adgjlsfhk1

7 ms·

91.
by adgjlsfhk1 5mo ago
the math doesn't work. a starlink satellite has ~10kw power consumption. A single ai optimized server rack (GB300) is 140kw. Starlink works because you get a massive benefit from putting networking in space for rural users. no one has
92.
by adgjlsfhk1 5mo ago
given that SpaceX is choosing what price they're charging starlink, there's a reasonable argument that starlink isn't profitable either
93.
by adgjlsfhk1 5mo ago
they aren't being specially punished. they are being made to follow the rules that quickly to every other company that IPOs. These rules aren't arbitrary. They exist because without them, retirement accounts would be vulnerable to
94.
by adgjlsfhk1 5mo ago
we're still at least 3 years too early for that. games usually are in a 5+ year dev cycle, so even if AI made gamedev 2x faster, we're still not at the point where the first opus 4.5 games are out
95.
by adgjlsfhk1 5mo ago
going from burning fuel 12 months of the year to 3 is still a 75% cut in fuel costs and emissions
96.
by adgjlsfhk1 5mo ago
Counterpoint: storage and ports should be upgradable. Ram and CPU both last forever, but drives die and ports wear out.
97.
by adgjlsfhk1 5mo ago
I think the clear demonstration of this is how small Apple's motherboard is for the neo (and other M series) compared to everyone else). It really seems like the PC makers don't understand the benefits of low power chips sufficien
98.
by adgjlsfhk1 5mo ago
it's not just factoring, but general discrete log over abelian groups
99.
by adgjlsfhk1 5mo ago
what about having sram that has extra bits due to claimed ecc that instead is used as extra ram?
100.
by adgjlsfhk1 5mo ago
> For example, the different NaN values could hold debugging information to track down errors. Unfortunately IEEE didn't bother specifying NaN propagation semantics so it ended up pretty useless.
101.
by adgjlsfhk1 5mo ago
yeah. a lot of the issue with Linux's approach is that until recently, the kernel was the one making the choice, and it doesn't know which processes matter. The part Linus does a lot better if not getting to oom in the first place
102.
by adgjlsfhk1 5mo ago
A lot of the remaining is multiples of 4, which you can either get from having a 2 in both factors or a 4 in one (multiples of 9 are similar).
103.
by adgjlsfhk1 5mo ago
they're a company that doesn't make any products. they just get actual hardware engineers to pay them to put a logo on their devices and to prevent Dolby from suing
104.
by adgjlsfhk1 5mo ago
because this is what causes Windows to use ~80% more memory than unixes
105.
by adgjlsfhk1 5mo ago
does landing a plane early due to a bomb threat seem reasonable? either there is a bomb, in which case landing early won't help, or there isn't, in which case landing early won't help
106.
by adgjlsfhk1 5mo ago
I think there's some chance we get a "proof of purchase" system where there is some entity that takes a $10 payment to give out a unique identity token that you need to present to visit most sites. if you have a revocation pr
107.
by adgjlsfhk1 5mo ago
Sorry, I have a patent on questioning whether open source codecs are parent encumbered. Venmo me $1000 or you will be speaking to my lawyers
108.
by adgjlsfhk1 5mo ago
WUFFS like approaches work better for algorithms like lz77 that are substantially bandwidth constrained. for something like a video codec, the computational intensity is much higher so you need better codegen to reach max speed
109.
by adgjlsfhk1 5mo ago
not really. they're mostly pure assembly and sandboxing assembly isn't really a things
110.
by adgjlsfhk1 5mo ago
Unfortunately, in absence of inlining, compilers mostly respect calling convention even when they don't have to.
111.
by adgjlsfhk1 5mo ago
Based on AV1's trajectory, hardware encode isn't necessary (though it is nice). The current encoder is a reference encoder. Now that the spec is finalized, expect significant speed improvements from production encoders (realtime l
112.
by adgjlsfhk1 5mo ago
cpus that aren't Intel are plenty fast on denormals. Intel is the only one where denormals are 100x slower. (and Intel has fixed that on their new cpus, but only on their e cores)
113.
by adgjlsfhk1 5mo ago
Unwillingness to embrace socialism?
114.
by adgjlsfhk1 5mo ago
It's fine (maybe even good) to allow everything through software, it's just that there needs to be hardware backups for a lot of it.
115.
by adgjlsfhk1 5mo ago
Distilling from a larger model is not only probably cheaper than from data, it's also likely higher quality. There's pretty strong support for the proposition that NNs learn a smoothed and regularized version of the data. The NNs
116.
by adgjlsfhk1 5mo ago
Courts can only handle cases that are brought.
117.
by adgjlsfhk1 5mo ago
How else do you expect them to get continual performance improvements with each generation?
118.
by adgjlsfhk1 5mo ago
I think that trying to make a large platform work with engines that aren't reliable and on a fuel system that you don't have experience with is a really dumb idea. The raptor engine is a pretty giant advance over the merlin engine
119.
by adgjlsfhk1 5mo ago
part of my argument is that if your test platform is 1/4th the size, you get much faster iterations. the path they took meant that every test destroys 39 engines, a huge amount of rocket fuel, and all of their construction. A falcon 9
120.
by adgjlsfhk1 5mo ago
no. the supreme Court got rid of that last year

More ›