I also asked ChatGPT itself, and it listed a number of approaches, and one that sounded good to me is to pin layers to GPUs, for example we have 500 GPUs: cards 1-100 have permanently loaded layers 1-30 of AI, cards 101-200 have permanently loaded layers 31-60 and so on, this way no need to frequently load huge matrices itself as they stay in GPUs permanently, just basically pipeline user prompt through appropriate sequence of GPUs.

So do they load all those matrices (totalling to 175b params in this case) to available GPUs for every token of every user?

That’s how llms work. When they say 175 billion parameters, it means at least that many calculations per token it generates

I don’t get it, how is it possible that so many people all over the world use this concurrently, doing all kinds of lengthy chats, problem solving, codegeneration, image generation and so on?

Should be “starting your own instance”, because otherwise you still have to conform to the rules of the instance you create your community/sub on.

It is what it feels like, but it’s not really 100% this way (yet). It is a bad self-reinforcing cognitive bias: we think “forums are dead, that’s why we stick to the sitename” instead of actually finding dozens of still alive forums and going there, in turn sitename gets more populated while forums feel more dead. But there are still plenty alive. Also, there are relatively new kinds of forums which sometimes work very well for their niche, like Discourse communities for example.

After you use ChatGPT for a bit, you will start recognizing its style of writing in posts and comments. I’ve seen dozens of obviously ChatGPT generated posts or replies on Reddit and Lemmy. Usually there will be a person who already replied to them something like “Thanks, ChatGPT”, because it is that obvious. This only happens with naive prompts though, if you ask ChatGPT to present its answer to your prompt in a different style (for example, mimic some famous writer, or being cheerful/angry/excited and avoid overly safe language), it will immediately start writing differently and there’s likely no limit on variety of writing styles you can pull out of it with enough effort of just asking it to write this or that way.

Yeah, got a bit carried away yesterday. Ultimately, there can’t be right or wrong, since the whole discussion is simply about individual understanding or even preferences for the term “slop”, nothing more. Some people will try to be as empirical as possible and choose a meaning based on how they’ve seen it being used in the wild, others will try to push the meaning they want for the term, it’s all good and subjective.

I’m not sure this is KDE Panel you’re talking about, but if it is, you can configure border radius in Panel Colorizer:

This is where the fundamental difference in attribution of connotations lies. From what you say, you perceive the term “slop” as a direct synonym to “low quality”, without any extras. I perceive it as something more of a synonym to “repetitive” but with extra connotations, the most accurate common divisor of which is “repetitive content produced at speeds suggesting low effort”.

I understand your point of view. But would you call something complex, high-quality, but repetitive, slop? And the same question, but if the person who produces it, does it extremely fast.

How would you use that term? Would you call “slop” something that was just mindlessly generated using AI in a single prompt and non-“slop” something that uses AI in more sophisticated/deliberate ways? What is the threshold of something being “slop” ultimately? Is this just result not looking decent enough or amount of effort combined with amount of knowledge and experience that was used to create that? I’m personally conflicted on this, because sometimes even mindless prompt may give great result, and sometimes a lot of manual effort may give shit result. I guess with “slop” I tend to gravitate towards “amount of effort combined with amount of knowledge and experience that was used to create” and perhaps also the amount of content that particular person produces and speed of its production. So if someone is really good with some tools (not necessarily AI) and figured some overpowered shortcuts that allow to produce results very fast with little effort, it also can be called “slop” just for the rate of production alone.

Damn, turns out it’s Blur effect strength that is killing transparency. Found a suggestion to turn it way down and now I actually see transparency working in window decorations. I never noticed it’s even possible to configure blur further in desktop effects. Thought it’s just an on/off flag.

I tried even more of manual installing, for example I gave a try installing aurorae windows decorations alone, and transparency never works. Turns out, transparency might be totally broken with Wayland on KDE / Plasma 6, at least there are many issues searchable on this topic and even AIs seem to have learned from somewhere that “In Plasma 6 on Wayland, transparency effects may not work as expected due to various bugs and issues, especially with certain themes and drivers. Users have reported that transparency works better in X11, and some have found that specific settings or themes can help achieve the desired effect.”

I also tried Maximized Window Gap (fixed metadata and manually specified margins), but it doesn’t seem to do anything on current KDE/Plasma.

Yes, it is not handler by Kvantum, which means if you install a Kvantum theme, it will not make your panel match that theme, so you still have to do that yourself. I found out about “Panel Colorizer” widget for KDE Panel, which looks very deep and promising, I’ll try and see, maybe I can come up with something good by combining multiple approaches.

As I mentioned in OP, I also tried Kvantum, but it didn’t work well because what’s the point of using it if can’t restyle taskbar. One of the most important things to customize for me personally. Ideally would like it semi-transparent blurred/glassy, like a shader or something.

I tried both through discover and manually by downloading. For example the most popular theme in the KDE Store called Sweet KDE, didn’t have Dolphin and some other things transparent+blurred like showed in screenshots. The only thing which has some transparency is console, but its transparency looks kinda bad, because it’s not blurred. I also tried re-enabling blur in effects system configuration, but it doesn’t affect it.

Cool idea, I almost forgot this feature even exists. I think I dismissed it the past when I realized it’s probably not going to be easy to switch VPN servers this way.

It does hurt, your VPN should support proper port-forwarding for soulseek to work well. In most cases, you will only be able to download files, but your shares will be inaccessible. It doesn’t seem to work with ProtonVPN for example, even when you built-in port-forwarding feature. And even if it did work, you would need to reconfigure and restart soulseek every time you reconnect the VPN, because their port-forwarding is randomizing the ports and there’s no way to turn that off.

Let me explain how Honkai Star Rail handles gearing. Every single character has six relic slots: head, hands, body, feet, planar orb, and planar ornament. These relics go from level 0 to level 15, and four of them have a randomized primary stat. They all feature four randomized secondary stats, and every three levels a random one of those secondary stats gets a bonus. Each relic also belongs to a set of relics, and characters benefit from having two or four pieces of a given relic set. That means for every character in your party, you need to get the right items at the maximum rarity, the right primary stats, the right secondary stats, and the right level-ups for those secondary stats.

This is min-maxer mindset and I would hope randomized systems like this will prevent it but unfortunately no: even here some people think they actually need to roll every dice exactly the right way. I don’t think it’s true that this is really necessary. And no, it is not necessary to do top 10 world parses; you can just beat endgame content on modest, casual difficulty and call it a day, rather than try hard to set a record.