Exercise 2: Why nlp1-6 and nlp1-7 Are Genuinely Different Pipelines Despite Sharing a Mechanism — Possible Solution
====================================================================

WHAT nlp1-6 AND nlp1-7 GENUINELY SHARE
------------------------------
Per nlp1-7's own explanation, both models are built from the exact
same underlying LSTM mechanism from nn1-8 — the same step-by-step
hidden-state computation, the same embedding-layer input. nlp1-7 was
explicit that the architectural change required to go from one to the
other was "smaller than it sounds": using every hidden state instead of
only the final one.

WHY THEY ARE STILL GENUINELY DIFFERENT PIPELINES
------------------------------
Despite that shared mechanism, per this chapter's own attribution
table, nlp1-6's classifier reads only the final hidden state through a
single sigmoid neuron, producing one binary label for the whole
sequence, while nlp1-7's tagger reads every hidden state through a
softmax classifier over multiple tags, producing one multi-class label
per token. These aren't just different settings on the same trained
model — they are different output layers (sigmoid vs. softmax),
different training objectives (binary cross-entropy over one label vs.
per-token classification loss over a whole tag sequence), and, in
practice, separately trained models, since a model trained to classify
whole-sentence sentiment has no mechanism at all for producing
per-token predictions, and vice versa.

WHY SHARING A MECHANISM DOESN'T MAKE THEM THE SAME PIPELINE
------------------------------
The LSTM mechanism they share is a general-purpose sequence-processing
tool, not a task-specific solution on its own. What actually determines
the pipeline's task is what surrounds that shared mechanism — which
part of its output gets used, what kind of classifier head reads it,
and what objective it's trained against. Two pipelines can reuse the
identical core mechanism and still be, in every practical sense,
different tools built for different jobs.

WHAT THIS REVEALS ABOUT PRE-LLM NLP
------------------------------
Per this chapter, "every task this course covered got its own
purpose-built, hand-assembled architecture. That's not a flaw in how
the course was taught — it's an honest reflection of how NLP actually
worked before LLMs." Even within a single course covering closely
related techniques, sentiment classification and NER required two
separately designed, separately trained systems. This is exactly the
"many task-specific pipelines" half of the contrast this chapter draws
against LLMs, where one pretrained architecture can instead be steered
toward many such tasks without being rebuilt from scratch for each one.

WHY THIS WORKS AS AN ANSWER
------------------------------
It identifies precisely what nlp1-6 and nlp1-7 share (the LSTM
mechanism) and precisely what makes them different pipelines despite
that (different output layers, objectives, and trained models), and
connects this directly to this chapter's own broader point about how
pre-LLM NLP required a new hand-built pipeline for each distinct task.