In general integer division is implemented using a form of long division, in binary. There is no base-10 arithmetic involved. It's a relatively expensive operation which usually requires multiple clock cycles to complete, whereas dividing by a power of two ("bit shifting") is trivial and can be done in hardware simply by routing the signals appropriately, without any logic gates.

The metric standard is to measure information in bits.

Bytes are a non-metric unit. Not a power-of-ten multiple of the metric base unit for information, the bit.

If you're writing "1 million bytes" and not "8 million bits" then you're not using metric.

If you aren't using metric then the metric prefix definitions don't apply.

There is plenty of precedent for the prefixes used in metric to refer to something other than an exact power of 1000 when not combined with a metric base unit. A microcomputer is not one one-thousandth of a computer. One thousand microscopes do not add up to one scope. Megastructures are not exactly one million times the size of ordinary structures. Etc.

Finally: This isn't primarily about bit shifting, it's about computers being based on binary representation and the fact that memory addresses are stored and communicated using whole numbers of bits, which naturally leads to memory sizes (for entire memory devices or smaller structures) which are powers of two. Though the fact that no one is going to do something as idiotic as introducing an expensive and completely unnecessary division by a power of ten for every memory access just so you can have 1000-byte MMU pages rather than 4096 also plays a part.

If it averages several instances, with enough signal you could decompose a linear combination (e.g. average) of different patterns back out into its constituent parts.

A smarter system won't just take the mean of the votes from different instances but rather discard outliers as invalid input (flagging repeat offenders to be ignored in the future) and use the median or mode of the remainder. The results should also be quantitized to avoid leaking details about sources or internal algorithms; only the larger trends need to be reported.

Of course you could always just keep the collected data private and only provide it to customers willing to pay $$$ for access, which handily limits instance operators' ability to reverse-engineer the source of the data. And nothing prevents you from using separate instances for public and private data sets.

The ubuntu:24.04 Docker image is only 77.30 MiB.

alpine:3.19.0 is 7.38 MiB.

Of course those sizes are without a kernel. Typical everything-included distro kernels are generally a few hundred MiB as they include drivers for everything that might be needed, but a custom build for known hardware can reduce that to just a few MiB.

Most of this is personal opinion and snobbery that I can't do much about except maybe ask that you examine how anarcho-capitalist your takes sound.

Objectivist, perhaps. They're the ones who obsess over controlling and monetizing free external benefits. There is no copyright in anarcho-capitalism (including "moral rights" etc.) so the GP doesn't sound at all anarcho-capitalist while arguing for infringement of others' real property rights to prop up their own artificial (non-rivalrous) "intellectual property" rights.

They didn't say it shouldn't have been developed. Improving the AI models so they can deal with this kind of malicious interference gracefully is a good thing.

The EULA also prohibits using Nightshade "for any commercial purpose", so arguably if you make money from your art—in any way—you're not allowed to use Nightshade to "poison" it.

It would be a nominal charge for storage, bandwidth, and indexing. Book stores carry public-domain titles, for profit, and most have no issue with that. You can always procure the same files somewhere else—they are public domain, after all. Those who pay are doing so for the convenience, not because they're forced to.

They could stick to public domain & indie titles. They won't, but they could.

It's a case of overlapping coverage. Her personal insurance company isn't disputing that the uninsured driver was responsible. They're arguing—not unreasonably—that the organizer of the event is more directly responsible for damages incurred while participating in their event (after the driver, naturally), so their insurance should cover the expense.

No one likes to be caught in the middle of something like this, but at the same time it would be irresponsible of the insurance company, toward both their investors and their other customers, to simply pay out without question when someone else should be paying.

You're restricting speech whether or not you confine your censorship to only AI-generated images.

Correction: Fortunately, not unfortunately. A rule like that would prohibit any form of public / street photography, news videos, surveillance videos, family photos with random strangers in the background... it's not reasonable at all.