The pricing is just insane. Its $450-500AUD for a 16GB pi5 (without storage), and I recently got a secondhand nuc with 16GB and storage for $200AUD.
Pis also don't seem to have a second hand market (which as a hoarder of pi3's I am part of the problem).
I had a pi4 booting from a USB ssd, but it was very flakey. Just a slight bump on the SSD and it would disconnect and never remount, so I wasn't super confident with it. Happened with a few different SSDs on different ports as well.
As for power, my pi4 + SSD was chugging down 15w, and the nuc that replaced it was also around 15w when limited properly, so it isnt as clear cut as it used to be. That was with identical services, homeassistant and a few local web services. I suspected the pi4 was never really able to idle.
Honestly, AI pricing or not, raspi just isnt a good time for a home server. I used to churn through SD cards on mine, and the software compatibility was quite annoying, particularly with docker images. Intel x86 minipcs have been far easier to deal with IMO. And $40 is a bargain.
Yeah, the longevity for tape backups feels entirely theoretical when the tape drives themselves are out of production, so you could end up with a perfect tape, and nothing can read it...
Yeah, same, but every time I look at second hand stuff its crazy expensive, or completely obsolete.
True in general, but I am not convinced its true for OpenAI or anthropic.
Their training data is probably already replicated across multiple datacenters, so tape archiving is less critical. And if tape archiving is critical, then surely they already have sufficient backups already right? And its not like the organic content of the internet is growing anymore, so there isn't going to be an explosion of new organic data they need to store.
So I don't buy that they need this. I think this is fujifilm trying to get in on the bubble, and nothing more.
RAT is Remote Access Trojan isnt it?
I buy the follow on effects rational, but I'm not convinced that tapes are a suitable alternative to HDDs. The random IO basically doesnt exist. But I guess if there are companies using HDDs for "cold" storage then using tape makes some sense for them.
My personal theory for this announcement is that its pure stock price manipulation. AI investing is full on fomo driven, so announce that your company is investing in X that will help AI, and watch the chuds hit buy in robinhood.
Okay, no need to name call. Tapes are for long term archiving. The access speed is minutes to hours, so these are definitely not for caching.
Thats exactly my point, why bother archiving when you need it accessible for training? Surely they would be putting the training data on a faster medium.
The ship has sailed on the front. But even if its for the corpus of training data, surely that can't be stored on tape, the access time would be way too slow.
But thats pointless archiving it on tape, the access time is going to handicap the training speed.
Seems like its been pinned there somehow, despite being 2 days old. Happens when logged out as well.
Those of you running behind, keep at it, but this is the leaderboard as it stands at the end of 2024.
Happy new year!
Well, that was a month. Congrats everyone who has reached the end, and thanks to everyone who has contributed solutions and advice.
Sometime in January I will create a megathread for visualizations. If anyone has any other ideas, happy to hear them, otherwise, take a well earned 11 month rest until next year :D
