What a Surprise

[00:02] That's the sound of AI coming in and

[00:04] non-consensually motorboating us all.

[00:07] Whether you want to admit it or not,

[00:09] whether you even recognize it or not,

[00:11] your life is already being affected

[00:14] deeply by AI. And I know I've yapped

[00:17] about it a lot, but I did just see The

[00:20] Atlantic's most recent report where they

[00:22] actually pulled down the pants on a lot

[00:24] of what AI models are being trained on

[00:26] here, aka what they're stealing to train

[00:29] their AI models on because they don't

[00:31] have permission for it. The word, the

[00:33] nomenclature is stealing. That's the

[00:35] scientific term. But because these

[00:37] multi-billion corporations are in the

[00:40] the driver's seat now at the helm, they

[00:42] can go ahead and rebrand that whole

[00:44] stealing thing into something entirely

[00:46] different. They're trying to argue that

[00:48] it's all by the book in fact. So, uh,

[00:51] The Atlantic now has this AI watchdog,

[00:53] which is just a open tool. Anyone can go

[00:55] in and they can just directly search

[00:57] what music AI is being trained on as

[01:00] well as like videos AI is being trained

[01:02] on. Now, a ton of huge artists have been

[01:04] made aware of this and have called out

[01:06] AI companies for just directly ripping

[01:09] their music. Some of it unreleased, by

[01:11] the way, in order to train their AI

[01:13] models on. Notably, Sunno has been

[01:16] getting called out with spitballs fired

[01:18] at it and Rotten Tomatoes thrown at it

[01:20] because they're pretty shameless and

[01:22] unapologetic about it. Anyone remember

[01:25] back in the day the whole you wouldn't

[01:26] download a car, one of the most iconic

[01:28] [ __ ] antipiriracy advertisements

[01:31] ever? arguably just one of the most

[01:33] well-known uh campaigns ever as well.

[01:36] Well, now the same people that used to

[01:39] run that [ __ ] are the ones that are

[01:41] literally just downloading everything,

[01:43] stealing everything they can. It It's

[01:46] It's pretty incredible the uh 180

[01:48] they've done there on that whole

[01:50] messaging. And I know what you're

[01:52] thinking. Yes, they have stolen my

[01:56] videos to train their AI models on. I

[01:59] searched my name in their data set. 67

[02:02] of my videos have made it into nine data

[02:05] sets used by AI corporations to train

[02:08] their models.

[02:10] Heaven help us. Lord have mercy on the

[02:14] absent souls of these AI models that are

[02:17] being [ __ ] clockworked orangeed

[02:19] having the eyelids pulled wide open to

[02:22] be trained off my videos here. Those

[02:25] have to be the dumbest AI models you can

[02:27] find. Just look at this poor bastard

[02:29] here from YT Temporal 180M who's got 221

[02:33] of my videos shoved down its throat

[02:36] being trained off of things like

[02:37] horrible chair where I'm just making fun

[02:40] of a dog [ __ ] chair. Some old game plan

[02:42] commentary like Doom four feathers where

[02:44] I'm playing a Doom mod where I'm a

[02:46] chicken shooting at warthogs from Halo.

[02:50] It's [ __ ] uh washing machine repair

[02:53] championship where I'm just commentating

[02:56] a washing machine repair competition

[02:59] between some of the highest quality

[03:00] athletes you can find in the washing

[03:02] machine circuit. Not to bismerch the

[03:04] good name there. A foot fetish wedding

[03:06] dildo circus where you know I did a lot

[03:08] of cool trick shots with dildos. Like

[03:11] this this [ __ ] AI model. It must be

[03:14] just sitting there drooling. Actually,

[03:16] let me learn a little bit about my son

[03:18] here, seeing as I taught him everything

[03:20] he knows. YT Temporal 180M. It's a

[03:23] collection of 5.4 million YouTube videos

[03:26] compiled by a team of researchers at the

[03:29] University of Washington and the Allen

[03:31] Institute for AI to train a multi a

[03:33] multimodal model called Merllo.

[03:37] It was released in 2021.

[03:40] Bro, is Merllo an idiot? Be honest with

[03:42] me. Is this the dumbest AI you can find?

[03:44] If it's being trained even partly on

[03:46] some of my videos, there's a chance you

[03:47] ask it who wrote the Declaration of

[03:49] Independence and it says dildo titty

[03:51] fart or something. Very fascinating.

[03:54] Very interesting. So, YT Temporal got a

[03:58] huge dollop of some of my incredible

[04:02] work like hintai survive. What the hell

[04:05] is this? I don't even remember this. As

[04:07] the 21st century continues to evolve,

[04:10] human sexual fetishes are evolving right

[04:12] there alongside it. Okay, you know what?

[04:14] That one might actually be somewhat

[04:15] educational. That that actually might

[04:18] help them out a little bit because it

[04:20] it's not wrong. Now, obviously, my

[04:22] videos weren't handpicked for these AI

[04:24] models or anything. They made it into

[04:26] those giant compilations a lot of these

[04:29] groups put together solely to train

[04:32] models off of to give like a huge sample

[04:35] size. I I understand that, but it

[04:38] doesn't make it any less garbage. It's

[04:40] so [ __ ] ridiculous because it is just

[04:43] stealing it. Same thing when they do it

[04:45] with music. It is just stealing all of

[04:47] that to train their models off of. And

[04:50] it's been a huge contentious topic for a

[04:53] while now when it comes to AI. And no

[04:55] amount of like trying to put makeup on

[04:58] it changes the truth that they are just

[05:01] stealing to train their models. Now, if

[05:04] you as just a normal person try and

[05:06] follow their footsteps and do the same

[05:08] thing of just taking a ton of artists

[05:11] music and videos in order to make your

[05:14] own product off of those works, you're

[05:17] going to get arrested and charged with a

[05:19] little something called theft. But that

[05:22] word doesn't exist once you reach a

[05:24] certain level of wealth. These

[05:26] billion-dollar corporations get to kind

[05:29] of skirt around that a little bit.

[05:31] They're able to tiptoe around that. and

[05:33] we don't need to worry about a little

[05:34] pesky thing called theft for them. It's

[05:36] very different rules they play by there.

[05:39] And I know there's not a soul on this

[05:41] planet surprised by this, but it's still

[05:44] something I think worth yapping about,

[05:46] especially now that it's so easily

[05:47] accessible to see how many things just

[05:50] get up by these AI companies to train

[05:55] their models off of. all of these huge

[05:57] compilations of work and IP that they

[06:00] steal to train their models on in order

[06:02] to sell it to the people that are now

[06:04] hooked on AI as everyone is in this

[06:06] giant [ __ ] gold rush, this whirlwind

[06:09] of the in the industry. It's just so

[06:12] bizarre how things have just accelerated

[06:16] to this point. Now, one thing I got very

[06:18] curious about is trying to think of like

[06:20] the most latigious company I could think

[06:22] of to see if their work had also just

[06:24] been stolen by a ton of these AI models.

[06:27] So, obviously, the first thing that pops

[06:30] in my noodle is Nintendo. And yeah,

[06:33] Nintendo is not exempt from this. 4,926.

[06:37] That is a pretty big chunk of their

[06:40] trailers being used to train AI models

[06:43] that I'm sure Nintendo didn't sign off

[06:45] on because Nintendo wouldn't just be

[06:47] giving away this for free. 1,000%.

[06:51] You would need to pay oodles of clams to

[06:55] have access to their material in order

[06:57] to train your own stuff off of to sell

[06:59] that product. They are extremely strict

[07:03] when it comes to their copyright. They

[07:05] rule that [ __ ] with an iron fist. They

[07:07] are judge dread when it comes to their

[07:10] copyright. And yet here we have Runaway

[07:12] Gen 3 that just shamelessly takes 4,772

[07:17] of their trailers. Again, Nintendo is

[07:19] extremely strict with their trailers.

[07:21] You can use their trailers in like

[07:23] YouTube videos if you follow a very very

[07:26] particular set of guidelines around it

[07:28] that is extremely transformative.

[07:31] Like there are in intense rules

[07:33] Nintendo. so brutal with this. I

[07:35] remember there was a couple YouTube

[07:37] channels that got taken down because

[07:38] they used Nintendo music. Like they they

[07:41] took Nintendo music from games they

[07:43] owned, put it in their videos, and they

[07:44] lost their whole channels for it. There

[07:46] was also that time where some streamers

[07:48] got banned because they watched Nintendo

[07:50] trailers during the direct on stream.

[07:52] Like the point is they take that very

[07:54] seriously. And now Runaway Gen 3 just

[07:57] comes in here from the top turnbuckle.

[07:58] Takes 4,772

[08:00] of their trailers, most likely

[08:03] completely for free without Nintendo's

[08:05] permission, didn't pay a dime for it, to

[08:07] train their model on. So, uh, let's see

[08:09] what this is. Runaway AI collected

[08:11] YouTube videos to train a

[08:12] videogenerating AI model released as Gen

[08:14] 3 in 2024. An internal company document

[08:17] that was obtained by 404 media lists

[08:19] 3,970.

[08:21] What? Wait, that can't be right. Oh, oh,

[08:23] channels. I thought I was talking about

[08:24] videos in general. I was like, "No, they

[08:26] have more than that from just Nintendo."

[08:27] That Runaway identified as sources of

[08:29] highquality video for training.

[08:32] Nintendo video game trailers. Okay. The

[08:35] spreadsheet contains comments describing

[08:37] what is desirable about some of the

[08:38] channels. For example, beautiful

[08:40] cinematic landscapes, high quality

[08:42] scenes from movies, only four videos,

[08:44] but they are really well done. Super

[08:45] high quality sci-fi short films, and the

[08:48] holy grail of car cinematics so far.

[08:52] That must be talking about like Mario

[08:53] Kart or something. It's not clear which

[08:56] if any videos Runway actually used for

[08:58] training its AI system. Our search our

[09:00] search tool includes all YouTube videos

[09:02] from the named channels that were

[09:03] published before May 17th, 2024, which

[09:06] is 1 month before Runway introduced

[09:08] their model. So, this is something that

[09:11] The Atlantic also made note of. It's not

[09:14] 100% confirmed that they used all of

[09:17] these videos, but the fact that they

[09:18] were compiled by these companies. I

[09:20] think you can make a pretty educated

[09:22] guess that it was likely used to train

[09:24] their models. And also another thing

[09:26] they mention is that there are likely a

[09:29] lot of other ones that even though

[09:30] they're not here, doesn't mean they

[09:31] weren't used to train their model. It's

[09:33] a very tricky and messy, sloppy thing to

[09:36] nail down exactly what and what isn't

[09:38] being used to train models. But I would

[09:41] wager a guess that since it's here in

[09:44] the data set, they probably used it.

[09:47] Much like all of these also most likely

[09:49] used it, like this company, which also

[09:51] used a lot of my YouTube videos, which

[09:54] that actually kind of gave me a giggle

[09:56] when I saw that Nintendo of America was

[09:58] part of the same data set as this one

[10:00] where my videos are in, such as

[10:03] Seaweed's [ __ ] cool or whatever that

[10:06] one was. I already forgot. So Nintendo

[10:09] and I, we're basically in the in the

[10:12] same ballpark now when it comes to

[10:14] quality. That's pretty cool. I bet

[10:16] Nintendo's thrilled about that. Uh but

[10:19] anyway, point is thanks to the Atlantic

[10:22] data sets here that you can freely

[10:23] explore. You can see so so so many

[10:28] things have just been taken by these

[10:29] companies and put in these data sets

[10:31] that are presumably being used to train

[10:33] their AI models on. It's just it's

[10:36] pretty egregious. I I wanted to yap

[10:38] about it a little bit. That's it. See

[10:40] you.