red light cameras - at least in australia - are stock standard canon DSLRs… they take images, but not video
there are some newer ones that do things like photos of people using their phones stopped at lights etc, but generally speed/red light and “single purpose” cameras will just be doing stills, and wouldn’t be too useful for anything other than a single photo when the sensor triggers it
yeah i remember that as well! considering the bandwidth netflix takes up i’m not surprised at all! i think it’s like 15% of global internet bandwidth or something crazy?
I’m guessing you dropped a zero or two on the user count
i was being pretty pessimistic because tbh i’m not entirely sure of the requirements of streaming video… i guess yeah 200-500 is pretty realistic for netflix since all their content is pre-transcoded… i kinda had in my head live transcoding here, but also i said somewhere else that netflix pre-transcodes, so yeah… just brain things :p
also added an extra zero to the wattage
absolutely right again! i had in my head the TDP eg threadripper at ~1500w - it’s 350w or lower
my numbers are coming from the fact that anyone who’s replacing all their streaming likely isn’t using a single disk… WD red drives (as in NAS drives) according to their datasheet use between 6 and 6.9w when in use (3.6-3.9w at idle)… a standard home NAS has 4-6 bays, and i’m also assuming that in a typical NAS setup they’re in some kind of RAID configuration, which likely means some level of striping so all disks are utilised at once… again, i think all of these are decent assumptions for home users using off the shelf hardware
i’m ignoring sleep here, because sleep for NAS drives leads to premature failure… this is why if you buy WD green drives for your NAS for example and you use linux, you wdparm to turn off sleep to avoid constantly parking and unparking the heads which leads to significantly reduced life (afaik many NAS products do this automatically, or otherwise manage it)
the top end of that estimate for drives (6 drives) is 41.4w, and the low end (4 drives) is 24w… granted, not everyone will have even those 4 drives, so perhaps my estimate is a little off, but i don’t think 30w for drives is an unreasonable assumption
again, here’s where data centres just do better: their utilisation is spread much more evenly… the idle power of drives is not hugely less than their full speed read/write, so it’s better to have constant access over fewer drives, which is exactly what happens with DCs because they have fewer traffic spikes (and can legitimately manage drive power off for hours at a time because their load is both predictable, and smoother due just to their scale)
also, as someone else in the thread mentioned: my numbers for severs were WAY off for a couple of reasons, but basically
Back of the envelope math says that’s around 0.075 watts per individual stream for a 150w 2U server serving 2000 clients, which looks pretty realistic to my eyes as a Sysadmin.
that also sounds realistic to me, having realised i fucked up my server numbers by an order of magnitude for BOTH power use, and users served
servers and data centres are just in a class of their own in terms of energy efficiency
here for example: https://www.supermicro.com/en/products/system/storage/4u/ssg-542b-e1cr90
this is an off the shelf server with 90 bays that has a 2600w power supply (which even then is way overkill: that’s 25w per drive)… with 22tb drives (off the top of my head because that’s what i use, as it is/was the best $/byte size) that’s almost 2pb of storage… that’s gonna cover a LOT of people with that 2600w, and imo 2600w is far beyond what they’re actually going to be pulling
that’s all irrelevant though… the rule is the rule and they got caught
people should be allowed to have awards for games which only use humans, and if a game is caught cheating they should be disqualified
if they want to compete for some awards, these aren’t the awards for them: there are others
an n150 mini pc - largely considered a very efficient package for home servers - consumes ~15w max without the gpu, and ~9w idle
a raspberry pi consumes 3-4w idle
none of that is supporting more than a couple of people streaming 4k like we’re talking about in the case of netflix
and a single hard drive isn’t even close to what we’re talking about… you’re looking at ~30w at least for the disks alone
as for internet cost, it’s likely tiny… my 24 port gigabit switch from 15 years ago sips < 6w… i can only imagine that’s pretty inefficient compared to today’s standards (and 24 port is pretty tiny for a DC, and port power consumption doesn’t scale linearly)
data centres are just straight up way more efficient per unit of processing than your home anything; it pretty much doesn’t matter how efficient your home gear is, or what the workload is unless you switch it off most of the time - which doesn’t happen in a DC
it’s kinda irrelevant to the make it to production part though: the rule is no gen ai used during development… there’s no ifs, buts, or maybes here: there definitively was, and nobody is denying that
it’s irrelevant whether you agree with the rule or not… the award is for games that didn’t use AI during development. the game should not have originally been in contention for the award
i tend to agree this is the right way to use AI assets, but this isn’t the award for them… it doesn’t matter if it was accidental, if it was removed before release, or anything else
self hosting is wildly less efficient… one of the biggest costs in data centres is electricity, and one of the biggest constraints is electrical infrastructure… you have pretty intense power budgets in data centres and DC equipment is pretty well optimised to be efficient
meanwhile a home server doesn’t likely use server hardware (server hardware is far more efficient), is probably about 5-10y or more out of date, and isn’t likely particularly dense: a single 1500w server can probably service ~20 people in a DC… meanwhile an 800w home server could probably handle ~5 people
add the fact that netflix pre-transcodes their vids in many different qualities and formats, whilst home streaming - unless streaming original quality - mostly re-transcodes which is a very energy-hungry process
heck even just the hard drives: if everyone ran their own servers and stored their content that’s thousands if not hundreds of thousands more copies of the data, and all that data is probably on spinning disks
i’d also say manufacturing the devices probably roughly doubles the carbon footprint (same with the car but we’re trying every trick in the book to figure out where the figure came from)
The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail.
this is not a rust problem… nor was the original problem of code writing entries to a file multiple times, and nor is the thing that made it worse: propagation of the poisoned file
the cloudflare issues were configuration… they have nothing even remotely relayed to any of this
github copilot is fantastic for exactly this reason… completes a few lines, auto corrects, automatic find and replace, automatically fills a 3 line function body that would otherwise be an extra dependency
and these browsers are specifically not that… these browsers are intended to do things like categorise tabs, complete forms, etc automatically without your interaction
of course they’ll ask before they do things they consider destructive, but what they consider destructive and what a malicious actor can use are very different things
some of that is certainly benign, but the point with prompt injection is that it can take benign things and make them plausibly malicious
imo anything firefox-based until servo or ladybird matures
i use waterfox, i know a lot of people use librewolf
remember that switching to anything chromium (like vivaldi, brave) still gives google power. on web standards at least, mozilla still fights for us (and is arguably the only one protecting us from shit like putting tracking stabdards directly into the web spec)
sometimes stores have sales that includes gift cards so you can get $500 of apple credit for 20% off
this is not something i’ve ever encountered, nor something that i’d ever expect from an LLM specifically… some kind of test-writing-specific AI? sure because its metric is just getting the thing to go green… but LLMs don’t really care about the test going green: they simply care about filling in the blanks, so its “goal” would never include simply making the test pass, and its training data has significantly more complete tests than placeholders
or fixing windows by only using WSL and reading the arch wiki
the zip file itself might also be generated (you can just tack random garbage into places in the zip format and it’ll be ignored - which is extremely quick to do), in which case the hash would change… the file itself is important in case it’s an exploit in the unzip program itself, but also the contents of the file is important
yeah they do certainly exist, but bog standard “red light cameras”… ie single purpose cameras are not that kind of problem… imo, as long as they’re deployed to combat actual issues they’re very much a beneficial tool
i think it’s important to differentiate these new kinds of cameras from the single purpose cameras so that arguments against them can be made independently