Exercise 2: Why Politeness Is Functionally Inert on a Diffusion Model But Not on Claude — Possible Solution
====================================================================

WHAT THE CHAPTER'S WARN-BOX SAYS
------------------------------
Per this chapter's own warn-box, "a prompt like 'please don't include
any text in the image, that would be great, thanks' performs no
better — often worse, from the extra noise words — than just `no text`
as a negative prompt. Politeness, framing, and instructional phrasing
are wasted effort on a system with no concept of being instructed."

WHY THE EXTRA WORDS ARE WORSE THAN NEUTRAL
------------------------------
This isn't just "no benefit" — the chapter specifically says
"often worse." Per Exercise 1's own reasoning, an image model treats
every word in a prompt as a conditioning signal pulling generation
toward statistically associated visual patterns, not as a parsed
instruction with some words carrying semantic weight (the actual
request) and others carrying only social weight (the please/thanks).
Words like "please," "that would be great," and "thanks" still get
fed into that same conditioning process — they add noise tokens that
compete for influence over the generation without contributing any
useful descriptive signal, which is exactly why the chapter calls
the extra phrasing counterproductive rather than merely neutral.

WHY THE SAME COURTESY GENUINELY HELPS WITH CLAUDE
------------------------------
Per this chapter's own comparison table, prompt1's own techniques
(clarity, context, constraints — prompt1-2) work specifically because
Claude "can genuinely shift tone and compliance" in response to framing,
since Claude was trained via RLHF to recognize and respond appropriately
to the full shape of a request, including its tone. A polite,
well-framed instruction can genuinely produce a different — often
better — response from an instruction-following system, because that
system has an actual mechanism (instruction-tuning) for treating
framing as meaningful. An image model has no such mechanism: per the
chapter, framing is "functionally inert — no comprehension to appeal
to." The same words that shift Claude's behavior have nothing to shift
in a diffusion model, because there's no comprehension process there
for the framing to influence.

WHY THIS WORKS AS AN ANSWER
------------------------------
It explains specifically why extra polite phrasing actively hurts an
image prompt (noise tokens with no descriptive content, per the
chapter's own diffusion-conditioning explanation from Exercise 1), and
contrasts this directly with the chapter's own stated reason politeness
can help an LLM (a real trained mechanism for treating framing as
meaningful, which a diffusion model simply doesn't have).