This is true of all OSes TBH. Issues arise when you start to mess with them.

Remember this if you switch to linux.

Even “big” open source models like DSV4 and Ling/Ring are very efficient. They’re big, but (seemingly) sparser than US models, so they’re cheap.

They run surprisingly well with hybrid CPU+GPU inference on desktops. And thats not even getting into the efficient attention mechanisms.

I can run DSV4 Flash, barely quantized, with ~1M context on my Ryzen desktop at ~11 tokens/s. If you told me that two years ago, I would not have believed you.

MS’s antics extend well beyond Windows, unfortunately. There are long blogs and writeups dedicated to it.

But what you said is also kinda true. 7 and XP worked reasonably well as long as you didn’t mess with them, push their bounds too much, or use them for anything “weird.” But you also probably missed a lot of dysfunction in the background.

I’m not trying to shill for Windows 7/XP era OSX or linux or anything either. But Windows has always been kinda notorious.

Like Scipitie said, sort of.

In the ChatGPT app, there’s tons of shuffling in the background, like context being injected, maybe sampling changed, agenic action, quantization… its “far” from the actual LLM, opaque by design, and as a result certainly not deterministic.

You can sometimes get deterministic output with the OpenAI API, but it’s also dependent on nothing changing on their end. And their end changes a lot.

But self hosting or using a more consistent provider will give you deterministic output.

Whoa nellie.

I mean, I guess it depends on your area and such, but there still great places to eat out if the situation allows for it. Of course groceries are cheaper, but some places are still affordable.

I wouldn’t touch any of the comic’s chains though. I think I went to a McDonald’s once since the pandemic, for a quick breakfast while on the road, and it was indeed disproportionately expensive. But (as an example) taco Tuesday at Rosa’s Cafe is still reasonably priced here.

We’re still talking about Microsoft, right?

They have a looooong history of enshittification and footgunning that predates “AI” by decades, no matter how much the world has depended on their stuff. So no, I suppose I wouldn’t think that, as that would break decades of precedence for MS. But I suppose I never trusted them in the first place.

Yeah.

Don’t get me wrong, I don’t want to touch projects vibe coded to this level, going solely on past experience with them.

But I am scientifically curious about its functionality.

Well… Microsoft is Microsoft.

This is what they do. They simply cannot help themselves.

Microsoft doesn’t treat it as a tool; that’s the problem.

…Which isn’t surprising. Not only because Microsoft is Microsoft, but they own like a third of OpenAI, so they have a vested interest in perpetuating the AI hype.

Yeah, in that case it’s sus. I’d be scared to look at the codebase.

Does it actually work, though?

I wish more would listen.

Many seem to interpret the headline as “Torvalds is an AI Bro now,” when its more like him yelling “If you all don’t treat these things as tools, shit is going to hit the fan.”

Technically, LLMs (and most ML models) are deterministic with the same input and same seed.

I get what you mean though.

Oh yeah… yeah, it doesn’t seem like it’s client/server or anything. It’s just a CAD program?

Doesn’t fit this at all.

I dunno if it fits the “vibe coded” label though, as it appears to work well. I tend to apply that to quick hacks, or to projects where the devs clearly have no idea what they’re doing (and end up with something really bloated and janky, yet hyped a ton).

Why is a post about a CAD program being downvoted so much?

I’m not againt LLMs. They’re usefull tools, and I’ve been tinkering with them since 2021, and generative ML before that. I have one loaded on my desktop pretty much all the time; DeepseekV4 0731 is running, right this second.

I use them all the time. I hack at them as a hobby; I know how useful they can be.

But you’re stopping just short of calling them “alive.” They’re not even close, not even in the same universe. They do tasks with intelligence or some variant of that, but they are not intelligent. And I’m not trying to be pedantic, that is a loaded term that implies cognition and adaptability they do not have.

They’re basically the same thing as weather prediction models, but for text output. They’re just a model.

And I think its very dangerous when people fall into the trap of thinking they are, fundamentally, more than that. That’s the fantasy Altman and such are selling the public, to con them.

Hence, the article above ^. What’s coming out of these leaders’ mouths does not match what people are getting, and they can tell. And that’s because the premise they’re being sold, of “artificial intelligence,” is a lie.

I mean… it’s an issue. OP’s system is swapping the SSD when it clearly should not be doing it. My system had horrible usability issues in some scenarios without a lot of reconfiguration, and thrashed my SSD for no benefit.

The default heuristic is not good for our systems, and OP’s situation is a pretty common scenario. That’s a fact. How is that not an issue?

And support for TRIM and whatever else was added for SSD support is not really the issue, here. That’s working fine.

I mean, if your argument is just semantics, that’s just… pedantic.

You could draw that line anywhere. The NPCs I play against in a strategy game could be “intelligence.”

You know what people mean. People are talking about ~anthropic cognition, which is clearly not what’s going on in current LLMs. They can do intelligent things, maybe they are an “intelligent” by some definition you define, but they are not even close to animalistic cognition.

I disagree with most here.

You need to reduce swappiness even more, and tweak some other variables (like the memory “threshold” Linux starts to swap, and disk cache aggressiveness) to stop Linux from swapping so opportunistically under such a light load.

IMO, Linux is configured for “old” systems by default: slow HDDs, and constrained RAM pools, where disk IO really, really needs caching, and where idle background processes take a large fraction of RAM.

You have superfluous RAM for your workload. And very fast disk IO that isn’t such a hindrance to apps anyway. And a disk you don’t want to wear. This is the opposite scenario: you don’t want Linux to swap unless it absolutely has to.

For reference, this is part of my config. It’s rather niche and you probably shouldn’t use it, but you should consider looking up the variables:

# Keep min reserve reasonable for 8GB usable space
vm.min_free_kbytes = 262144        # Lower absolute minimum to 256MB

vm.watermark_scale_factor = 10     # Lower to 0.1% (which is ~128MB on 128GB)

# Disable watermark boosting completely
vm.watermark_boost_factor = 0

# Normal-ish metadata pressure so desktop doesn't stutter on disk reads
vm.vfs_cache_pressure = 120

# Allow reasonable swapping of inactive anonymous desktop pages, could be lower
vm.swappiness = 10

# Dirty bytes limits to limit caching
vm.dirty_background_bytes = 67108864
vm.dirty_bytes = 268435456

# Disable compaction & proactive scans to stop freezing with large portions of RAM mlocked
vm.compaction_proactiveness = 0
vm.compact_unevictable_allowed = 0

vm.page-cluster = 0 # 4kb pages for SSD

My system still uses RAM as disk cache with this config, it just won’t go out of its way to swap just to keep that cache, especially I lower swappiness to 1-3.

I also have a 1GB zram pool, prioritized over ssd swap. But you should make yours even larger (maybe 4GB?). This will intercept anything that does swap first.

I also start some applications with systemd-run and specify memory caps and swap limits (often forbidding them from swapping entirely).

It makes a night-and-day difference for some workloads on my system, that would otherwise swap pointlessly, just tank performance and even de-stabilize the system.

It’s “free ram” until stuff actually goes in swap.

It depends on the scenario, of course. Maybe you have huge, totally idle processes, and active ones working from an HDD that need cache desperately, in which case this is good.

But my experience with Linux is that is waaay to aggressive with this, in most configs, and I get far better performance from configs that avoid swap until it absolutely has to be used.

The DeepseekV4 architecture was the eye opener to me.

Qwen 3.8 27B too. Minimax M3…

They are sooo dirt cheap to run, and “good enough.” And they’ve barely picked low hanging fruit of inference optimization.

No matter how anyone feels about the utility of LLMs, one just can’t argue they’ll need all these datacenters, if you look at the architectural trends. It makes no sense.

So once the music stops… what happens?