Even lemmy.world is in this bucket (not sure about other instances). See: https://lemmy.world/post/28304534

This is my second time bookmarking this meme.

…I can’t find the first.

I’m talking theoretically, heh, I don’t think anyone actually does that yet.

And I am just talking edge cases where existing blockers fail and there’s no manpower to figure out a customization.

Right? It makes Afganistan (the Soviet and US one) look like nothing. A calamity doesnt even begin to describe it.

This ^.

I think people forget the fabled “old” internet was actually a pile of trolls where one had to double check what they read.

Basic sanity checks really aren’t that hard. But its a forgotten habit, I guess.

back-to-basics apps and services.

I think these do exist, but they’re in such a sea of shit that most users scrolling on their phones can’t find them. Shameless apps have an intractable engagement/marketing advantage over them, as do the ‘lets get acquired by Big Tech’ ones.

I guess big companies could engage in this, but… shrug.

Hassle-free self hosting is hard, yeah, AI or not. Not going to argue with that one bit.

There are a few that are “truly” open like IBM Granite, and a handful of others over the 7B range.

Blocklists are a much more efficient way to do this, and TBH many “traditional” adblockers are still huge performance hogs. Ublock is an exception in this regard due to webassembly and its explicit dedication to lightness.

Vision models are a pretty good way to build sponsorblock/adblock databases though, and maybe even engineer HTML workarounds automatically. It would be cool if you, say, encounter an ad or a dysfunctional web page, and you can opt-in to automatically contribute a fix with your own compute.

some bits related to its training data

AKA ANY details about its training data, and its training hyperparameters, and literally any other details about its training. An ‘open’ secret among LLM tinkerers is that the Chinese companies seem to have particularly strong English/Chinese training data (not so much other languages though), and I’ll give you one guess on how.

Deepseek is unusal in that they are open sourcing the general techniques they used and even some (not all) of the software frameworks they use.

Don’t get me wrong, I think any level of openness should be encouraged (unlike OpenAI being as closed as physically possible), but they are still very closed. Unlike, say, IBM Granite models which should be reproducible.

I think Lemmy’s userbase is a bit predisposed to that. Unfortunately, that sentiment isn’t common enough, and while most people don’t want to be monetized if asked, with the convenience the reaction is a collective shrug.

But another thing we are predisposed to is dev bugs, and I think the average person won’t like how unreliable many such features are.

Ugh, you said that way too well.

Still (again, rolling with the reasoning), it’s an existential problem, no matter who’s fault it is or how much of a victim they are or how much fatalism they’ve accepted…

I can understand the mistrust of the system, but I can’t understand the conclusion that some anonymous yahoo with no inside perspective somehow has the resources to have figured it all out.

There was always some percentage of the population that believes in that stuff, for a variety of understandable reasons.

That’s… fine. It kinda worked.

The problem is social media has amplified those voices by orders of magnitude because they’re engaging (hence, profitable). The missing panel above is the FB post, YT segment, podcast (all recommended algorithmically) or whatever that led that guy down the rabbit hole.

In other words, It’s not a content problem, but an engagement one.

Not according to a twenty second check, assuming you trust the World Economic Forum:

https://www.weforum.org/publications/global-risks-report-2025/

However, the top risk in 2027 is Misinformation and disinformation, for the second year in a row (Figure C). There are many ways in which a proliferation of false or misleading content is complicating the geopolitical environment. It is a leading mechanism for foreign entities to affect voter intentions; it can sow doubt among the general public worldwide about what is happening in conflict zones; or it can be used to tarnish the image of products or services from another country.

This is how the fabled “old” internet worked IIRC. BS, bots and trolls were everywhere, hence clicking through hyperlinks or a quick search as a sanity check were pretty standard fare. And it worked reasonably well. And like the old internet, this random substack could be fishy, but it doesn’t make the source wrong.

What I am getting at is that dismissing something as “possibly AI” as a first impulse isn’t good information hygiene.

I mean, billions of people would disagree with you. Many would say Ukraine and their western backers are at fault.

But theoretically, I am struggling to picture how Russia’a war goals (again, accepting every reason for its continuation from Russia) could possibly be beneficial to Russia, and the um oppressed Ukrainian people, in light of this measurable effect on demographics. Whatever the reason for Russia staying in Ukraine is, even with the most extreme sympathetic view of Russia and their position, it is clearly not worth it, and they should just capitulate an pull out for their own benefit.

…Wow.

So what’s the tankie angle to this? From that perspective, the war is objectively awful for Russia, even if all the stated war goals are true.

Yeah. But what I’m getting at is the economics may hint at what valve is planning.

Maybe AMD isn’t making a “specialized” monolithic die like Van Gogh? Perhaps Valve is simply customizing blocks of AMD’s existing product (die) stack, which is more financially plausible.

AFAIK one of the current issues with Strix Halo for a handheld would be high idle power, but maybe the next generation is better in that respect.

…once the tech has advanced enough that putting one together makes for a substantial boost in what you get for the same price and power envelope.

It already does.

Its more a question of economics of scale. Taping out a single custom chip is extremely expensive, like hundreds of millions of dollars before a single chip is sold.

AMD could make a custom Strix Halo SKU for Valve (think a 6-core X3D CCD, a 32-40CU GPU clocked low for efficiency) for much less. Perhaps something like that (a custom multi-die configuration of Strix Halo’s successor?) is what Valve opted for.

Valve is getting first dibs at the next AMD SoC as far I’ve heard

This is huge if true, as Van Gogh (the Deck chip) was a seperate “line” than all the overly CPU-heavy laptop chips other handhelds are using at the moment.

You can squeeze a lot more performance out with a newer framework and a model tailored for your GPU and task.

I’d recommend:

Kobold.cpp rocm, follow the quick-install guide here: https://github.com/YellowRoseCx/koboldcpp-rocm/?tab=readme-ov-file#quick-linux-install

Download this quantization, which fits in your VRAM pool nicely and is specifically tuned for coding and planning, select it in kobold.cpp: https://huggingface.co/mradermacher/Qwen3-14B-Esper3-i1-GGUF/blob/main/Qwen3-14B-Esper3.i1-IQ4_NL.gguf

Use the “corporate” UI in kobold.cpp in your browser. If that doesn’t work well, kobold.cpp also works as a generic OpenAI endpoint, which you can access from pretty much any app, like https://openwebui.com/

Given the lack of regulation from anywhere else, I agree with my state here. This is a good thing (even if its made for the wrong reasons).

And phones are so freaking locked down, and apps so centralized, that (on aggregate), it’ll be pretty enforcable. If it pushes a small fraction of kids to sideloading, smaller/web platforms, VPNs or even hacking, good.