how on earth does openssh not only have incredibly fucked versioning but is the only package i've ever seen to mention a "privilege separation" chroot path which actively breaks non-root builds. are you fucking kidding me bruh. "privilege separation chroot" defaulting to /var/empty and it immediately errors if it can't create it????? why are your version strings like that bruh?????
241
Replies
·
0
Boosts
@aeva so we hit the 90s and computers have completely revolutionized the financial industry. why? because you can make decisions so much faster! and that means more decisions, which means more money.
the internet is sometimes slow? sometimes down? netscape to the rescue! the internet is an application platform! the application is on the internet! what do you mean down? i'm on the internet right now!
@hipsterelectron ok so this is fascinating but why is SIMD evil?
@hipsterelectron like was that remark specific to avx512 being something of an "expensive planes that don't fly" situation or was your meaning that there is something fundamentally bad about vector processing?
@aeva there is something bad about ARM's vector processing standards, which are different in many ways. the latent potential evil in in vector instructions is because they are in fact motivated, in part, to achieve what LLMs are hoped to
@aeva that if you can pay more money for hardware you can leverage that technical advantage for corresponding market dominance. this mindset is the monopoly shit reaching its apex with LLMs
@aeva the reason i personally have bad vibes about them is because the bigger and bigger ones have ridiculous pipeline latency issues and are progressively more difficult to use within that sweet spot where you can achieve a useful application result without hitting the redline too hard and stalling out
what this means is (unless you are geoff langdale + hyperscan team and spend immense amounts of effort developing novel mathematical frameworks) they're literally just accelerators for extremely simple straight-line shit
@aeva [i promise this concludes immediately]
@aeva and literally all i need to do is exactly two things. my fucking kingdom. two things:
- indices of individual bytes matching this set of bytes in a contiguous memory region
- asynchronous memcpy queue that doesn't interface with any part of the fucking x86 CPU
@aeva SIMD instructions largely don't even need 256 wide vectors to do cool shit because the "SIMD" acronym was coined by 80s sloppers
@aeva i guess technically a scatter and gather are "SIMD"? but masked loads for example are actively reducing the M because human intelligence can compose the I
@aeva so along with "cryptographic hash" i guess i'm also immediately ceasing my use of "SIMD" because "vector" is the cool as shit fucking alligator in shadow the hedgehog
@aeva and i plan to make "cryptographic hash" understood as a buzzword for suits
@aeva oh to be clear i absolutely do not believe processor vendors have any idea long vectors are just riemann approximating actually useful instructions. making AVX-512 expensive means you have to be the kind of person to buy expensive CPUs because number go up--these people are incapable of understanding the reason these instructions are not generalizable, let alone performing reproducible experiments to demonstrate such a thing.
@aeva but pavan was, again, actually doing dense numerical operations that could benefit from that, and was excited because he knew the standard tradeoff but thought he could do something nobody else had tried.
unfortunately, brilliant and creative practictioners are rare, and anything like that was immediately drowned out by LLM sloppers a few years later.
SIMD is not evil. we will win.
