Mirodir @ Mirodir @discuss.tchncs.de

Posts

1

Comments

186

Joined

2 yr. ago

Yeah, I've noticed that too—there's a distinct 'AI vibe' that comes through in the generated responses, even if it's subtle.

It's not as accurate as you'd like it to be. Some issues are:

• It's quite lossy.
• It'll do better on images containing common objects vs rare or even novel objects.
• You won't know how much the result deviates from the original if all you're given is the prompt/conditioning vector and what model to use it on.
• You cannot easily "compress" new images, instead you would have to either finetune the model (at which point you'd also mess with everyone else's decompression) or do an adversarial attack onto the model with another model to find the prompt/conditioning vector most likely to create something as close as possible to the original image you have.
• It's rather slow.

Also it's not all that novel. People have been doing this with (variational) autoencoders (another class of generative model). This also doesn't have the flaw that you have no easy way to compress new images since an autoencoder is a trained encoder/decoder pair. It's also quite a bit faster than diffusion models when it comes to decoding, but often with a greater decrease in quality.

Most widespread diffusion models even use an autoencoder adjacent architecture to "compress" the input. The actual diffusion model then works in that "compressed data space" called latent space. The generated images are then decompressed before shown to users. Last time I checked, iirc, that compression rate was at around 1/4 to 1/8, but it's been a while, so don't quote me on this number.

edit: fixed some ambiguous wordings.

If someone wants to read one of those papers, I can recommend Extracting Training Data from Diffusion Models. It shouldn't be too hard for someone with little experience in the field to be able to follow along.

Understanding the math behind it doesn't immediately mean understanding the decision progress during forward propagation. Of course you can mathematically follow it, but you're quickly gonna lose the overview with that many weights. There's a reason XAI is an entire subfield in Machine Learning.

I think it's much more likely whatever scraping they used to get the training data snatched a screenshot of the movie some random internet user posted somewhere. (To confirm, I typed "joaquin phoenix joker" into Google and this very image was very high up in the image results) And of course not only this one but many many more too.

Now I'm not saying scraping copyrighted material is morally right either, but I'd doubt they'd just feed an entire movie frame by frame (or randomly spaced screenshots from throughout a movie), especially because it would make generating good labels for each frame very difficult.

I haven't personally used it but from what I can find: if you're using torrents with Stremio (e.g. the ones found with torrentio) you are totally uploading parts of what you're watching to others.

Even real Pokémon has things like humans marrying Pokémon in its lore. Not at all a surprise that the "Edgy Pokémon" would push the boundary a little further.

Scarlet and Violet sold 10 million copies during the launch weekend, so I think they'll be pretty content to keep churning out mediocre games at a much faster pace.

In everything but aesthetics, it's the most similar to Ark I'd say.

It's still a bit rough around the edges, but I've had plenty of fun with it so far.

Or into a horse.

It usually happens a lot faster in video games than 3 sessions in. If it happens later in a video game, it's usually a very short, very temporary scene of depowerment.

I had a whole paragraph typed out on my phone but didn't like most of it. By now many other players said most of what was in there already before I had the chance to proofread and reword it. The gist of it was though: Don't alter player characters or take their power away without at least one of those three being true:

• The player agrees beforehand and is aware it will happen.
• The player character has done something so horrendously stupid that it could've easily been their death so e.g. them losing a limb and now having a pegleg is them being lucky.
• It is very temporary, I'd say max 1 in-game day/1session and the player (not necessarily the character) is aware of that.

You might argue that picking that fight that would get them sent to hell would qualify as #2. But with you planning it out ahead of time it's less them doing something dumb and more the DM guiding them to do something dumb.

Giving you the benefit of the doubt of only 3 hour sessions and ignoring the time they planned out their characters, you let them play with their characters for around 6 hours by now and it'll probably be another hour or two until they "die". This might sound harsh but even with you backtracking on this, seriously entertaining this idea in the first place worries me about what else you might have in store.

Regarding the "OP staff of fire" one of your players has: Did you talk to that player about it in private? I find that usually players respond well to the DM being open about something being so overpowered it warps the entire campaign to the point where you have to design every encounter around it. I'd recommend approaching them about it in private, and not at the (virtual?) table when everyone's eager to play already.
Maybe you could just get the player on board to trade the item in for something less disruptively powerful. Essentially nullifying their magic item by being in hell where every enemy is fire-immune while everyone else still has some useful, fun magic toy feels uncool too after all.

Edit: and a player who wouldn't agree to "Hey, your item is so strong I have to design everything around it so you don't just steamroll everything. Can we, for example, have you meet a merchant where your character trades it for something else?" would react HORRIBLY to having it and all levels taken by force to the point where they'll just quit.

no where near Reddit yet on niche subjects

I'm always saddened by how not-active some of those subjects are. For example: Even many large games struggle to have dedicated, active communities on Lemmy (assuming I'm not terrible at finding them, which is sadly also possible). Even some of the largest games have only completely dead communities here. A huge draw of Reddit for me was to be able to talk about the games I play with other people who do too. And mostly, the games I'd love to talk about aren't in the top 10 most played games list.

Now I could try to (re)vitalize those communities I would love to see around, and I have done so shortly after the exodus (on my previous account that died with the instance it was on). However, there's only so much talking into the void I can do until it gets boring.

I also feel like that might be a big issue for people coming over. After I manage to explain to my friends how federation works, they ask me to help them find the [topic of their interest] community, and all I can show them is a community with 10 threads, all over 3 months old and with 0 comments. Sadly it shouldn't surprise anyone they're not sticking around after that.

From everything I've seen, they'll hear it, take it as an excuse to claim you weren't complying, bust down your door based on that and shoot your understandably agitated dogs.

Good luck.

Ah yeah. I can see how anyone could understand it that way.

And same.

Wait, that's what I meant by "clipped".

Now I'm curious in what other way my comment could be taken.

I was curious too and checked the article but skimming it, instead of a total, I found this:

A new analysis from MUSO, a U.K.-based anti-piracy analyst [...]

With the study being done by a clearly biased person/group and that large omission, I think it's fair to assume that the % of total web traffic going to pirates might not have gone up all that much, maybe it even went down.

Why did they clip the XKCD joke to the point where it makes no sense anymore?

Suddenly you have a 26+ character password that you don’t forget and doesn’t compromise you across other services because each is different.

It depends on what is compromised and how the attacker operates. If the attacker has your plaintext password instead of just a (hopefully salted) hash AND targets you individually instead of just having your password in a giant list of login-info to automatically try on other services then it's trivially easy to guess that e.g. your Spotify password is

<Spotify>

</Spotify>

This exact image (without the caption-header of course) was on one of the slides for one of the machine-learning related courses at my college, so I assume it's definitely out there somewhere and also was likely part of the training sets used by OpenAI. Also, the image in those slides has a different watermark at the bottom left, so it's fair to assume it's made its rounds.

Contradictory to this post, it was used as an example for a problem that machine learning can solve far better than any algorithms humans would come up with.