AI Is a Black Box. Anthropic Figured Out a Way to Look Inside
...Concerning our earlier disagreement about the inner workings of large language models and whether there are 'concepts' stored inside...
- Posts
- 3
- Comments
- 218
- Joined
- 3 yr. ago
- Posts
- 3
- Comments
- 218
- Joined
- 3 yr. ago
Ah, you're right. Nostr uses relays. Now I know what the name stands for. Sounds a bit like your proposal in extreme. The "servers" get downgraded to relatively simple relays that just forward stuff. The magic happens completely(?) on the clients.
I'm still not sure about the application logic. Sure I also like the logic close to me (the user.) The current trend has been towards the opposite for quite some time. Sometimes the explanation is simple: If you do most things on the server, you retain control over what's happening. That's great for selling ads and controlling the platforms in general. On the other hand it also has some benefits for power efficiency on the devices. I'm not talking about computing stuff, but rather about something like Google Cloud Messaging which has the purpose of reducing the amount of open connections and power draw and combine everything into a single connection for push messages. In order to do decide when to wake a device, it has access to to the result of the filtering and message priorization. Which then needs to be done server-side.
I'm also not sure with the filtering of hashtags. I mean if you subscribe to a hashtag. Or want to count the sum to calculate a trend... Something needs to work through all the messages and filter/count them. Doesn't that mean you'd need all Mastodon's messages of the day on your device? I'm sure that's technically possible. Phones are fast little computers. And 4G/5G sometimes has good speed. But l'm not sure what kind of additional traffic you'd estimate. 50 Megabytes a day is 1.5GB for your monthly cellular data plan. A bit less because sometimes people are at home and use wifi... But then they also don't just use one platform, but have Matrix, Lemmy and Mastodon installed. And you can't just skip messages, you'd need to handle them all to calculate the correct number of upvotes and hashtag use. Even if the user doesn't open the app for a week.
I don't quite "feel it". But I also wouldn't rule out the possibility of something like a hybrid approach. Or some clever trickery to get around that for some of the things a social network is concerned with...
Or like something I'd attribute more to edge computing. The client makes all the decisions and tells the edge (router) exactly what algorithm to use to do the ranking, how to do the filtering and when it wants to be woken up... That device does the heavy lifting and caches stuff and forwards them in chunks as instructed by the client.
Hmmh. But how would that then change Mastodon not displaying previous (uncached) posts? Or queries running through the server with it's perspective?
And I fail to grasp how hashtags and the Lemmy voting system is related to a client/server architecture... You could just implement a custom voting metric on the server. Sure you can also implement that five times in all the different apps. But you'd end up with the same functionality regardless of where you do the maths.
And if people are subscribed to like 50 different communities or watch the 'All' feed, there is a constant flow of ActivityPub messages all day long. Either you keep the phone running all day to handle that. Or you do away with any notification functionality. And replicating the database to the device either forces you to drain the battery all day, or you just sync when the user opens the App. But opening Lemmy and it takes a minute to sync the database before new posts appear, also isn't a great user experience.
I'd say we need nomadic identity, more customizability with the options like hashtags, filters and voting. Dynamic caching because as of now Fediverse servers regularly get overwhelmed if a high profile person with lots if followers posts an image. But most of that needs to be handled by servers. Or we do a full-on P2P approach like with Nostr or other decentralized services. Or edge-computing.
I don't quite get where in between federated and decentralized (as in p2p) your approach would be. And if it'd inherit the drawbacks of both worlds or combine the individual advantages.
And ActivityPub isn't exactly an efficient protocol and neither are the server implementations. I think we could do way better with a more optimized, still federated protocol. Same with Matrix. That also provides me with a similar functinality my old XMPP server had, just with >10x the resource usage. And both are federated.
Because with all of that, messaging, email, xmpp, matrix and ActivityPub most of the magic happens on the server. Take email for example. The server takes care to be online 24/7. It provides like 5GB of storage for your inbox that you can access from everywhere. It filters messages and does database stuff so you can habe full text search. Same with messaging. Your server coordinates with like 200 other servers so messages from users from anywhere get forwarded to you. It keeps everything in sync. Caches images so they're available immediately.
That allows for the clients/Apps to be very simplistic. It just needs to maintain one connection to your server and ask if there's anything new every now and then. Or query new data/content. Everything else is already taken care for by the server.
OP's suggestion is to change that. Move logic into the client/App. But it's not super easy. If you now need to be concerned on the client with maintaining the 200 connections at all times instead of just 1 to see if anyone replied... Your phone might drain 200 times as much battery. And requiring the phone to be reachable also comes with a severe penalty. Phones have elaborate mechanisms to save power and sleep most of the time. Any additional network activity requires the processor and the modem to stay active for longer periods of time. And apart from the screen thats one of the major things that draws power.
That's a nice idea but has some pretty obvious technical drawbacks that aren't discussed in the blog article:
The complexity of most networks grows about exponentially with the number of connections between the entities. It gets immensely more computationally expensive that way and you're bound to use lots of additional network traffic and total cpu power that way.
And some (a lot of) people like using social media on their phones instead of a computer. You're bound to drain their batteries real fast by moving application logic there.
Other than that I like the general idea. The Fediverse should be more dynamic. Caching and discovery have some big issues in the current form. That should be tackled and we need technical solutions for that. And the current architecture isn't perfect at all.
Furthermore, if talking about the edge where networks are smarter... Why then move it into the browser which isn't at the edge? Wouldn't that be an argument to invent edge-routers like in edge computing? I mean with c2s you have a server on the one side and a client on the other side with the edge somewhere in between. If you now flip it you end up in a different situation. But there's still nothing at the edge where you could introduce some smarts...
As far as I know you want a web application firewall to block attacks. A reverse proxy is just to proxy requests and doesn't necessarily care if it forwards legitimate traffic or attacks.
Maybe you can find a guide/tutorial on how to set it up?
Usually you need the correct packages installed on your system to enable something like VAAPI or QSV. Then you need a version of ffmpeg with that enabled. And then configure it in Jellyfin correctly.
I don't have any specific insights on how to do it with Fedora. I suppose it's very similar to how it's done on other Linux distros.
Hmm. There is value in both. When I started out with NixOS I read lots of wiki articles. And we all know there is some room for improvement. And I also read several configs of other people to see how things tie together. And to look up things that aren't documented. Nowadays I just put in what I'm looking for and "language:nix" into Github. So there's lots of personal configs that turn up. Sometimes with useful stuff. So I think anything is better than nothing. But obviously if you have kids, prefer them and let other people come up with the detailed wiki articles 😆
Fair enough. I personally think someday someone will have the same niche issue I've already tackled and be happy to stumble over my code while googling it. So I just drop most things I do somewhere for other people to find. Regardless.
But concerning NixOS, I also still need to switch over a few things to agenix and generalize parts of my config before publishing it.
Just put it on Codeberg or Github. Having other people's config for reference is always nice. Especially for beginners.
I'd recommend YunoHost, too. It's pretty beginner friendly and you'll probably get some positive results without learning all at once. I mean you have quite something on your plate if you're learning Linux, Docker, Docker-Compose and maybe networking and Dev-Ops all at the same time.
Out of curiosity: Did you measure the idle power consumption?
A second-hand used laptop. Or an used Intel NUC.
I'd say it's difficult to buy anything new for $100 that's actually worth spending that money.
I'd recommend one if the Mini PCs like
But that's about twice your budget with a decent amount of RAM and some storage. (And way faster than a RasPi.)
Goodbye secularism, goodbye constitution and whoever still likes what's been the original idea behind the foundation of the United States of America.
But actually following Jesus' teaching would be way too progressive. As far as I remember he was basically a hippie, advocating for love, helping each other out and the poor, and strongly against hate and capitalism. And he didn't quite like the old traditions. So I think as a christian as of today you definitely need some counterbalance and some other book to point at to defend your conservatism, egoistcal behaviour and hate towards people who aren't 100% like yourself.
Hmm. I wasn't trying to recommend privileged or non-privileged mode, just trying to use that to single in on the actual issue.
Alright, if it's just av1, maybe try to use a tool like
vainfoto find the supported codecs. I think ffmpeg fails if an unsupported codec is explicitly specified. But take care if Encoding is mentioned. Some hardware has decoding capabilities only.It's a complicated topic. And it also took me 2 whole evenings to get the permissions and everything right. I'm using systemd-nspawn, so my experience doesn't directly translate. And it's not any easier than docker.
For video acceleration I found the Arch wiki somewhat helpful. But it's lots of info and not specific to Docker. Maybe it helps anyways: https://wiki.archlinux.org/title/Hardware_video_acceleration
Hmm. I mean now reading your first output in privileged mode properly, I don't see any errors, or am I missing something... Seems it loaded vaapi sucessfully?!
Have you tried verifying it's not the group permissions? You could
preliminarilytemporarily set permissions withchmod 666 /dev/dri/render128My older Skylake processor has a bit worse video quality (occasional artifacts) with QuicSync.
Thanks. I've learned a lot.
In the end I still don't understand that specific culture. I've scrolled through a few of the hashtags and links you gave. Some of them I'd shorten to half the length. That some bubbles in an infographic have different color is completely useless information without telling what they're trying to convey with the color and how that connects things. Other images I think they describe the details that are just fluff. Those details are irrelevant because they just set the atmosphere. Just say what the armosphere is, then. I think that's making the text too long and all over the place. Making it difficult to focus on what's really going on in the picture, what's important, because there's so much noise added.
But some of the descriptions are really next level good. I wouldn't have expected that. I think I need some more time to familiarize myself with that culture. I can't tell if it's some people being ultra good at it and some people mimicking it without really understanding its purpose... Or it's me not grasping the concept / culture.
If you say you're already adding a concise description and a long one and adding that to the body text... Seems I've arrived with my reasoning somwhere near what you've already been doing.
I see now why you'd like to talk about the Fediverse as you originally said. Seems to me like a matter of the Fediverse not interconnecting the way you'd need it to. And I see a fundamental problem here. I got that you're using Hubzilla. But we've got to think about the perspective of a Mastodon user as long as most of your audience is there. And that platform is meant for short chunks of text. The whole platform and interface is designed to cater to that. And you're doing long blog posts. There is a fundamental split between the two. Yet the platforms interconnect. I don't see a way to make messages short and long at the same time. And the Fediverse is about connecting a diverse set of platforms. There is bound to be some difficulty and I don't know if there is a good solution.
And your perspective might be a bit spoiled. Since you're on Hubzilla and that's meant for a wide variety of tasks. And Mastodon on the other side is meant to narrow things down to the use-case of microblogging... It's kind of per design that your content falls through in the process of narrowing it down. And lot's of Fediverse platforms are meant for one task only. Either pictures or videos or threaded conversations like here. That also doesn't translate to other platforms and looks weird on Mastodon. The users of "all-in-one" platforms like Hubzilla or Friendica etc get it all. But then it get's problematic when interconnecting to users of "narrower" platforms. It's always been that way. And I don't see a way around that. At least fundamentally.
And this manifests in the smaller issues you're having. Like alt-text and culture that's different amongst platforms. It's all consequence of connecting diverse places. With your added explanation, I think I've now homed in on your issue...
Lemmy seems to be the wrong place to discuss it. I don't see the users here have and particular knowledge about such topics. And Lemmy doesn't federate in any unique way that'd make it stand out concerning this. It's a good place for discussion, though. Mastodon's choice to narrow down social media is valid. So if they like not to have long text, it's their choice. And I applaud them for developing their own culture. I'm not sure if there is a good place to discuss this. Maybe within the "all-in-one" platforms like Hubzilla. You're bound to find more people with the same struggles there. But you also want to reach us and the Mastodon users. I mean these places are also about linking external content and blog posts. So linking a Hubzilla blog post starting a discussion about this is the best thing I can come up with. But you need to lay down the groundworks properly. I mean it also took me several back and forths to understand the core of the issue. And it's kind of a niche topic in a niche. So brace for little engagement or interest.
Meh. No Flatpak, no worries. And no updates, no new software or security patches.