Nodja's recent activity
-
Comment on Why is AI pornifying Asian women? in ~tech
-
Comment on DALL·E now available in beta - with full commercial usage rights in ~tech
Nodja (edited )LinkIf you're interested in the open source implementation of the DALL-E 2 model, it is going very well. The big hurdle right now is the training dataset. These models are trained on 400 million...If you're interested in the open source implementation of the DALL-E 2 model, it is going very well. The big hurdle right now is the training dataset. These models are trained on 400 million images, that means that even if you have a crazy fast GPU that can process 50 images per second it would take 3 months just to train the model and test it, the good news is that these models can be easily parallelized and if you have 4 GPUs it takes roughly 1/4 of the time (a little bit more due to some sync overhead). But if there's some bug in the code that isn't apparent until you train on a lot of data, you gotta start everything over...
What this means for the end user is that while these models are considered huge and a bit of a pain to reach the quality stated in the papers, after anyone trains a model and releases it publicly, anyone with a decent GPU can run it. The memory requirements for running a pre-trained model are significantly lower, roughly 1/2. So any GPU with 12GB of VRAM should be able to run these models, even less memory required if you apply some techniques that trim down the model (at the cost of slightly lower quality). So if you're shopping for a GPU and plan on running these models on your own hardware, don't skimp on that VRAM!
I also want to take this chance to say that DALL-E 2 is not the best text to image model around. It's just the one with all the publicity since it's the only one people can check the results publicly. Google's Imagen model overtakes it, and it can actually spell text properly. I'm following the open source implementations of these on github/LAION discord and all I can say is that the simplicity of the Imagen model is what stands out for me, it's a simple chain of diffusion models, this enables you to easily implement improvements made on the diffusion field (which is in it's early stage btw, think of GANs pre-2017), a recent example is speeding up the generation of images to a small fraction of diffusion steps (went from 1000 steps to 20-40 steps), a future possible example is using the model to generate video instead of single images, other models could also be adapted to do this, it's just that the simplicity of Imagen makes it much easier to implement.
You might recently have heard of Parti, also made by google, which performs similarly to Imagen, but at a great cost. Parti uses 20B parameters vs Imagen's 3B. This alone puts it out of scope of consumer hardware and therefore out of reach of most researchers. It's mostly an experiment of scaling a previous model and see what happens, like openAI did with GPT-3, but this time I think most researchers consider it a dead-end since it's an outdated architecture and scaling diffusion models will probably yield much better results.
-
Comment on Changes to ad-supported viewing in ~anime
Nodja Link ParentI agree that some older fansub groups were obnoxious and tried to shove as much Japanese culture in your face in the subs themselves or by adding dozens of translator notes. I recall some old...I agree that some older fansub groups were obnoxious and tried to shove as much Japanese culture in your face in the subs themselves or by adding dozens of translator notes. I recall some old comedy/slice of life anime (School Rumble maybe?) where the fansub would explain all the jokes in long ass translator notes where you had to pause the anime to read, on top of adding a bunch of explanations at the end for a bunch of details that didn't involved dialogue. But while I didn't have much choice which fansub I watched back then (no fast speed internet), I later learned what were the quality fansubs and which ones were not.
Pro subs are not better or worse than quality fansubs.
I'm not sure about now, but Crunchyroll translators used to be former fansub translators. The site was formerly a
anime pirate website after all, they're just considered pro translators because they're being paid for their job now.While pro translators are hand-picked and at least follow minimum guidelines, the quality often suffers due to the time and logistics constraints imposed on them. I understand Japanese at an beginner-intermediate level (~1/2 of N3) and I've seen anything from lines not matching what was said at all--which I suspect the script changed and the subs timer didn't notice--to translations that probably made sense when the script was handed in but don't in the context of the visuals. While these errors are eventually fixed when people spot them, sometimes the pro subs make more subtle mistakes that aren't easily noticed that can have big consequences.
For example for Pretty Boy Detective Club people highly recommend you don't use the pro subs as a lot of meaning is lost/changed in translation to the point that it affects the themes of the show, for reference the show is mostly about themes with little plot. A proper fansub was later released that used the source material as reference to provide a more accurate translation that is closer to the author's intent. I heard more recently that Sonny Boy suffered from similar issues.
Doesn't help matters that Crunchyroll pays 1/3 the rate of other services to their translators. So I doubt they're able to retain the high quality personnel they need.
-
Comment on Changes to ad-supported viewing in ~anime
Nodja Link ParentHorribleSubs and SubsPlease are the main ones. They do what you would call web-rips on the TV scene. They rip off websites like crunchyroll/netflix/amazon/hulu/etc. Just to note that HorribleSubs...HorribleSubs and SubsPlease are the main ones.
They do what you would call web-rips on the TV scene. They rip off websites like crunchyroll/netflix/amazon/hulu/etc.
Just to note that HorribleSubs disbanded a while ago and SubsPlease came to fill the void. So it's SubsPlease for anything the past ~2 years and HorribleSubs for anything before that.
-
Comment on Proton vs. Native: Is there really a difference? in ~tech
Nodja Link ParentMost game developers don't have the time or political influence on their publisher to target linux directly. Most games are released with known bugs and multiple gigabytes of day 1 patches. Even...Most game developers don't have the time or political influence on their publisher to target linux directly. Most games are released with known bugs and multiple gigabytes of day 1 patches. Even if they only target the steam linux runtime, linux will not be as well tested and you're basically asking the devs to at the minimum double the testing surface, i.e. now the multiple machine configs need to be tested across both windows and linux, with varying levels of graphic drivers, etc.
Adding all that extra work for a small userbase that is more likely to add bad reviews for a bad port, or polute your bug tracker with linux support issues is unrealistically to expect from developers that already delay games and suffer from crunch culture. Don't forget that linux is one of the more elitist and toxic cultures that exists in the tech world, most devs don't want to deal with that shit.
Would it be nice if companies would release linux native ports of their games? Absolutely. But it simply ignores the reality that is the game dev world.
Proton is not a crutch, but a bridge for game devs to get the linux userbase to try their games. Wine is an amazing piece of software, and while I'm not familiar with the internals like a wine dev is, I used to follow the wine-devel mailing list back in the mid-2000s and I'm aware at a surface level what the devs need to do to get stuff running on linux. It's much less emulation or translation, and more like teaching linux how to speak Windows, in the sense that while it's not it's main tongue, it can get pretty damn close.
-
Comment on Netflix suspends trans employee who tweeted about Dave Chappelle special in ~lgbt
Nodja Link ParentThe generation names have changed meaning in certain internet cultures. Boomer is just shorthand for "old people that don't get my generation" and zoomer is just shorthand for "spoiled young...The generation names have changed meaning in certain internet cultures. Boomer is just shorthand for "old people that don't get my generation" and zoomer is just shorthand for "spoiled young people that never suffered and are addicted to their devices".
For example, if you watched an old movie and liked it you say "I've watched this boomer movie the other day and..." even tho the movie is probably only 20 years old. i.e. the matrix is a boomer movie. Or if there's a new trend/show that you don't get why it's popular you say "I don't get these zoomers and their tiktoks and vtubers".
Of course this is all colloquial online discourse, the change in the terminology always takes a bit to catch up on serious discussion/articles. But the nicknames for the generation are no longer direct references to the generation, but to a certain way of thinking.
-
Comment on TikTok overtakes YouTube for average watch time in US and UK in ~tech
Nodja LinkI wonder how much the algorithm each platform uses affects usage. YouTube's algorithm has failed to introduce me valuable new content/channels and most of my discovery has been done externally....I wonder how much the algorithm each platform uses affects usage. YouTube's algorithm has failed to introduce me valuable new content/channels and most of my discovery has been done externally. YouTube's algorithm seems to find your interests and stick with them giving you a very narrow selection of channels that get recommended and never expands. While tiktok while mostly giving me things for my tastes, also throws me random stuff from creators that I oftentimes just skip but sometimes enjoy.
YouTube also really insists on recommending me videos that I've already watched, right now my feed has 4 videos that I watched fully yesterday (the have the full red bar under them), and they'll usually come back for a while until I click "not interested" and no matter how many times I mark videos as "Not Interested -> I already watched video" it consistently keep recommending me videos of stuff I watched before.
On tiktok every video I watch is fresh, well some of them are duets/stiches of videos I've watched before but you get the meaning. The value of YouTube used to be it's content discovery platform, and I fear that if you've been on YouTube for too long, it won't recommend you new stuff, unless that new stuff is from the same circle of creators you watch. On mobile I get a button that says "New to You" and it's kinda what I want, but why is mobile only and hidden away? It doesn't even show up sometime having to restart the app for it to show up. It's also mobile only so when I'm most likely to use the feature- sitting on the sofa watching random stuff or on the pc looking for background noise to listen to- the feature is not available to me. Why isn't the "new to you" feed incorporated into the recommendations? I would make more sense if you had 10% of recommendations be "new to you" videos than you having to jump through hoops to find new stuff.
-
Comment on YouTube experimenting with removal of public dislike count in ~tech
Nodja LinkTitle is wrong, it's removal of dislike count. you can still dislike it, you just won't know the like/dislike ratio.Title is wrong, it's removal of dislike count. you can still dislike it, you just won't know the like/dislike ratio.
-
Comment on Does reformatting an ext4 partition fix bad sectors, and what are they anyway? in ~comp
Nodja LinkAs other people said this is most likely an hardware failure. But before you throw the drive away consider that the failure may not be coming from the SSD itself. This is much less unlikely to be...As other people said this is most likely an hardware failure. But before you throw the drive away consider that the failure may not be coming from the SSD itself. This is much less unlikely to be the case, but worth spending the extra 5 minutes.
If this is a SATA drive, try using another cable and plugging it into another port (a different colored one if available). Faulty cables are pretty common cause for read errors, SATA controller failures less so, but doesn't hurt checking (different color port usually means different controller).
If it's an m.2 drive (SSD that slots directly into the motherboard), then try a secondary m.2 slot. Very unlikely to do anything but again, doesn't hurt.
-
Comment on Is there a video downloader extension/program that still works? in ~tech
Nodja LinkYoutube recently removed the old layout for pages which made some downloaders that rely on them to fail. You have to update them. Youtube-DLG is a simple frontend for youtube-dl, you need to...Now neither is working
Youtube recently removed the old layout for pages which made some downloaders that rely on them to fail. You have to update them. Youtube-DLG is a simple frontend for youtube-dl, you need to update youtube-dl itself by going to the gear on the top right and clicking update.
If that fails, download the executable from https://youtube-dl.org/downloads/latest/youtube-dl.exe and place it in
%appdata%\youtube-dlg(ctrl+v this into the start menu) replacing the version that's there. -
Comment on YouTube Vanced: A privacy-friendly YouTube app for Android with ads and telemetry stripped out in ~tech
Nodja Link ParentIIRC Youtube changed some internals and no longer serves the old version of pages, you could get the old version of youtube with disable_polymer=true url param. This affected mostly channel pages...Oh then the hit and misses is something else. Sad I hoped it was that that caused the crashes since they cropped up at the same time. A shame then the issues are theirs instead, bummer
IIRC Youtube changed some internals and no longer serves the old version of pages, you could get the old version of youtube with disable_polymer=true url param. This affected mostly channel pages and playlists.
-
Comment on YouTube Vanced: A privacy-friendly YouTube app for Android with ads and telemetry stripped out in ~tech
Nodja Link ParentYes. When anyone with youtube premium watches a video, the video is flagged as "watched by youtube premium member" and goes into a pool. At the end of the month, 55% of the money of all youtube...Yes.
When anyone with youtube premium watches a video, the video is flagged as "watched by youtube premium member" and goes into a pool. At the end of the month, 55% of the money of all youtube premium subscriptions is split among all content creators with videos on that pool, the split is made by total watch time (not video length). Note that it's not total watch time by youtube premium members, it's watch time by everyone that watched it, being a youtube premium member just flags videos for inclusion. I assume this is done this way so that the millions of videos on youtube with 1-5 views don't dilute the pool (90% of videos don't reach 1000 views).
A youtube creator that I follow on twitch said once that half his youtube income is from premium. He makes long form game critiques.
-
Comment on After 352 days since PC release, Red Dead Redemption 2 was finally cracked this morning in ~games
Nodja Link ParentDisclaimer: I've done reverse engineering, but never reverse engineered DRM, my knowledge is purely based on what I've read online. I don't believe that's how this new protections work. They use...- Exemplary
Disclaimer: I've done reverse engineering, but never reverse engineered DRM, my knowledge is purely based on what I've read online.
I don't believe that's how this new protections work. They use an online component for authentication, but it has nothing to do with how the software is protected.
These new protections usually work by converting x86-64 bytecode into an internal representation that runs inside a VM. So say, if you want to protect a function, you can mark it to be protected by Denuvo/VMProtect/whatever and instead of getting x86, you get some obscure bytecode made for a custom VM that you need to at least be able to understand so you can start removing the protections, this new VM is not made sensibly, it's made to be as obtuse and quirky as possible so that it's very hard to reverse engineer.
Note that this is the main reason people say that the DRM slows down the game, it's technically true, but developers are smart enough to not protect function that run in the main game loop. So most of the games are not affected by such protections, DRM will mostly add some extra seconds to the loading screens.
Note that the goal of these protections is not to protect the game fully, but to annoy reverse engineers sufficiently that it'll take them too long to reverse the protections. On top of the VM stuff there's also tons of obfuscation and dead code among other more classic techniques. Combining all of this makes it so that it'll take weeks/months of work/hours to be able to remove the authentication checks on protected functions. Some cracking groups probably have written tools/compilers to convert denuvo-vm back to x86-64 or deobfuscators, etc. So their releases are usually faster, but as the cracking groups techniques evolve, so do the DRM protections.
Online is still required for these protections mind you, but it's just a token system. The DRM servers basically check if you own the license and issue a token with an expiration date which the game stores locally, the protected functions just check if the token is valid for your machine and is within the expiration date.
-
Comment on Alternative to TeamViewer for LAN? in ~comp
Nodja LinkTeamViewer does not require an account for this. Setup a password for unattended access: Extra -> Options -> Security Enable direct IP address from LAN: Extra -> Options -> General -> change...TeamViewer does not require an account for this.
- Setup a password for unattended access:
- Extra -> Options -> Security
- Enable direct IP address from LAN:
- Extra -> Options -> General -> change "Incoming LAN connections" dropdown to accept
- Make sure TeamViewer starts with windows:
- Extra -> Options -> General -> make sure "Start TeamViewer with Windows" is ticked
With all these, you should be able to access the computer through TeamViewer, even if the computer has no internet access, as long as windows is booted (even just the login screen is fine). To connect to it just use the LAN IP address instead of the TeamViewer ID (TeamViewer ID also works if the computer has internet access), and use the password you setup for unattended access instead of the generated one.
edit: formatting
- Setup a password for unattended access:
-
Comment on Is there a Google-free future for Firefox? in ~tech
Nodja Link ParentYou can remap the 0 key to the o key, assuming you're on windows Download and install autohotkey Download this autohotkey script (ctrl+s on the page). Double click the file This will make 0 type o...You can remap the 0 key to the o key, assuming you're on windows
- Download and install autohotkey
- Download this autohotkey script (ctrl+s on the page).
- Double click the file
This will make 0 type
oand alt+0 typeO.
To disable the mapping just right click the tray icon that shows up when you load the file. -
Comment on Passing the Same Parameters to Multiple Functions in ~comp
Nodja Link ParentThe only change I would make would be to change def distances(sentence) to def distances(sentence, window=window, scale=scale), this makes it explicit what data the function is using. When reading...The only change I would make would be to change
def distances(sentence)todef distances(sentence, window=window, scale=scale), this makes it explicit what data the function is using. When reading functions I like to know what kind of data they're working on, and using "pseudo-globals" like that can be surprising, even to yourself down the line. -
Comment on A first look at Unreal Engine 5, including a real-time demo running on PS5 in ~games
Nodja Link ParentMS has come out and said they'll support slower media for old games (anything before XSX). On Sony's side they'll only support m.2 SSD's that are at least as fast has the one the PS5 has (5.5GB/s)MS has come out and said they'll support slower media for old games (anything before XSX).
On Sony's side they'll only support m.2 SSD's that are at least as fast has the one the PS5 has (5.5GB/s)
-
Comment on A first look at Unreal Engine 5, including a real-time demo running on PS5 in ~games
Nodja Link ParentWhile internet speeds have improved, you still have to wait 3-5 hours for the download to finish for some of these games (on a 100Mbit/s connection). Also possibly blowing 1/5 of your monthly...-
While internet speeds have improved, you still have to wait 3-5 hours for the download to finish for some of these games (on a 100Mbit/s connection). Also possibly blowing 1/5 of your monthly internet cap in one day, not to mention rural areas that don't even have 10Mbit/s and have no access to the game for several days because the game requires a several dozen GB patch every other week.
-
Future consoles will require SSDs, which are currently ~$100/TB for a version that is at least 5x slower than what's shipping with the consoles.
Merge 1 and 2 together and you've got a cumbersome system where you have limited storage space and can't really just delete and redownload because it takes many hours to do so.
So I think that yes. consumers should care that games are blowing up in size and ask developers to tone it down a notch for next gen.
-
-
Comment on What are your unpopular movie opinions? in ~movies
Nodja Link ParentDoesn't it double the cost and time of a render? I heard VFX companies have huge render farms, so it can't be cheap. And I assume some VFX stuff is also handled manually frame by frame, so this...The decision to use 24fps is laziness and habit
Doesn't it double the cost and time of a render? I heard VFX companies have huge render farms, so it can't be cheap. And I assume some VFX stuff is also handled manually frame by frame, so this would also double manual work.
-
Comment on "#GameofThrones was basically an expensive film school [for us]"—David Benioff & D.B Weiss in ~tv
Nodja Link ParentYeah, it seems to me that there was a lot of talent involved in the TV series, even G R R Martin himself was involved in the show's production and he has plenty of experience writing for...Yeah, it seems to me that there was a lot of talent involved in the TV series, even G R R Martin himself was involved in the show's production and he has plenty of experience writing for television. It's more of a case that when the ship's crew knows what to do, all the captain has to do is to not pilot the ship towards the rocks.
Sure, but the prompts you given it aren't example of that. Google images for "a city without cars" and you get multiple examples of cities without cars, mixing concepts like "a city without cars" and "new york city" isn't hard for the model either, this is why the "avocado chair" example astonished people when these models were first coming out. Googling "woman lying down after workout" also wields plenty of results, so I don't know why you think the models have never seen that.
You're just misunderstanding how it works. The text to image model you used doesn't use an LLM as the backend for the text portion, so it's understanding of text in more akin to word salad. the vectors for the prompt "a city without cars" and "a city with cars" are virtually the same. There's a paper explaining this phenomena in the CLIP text encoder if you want to know more.
If you want your results you need to be more creative with your prompting, try "new york with empty roads" and you'll have more luck, putting "car" in the negative prompt also helps. My GPU will be busy for the next couple days so I can't show you SDXL results, but here's what novelai can do (SDXL based anime model).