a_fancy_kiwi

joined 3 years ago
[–] a_fancy_kiwi@lemmy.world 1 points 13 hours ago* (last edited 13 hours ago)

I can’t tell if you’re actually asking or if your question was rhetorical but I’ll go ahead and answer. This is over simplified but this is basically what is being reported:

The RAM companies know how much RAM each of their factories can make in a given hour, day, month, year, etc.

To keep math simple, let’s say the RAM company owns 1 factory. They can make 100 chips in that factory per year. In January 2026, Company A orders 50 chips, Company B orders 50 chips, Company C orders 50 chips, and Company D orders 50 chips. It’ll take 2 years to complete those orders. Maybe each company gets some chips every month for 2 years or maybe it’s first come first serve. Regardless, 2 years. The factories are at max capacity, they physically can’t make them faster.

The only way to make more chips is to build another factory which takes time. But what if you make a lot of money overcharging customers? What if you think AI might be a bubble and demand will go down? Speaking as a company, building a new factory might not be worth the investment.

[–] a_fancy_kiwi@lemmy.world 1 points 1 week ago* (last edited 1 week ago)

Sure, all articles should be read with skepticism.

It’s not a huge logical leap to go from identifying faces in photos at some confidence level to identifying some diseases in imaging at some confidence level. Your first link mostly talks about the human factor and specifically states that with bad ai models, the radiologist is less effective but with good ai models, the radiologist is more effective which suggests the technology works. That link isn’t commenting on the technology, it’s commenting on the human using it.

Same source claiming the tech does work: Source1 Source2

Your last link is talking about people being able to forge medical documents. To be honest, I don’t get the point you’re trying to make. I agree that, like with most technologies, people can use AI nefariously. However, generating a photo with AI to forge a medical document and using AI as an assistive tool to scan medical imaging for diseases are two different applications.

Those articles you’re reading are written from the press releases of for-profit companies

Sorry FBI, you’ve hacked into the wrong mainframe

[–] a_fancy_kiwi@lemmy.world 2 points 1 week ago (2 children)

Articles come out all the time that mention AI being able to detect different diseases months or years before symptoms present. It can be trained to sift through lifetimes worth of data and recognize patterns. There’s no way that’s not useful.

The wider social, economic, and environmental impacts are definitely issues that should be looked into but the tech itself is a marvel. The AI hate, I think, mostly comes down to chatbots and the companies pushing them.

[–] a_fancy_kiwi@lemmy.world 5 points 1 week ago* (last edited 1 week ago) (4 children)

Corporate AI as a service is cancer. But I don’t completely discount locally hosted AI for certain tasks. If you know it’s capabilities and trade-offs, it can be used as one of many tools in a toolbox.

For example, vibe coding entire projects is shit but I don’t think there’s anything wrong with a locally hosted AI spitting out ~6 lines of boiler plate code that you can edit for your needs. It’s a huge time saver when you don’t code for a living but enjoy it as a hobby. I don’t have unlimited time for hobbies and my local chatbot does let me accomplish some tasks faster.

Edit: for those of you who disagree, I’m curious to know at what scope you’re thinking about AI at. Are you just thinking about the tech and/or larger social consequences and/or environmental impacts etc. If we’re just talking about the tech, the tech is amazing and has its uses but if you’re thinking about it more broadly then I get it.

[–] a_fancy_kiwi@lemmy.world 1 points 2 weeks ago

I do it so infrequently that I just do it manually. Sometimes my friends send too many YouTube links in our group chat. I care enough to respond to them but not enough to watch them all if it’s going to take up a good chunk of my day to do so.

[–] a_fancy_kiwi@lemmy.world 3 points 2 weeks ago* (last edited 2 weeks ago) (2 children)

I use yt-dlp to download the audio from a YouTube video and then use MacWhisper to transcribe, and summarize the audio. Not sure what a windows or linux equivalent of MacWhisper would be

[–] a_fancy_kiwi@lemmy.world 4 points 1 month ago* (last edited 1 month ago)

I have 4x 8TB WD Red Plus drives in my NAS. I started with 2 and added 2 later. 1 of the original 2 failed a little under the 4 year mark. I replaced it with another 8TB WD Red Plus drive. I'm happy with them. I wouldn't say they get particularly loud unless I'm writing a bunch of data to them.

SMART Power_On_Hours (converted to years):

sda = 5.42
sdb = 1.58
sdc = 3.85
sdd = 3.85
[–] a_fancy_kiwi@lemmy.world 18 points 1 month ago (1 children)

Damn…I feel bad doing this but maybe you can cut your losses now and invest time in another project :/

https://immich.app/

[–] a_fancy_kiwi@lemmy.world 13 points 1 month ago

I’ve been pretty happy with Kavita as the server. You can read the books directly from the web app but if you think you might end up in an area with no signal, you can download the book from Kavita, and use a separate eReader app to read the file locally.

[–] a_fancy_kiwi@lemmy.world 5 points 2 months ago

And file away your scanned papers separately

This^. No matter how many layers of backups I have for paperless, I'm still keeping the most important physical documents in a file cabinet.

31
Raid Z2 help (lemmy.world)
submitted 5 months ago* (last edited 5 months ago) by a_fancy_kiwi@lemmy.world to c/selfhosted@lemmy.world
 

tldr: I'm going to set up raid z2 with 4x8TB hard drives. I'll have photos, documents (text, pdf, etc.), movies/tv shows, and music on the pool. Are the below commands good enough? Anything extra you think I should add?

sudo zpool create mypool raidz2 -o ashift=12 /dev/disk/by-id/12345 ...

zfs set compression=lz4 mypool #maybe zstd?
zpool set autoexpand=on mypool
zpool set autoreplace=on mypool #I might keep this off. I can see myself forgetting in the future
zpool set listsnapshots=on mypool

With ai raising hard drive prices, I over spent on 3x10TB drives in order to reorganize my current pool and have 3 hard drives sitting on a shelf in the event of a failure. My current pool was built over time but it currently consists of 4x8TB drives. They are a mirrored stripe so a usable 16TB. If I understand it correctly, I can lose 1 drive for sure without losing data and maybe a second drive depending on which drive fails. Because of that, I want to move to raid z2 to ensure I can lose 2 drives without data loss. I'm going to move data from my 4x8TB drives, to the 3x10TB, reconfigure the 4x8TB, and move everything back. I run Immich, plex/jellyfin, and navidrome off the pool. All other documents are basically there for long term storage just in case. What options should I use for raid z2 when setting it up?

I know I can look this stuff up. I have been and continue to do so, I was just hoping for some advise from people that are more knowledgeable about this than me. The move from the 4x8TB drives to the 3x10TB is going to take ~3 days so I really don't want to mess this up and have to start over 😅

Edit:

After looking up each property, this is the command I will probably end up using to create the raid z2 pool, thanks Avid Amoeba:

~~sudo zpool create
-o ashift=12 -o acltype=posixacl -o xattr=sa
-o compression=lz4 -o dnodesize=auto -o relatime=on
-o normalization=formD
raidz2
mypool
/dev/disk/by-id/12345 ...~~

Edit2:

Above command didn't work on my machine. The order and uppercase "O" matters. Had to do this:

sudo zpool create \
  mypool \
  raidz2 \
  -o ashift=12 -O compression=lz4 \
  -O normalization=formD -O acltype=posixacl \
  -O xattr=sa -O dnodesize=auto \
  -O relatime=on \
  /dev/disk/by-id/12345 ...

Edit3:

And finally, after all this, I set up my tmp pool of 3x10TB disks as a raid z2 instead of raid z1. Spent a day and a half transferring before I finally saw my mistake after running out of space 🫠

 

My friends are open to leaving Discord which has finally given me a reason to look into Element/Matrix. I found the install instructions and am immediately put off. Is this it? No official docker compose? 😞

 

I recently noticed that htop displays a much lower 'memory in use' number than free -h, top, or fastfetch on my Ubuntu 25.04 server.

I am using ZFS on this server and I've read that ZFS will use a lot of RAM. I also read a forum where someone commented that htop doesn't show caching used by the kernel but I'm not sure how to confirm ZFS is what's causing the discrepancy.

I'm also running a bunch of docker containers and am concerned about stability since I don't know what number I should be looking at. I either have a usable ~22GB of available memory left, ~4GB, or ~1GB depending on what tool I'm using. Is htop the better metric to use when my concern is available memory for new docker containers or are the other tools better?

Server Memory Usage:

  • htop = 8.35G / 30.6G
  • free -h =
               total        used        free      shared  buff/cache   available
Mem:            30Gi        26Gi       1.3Gi       730Mi       4.2Gi       4.0Gi
  • top = MiB Mem : 31317.8 total, 1241.8 free, 27297.2 used, 4355.9 buff/cache
  • fastfetch = 26.54GiB / 30.6GiB

EDIT:

Answer

My Results

tldr: all the tools are showing correct numbers. Htop seems to be ignoring ZFS cache. For the purposes of ensuring there is enough RAM for more docker containers in the future, htop seems to be the tool that shows the most useful number with my setup.

 

This is a continuation of my other post

I now have homeassistant, immich, and authentik docker containers exposed to the open internet. Homeassistant has built in 2FA and authentik is being used as the authentication for immich which supports 2FA. I went ahead and blocked connections from every country except for my own via cloudlfare (I'm aware this does almost nothing but I feel better about it).

At the moment, if my machine became compromised, I wouldn't know. How do I monitor these docker containers? What's a good way to block IPs based on failed login attempts? Is there a tool that could alert me if my machine was compromised? Any recommendations?

EDIT: Oh, and if you have any recommendations for settings I should change in the cloudflare dashboard, that would be great too; there's a ton of options in there and a lot of them are defaulted to "off"

 

tldr: I'd like to set up a reverse proxy with a domain and an SSL cert so my partner and I can access a few selfhosted services on the internet but I'm not sure what the best/safest way to do it is. Asking my partner to use tailscale or wireguard is asking too much unfortunately. I was curious to know what you all recommend.

I have some services running on my LAN that I currently access via tailscale. Some of these services would see some benefit from being accessible on the internet (ex. Immich sharing via a link, switching over from Plex to Jellyfin without requiring my family to learn how to use a VPN, homeassistant voice stuff, etc.) but I'm kind of unsure what the best approach is. Hosting services on the internet has risk and I'd like to reduce that risk as much as possible.

  1. I know a reverse proxy would be beneficial here so I can put all the services on one box and access them via subdomains but where should I host that proxy? On my LAN using a dynamic DNS service? In the cloud? If in the cloud, should I avoid a plan where you share cpu resources with other users and get a dedicated box?

  2. Should I purchase a memorable domain or a domain with a random string of characters so no one could reasonably guess it? Does it matter?

  3. What's the best way to geo-restrict access? Fail2ban? Realistically, the only people that I might give access to live within a couple hundred miles of me.

  4. Any other tips or info you care to share would be greatly appreciated.

  5. Feel free to talk me out of it as well.

EDIT:

If anyone comes across this and is interested, this is what I ended up going with. It took an evening to set all this up and was surprisingly easy.

  • domain from namecheap
  • cloudflare to handle DNS
  • Nginx Proxy Manager for reverse proxy (seemed easier than Traefik and I didn't get around to looking at Caddy)
  • Cloudflare-ddns docker container to update my A records in cloudflare
  • authentik for 2 factor authentication on my immich server
 

I've been interested in building a DIY NAS out of an SBC for a while now. Not as my main NAS but as a backup I can store offsite at a friend or relative's house. I know any old x86 box will probably do better, this project is just for the fun of it.

The Orange Pi 5 looks pretty decent with its RK3588 chip and M.2 PCIe 3.0 x4 connector. I've seen some adapters that can turn that M.2 slot into a few SATA ports or even a full x16 slot which might let me use an HBA.

Anyway, my question is, assuming the CPU isn't a bottle neck, how do I figure out what kind of throughput this setup could theoretically give me?

After a few google searches:

  • PCIe Gen 3 x4 should give me 4 GB/s throughput
  • that M.2 to SATA adapter claims 6 ~~GB/s~~ Gb/s throughput
  • a single 7200rpm hard drive should give about 80-160MB/s throughput

My guess is that ultimately, I'm limited by that 4GB/s throughput on the PCIe Gen 3 x4 slot but since I'm using hard drives, I'd never get close to saturating that bandwidth. Even if I was using 4 hard drives in a RAID 0 config (which I wouldn't do), I still wouldn't come close. Am I understanding that correctly; is it really that simple?

view more: next ›