Exercise 1: Why Generation Can Start From Pure Noise and Still Produce a Coherent Image — Possible Solution
====================================================================

WHAT TRAINING ACTUALLY TEACHES THE NETWORK
------------------------------
Per this chapter, training works by "taking a real image, adding noise
in increasing amounts across many steps, [and training] a network to
predict and remove the noise added at each step." Crucially, this
training happens across the FULL RANGE of noise levels — from images
that are only slightly noised all the way up to images that are
essentially pure noise. The network isn't just learning "how to clean
up a mostly-clear photo" — it's learning, for every possible noise
level from mild to total, what a slightly-cleaner version of that
noise level would look like, conditioned on the text embedding.

WHY THIS MEANS "ALMOST PURE NOISE" IS JUST ANOTHER TRAINING CASE
------------------------------
Because the forward process trains the network across that entire
range, "pure random noise" isn't actually an unfamiliar starting point
for the network — it's functionally identical to the final, most-
noised step the network already saw thousands of times during
training. The network doesn't need to recognize "is this a real photo
that's been noised, or synthetic random noise" — per this chapter, it
was never taught to make that distinction in the first place; it was
only ever taught "given noise at this level (and this text embedding),
predict what a slightly less noisy version looks like."

WHY REPEATED SMALL STEPS ADD UP TO A COHERENT IMAGE
------------------------------
Per this chapter's own generation description: "run the trained
network's denoising step repeatedly, watch a coherent image emerge
from static." Each individual step only needs to make one small,
locally plausible improvement — nudge the noise slightly toward
something more image-like, steered by the text embedding (covered in
the chapter's next section). Because that same small improvement is
applied hundreds of times in sequence, and each step's output becomes
the next step's input, small local improvements compound into a fully
formed, globally coherent image by the final step — without the
network ever needing to "know," at any single step, what the final
image will look like.

WHY THIS WORKS AS AN ANSWER
------------------------------
It explains, using the chapter's own description of the forward
training process across the full noise range, why starting from pure
noise isn't actually a special or unfamiliar case for the network, and
explains how repeated small denoising steps compound into a coherent
result without requiring any single step to "understand" the whole
image.