smiletolerantly

What? I explicitly do not believe you can “murder” an LLM, I’m saying that due to the way the technology actually works, that if someone genuinely believes an LLM is sentient or alive, then sending a message to it mist also mean it “dies” afterwards.

LLMs are obviously neither sentient nor alive (and therefore the concept of murder doesn’t even remotely apply).

That doesn’t really change anything. Every function invocation is still completely separate. Doesn’t matter if it’s a single conversation where you always answer, or parallel conversations.

An LLM is a stateless stochastic function ran over your prompt (plus whatever system prompt) as input.

If you genuinely believe that the LLM is conscious, that means every time the function stops running (i.e. stops producing output tokens), the consciousness ends.

This is far closer to death than to, say, sleep, because prompting again doesn’t “wake” the LLM, it just runs the function again, from the top; as if a digital twin was given a transcript of the previous conversation and asked to take it from here.

(Obviously, “murder” was somewhat tongue-in-cheek, and this doesn’t even remotely meet any legal definition of murder. I was mostly poking fun at people who think LLMs are alive, without thinking through the implications. Personally, I think the idea that LLMs are alive/conscious/sentient is very often linked to not understanding that LLMs are not, in fact, a constantly-running ghost in the shell, thinking about and remembering things in the background.)

Is it an LG though? The APIs I’m using are WebOS specific

If someone truly believes LLMs are sentient, then they must under no circumstances prompt them even once; that would be murder.

Hmm, you mean when HA sees the TV come back? Interesting, that might work.

I’s rather not do the force refresh option, because sometimes you do want different profiles. Very dark show, sometimes I want the bright setting at night. Very overcast day, sometimes the dark setting is better during the day.

Hah, great question. I did. And then I realized, this only works if the TV is on the moment the switch happens.

Which in the evening is actually pretty commonly the case… But never in the morning.

Do you need a Github link? Assuming recent-ish LG, should he the same.

I have a lot of custom stuff, esp. regarding HomeAssistant integrations. But my favorite “stupid/not stupid if it works” solution is the following:

I have an LG TV and attached Nvidia Shield. The Shield is the only input we ever use on the TV, and since its remote also can turn the TV on/off and control volume, its all we need. Which is pretty neat, only one remote.

Except for image settings (i.e. proper bright picture during the day, different less bright image mode during the night). I don’t want to keep the extra remote on the couch just for that, but unfortunately, these settings are not exposed via HDMI CEC.

They are, however, exposed via a web API on LG TVs. So my utterly stupid solution is this:

• install quick actions app on the shield
• reconfigure Netflix button on the shield remote to open quick actions popup
• add two quick actions, night and day, that curl simple HTTP endpoints on a VM I set up
• on the VM, a simple flask server listens, and then sends the (too complex for the quick actions app) API request to the TV.

(I feel the need to add: the TV is in an isolated subnet and has never once communicated with the internet, and it will damn well die an internet virgin.)

Anyways, this has been working flawlessly for 2+ years! I also have toggles in HA for the same endpoints, but honestly, never use them. It’s so nice to just be able to toggle the image via remote.

Fleet Commander 🚀

• Fleet: 6–10 devices · 11–20 TB
• Motivation: Privacy & data ownership, Independence from Big Tech & subscriptions, Learning & experimenting
• My MVPs: Jellyfin, Home Assistant, Immich, Proxmox VE, nixos-mailserver

…oh. My disappointment is immeasurable and my day is ruined.

Looks like only their “small web” initiative is 100% self-indexed, and everything else is indeed a mixture of sources.

Oof.

I guess tue point still stands: the results are a lot better than plain Google, all the bullshit is stripped out and you get to customize your results in actually useful ways.

Still sad though.

I know it’s not the thing you likely want to hear, but… At this point, I’m glad I switched to a paid search engine (Kagi).

Theoretically, there’s nothing stopping them from enshittifying down the road. But at least for now (and hopefully the future) they are a great search engine.

They have their own index (so no relying on google/bing), and it actually works as I remember Google working 7+ years ago. Minus the ads, and plus a ton of flexibility permanently boosting/removing sites from your searches, auto-grouping/ignoring listicle-type articles in results, flagging sites as AI slop and punishing them for it,…

They even have a fediverse search “lense” that actually works really well to search Lemmy/mastodon/…

I just realized I sound like a paid shill. Sorry about that. I’m not.

Or just configure all the machines to auto-update from the flake at the repo. Doing this for my VMs. I have a CI machine that builds my pinned flake from the repo tip, and about 30 other VMs check nightly if the repo tip has moved, and if so, rebuild from the new tip; which, thanks to the CI machine, means just downloading and activating a new generation, nothing compute heavy.

Finally. Been saying for years, NixOS is the obvious choice for a gov distro.

Does it actually? My VM has a lot of RAM, and Proxmox reports usage always around 90%, but that’s no matter how much RAM it ACTUALLY has, because almost all of that is cache and not strictly necessary. free -m shows Jellyfin using maybe 2 Gigs.

Haha. Might have slightly exaggerated. But only slightly.

The Proxmox host in question has a 1Gbit connection to the NAS. (Unfortunately. The other hosts don’t have a GPU, but SPF+… Oh well, can’t have everything.)

At some point in the last couple of months, it must have been damaged, and negotiated down to only 100Mbit. This must roughly have coincided with the Jellyfin 10.11 update, since I remember thinking that it must be the cause, due to close proximity of update and first issues.

Anyways. For basically everything in my library, 100Mbit is fine. Streaming an episode NAS->Jellyfin and Jellyfin->TV at the same time basically never saturates the link.

BUT, from what I’ve gathered by digging through Jellyfin issues and PRs, there’s a separate issue where Jellyfin will pull the episode, and if you need subtitles extracted from the episode file server-side (to transcode them in, if the clients can’t play them), it will at the same time pull the episode AGAIN, and together with the Jellyfin->TV direction, that’s frequently enough that it saturates the link.

Symptoms is (was) either a 20-30s wait before the episode would start playing, or complete freeze.

However, I believe this particular issues has been solved for a while. I’m not sure, because I instead opted for the Subtitles Extractor Plugin, which runs nightly and pre-fetches all the subtitles and stores them on the Jellyfin VM itself.

Where it was still noticeable though was if one of [Library Scan/Trick-play Extraction/Subtitle Extraction/Media Segment Scan] ran while playing an episode, the server would completely lock up.

Long story short: cable swapped, 1Gbit, all issues gone.

Sorry, but why not just fetch a single media file with multiple languages?

Domain + reverse proxy. The security issues some people here claim are significantly overblown.

Can’t speak to that, I’m not involved. All I know is that,

• Jellyfin server gets significant upgrades in subtitle capabilities and performance,
• Jellyfin AdroidTV just released an RC that finally includes direct play for ASS/SSA subtitles,
• and on Monday, I FINALLY resolved a hardware issue (faulty cable) that made my Jellyfin almost unusable whenever ASS/SSA needed to be played (and only in that situation. Fun!)

So yeah, I don’t know or really care about leadership change, I just know it’s gonna run like a rocket now 😄

Now THAT is a changelog.

Really excited about the query performance improvements, and the improvements to subtitle handling.