So, if the AI bubble pops, it'll be a great time to build a PC
post
What else are these data centers going to hoard?
Jobs, GPUs, water, hard drives
Electricity
Ironic Microsoft doesn’t have enough electricity to power their hoarded gpus.
Remember a year or so ago when they all spun down production so they could charge more money for drives? I do.
AI crap. Infesting everything. Search of all kinds, photo management, telephone menus, who knows what else. And it does none of it well.
I don’t have issues with local AIs, for things like searching your local immich instance, or controlling your local Home Assistant devices. That photo of a bird you took 3’ish years ago? Yeah, you can find it in like three seconds with a local AI search. Want to turn the lights on with a voice request? AI is one of the easiest ways for a layman to handle the language processing side of things. All of that is a drop in the ocean.
But corporations have been trying to cram it into everything, even when it’s not a good fit for what they want to do. And so far, their solution to making it fit hasn’t been to rethink their usage and consider whether or not it will actually improve a product. Instead, their approach has simply been to build more and bigger data centers, to throw increasing amounts of processing power at the problem.
The technology itself isn’t inherently harmful on the small scale. But it has followed the same pattern as climate change. Individual consumers are blamed for climate change, and are consistently urged to change their consumption habits… When it’s actually a handful of corporations producing the vast majority of greenhouse emissions. Even if every single person drastically changed their emission habits, it would barely make a dent in the overall production. It was all because of massive astroturfed PR campaigns to shift the blame away from those companies and onto individuals. And we’ve seen that same thing happen with AI, where individual users have been blamed for using AI, instead of the massive corporations.
Think of all the cheap hardware being resold when the AI bubble pops.
There wasn’t as big of a price drop as I thought there would be when the crypto mining switched to ASIC from GPUs. Don’t know if all that hardware just got dumped or is sitting in a rack rotting somewhere. Hope that we get cheaper prices when the bubble pops, this artificial scarcity sucks.
Hope that we get cheaper prices when the bubble pops, this artificial scarcity sucks.
Not likely. Why would they give up money?
I think your sentiment and the back end requirements of AI is a big downfall of it, as while your sentiment has validity in many public facing deployments of it there are some things it is actually succeeding at. I speak from experience having used it for several specific use cases which it excels at, but you and others probably don't have time nor care that this is true. And again marketing idiots out weight the deliberate approach that engineers and others might want, much less the economy might need.
Preaching down to folks who already have a good reason to hate your product is a good way to join that pile
Having done my research, and tried it, I'm not an ignorant F*** as I read and engage at least. You just like being a troll.
Man go fuck yourself
HaViNg DoNe My ReSeArCh!
What's the over/under that they're a Libertarian?
cases which it excels at
Uhuh
Like how my colleagues often come to me saying they fixed it with Cursor and when I check the UI where the bug was, the page doesn't even load at all now.
I just had someone tell me they did something with AI and when I checked they didn't even get right the very basic thing of coding around the right controller names. The fucking names were wrong. They didn't even check the feature, they just shipped, called it fixed, and told me Cursor figured it out really quickly.
I'm tired of this. I'm REALLY tired of this, man...
I posted this as a perspective on AI that is not given by AI, nor by someone who believes it will stay this way, but nor am I promoting it. I believe it's more nuanced that just being crap, although it is taking over many things in life. I have used it, I know how to use it for good (keeping it private, local, and to help teach reasoning as well as do the thing that we need done (like dishes, bills, and other bullshit). I'm fully aware it's a bubble (14 billion to 1.4 trillion for OpenAI alone), dislike it and hate the energy waste. You all just seem to want to keep up the ignorant web user stereotype.
Have fun down voting something you don't really understand.
First outrageous DDR5 RAM prices now ssd's.
Welp. Won't be upgrading my pc for the next few years I see
RAID5, don't fail me now!
As a PSA: depending on your needs (bcs I see a lot of simple homelabs RAIDing), consider second server or even second location simple snapshot backup vs same-machine duplication.
Like, which would you want to set up first depending on your specific risks profile.
The later is like like a proper backup, the former more of a low-downtime strat (that most smol homelabs can do without - "just wait a day, mom")
Eg as the most basic example - instead of local duplication you could have a small PC (even like some old Pi), big HDD, and rsync once a day.
(Ofc, with duplicated serves, finances permitting, you can also provide backup services which helps that downtime issue. And by finances I mean the rest of your server group stuff, which in my cases is mostly HDD cost anyways.)
Ram shortage too. But most users should avoid QLC.
This collapse when we only need 1/100 of what they are planning. Because not everyone is going to need a massive data center. Just like cars, computers and every other technology. It starts out diverse than shrinksssssssss.
I noticed the price doubled since the last time I ordered a drive, like a year ago.
Fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you, fuck you
Oh this must be for training stolen data
When it all comes crashing down, at least the Internet Archive could have easy & cheap access to it all. I trust them to handle it more responsibly than the AI bros.
Oh for fucks sake… I wanted to expand my NAS… Crypto was (is still responsible) for the shit state of the GPU Market and now it’s the next scam.
Just got 12TB IronWolf last week for an average price if 260euro. Delivery is long but other than that, price is average.
Checked just right now - average 310euro for the same HDD. From the place I picked mine - 293. SCORE!
Thinking about this for a minute - that's probably pre-Black-Friday prices. We'll see
Imagine how I feel right now, I bought 16 used 20tb Seagate drives around September last year for $200 each, from Amazon, and finally built my dream NAS after a decade of waiting and planning. I don't understand how I got so lucky but I did. They're all working just fine, and I somehow avoided having to wait another decade to finally build my NAS due to random capitalism gold rush bullshit. Praise be.
Is IronWolf a website?
Note: Posted by the same media outlet that reported last week about the 9700X3D with zero fact checking
TL;DR
QLC drives have fewer write-cycles than TLC and if their data is not refreshed periodically (which their controllers will automatically do when powered) the data in them gets corrupted faster.
In other words, under heavy write usage they will last less time and at the other end when used for long term storage of data, they need to be powered much more frequently merelly to refresh the stored states (by reading and writting back the data).
So moving to QLC in cloud application comes with mid and long terms costs in terms of power usage and, more importantly, drive end-of-life and replacement.
--
Quad Level Cell SSD technology stores 4 bits per cell - hence 16 levels - whilst TLC (Triple Level Cell) stores 3 bits - hence 8 levels - so the voltage difference between levels is half as much, and so is the margin between levels.
Everything deep down is analog, so the digital circuitry actually stores analog values on the cells at then reads them back and converts them to digital. When reading that analog value, the digital circuit has to decide to which digital value that analog value actual maps to, which it does by basically accepting any analog value within a certain range aroun the mathematically perfect value for that digital state.
(A simple example: in a 3.3V data line, when the I/O pin of a microcontroller reads the voltage it will decide for example that anything below 1.2V is a digital LOW (i.e. a zero), anything above 2.1V is a HIGH (a one) and anything in between is an erroneous value - i.e. no signal or a corrupted signal - this by the way is why if you make the line between a sender and a receiver digital chip too long, many meters, or change the signals in them too fast, hundreds of MHz+, without any special techniques to preserve signal integrity, the receiver will mainly read garbage)
So the more digital levels in a single cell the narrower the margin, the more likely that due to the natural decay over time of the stored signal or due cell damage from repeat writes, the analog value the digital circuitry reads from it be too far away from the stored digital level and be at best marked as erroneous or at worse be at a different level and thus yield a different digital value.
All this to say that QLC has less endurance (i.e. after fewer writes the damage to the cells from use causes that what is read is not the same value as what was written) and it also has less retention (i.e. if the cell is not powered, the signal decay will more quickly cause stored values to end up at a different level than when written).
Now, whilst for powered systems the retention problem is not much of an issue for cloud storage (when powered, the system automatically goes through each cell, reading its value and writting it back to refresh what's stored there back to the mathematically perfect analog value) with just a slightly higher consumption over time for data that's mainly read only (for flash memory, writting uses way more power than reading), the endurance problem is much worse for QLC because the cells will age twice as fast over TLC for data that is frequently written (wear-leveling exists to spreads this effect over all cells thus giving higher overall endurance, but wear-leveling is also in there for TLC so it does not improve the endurance of QLC).
I've been waiting over a month for my MicroCenter to restock Western Digital 14-16 TB HDD to upgrade my NAS. I finally caved and ordered direct from Western Digital. It still took them 10 days to ship the order.
Seagate drives continue to remain in stock but i'd like to reduce my chances of premature drive failures.
Not sure I trust WD after the whole lying label NAS thing, the wiped drive accident, and more.
those are interesting links. I just went off of backblaze reports for hdd failures, and it looked like Western Digital had some of the best results
Why HDD’s?
I thought LLMs ran on a fuckload of VRAM and thats pretty much it. So the GPU market was the main affected?

top 97 comments