Those Gray Beard Hackers And Their Tall Stories

That only works if you're building stuff for yourself. And then only if it's for a hobby. Otherwise you find yourself trying to (e.g.,) stuff a complex and somewhat safety-critical(1) control loop with 16-bit fixed-point division implemented in assembler into a processor with 1200 bytes of memory.

(1) phone calls from the irritated client that your design caught fire and would have burned down his lab if he hadn't noticed the smoke at 5PM are generally not a good thing...

I don't get to build as many of them as I once did, but they're incredibly fun little puzzles to solve. Cramming largish photographs, animation, copy (including fonts, which usually result in the bulk of filesize), and animation into such a tiny package is no small feat, and requires a lot of time (the productivity tradeoff Jacques referred to) and tons of little optimization tricks you need to spend a few years collecting and perfecting, like:

* becoming a vector artist. Sometimes its necessary to have a partially transparent image, which means 32 bit PNGs, which are super expensive. Flash allows you to create vector masks (vector illustrations are incredibly size efficient), which means you can use a PNG-8 or JPG and mask out the areas that need to be transparent. Which means getting really good, and fast, at tracing outlines. The next level being

* learning how to do photo-realistic illustrations. In the old days of 12k for a banner, I did a lot of car banners, which meant a lot of car illustrations that needed to look exactly like the photos. Even then, you'd have to reuse your assets a lot, in creative ways, in order to make the filesize. (Like using a tire illustration for the windshield [by scaling and masking]).

* writing really tight code. In the early days, our basic animation engines were a few KB, but you could rewrite your own using only the chunks of code you needed, and save a few K. My newest library is about 800 bytes and is powerful enough to use on most of my banners.

* learning that people really aren't as detail oriented as you might think, and blurs, rotations, tints, and speed are usually all the tools you need to make a convincing animation

That's the skill set of which I speak. Getting a shop in China to actually produce a working product for you should be a major of its own at MIT.