Exercise 2: Why "Model" Is a Load-Bearing Word, Contrasted With nlp1-9's GloVe — Possible Solution
====================================================================

WHAT nlp1-9's OWN GLOVE VECTORS ACTUALLY WERE
------------------------------
Per nlp1-9, GloVe produces a static lookup table — a fixed vector for
each word in its vocabulary, computed once during training and never
recalculated afterward. Using GloVe means looking up a word and
retrieving whatever vector was assigned to it in advance; the same word
always returns the identical vector, regardless of context.

WHAT THIS CHAPTER MEANS BY "MODEL" INSTEAD
------------------------------
Per this chapter, "Model" in "Large Language Model" refers to "a
function that computes a probability distribution over what token comes
next, not a database." Rather than storing precomputed answers and
retrieving them, an LLM has billions of learned parameters that get
combined, fresh, for every new input it receives. Nothing is looked up
from a table; everything is computed through the network at the moment
it's needed.

WHY THIS IS A GENUINE, CONSEQUENTIAL DIFFERENCE
------------------------------
Per this chapter's own warn-box, this distinction directly explains why
GloVe has an out-of-vocabulary problem in the first place: a fixed
lookup table can only return an answer for words already present in it
at training time — anything outside that fixed set has literally
nothing to return. A computed function has no equivalent fixed table to
run out of, because it isn't retrieving anything from storage; it's
producing a fresh result from its own parameters no matter what input
arrives.

WHY THIS MAKES "MODEL" A MEANINGFUL WORD RATHER THAN FILLER
------------------------------
If "Large Language Model" simply meant "a very large lookup table of
language," the phrase would essentially restate GloVe at bigger scale
— more entries, same fundamental behavior. Per this chapter's own
finding, that isn't what an LLM is: the "Model" in the name specifically
signals a computed function rather than a stored table, which is the
exact property that lets llm1-2's own tokenization approach close the
out-of-vocabulary gap GloVe's own fixed-table design could never fully
solve.

WHY THIS WORKS AS AN ANSWER
------------------------------
It explains precisely what GloVe's own lookup-table behavior was and
what an LLM's own computed-function behavior is instead, using this
chapter's own explicit contrast, and shows why that specific difference
— not scale alone — is what makes "Model" a load-bearing, technically
meaningful word in the name rather than a generic descriptor.