• le mois dernier
Transcription
00:00All right, so this is simultaneously really impressive and really frightening at the same
00:10time and it's hitting me in ways that I didn't really expect.
00:12So do you remember Will Smith eating spaghetti?
00:15Do you remember when this was what AI generated videos look like?
00:18Remember when we said, okay, this AI stuff is cool and all, but clearly there's a long
00:22way to go before there's any need for concern.
00:25Well, welcome to the future people, because this is also an AI generated video.
00:32And so is this, completely synthesized out of thin air by computers.
00:36This one too.
00:37This is not real.
00:39Absolutely ridiculous how far we've come in literally one year.
00:43This does feel like another ChatGPT Dolly moment for AI.
00:49And maybe I'm overreacting because, okay, I'm a video creator.
00:53So an AI that's actually doing my job, maybe that feels a little more threatening.
00:58So I'm particularly impressed by it, but also this, this stuff is really good.
01:02So today Sam Altman and OpenAI announced a new model called Sora, and it can generate
01:08full up to one minute video clips from just text input.
01:13So the same way Dolly was able to understand our text input and turn it into a photorealistic
01:19or stylized image or whatever you want.
01:22Same thing with Sora.
01:24But now since it's videos, it also needs to understand how all these things like reflections
01:30and textures and materials and physics all interact with each other over time to make
01:36a reasonable looking video.
01:38And of course, right away, there's a bunch of examples on their website that are crazy.
01:41Now, before I show you these, I just need you to keep this in mind.
01:44You're about to watch a bunch of AI generated videos, and you know that you're about to
01:49watch a bunch of AI generated content.
01:51So your brain, you're already looking for this stuff, and it's not perfect, you will
01:55find imperfections.
01:57But not everybody who sees AI generated content on the internet knows to be looking for that.
02:04So also keep that in mind.
02:06This is also the worst that this technology is going to be from here on out.
02:10So okay, here's one of the videos.
02:12There's no audio to any of these clips.
02:13But the prompt for this one is, a stylish woman walks down a Tokyo street filled with
02:19warm glowing neon and animated city signage.
02:22She wears a black leather jacket, a long red dress, and black boots.
02:27This video is already miles ahead of where we were.
02:31It has accurate lighting, it has materials, it has skin tones, movements, even has reflections
02:37all over the place.
02:38Now, of course, if you look at it for more than about 10 seconds very closely, there
02:42are lots of giveaways.
02:44Like this dude in the background kind of looks like he's gliding in a weird way.
02:48The frame rates and the reflections in the water are for some reason lower than the
02:51rest of the video.
02:52The camera movement overall is just a bit inconsistent.
02:55And it just, I don't know, just kind of feels a little bit off.
02:59But then again, this is where we were one year ago.
03:02So just keep that in the back of your head for all this.
03:05Okay, how about this one?
03:06This is another one, which has a long prompt about a camera following behind a white vintage
03:11SUV with a black roof rack as it speeds up a steep dirt road.
03:16This is also, again, really good.
03:20It kind of looks a little more video gamey because of how rock solid the drone footage
03:24is, but clearly very usable.
03:27Here's another one.
03:28A litter of golden retriever puppies playing in the snow.
03:31Their heads pop in and out of the snow covered in it.
03:34It's so good.
03:35It feels like the physics of the fur and the ears and everything with the snow flying
03:39around in slow motion is incredible.
03:42I've looked through all the sample videos on OpenAI's website.
03:45And clearly these are the handpicked best ones that they chose to share where they
03:49just put in some text and then get a video and don't modify it.
03:53But there's really impressive stuff in there.
03:54Some of it has humans, some of it doesn't.
03:56Some of it is more realistic feeling like the truck driving one, but some of them are
04:00more video gamey or more stylized.
04:02A lot of it is slow motion.
04:03I just have to say how insanely fast these models are improving is genuinely, like that's
04:10the shocking part.
04:11I remember not even that many months ago, Dolly 3, really, really high end.
04:17And you could always still find something off about it.
04:19Like especially if you ask it for something like a photorealistic image of a human, something
04:26about like the hands or the ears would always just be a little bit off.
04:30Nevermind the physics.
04:31But even this video here is crazy at first glance.
04:36The prompt for this AI generated video is a young man in his twenties is sitting on
04:40a piece of a cloud in the sky, reading a book.
04:43This one feels like 90% of the way there for me.
04:48Like it's beyond the uncanny valley of like Apple's personas, which are actually based
04:52on humans.
04:53This is a made up person.
04:54I mean, his eyes are kind of weird.
04:56And the motion of the pages in the book are kind of odd.
04:59And yeah, obviously he's in a cloud and that's a giveaway, but like the lighting and the
05:03shadows and the skin tones, and then all the realism of the textures on the shirt and
05:07the way that the shirt and the pants move and the hair, they're all really impressive.
05:12And then for this one, they typed in a movie trailer featuring the adventurers of the 30
05:17year old spaceman wearing a red wool knitted motorcycle helmet, blue sky, salt desert,
05:25cinematic style shot on 35 millimeter film.
05:28And the closeups of his face, the fabrics on the helmet, the film grain through every
05:34shot and the cinematic style.
05:36This is one of the most convincing AI generated videos I've ever seen, minus maybe the weird
05:42physics of that dude walking kind of in fast motion.
05:45So Sam Altman, if you follow him on Twitter, he's going through a whole bunch more of
05:48like people's requests and posting a bunch more generated videos.
05:51And so if you want to check out his profile, you can see those.
05:53But here's the thing about these AI generated videos now, as good as they've gotten to this
05:59point, they can and will pass as real videos to people who are not looking for AI generated
06:09videos.
06:10Now that is obviously insanely sketchy during an election year in the US and also terrifying
06:15for a bunch of other internet related reasons.
06:18But it's also perfect for stock footage.
06:22Like there are already all kinds of presentations and advertisements and then PowerPoints that
06:28are in need of oddly specific stock videos.
06:33And these AI generated videos are already good enough to 100% pass for that purpose.
06:41Like look at this one, this one with the waves at Big Sur, this drone shot.
06:45Honestly, if I saw this on Twitter, I wouldn't even think twice.
06:48I'd be like, oh, nice drone shot, dude.
06:51Wouldn't even think about AI if I wasn't pixel peeping at like the way the water was moving
06:55like this.
06:56This is a totally usable video in an ad for some California based product.
07:01And that has all sorts of implications for the drone pilot that no longer needs to be
07:06hired for all the photographers and videographers whose footage no longer needs to be licensed
07:12to show up in that ad that's being made.
07:15It's already that good.
07:16There's other stuff like this wall of TVs, which would be a totally expensive and difficult
07:21thing to shoot with a camera and all these old expensive props.
07:25But if you can just generate it this well with reflections and the environment and
07:30everything else around it, I mean, why do it any other way?
07:34It's also very capable of historical themed footage.
07:37So this is supposed to be California during the gold rush.
07:41It's AI generated, but it could totally pass for the opening scene in an old Western with
07:45the right music over it.
07:46How long until an entire ad, every single shot is completely generated with AI?
07:51Or what about an entire YouTube video or an entire movie?
07:54I'm tempted to say like we're a long way away from that because, you know, this still
07:58has flaws clearly and there's no sound and there's a long way to go with the prompt
08:02engineering to iron these things out.
08:05But then again, the spaghetti was like a year ago.
08:09Now, actually, like that open AI on their website, they show some of the downfalls too
08:13of this particular model and because who would know better than the people who have been using
08:18it?
08:19This is a very private tool, by the way, right now, it's in super limited access.
08:21So it's in the hands of red teamers, which basically means people testing it, pushing
08:26the limits, trying to break it and a few trusted creators.
08:30But they have found plenty of weird edge stuff like this clip here of a bunch of gray wolf
08:36pups looks normal at first, but then it's pretty clear that something's kind of off
08:39with the way they're just kind of appearing out of nowhere and walking through each other.
08:44That's kind of weird.
08:45Or this clip of a guy running on a treadmill, which I mean, I don't really have to say much
08:49more about why this one's weird, but this is my favorite one again.
08:52So again, just try to try to put yourself in the mind of someone who's not expecting
08:56AI.
08:57You're just scrolling through Facebook or Twitter or something, right?
09:01So you just see this video.
09:02So first, I just want you to watch this clip as if it's just a stock video you found of
09:05a grandma celebrating her birthday and just try to try to think like what I wonder what
09:10birthday she's celebrating.
09:11Right?
09:12I don't know.
09:13How old do you think she is?
09:1460?
09:1565?
09:16Maybe it's the big 7-0.
09:17She seems to really like that cake.
09:18Did you see it?
09:19Did you catch that?
09:20I'm going to play it again, but this time, watch the video knowing that AI-generated
09:26photos and videos have trouble accurately doing hands.
09:32I'll play it again.
09:33And now it feels super obvious.
09:36Like every time you watch it, watch a different set of hands.
09:38It gets weirder and weirder.
09:40You can watch it like five times and there's dead giveaway after dead giveaway, not even
09:44mentioning the weird inconsistencies with the direction of the wind on the candles.
09:48But even as I'm saying all that, even as it's coming out of my mouth, I can't help but remember
09:53that 12 months ago we were critiquing this.
09:58So what does this all mean?
10:00Well, I mean, there's what it means now and there's what it means for the future.
10:05Now, Sora, this thing that they've made is clearly a really impressive video generation
10:11AI tool that is both going to fool people and also be very useful.
10:18There's also a watermark in the bottom corner of every video generated by it.
10:22So if you see one of those videos and ideally it hasn't been cropped out, then that's at
10:26least a pretty clear indicator that it's AI-generated.
10:29It's a Sora video.
10:30But also, I do think they're going to have to be very careful with this.
10:34They're going to have a whole bunch of safety stuff to keep in mind.
10:36I think they'll probably have to be even more safe than Dolly.
10:39You shouldn't be able to generate people's likenesses, like you shouldn't be able to
10:43make a politician look like they're doing something on video, especially this year.
10:47You probably won't be able to make Will Smith eating spaghetti.
10:50But it also definitely means stock video generation is absolutely going to take a dent out of
10:59video licensing.
11:00Like I can basically guarantee that.
11:02Like logistically, why would anyone making something pay for footage of a house in the
11:07cliffs when they can generate one for free or for a small subscription price?
11:12Like that is the real scary part of what this tool implies.
11:15But in the future, it gets pretty existential, man.
11:20I mean, okay, if this is trained on all videos that have ever been made by humans, then surely
11:25it can't be innovative or creative in ways that humans haven't already been, right?
11:32I don't know.
11:33Either way.
11:35I'll have all the links below for all the Sora stuff, for OpenAI stuff.
11:38And I guess I'll talk to you next year when we look back and go, remember that first version
11:43of Sora and how bad those wolf pups looked when they spawned out of nowhere?
11:49Just remember, this is the worst that this technology is going to be from here on out.
11:54Thanks for watching.
11:55Catch you in the next one.
11:56Peace.