They already have started these layoffs. My cousin got hit yesterday.

Spinning media is slower than solid state NAND storage

Generalizable or not, the issue is that as statistical machines LLMs are really only capable of producing likely output. To some degree this firmly limits the output to being unremarkable.

No one can help tell you that, unfortunately. The hype squad will insist we’re at the inflection point of exponential growth and point to Anthropic using Claude to come up with novel optimizations for training Claude. The haters will insist that a statistical word model will never be useful because it’ll never be accurate, and they’ll never find a way to make inference cheap enough to be viable.

Turns out if your job is selling 95% correct words (generic business people, mediocre writers, etc), then a word guessing machine that does a 95% good enough job may be sufficient to have value. But if your proposition is in being actually good at something, then you’ll continue to outcompete the machines. It’ll never be able to displace classics, but I can’t imagine a world where it doesn’t displace human-slop erotica, murder mystery, etc. where the quality was less important than volume.

As to whether it’ll plateau, I’m not sure. It’ll obviously never be anything other than a word guessing machine, but it turns out there’s a lot of interactions with the world that can be described with words. About 3-6 months ago people generally noticed a big change in how well things like Claude code worked, and it had nothing to do with model improvements. Maybe the same model + new secret sauce can yield yet another improvement.

Anthropic and OpenAI will never recoup their burned CapEx but that doesn’t strictly mean that inference will never be cheap enough to be viable. If the big model shops explode from bad economics, it’ll still leave behind open source models that could be “good enough” that could run on your local machine. Is that going to be sufficiently “good enough” for the 95% correct use cases? I dunno. No one knows.

So… I dunno. For you, all this means is we’re gonna have a big economic reckoning at some point and if you’re good but not great at your job, your job involves selling words, and your job doesn’t require 100% precision, I’d reckon there’s a chance LLMs stick in your world, to some degree.

And before I get banned for this take, please note I’m not justifying the massive ecological and economic damage the process has and will do on the way to whatever end it reaches. Reactivating coal plants for power to feed an economic bubble that also itself destroys the environment is almost certainly too high a price to pay to replace human slop with AI slop.

Are you running gluetun on docker desktop for windows by chance?

If so, the networking layer for DD is fucked and every new release fucks it further and differently. Make sure you’re not running any other VPNs on the host, restart the docker daemon and pray. And then move off windows before it’s too late.

Source: made the mistake of starting an *arr stack on DD for Windows cause I didn’t wanna figure out GPU pass-through on VMs, and I suffered with this for a year.

claude performs acceptably at repetitive tasks when I have an existing pattern for it to follow. “Replicate PR 123, but to add support for object Bar instead of Foo”. If I get some of this busy work in my queue I typically just have claude do it while I’m in a meeting.

I’d never let it do refactors or design work, but as a code generation tool that can use existing code as a template, it’s useful. I wouldn’t pay an arm and a leg for it, but burning $2 while I’m in a meeting to kill chore tasks is worth it to me.

Neither do I - I use either my phone, or my smart TV, or my fire stick. SSO works fine there, or you can use the QR based session transfer to SSO on your phone and then “sign in on another device” or whatever by scanning the QR your other device is showing. I think they call it quick connect or something.

It does what you want.

And if you think Grandma can’t figure out scanning a QR code, Grandma is also not gonna figure out MFA lol.

You can run the OIDC version and use SSO and implement MFA on the IdP. I use Keycloak for SSO w/ MFA and users sign into my Jellyfin via Keycloak. Just disable username/password auth and leave it SSO only.

The only benefit Plex really has is the relaying, but I was able to sync watch with 3 people basically as far across North America as you can get from me and it worked without issue so…

Of course it is, LLMs are inherently regurgitation machines - train on biased data, make biased predictions.

You can export images to tarballs and import them to your local docker daemon if you want.

Not sure how podman manages local images.

Idea:

1. Setup inotifywait on your sources directory and rsync to keep directories in sync, and on change, build and run latest image
2. Setup inotifywait on your image tarball directory/file and on change import and run latest
3. Mount your source directory to a docker server image that supports hot reloading.

Per the email text and Plex’s policy, they are correct - only Server Owners need the Pass.

That said I moved to Jellyfin months ago when they announced it.

My humble experience so far: https://kcross.engineering/blogs/matrixandmautrix/

Biggest thing so far is “go slow on federation”. Large federated servers are where you get into trouble with resource requirements and needing to spin up workers, etc. Small, private servers are relatively easy.

My humble experience so far: https://kcross.engineering/blogs/matrixandmautrix/

Biggest thing so far is “go slow on federation”. Large federated servers are where you get into trouble with resource requirements and needing to spin up workers, etc. Small, private servers are relatively easy.

I’m messaging Facebook users over Matrix via the bridge.

I can try to write some stuff up, it’s not super complex. Core requirement for my setup is Docker + a domain. I recommend Linux host but you can make Docker Desktop work.

Let me write some stuff down this week.

I’m running a Matrix server with a FB Messenger bridge via mautrix-meta and that makes it a clear winner. Half my group chats have migrated entirely since I’ve set my close friends up with accounts in my server and they also use the bridge. The fact that people can slowly migrate chats without losing messages or groups is killer for adoption imo.

Yea I’m gonna do zfs or something when I get set up properly again. I’ve got 2 16TB HDDs and Storage Spaces won’t let me pull a drive out :v

I think I’m gonna have to make a new Storage Space and slowly grow that one and shrink the other as I basically shift the extra storage budget between the two until the data is on just one of my drives without redundancy, and then I’ll pull that drive, dual boot Ubuntu or something, format, get everything prepared, and then mount, copy, start services, and then go back and kill the old storage spaces and then never run Windows for anything meaningful again.

I need to migrate off Docker Desktop for Windows and Storage Spaces but I fear the process will be difficult due to my data volume and the stupidity of Windows. I should never have gone Windows, but I wanted to use Steam Big Picture off the media PC and didn’t want to deal with getting that functional on Linux.

But Docker Desktop for Windows keeps crashing WSL and bricking the network devices randomly, and also continuously grows memory consumption until the machine reboots. Piece of shit.

This has also been my personal experience with their brand.

If I’m shopping in a new space where I can’t find a clear winner in reviews, I consider Anker to be the safe bet for a good product. Let’s hope they never logitech themselves.

“A Lightweight VPN Built on top of IPFS + Libp2p”

Seems like both at a glance