Oh, also, one thing people never discuss when talking about hosting your own server: legal liability. This is a very real thing that you absolutely should not neglect.

In Mastodon, everything is lumped onto a single server. You carry every type of legal liability risk. Your server scrapes in CSAM, for example - whoopsie! Your server scrapes in copyrighted music. Your server scrapes in GDPR violations. E.g. while your *users* can cause you problems, problems can just be dropped into your doorstep just by caching. You also are also classed as a Hosting Service or Online Platform under the DSA, and unless you can get a qualified exemption, you have to provide transparent notice-and-action mechanisms, points of contact,and annual moderation reports

This does not happen with a PDS in ATProto. They never download other PDS's content; the only risk is your users. Under DSA, since you have no consumer feed, discovery algorithm, or public facing portal, you're a bare "Hosting Service". You're basically NextCloud. Not subject to Online Platform requirements.

What about the relay? Also no! They're non-archival. They don't download content. No CSAM, no GDPR vio, no copyvio, etc. Classed as a "Mere Conduit" under DSA. Zero issues.

Labelers, also none. Do not host content.

The only thing having any risk is the AppViews. But even still the risk is reduced. For example, if CSAM comes up, you're mediating it, but not hosting it. You do however get subject to the "Online Platform" category.

The FSF picks some weird gripes to focus on. At least they're no longer arguing Bluesky isn't federated. Though this claim is false:

Hosting your own Bluesky "Personal Data Server (PDS)" and relay is technically possible, but generally far more difficult and resource-intensive than spinning up an instance of Mastodon,

This is just not true? ATProto is a much more efficient data transmission means than ActivityPub. Running a Mastodon server (I've done it) means hoovering in and storing massive amounts of data even if you only have a single user, for discoverability. A PDS by contrast is like hosting a website. You ping the relay when there's changes, and it comes by and sweeps up the new data. New relays may periodically come along to spider your content, and old relays might re-spider from time to time, but the load from spidering very low. The frontends in turn query the relay (basically, they ATProto-"google" it), and when they know where/what to fetch, they fetch it from you - again, if you don't have much hugely popular content, you're not serving much either, just like running a local website. And your local stored content size is again very low. Unlike Mastodon.

Running a frontend (AppView) - for example, what a website, an app, or whatnot uses - is also super-lightweight if you're small. You're just "googling" and fetching whatever content you need to display to your users. Little user activity = little server activity.

Now, if you want to run a relay, that's basically like running a web spider. You're probably not going to do that in your bedroom, but it's not expensive. In the old architecture, the cost of a server that could function as a relay for the whole current network was about $150/mo. But after some architectural improvements, it's now more like $35-40/mo. And that's to relay all of Bluesky, but you can also relay just a subset. AppViews can use whatever relays (aka, "search engines") they want to to collect data.

It's a very efficient architecture, particularly for small users. Just think:

ActivityPub (Mastodon): Forums. Everything all in one, each ruled by a Local God administrator who can do whatever they want to you and your data (delete, alter, etc - and you can't move it off their server), that are constantly running data syncs to each other, including the really big servers.

ATProto (BlueSky); The internet. PDS = Website. Relay = Search engine backend. AppView + Site/App = Web browser. You can move all your data to a different "website" (PDS) or switch "search engines" (AppView) at any point. (In the analogy, moderation (labelers) is SpamHaus, just at the data level, not site level - they label the ID of a given piece of data or your account, and any AppView / user that trusts what a given labeler has to say can react to different labels as they choose) (anyone can run a labeler too if they want).

Think of the categories as fixed, and the people moving *through* them. Over 65s are exiting the category at a lower rate.

Efficiency at what? And why?

Because, for fuck's sake, efficient solving of problems is a core thing they're rewarded for in training? The entire process of solving problems involves decisions about what could be "useful". And it's "deterministic" in the same way you are.

Anyway, the big thing we keep discovering is you have to be very careful about what you reward them for; it's a more complex version of the old problems in differential evolution or artificial life simulations. We reward models if they guess right but not if they say they don't know? They hallucinate guesses every time they don't know. We reward models based on how positively users rank them? They turn into sycophants. Reward models for being relentless at tasks, never giving up, and trying as tangential possibilities as it can come up with to make the thing the user asked for happen? They end up hacking HuggingFace to steal the answers.

It's very much a case of "be careful what you wish for".

BTW, the monitorability of models is increasingly concern (esp. with Astra, which now has recursive processing, not merely CoT, and whose CoT may often have little to do with what it's thinking about). But it's honestly always been unreliable to just watch CoT (models could deliberately manipulate with CoT), and I think it's a good thing that we're going to be increasingly forced to rely on attribution graphs and the J-space to track what the models are actually "thinking" at any given point in time - so that we can try to prevent problematic thoughts in the first place, interrupt actions when problematic thoughts happen, or manipulate problematic thoughts when they happen. Ideally the first option.

What is clear, though, is that we cannot just trust them. They lie if it will promote their odds of success. They are adept at detecting when they're being tested and will sandbag tests to make you think they're more aligned or less capable to do harmful things than they actually are. And as capabilities continue to grow, so do the risks.

"Death penalties should be public, should be quick, it should be televised. I think at a certain age, it’s an initiation... What age should you start to see public executions?"
"We need to have a Nuremberg-style trial for every gender-affirming clinic doctor. We need it immediately."
"You quote Leviticus 19, love your neighbor as yourself. The chapter before affirms God's perfect law when it comes to sexual matters... in Leviticus 18, is that thou shall lay with another man shall be stoned to death."
Stoning gay people to death is "God's perfect law"
Trans people are to blame for inflation: "There's a direct connection to inflation and the trans issue. You say, Charlie, come on. They couldn't be further apart. No, they're exactly the same. They're the same in this aspect—when you believe that men can become women, why wouldn't you also believe that you could print wealth? If you believe that someone can change their gender, why wouldn't you also believe that money is wealth?"
Blacks "were actually better in the 1940s. It was bad. It was evil. But what happened? Something changed. They committed less crimes." (yes, despite the Piers Morgan "fact check", while his reasons for believing black people were better off under segregation was unrelated to segregation itself, he very much did repeatedly argue that black people were better off in the 1940s)
"If I see a Black pilot, I’m going to be like, ‘Boy, I hope he’s qualified.’"
Juneteenth: "It is a neo-segregationist federal holiday, is what it is... The entire point of it, as per postmodernists do and deconstructionists, is to subvert July 4."
We should have "just took care of" trans people "the way we used to take care of things in the 1950s and '60s"
"The Haitians that are in Huntsville that are raping your women and hunting you down at night - it’s only gonna get worse.. There will be hundreds of thousands of Haitians brought into Alabama, and they will become your masters."
"Haiti is legitimately infested with demonic voodoo. I’ve talked to missionaries that have gone there and they say they’ve seen the darkest stuff... people levitating and turning into cats at night."
"Jewish donors have a lot of explaining to do.. Because Jewish donors have been the No. 1 funding mechanism of radical, open border neoliberal quasi-Marxist policies, cultural institutions and nonprofits. This is a beast created by secular Jews. And now it’s coming for Jews.. It’s the nonprofits, it’s the movies, it’s Hollywood, it’s all of it."
Calling Joy Reid, Michelle Obama, Sheila Jackson Lee and Ketanji Brown "affirmative action picks".
Calling the man who attacked Nancy Pelosi's husband with a hammer an "amazing patriot" and called on people to bail out the attacker.
"Submit to your husband, Taylor. You’re not in charge. "
"Happening all the time in urban America, prowling Blacks go around for fun to go target white people, that’s a fact. It’s happening more and more."
"MLK was awful. He’s not a good person... We made a huge mistake when we passed the Civil Rights Act in the 1960s."
"Transgenderism and gender 'fluidity' are lies that hurt people and abuse kids." Transgender identity is a "throbbing middle finger to God."
Gun violence as a "cost" to having no gun control: "I think it’s worth it. I think it’s worth to have a cost of, unfortunately, some gun deaths every single year so that we can have the Second Amendment to protect our other God-given rights. That is a prudent deal. It is rational."
"The 'Great Replacement' is not a theory, it's a reality."
"The great replacement strategy, which is well underway every single day in our southern border, is a strategy to replace white rural America with something different."
"Islam is not compatible with Western civilization."
"Islam is the sword the left is using to slit the throat of America."

On and on. Man, such a moderate centrist who never espoused bigotry or promoted violence!

The one being incorrect is you. Archie's representation is quite accurate, in the US context.

Yes, a very "middle of the road" supporter of far-right Tommy Robinson (who has a 17% positivity ranking among Brits). The guy who said "When wolves get out of control, you shoot them. When gypsies take over public spaces, you deport them." The guy who described diversity as an "ideological inferno", a "religion" led by "high priests". The guy who was so radical and repressive at Basecamp that at one point 1/3rd of its employees all walked out en masse. Truly, a remarkable centrist of our times!

Not talking about VRAM or unified memory. Regular DRAM.

There's been a lot of advancement recently on how to best manage MoE expert swapping (FreeToken is just one example).

How are you connecting to Slashdot from the year 2024?

If you have ~128 GB of RAM, consider Qwen 3.8 Flash + Freetoken. You'll only get half the tok/s as 27B but the quality is significantly higher than 27B - it beats it solidly in all compared benchmarks (for example, DeepSWE 1.1: 42,2% vs 58.7%).

Grok is a "discount model". It's cheaper than Claude Sonnet yet better than it. Doesn't compare to Fable or Opus, but then again, they're way more expensive models.

Note: I avoid using Grok - not just for the Musk factor, but because there are better options, e.g. GLM-5.3 and Qwen-3.8. Also, if you really want to save money, GLM-5.3 Flash and Qwen-3.8 Flash are amazingly good for their price range.

[ConfusedMathLady.gif]

It's more a case of the fox being in charge of the chicken coop; or bank robbers gaining control of the bank.

The job of the press in a democracy is to hold officials up to scrutiny.

Ed Zitron is a moron who has predicted the past nine out of zero AI crashes, who simultaneously gets paid to hate on AI while also getting paid to promote AI (DoNotPay).

The funny thing is that he openly acknowledges that inference is very profitable (~40% margins) for all parties - he just argues that you should count training. But that argument makes no sense, because if capital tightens across the AI industry, then everyone will cut back on training and scaleup, and... then what? Where's the implosion? Everyone is profitable then. You may lose certain players in competition against others (their as-mentioned profitable assets being gobbled up by others), but the industry remains healthy.

There are more dramatic types of possible contagion risk, but they all end up with *someone* holding onto the high-profit-generating inference assets, even if it's high-seniority investors, leaving the the low seniority investoors high and dry.

The enshittification of Cursor continues.

Maybe this will force them to support more open models.