Dran

I think the debate is about what a reasonable class is. I don’t think that an appendage, or identity for that matter, is a reasonable proxy for capability class. In my mind you really have to go one of two ways.

You either make everything class-less (think UFC 1) where all weights, sizes, abilities, genetics compete for a singular title

Or

You make science-based classes, based around whatever the best proxy for capabilities are (testosterone, chromosomes, height, weight, body fat percentage, some combination of the former, etc)

If you use nothing as a proxy, there would be a lot of people unable to compete but it would at least be unequivocally “fair”. If you use science-based capability classes you would have a wider range of “fair-ish” competitions, but there might be some weird overlap where some men, some women, and those in-between bridge accepted norms.

do you know why they’re illegal? is there some danger to them?

I’m sorry but this is just a fundamentally incorrect take on the physics at play here.

You unfortunately can’t ever prevent further breakdown. Every time you run any voltage through any CPU, you are always slowly breaking down gate-oxides. This is a normal, non-thermal failure mode of consumer CPUs. The issue is that this breakdown is non-linear. As the breakdown process increases, it increases resistance inside the die, and as a consequence requires higher minimum voltages to remain stable. That higher voltage accelerates the rate of idle damage, making time disproportionately more damaging the more damaged a chip is.

If you want to read more on these failure modes, I’d recommend the following papers:

L. Shi et al., “Effects of Oxide Electric Field Stress on the Gate Oxide Reliability of Commercial SiC Power MOSFETs,” 2022 IEEE 9th Workshop on Wide Bandgap Power Devices & Applications

Y. Qian et al., “Modeling of Hot Carrier Injection on Gate-Induced Drain Leakage in PDSOI nMOSFET,” 2021 IEEE International Conference on Integrated Circuits, Technologies and Applications

+1 for cmk. Been using it at work for an entire data center + thousands of endpoints and I also use it for my 3 server homelab. It scales beautifully at any size.

The “problem” is that the more you understand the engineering, the less you believe Intel when they say they can fix it in microcode. Without writing an entire essay, the TL/DR is that the instability gets worse over time, and the only way that happens is if applied voltages are breaking down dielectric barriers within the chip. This damage is irreparable, 100% of chips in the wild are irreparably damaging themselves over time.

Even if Intel can slow the bleeding with microcode, they can’t repair the damage, and every chip that has ever ran under the bad code will have a measurably shorter lifespan. For the average gamer, that sometimes hasn’t even been the average warranty period.

Are you maybe thinking of https://distr1.org/ made by the i3 guy?

I’ve heard anecdotally that some 911 services were down in my area, but I can’t speak to how wide that was.

Generally the lifecycle with this sort of thing is old_thing becomes an alias to new_thing, and eventually old_thing gets dropped as an alias down the line.

It’s still decent advice to learn dnf native calls and to update scripts using yum to those native calls.

I believe that one was patched a while ago

I’m a big fan of tiling window managers like i3 or awesome (awesome wm). Awesome is the one I use. It’s tiling and the entire interface is built from scripts that they encourage you to modify. Steep learning curve but once you get it how you like, there’s nothing like it.

TL/DW for those of us who don’t learn well from video content?

You would expose a single port to multiple vlans, and then bind multiple addresses to that single physical connected interface. Each service would then bind itself to the appropriate address, rather than “*”

Always happy to try and productively add to someone’s learning.

I guess I just misunderstood what you were arguing then. For posterity: I believe datasets containing children is fine, datasets containing csam is not, and the legality of generating csam should be left up to psychologists on whether or not it is a societal net benefit. Whichever way is better for children that exist is my vote.

I’m not arguing whether or not it should be legal, I was just offering my first hand experience in regards to the capabilities of these local models since people seem to be confused as to how this actually works.

That would be true, it’d be pretty difficult to build a model without any pictures of children at all, and then try and describe to the model how to alter an adult to make a child. Is anyone asking for that though? To make it illegal to have regular pictures of children in these datasets?

I’m not going to say that csam in training sets isn’t a problem. However, even if you remove it, the model remains largely the same, and its capabilities remain functionally identical.

You should consider reversing the roles. There’s no reason your homelab cannot be the client, and have your vps be the server. Once the wireguard virtual network exists, network traffic doesn’t really care which was the client and which was the server. Saves you from opening a port to attackers on your home network.

It doesn’t need csam data for training, it just needs to know what a boob looks like, and what a child looks like. I run some sdxl-based models at home and I’ve observed it can be difficult to avoid more often than you’d think. There are keywords in porn that blend the lines across datasets (“teen”, “petite”, “young”, “small” etc). The word “girl” in particular I’ve found that if you add that to basically any porn prompt gives you a small chance of inadvertently creating the undesirable. You have to be really careful and use words like “woman”, “adult”, etc instead to convince your image model not to make things that look like children. If you’ve ever wondered why internet-based porn generators are on super heavy guardrails, this is why.

There is also the argument that it’s more complicated under the hood and harder to troubleshoot, particularly because of it’s inherent parallelism and dependency-tree design, whereas initv was inherently serial. It was much more straightforward to pick the order in which services started and shut down on an initv system.

For example, say I write a service and I want it to always be the first service stopped during a shutdown, and I want all other services to wait for it to stop before shutting down. That was trivial to do on an initv system, it’s basically impossible on systemd.

For those wondering, yes I did run into this situation. My solution was clobbering the shutdown, poweroff, and restart binaries with scripts earlier in path search that stop my service, verify that they’re stopped, and then hook back to systemd to do the power event.