I'd say one of the things llms can do well is language. (Or are you arguing that humans deserve to have the proofreading job instead, which I can agree with)
Do you think local llms or community hosted ones are still as bad? Because most of those concerns seem to be more with the corporate ownership of ai, which is definitely a bad thing.
I also can find phrasing stuff hard at times, so being able to send a word salad to a chatbot and get a proper sentence is useful. Or suggesting words.
Sometimes the duckduckgo ai summary of results can include a source I was looking for that was buried in the results (I don't trust the ai much since it literally cited Ireland's Nazi party uncritically once, but it does link to sources). Formatting text is also really useful, or turning something from csv to a markdown table and vice versa.
Its also able to extract text from images a lot better than "dumb" ocrs (which I still use for basic images), and can format it in a certain way (e.g., take this screenshot from an ebook and format it as a quizlet).
I have try seeing what they're like for programming every once in a while, and my verdict is still shit. They can do very basic stuff thats basically regurgitating functions that were written before, but not original stuff (they're very poor at making regexes). They're similarly bad at debugging, though they can sometimes point you in the right direction.
The response from ai-bros for the programming thing is usually something along the lines of how you should try Claude opus, but I am never paying a cent on any ai thing. At that point it's easier to just use my brain.
answered with lists of theoretical protocol features instead of engaging with how the network actually operates
I'm not just listing theoretical features, these are things that happen in the network right now. Is there anything I mentioned in my comment that I forgot to give an example for?
Doesn’t matter whether migration is technically possible under ideal conditions because if you’ll need it they won’t be ideal.
I don't see why so many people say migration is only "technically" possible, migration can be done today. If there is more demand for third party servers, say, if Bluesky starts fucking up with moderation more, more third party servers will pop up, because right now the user concentration isn't a technical problem or fault of the protocol. I don't disagree that it's a problem.
And ATProto’s architecture, particularly the cost and complexity of running the more demanding components that need to have a global view of the network, structurally favors concentration at those layers.
It's not necessary to have a global view of the network to participate in the network.It is possible to have a global view of the network without a relay using constellation, constellation instances are very cheap to run, and work by indexing backlinks. It's what powers reddwarf and recently wafrn (wafrn optionally supports relays as well).
Atproto isn't significantly more complicated than AP, it's just different.
Blacksky doesn't just run a relay, they run an appview (way more expensive than a relay) and pds (admittedly pretty cheap).
The point of atproto isn't to have many different groups running the entire stack, you can use an appview by one group, powered by a relay by another, while using a pds by a third.
A relay I listed in the comment is a real-world one that is currently only costing the creator $30/month, which is ingesting all PDSes, and being used by a lot of apps.
true, although no one said the contrary
While the article itself didn't say it, the overall attitude of most people on the fediverse is that.
I do agree with you that users aren't exposed enough to third party infrastructure, and that most users using bluesky's servers is a problem, but the alternative is the jankiness of the fediverse, which completely puts new users off.
It feels like you haven't read my comment thoroughly.
To start, relays do not require large capital to run. This has been a misconception from the very beginning. I linked to this blog post, where a bluesky engineer runs a relay for ~$34 a month. If relays really had astronomical costs to run, I doubt Bluesky would run a whole separate one.AppViews aren't limited to one relay, most I know point to blacksky's one as well.
technically, users can leave. Technically, you can self-host. Technically, you can run your own relay. The capability exists at every layer.
There's no need to self host as there's already public third party instances you can switch to. The alternatives already exist at each layer.
I do agree that too many users are on bluesky's servers, but that's not a fault of the protocol, and it's not something the fediverse is immune to either.
They never have with any protocol. Not email, not RSS, not XMPP. The default wins. Always.
This is just incorrect. RSS is probably one of the least centralised protocols right now, it's not even federated, which makes me question why the author even included it as an example.
If anything, this reads as an argument against federation, rather than an argument for the fediverse.
I'd say one of the things llms can do well is language. (Or are you arguing that humans deserve to have the proofreading job instead, which I can agree with)