00:00The other thing I alluded to before, you saw an AI company this year, Interpositive, to Netflix.
00:07I'm a moron.
00:09Could you briefly explain to me what this thing does and why Netflix wanted it?
00:15Sure.
00:15I've always been kind of into computers since I was young.
00:20And then when film started to move from film, analog film, to digital,
00:25I became more interested in that aspect of it.
00:27And the visual effects workflow, for many years, has included machine learning.
00:31So I can write pretty shitty Python scripts and stuff like that.
00:35Because with convolutional neural networks, which are the sort of precursors to what the transformer can do,
00:41which is just much more computation simultaneously,
00:43that you would do things like look at what's called a tensor,
00:46which is just the numerical translation of a visual image in numbers,
00:50like the batch number, the frame number, the red, green, and blue values of each pixel in each frame.
00:56It's just that simple, right?
00:58That numeric is called a tensor.
00:59And a tensor, you'd use a convolutional neural network to identify patterns in that
01:03that would reveal what's called edge detection or feature extraction,
01:08which is just identifying patterns enough to know this is where the window ledge is
01:14so we can more easily take the green screen image out and replace it with something.
01:18So that was sort of familiar to me.
01:20Early on, because I, prior to Arsequity, had a small visual effects company,
01:25I've worked with GPUs a lot too.
01:27And the visual effects guys said, hey, you should see, there's a couple of Google
01:30and its other company opening hour doing really interesting stuff with like transformers in video.
01:35So I have learned to be able to sort of like, I'm like, oh, I can actually just call up
01:40and go,
01:41hey, it's Ben Affleck.
01:41Can I come and see what you're doing?
01:42And sometimes people say yes to my astonishment.
01:45And so I did that.
01:46And then I had this moment.
01:47I remember it was I called Matt and I was like, dude, we're fucked.
01:50Like you can just, with a few keystrokes, like generate very realistic looking video.
01:56You know, it was like a day or two where I really was thought like, oh,
01:59should we better do all the movies we can now because it's going to run out.
02:02Then I went back and sat down with a bunch of people, engineering teams,
02:07actually looked at the data, how it was captioned.
02:09But I started going, well, hold on, this doesn't work.
02:11You can't really control this.
02:12It's not really consistent.
02:13And if you say move the camera, well, it just generates a camera in the image.
02:15And let's look at how the data is captioned and how this,
02:19because I understood that, you know, machine learning and generative AI,
02:23it's just functions of back propagation.
02:25Like you have to have paired data.
02:27You have to have the question and the answer, essentially,
02:29in order to teach it to reverse engineer how it arrived at that answer
02:32across every vector and transformer.
02:34And so in looking at that, it became clear to me that this was a brilliant technology
02:41developed by people who were brilliant at, you know, these researchers and scientists
02:44who understood and built something new, but that they had no understanding
02:48of the domain expertise of filmmaking, to the point where they didn't even understand
02:53that, like, native filmmaking, unless you're scraping, like, H.265 clips from the internet,
02:59which is how most of these were trained, actually comes with a very specific
03:03and ordered data set.
03:04Like native, you don't even have to caption it, it exists, right, in the raw format.
03:09So once I saw that and sort of argued with and badgered some of these folks,
03:15and I only know enough to, like, sort of think I know enough to get into these conversations,
03:20which is great.
03:20I appreciate the way you're talking to me, like, I understand any of it,
03:23and I'm really flattered and honored by that, so thank you.
03:26I continue.
03:26Well, I'm doing a very good job.
03:27But the point being, I said, there's a different way to do this than those guys that relies
03:32on the understanding of, like, filmmaking and how an image is actually created in the world.
03:36And they said, no, it'll generalize, it'll just learn that with enough data.
03:39I had the experience, like, I started a company called Live Planet in 1999,
03:43where we did, you know, the internet and movies and television, like Project Greenlight,
03:48and it was an algorithm for your screenplay contest and your director's video that you would upload.
03:53And going into ancient history, I had some experience with that,
03:55and I realized when I got some offers, like, hey, well, why don't you work with us
04:00and show us what you're talking about, and we'll do the AI for you.
04:03I thought, well, then all the value, you'll just take all the value.
04:06And these are companies that aren't in Hollywood, that don't care about rights and likeness
04:12or any of that stuff.
04:13And I thought, like, well, I don't want this technology to develop somewhere else
04:16and not be available to anyone who's a filmmaker who wants to make their movie,
04:20because then I realized it's never going to, you're never just going to type your movie out, Scott.
04:23That's just never going to happen.
04:24This is going to be the last thing that gets disintermediated by AI
04:27for reasons that are a lot of complicated and I'm happy to go into.
04:31But I firmly know that that's not the case.
04:34One thing is enormously expensive, actually, which people don't talk about,
04:37but it is to generate video.
04:40The scale of compute is a lot more than language models on language models.
04:44Now, I've gotten extremely large and very costly to train aimed to run inference on.
04:49So I basically went, used some money, and then I had sort of access to capital.
04:54I had started my own business.
04:55I drew some capital down.
04:56I started a second company, and I went and created a data set, because they were like,
05:02can we buy your data?
05:03And I was like, you know, we build data.
05:05Like, you don't have to just buy it.
05:06You can make it.
05:08And, oh, that's not enough.
05:09And I said, but it will be if it's with, no, it won't.
05:11But I went and did it anyway and made, you know, for months and months and months with
05:15a lot of cameras and some other equipment and some patents, like using LiDAR and location-specific
05:22devices and incorporating Unreal and volume stages, built a very robust data set that was
05:27essentially a way of creating AI images that didn't rely on training on some work that
05:35one of my peers did, that I'd have to tell them, oh, yeah, well, part of the intelligence
05:39native to this is something that was learned, you know, in a tiny fraction, but like, nonetheless,
05:44from something you did.
05:45I was not comfortable with that, and I also realized you don't have to do that.
05:48But there was this mad rush, and people didn't want to pause and stop.
05:51So doing that for one specific use and thinking about what is that use going to be, well, the
05:56applied side, right?
05:57Like, which is, how is it actually going to be used?
06:00Not just, oh, look, I can create a fire, right?
06:03But like, okay, well, what do we want to do?
06:05We want to cook food.
06:05We want to heat our homes.
06:06We need a furnace, you know, that kind of stuff.
06:08Built out this data set, hired some really, really great engineers, inference specialists,
06:14and data architects and visual effects folks, started this company, the entire goal of which
06:19was to say, like, to capture this technology for filmmakers and for artists that they could
06:23control, feel comfortable with, and then socialize that.
06:26And go, look, we're not making up movies and replacing everybody.
06:30That's all bullshit.
06:31There's no, that's just not going to-
06:33I mean, I was going to ask you, you know, I think, you're surely aware, like, that,
06:36I think societally at this point, definitely in the creative community, a lot of ambivalence
06:41about AI as a technology.
06:42Yeah, ambivalence, a lot of, like, terror and rage.
06:45Yeah.
06:45And I understand why.
06:47But on the one hand, the only real voices that you've heard, by and large, are constantly
06:53sort of, whether they're puffing up their chest or advertising their product or whatever they're
06:57doing, they're doing it in a way by saying, and it's so powerful, it's going to be wipe
07:01everything out, which is why we need to control it, or which is why you need to download our
07:04model, right?
07:06And there are dangers in AI, to me, that have to do with, yes, because, you know, you have to
07:14use it
07:15responsibly.
07:16And I wouldn't just give an unpredictable stochastic technology the controls over something that had a
07:22life or death impact on people.
07:24Is it going to turn into Skynet and take over the world and be the paper?
07:28No, that's not the case.
07:29Oh, we're good.
07:30Yes, that's not going to happen.
07:31But what might happen is somebody might say, oh, we want to let AI target all of the targets,
07:38you know, for our war machines or do security, right?
07:43Like, and if you come up, like, it'll lock you in or shoot you or something.
07:47If you're going to do that, you're going to have accidents.
07:48It's very different from deterministic computing.
07:50Like, that's what we're used to.
07:51We're used to computers that are like calculators.
07:53It gives you the right answer every single time.
07:55In order for it to do what it does, it has to work fundamentally different from that,
07:59which is why it's called generative or stochastic, which just means, like, to aim.
08:04It doesn't mean exact, right?
08:06It means good at aiming.
08:08Inference, which is, like, running information through the attention layers and the, you know,
08:15the perceptron layers and getting on and then the outputs and generating an output is called
08:22inference because you're inferring.
08:24Like, so you keep, not only can you not rely on it, only, we all know that.
08:28So you'd be a fool to say, yes, I'm going to hand the keys over to this in this very
08:33important matter.
08:34So those are societal concerns.
08:36Do you think the creative community is right to be as concerned as they are?
08:39Anybody who's told, hey, this is going to replace you in your livelihood is fucking concerned.
08:44Like, the notion that that's somehow a surprise is baffles me.
08:48Like, going out and saying, we're going to wipe out all half of white-collar work does little to stir
08:53enthusiasm to adopt your product.
08:56However, look, I think that you would evaluate a person's ability to predict the future in large part based on
09:03how well their past predictions have turned out.
09:06And in this case, already, you have people with a record of, like, wildly off-base predictions.
09:12And there's a lot of power in being the authority who can tell you whether you're going to live or
09:16die.
09:17And there's a lot of click value in being like, are we doomed, right?
09:21That's the first thing you want to click on because as human beings, evolutionarily, we are tuned to pay attention
09:27the most
09:28and remember the most sharply things that could threaten our lives, which is why you have PTSD.
09:32Because your body is still, you know, having these flashbacks or associating a car backfiring with a weapon going off.
09:39Because it's, we have to know what a lion sounds like.
09:42We pay less attention to the good things or the safe things.
09:45What I would think is that you would want to take, like, a measured, thoughtful, responsible approach.
09:51And one of the best tools we have to ensure that is that, like, if you have a company that
09:54hurts people, you're liable for that.
09:57And it's one of the counterweights to the, I think, the desire to accrue more money that's behind the sort
10:03of frenetic race
10:05to just improve regardless that will be helpful in mitigating some of that.
