(Something I ran into in 2008 when I was writing a multilingual book and heard about LuaTeX, only to find out it was still extremely incomplete and was roadmapped to not be done until 2012. And I couldn't wait 4 years)

So, no: it's nothing like what you replied with and that was a rather dumb thing to say.

Yes, databases could have any extension. No sane dev team would accept code that doesn't use an object extension for a sqlite database.

Yes, databases can contain PII but no sane product manager would go "yes, that's a good use of sqlite".

Yes, you can trivially copy database files, but no sane product needs to in the same way that no sane product should require folks to just clone the db just to do some work.

Pretty much every reason a company has for banning sqlite is a red flag for working there.

The file might be 4GB but the transfer sure as heck wasn't, so what are we even talking about? How much data is actually transferred? Can someone just grab that weights.bin file and zip it up with max compression and report a more realistic number that we can do the math with, if the number is even worth doing the math for?

Practical example: I have a bird that's being chased by another bird, and they overlap in the shot. There's weird lighting on the bird that's further away, so I need to grade them differently. But they overlap so now I have a challenge. I could try to do this using layers and masks: mask both birds in a way that the masks don't overlap, while perfectly tweaking the mask feathering so that there's minimal bleed on their overlap, then tie each mask to an adjustment layer.

But if I have graph based adjustments available, I first split my input into separate nodes for the background and each bird, then for each of those, I can send them through a node that masks them appropriately without worrying about mask overlap. I can then chain adjustment nodes to grade all three and I can save those grades separately, too so I can use them on other shots from the same series, then I can send each chain into a muxer that turns the three elements back into a single composition.

I could do that with layers, where I clone the full image several times, create my adjustments in groups, then render each group to a new layer, hide everything else, and mix those layers, but now what do I do if I want to tweak the grading? Delete my layer, unhide the group, tweak the adjustments, rerender the group, mix the new "final" layer in, and holy crap how many things I did I just need to do that weren't "making my adjustments"? Whereas with a node graph you just make your adjustment. Done, your change simply cascades through the graph.

There's a lot that you can do with layers, but layers are just a linear graph: you can do more if you can branch and merge your graph.