. Training data-vo' emergent.
  • 2D Space Example: Dinosaur axis + roller coaster axis. Documents plotted. Question plotted. Nearest k items = relevant results.
  • More Dimensions = More Topics: 4D, 1024D — math works! Spatial picture stops, calculation continues. Real {embeddings} ~1000 dimensions use.
  • Black Box Axes: Dimension 847 = what? No one decides. Emergent training-vo'. Human-readable labels pagh.
  • {Embedding Model} = Text → Coordinates: "work from home" → [0. 23, -0. 87, 0. 41, ... ] 1024 values. Input length irrelevant; output always same length.
  • {Cosine Similarity}} ≠ Distance: Vector direction matters, not position. Opposite vectors = -1. Identical = 1. Unrelated = 0.
{flashcards} 8 {cards}
{question}
yI'ang · ←/→
{answer}
{flip}
Anki (.tsv) ↓
{knowledge check} 6 {questions}