Natanael
Cryptography nerd
- 0 Posts
- 89 Comments
Already broken
https://en.wikipedia.org/wiki/8K_resolution
8K resolution refers to an image or display resolution with a width of approximately 8,000 pixels. 8K UHD (7680 × 4320)
The number refers to the horizontal resolution. FHD is nearly 2K pixels wide, just as 4K resolutions are nearly 4K pixels wide, although FHD is the typical term for the resolution and QHD is more commonly called 2K instead than FHD
Humans learn a lot through repetition, no reason to believe that LLMs wouldn’t benefit from reinforcement of higher quality information. Especially because seeing the same information in different contexts helps mapping the links between the different contexts and helps dispel incorrect assumptions. But like I said, the only viable method they have for this kind of emphasis at scale is incidental replication of more popular works in its samples. And when something is duplicated too much it overfits instead.
They need to fundamentally change big parts of how learning happens and how the algorithm learns to fix this conflict. In particular it will need a lot more “introspective” training stages to refine what it has learned, and pretty much nobody does anything even slightly similar on large models because they don’t know how, and it would be insanely expensive anyway.
Yes, but should big companies with business models designed to be exploitative be allowed to act hypocritically?
My problem isn’t with ML as such, or with learning over such large sets of works, etc, but these companies are designing their services specifically to push the people who’s works they rely on out of work.
The irony of overfitting is that both having numerous copies of common works is a problem AND removing the duplicates would be a problem. They need an understanding of what’s representative for language, etc, but the training algorithms can’t learn that on their own and it’s not feasible go have humans teach it that and also the training algorithm can’t effectively detect duplicates and “tune down” their influence to stop replicating them exactly. Also, trying to do that latter thing algorithmically will ALSO break things as it would break its understanding of stuff like standard legalese and boilerplate language, etc.
The current generation of generative ML doesn’t do what it says on the box, AND the companies running them deserve to get screwed over.
And yes I understand the risk of screwing up fair use, which is why my suggestion is not to hinder learning, but to require the companies to track copyright status of samples and inform ends users of licensing status when the system detects a sample is substantially replicated in the output. This will not hurt anybody training on public domain or fairly licensed works, nor hurt anybody who tracks authorship when crawling for samples, and will also not hurt anybody who has designed their ML system to be sufficiently transformative that it never replicates copyrighted samples. It just hurts exploitative companies.
Remember when media companies tried to sue switch manufacturers because their routers held copies of packets in RAM and argued they needed licensing for that?
https://www.eff.org/deeplinks/2006/06/yes-slashdotters-sira-really-bad
Training an AI can end up leaving copies of copyrightable segments of the originals, look up sample recover attacks. If it had worked as advertised then it would be transformative derivative works with fair use protection, but in reality it often doesn’t work that way
See also
Just keep building housing until there’s enough available that nobody’s left without an affordable home. Enough that investors can’t buy it all (without simultaneously killing the value, because TOO MUCH empty housing kills the value of the area). If that requires municipalities investing in building then let’s have them do that.
Not really.
Since handles are domain names (and your own DNS entry points to your account public key, DID) they have a referral deal with a registrar to let people easily get a custom domain and set it as a handle (otherwise your handle is a subdomain on the bluesky domain). But future plans are uncertain
As said many times before, Jack is now AWOL and left for nostr, he even deleted his bluesky account because the crowd didn’t like him there. He doesn’t have a majority on the board and don’t own any majority stake either.
The motivation for a new protocol is there’s architectural limits to activitypub. It’s essentially email over http, it really behaves like public mailing list archives, as servers push each interaction as a message. This is part of why there’s often a discrepancy between visible replies across servers because retries are limited. Account portability is also very limited as accounts and posts are tied to a server.
Bluesky switches to a content addressing model plus user ID based on a public key, allowing you to more easily move across servers as well as syncing data between servers such as thread replies, it’s very much like git (user data is held in personal repositories signed by your key) with a shared CDN/cache (relay servers, previously called BGS) and “worker agents” (mostly driven by the “appview” which is the api endpoint for your client + feed generator servers). You post to your repository via your appview, it sends a ping to other servers and they sync new relevant entries.
They already have federation with 3rd parties in a sandbox network and the official server just switched on “internal federation” (used to be a single shared server, now there’s 10 using the same protocol that open federation will later use)
The code is already open source, several servers in the sandbox is 3rd party reimplementations
FYI the bluesky protocol is open and there’s plans to standardize. It’s also federated (the sandbox network is open to 3rd parties)
There’s lots of new privacy techniques from cryptography, stuff like differential privacy and MPC could help a lot with making it easier and safer to use collaboration tools.
I don’t see how that’s a negative, all the choices attached to Mastodon hosts are distributed to multiple services in bluesky which optionally could be served by the same entity, but doesn’t need to be. A PDS can run its own moderation services, or subscribe to another, or leave it to clients. A PDS can run their own feeds, or leave it to others. Clients can choose to use the services provided by the PDS, or to use others.
I don’t see where the centralizing forces are (other than economy of scale stuff). Having the most users doesn’t mean much when it’s trivial to substitute your service, regardless if that’s a moderation labeler service, a collection of feeds, or whatever else. It’s really just the most popular client apps which have disproportionate power, but that’s true for every protocol.
Edit: I also want to point out that the PDS by default controls a bunch of stuff for the client via the appview service, that’s the service which the client talks to and it assembles your home feed and assemble post views (where it control sorting, etc) and it apply blocks and mutes and applies the PDS’s own moderation, and it forwards moderation labels on posts (like NSFW tags) to the client.
In bluesky I think those effects mostly lie on the side of which client people use.
The protocol is extensible and you can add new post types and formatting options by creating a new schema/lexicon, but these would only be readable by other clients which supports it. I hope they’ll be able to add some general “category template” lexicons so a graceful degradation scheme can be implemented to support compatibility without hindering 3rd party development.
To protect against a PDS server going bad the client could assist with automated account migration (the new PDS doesn’t need to understand the lexicon of your posts to be able to migrate them intact), even if the old PDS won’t cooperate (the client could maintain backups for you to make migration quick). But if you don’t control your keys separately then a bad client update could make your account unrecoverable, similarly to a mastodon server going bad.
Yeah, with no strong central control the best you can do is to persuade PDS account servers and client developers to put in good moderation filters by default, so that the average user won’t have to see that stuff assuming they land on a client/server which filter it. You can’t stop it from existing in the network, but you can coordinate ways to inhibit reach. And users who need even better tools can deploy them without having to move.
On the other hand, the work on private profiles haven’t started yet, and you can’t currently prevent yourself from getting visible to others.
On Mastodon the options are essentially just finding a server with a good moderation team and importing block list files manually, as well as keyword filters. And that’s pretty much it. The server features and moderation quality are part of the same bundle.
https://slrpnk.net/comment/3996311
You’re missing details
The Mastodon fediverse have stronger network effects because big servers can enforce policies on other servers to stay federated. It’s complicated for users to move servers.
In Bluesky you have plenty more options, including using 3rd party moderation, using clients which can pull censored posts from other servers and cleanly render them into threads, and you can move servers much more easily even if the server operator don’t want to let you.
The “reach” layer is a mix of relay servers (BGS) and 3rd party feeds (which already are operated independently)
Blue Sky is a for-profit corporation. How do they plan to make money?
🤷
They use domain names for handles, they do have a partnership with one registrar for integration for users who want custom domains for handles (commission model). Other than that, to be seen.
Who controls access to the network?
Once full federation is live, nobody. Anybody could create a relay server (BGS, shared cache server like a CDN), and anybody can run a PDS (account hosting server).
3rd parties already run feeds on their own servers and 3rd party clients exists, and the sandbox network for federation testing has 3rd party PDS servers too.
For user account lookups, if you use the web-DID type then you’re not dependent on bluesky servers at all.
Account portability and the ability to mix and match services and switch quickly are the biggest enshittification protection mechanisms. You can’t really lock in users in this model. You can’t even prevent users from ditching your PDS account host if they kept a backup of their data and held their own keys.
I’ve already explained how they have designed it so they can’t control it. Just use web-DID and your own domain handle on your own PDS, and then you can connect to a 3rd party relay (BGS, the CDN like cache) and whichever feeds and moderation tools and filter subscriptions you like. You don’t need to touch the official servers at all.
The network doesn’t use any cryptocurrency technology. There’s no blockchain, etc.
CF has multiple options, you can use them as just a load balancer/firewall while handling your own TLS cert. I think most let them hold the cert so they can get CF caching services though