Select text to annotate, Click play in YouTube to begin
00:00:00
so there's a lot of stuff Brewing behind the scenes at open AI as you'll see open AI might be thinking multiple steps ahead steps that we can't even see keep this image in mind what opening ey is showing us is just the tip of the
00:00:12
iceberg what's available behind the scenes is far bigger and greater than we know who would have predicted this that open AI would unleash something like Sora far in way the best AI video
00:00:24
generation model we're getting very close to the point where AI video will be indistinguishable from reality but as you see in a second this is just what's showing what's behind the scenes is far
00:00:35
bigger in this video Let's see why Sora is different what do you think you're looking at right now Minecraft well not quite this is generated by Sora why does it look like a 3D World in a video game
00:00:49
why does it seem like the physics are incredibly well generated with something that's supposed to be just a video generation platform the glass doesn't shatter but notice ice cubes notice the water what does that look like to you in
00:01:03
this video let's talk about video generation Sora AI video models as world simulations and what does the unreal game engine have to do with all of this let's dive in so I'll link below to your
00:01:15
previous video where we went over all the various videos that are available to see that are made by Sora so here we're not going to be actually showcasing all of them we're mainly going to be digging into how the heck did they manage to
00:01:27
create this and also specifically what this means what does this mean what does this mean what I think it means is that opening eyes building at light speed things that we can't even see
00:01:40
let's start here this is Dr Jim fan uh one of the senior AI researchers at Invidia covered a lot of his papers here so he's saying this if you think opening ey Sora is a creative toy like Del think
00:01:52
again Sora is a datadriven physics engine it is a similation of many worlds real or fantastical the simulator learns intricate rendering intuitive physics long Horizon reasoning and semantic
00:02:05
grounding all by some denoising and gradient maths I won't be surprised if Sora is trained out lots of synthetic data using Unreal Engine 5 it has to be and he's not the only person mentioning
00:02:16
that multiple people have pointed out that a lot of the footage does have some like video game qualities Dr Jim fan continues here let's break down the following video prompt photo realistic close-up of video of two pirate ships
00:02:29
battling each other as they sail inside a cup of coffee which by the way I mean if you take a look at this video you know exactly what the prompt was I mean you can see the cup you can tell this is
00:02:43
coffee this doesn't look like water that looks like coffee with the froth on top and everything else you know those are two pirate ships and you can tell that they're miniature pirate ships cuz again they fit into a cup of coffee he
00:02:56
continues the simulator instantiates to exqu isit 3D assets pirate ships with different decorations Sora has to solve text to 3D implicitly in its latent space if some of these words make no
00:03:09
sense to you it's going to make a lot more sense once we start unpacking it I'll show you some research that shows why this what you're seeing here why the thing that produces this the digital brain that produces it is is probably
00:03:20
far Stranger Than You might know you continues the 3D objects are consistently animated as they sail and avoid each other's paths that's important I mean this is a
00:03:32
similation of a 3D environment with physics that are carrying these things in this sort of uh melstrom Whirlpool fluid dynamics of the coffee even the foam that forms around the ships fluid
00:03:45
simulation is an entire subfield of computer Graphics which traditionally requires very complex algorithms and equations photo realism almost like rendering with Ray tracing the simulator
00:03:58
takes into account the small size the cup compared to oceans and applies tilt shift photography to give a minuscal Vibe the semantics of the scene does not exist in the real world but the engine
00:04:10
still implements the correct physical rules that we expect next up add more modalities and conditioning then we have a full data driven uee UE I'm reading that as Unreal Engine here that will
00:04:22
replace all the hand engineered graphic pipelines so un real engine if you're not aware of it it's this probably one of the best sort of game engines that people use to create games it's built by Developers for developers and it
00:04:36
basically enables people that develop games to have a lot of like the tools instead of coding up everything from scratch and creating every single part of the from scratch this kind of provides you a workshop where you can
00:04:48
use some existing common tools and then build on top of them to create your own game however you want to so they're saying this is the world's most advanced realtime 3D creation tool so this is Unreal Engine 5 and it started out kind
00:05:01
of looking like a pixelated video game and now a lot of the things that you can do with it are more or less indistinguishable from reality so the images you see here this is a 3D
00:05:14
creation tool that created this onreal Engine 5 so as you can see here I mean looks very real almost in a certain way more real than reality and and it creates these massively detailed worlds where every single thing is a 3D object
00:05:27
with the Shadows Fall Etc or as they put Dynamic Global illumination and reflection so by rendering the 3D space with the lightings and everything else you're able to create incredibly realistic looking images and last point
00:05:40
about the Unreal Engine is that this shot here can be moved around you can pan left right up and down each building is its own 3D asset so you can circle around it zoom in zoom out Etc so let's
00:05:54
get back to Sora this I think is where open a ey is playing on a much deeper level level on a much more advanced level than anyone else they are using as far as we can tell synthetic data to train Sora now of course we don't know
00:06:07
that for a fact here Dr Jim fan continues he's saying apparently some folks don't get data driven physics engine so let me clarify Sora is an entend diffusion Transformer model it inputs text SL image and outputs video
00:06:20
pixels directly Sora learns a physics engine implicitly in the neural parameters by gradient descent through massive amounts of videos Sora is a learnable simulator or World model of
00:06:33
course it does not call ue5 explicitly in the loop but it's possible that ue5 generated that so this is again Unreal Engine That 3D sort of asset World Builder thing we were talking about but it's possible that ue5 generated text
00:06:46
video pairs are added as synthetic data to the training to the training set so what he's saying here is that it's well possible or at least this is the guess that maybe Unreal Engine 5 was used to
00:06:58
generate many many different images that during the training of the Sora model they've used many different videos that were created in Unreal Engine 5 a camera flying through the city zooming in on
00:07:11
objects flying through the streets like a drone Etc and each had some text description of what was happening in there and that combination of text and video those pairs were added as synthetic data to the training set now
00:07:22
the synthetic data part is big and we've talked about it quite a bit on this channel so in the last few years a lot of people have question how far we're going to be able to scale these AI neural Nets these AI models because
00:07:35
they've already consumed such a massive amounts of data that was generated by humans that we figured hey we're approaching the end of that like they already read all of the books all of the internet so it seemed unlikely that we're going to be able to you know to
00:07:49
increase the amount of text that we have available to you know 10x or 100x right because if you gave me you know all the textbooks and all the books every written and everything written online they said okay and that would give me you know 100 times that amount that
00:08:02
you'd be hard pressed to find that text elsewhere and yet that's what's needed to keep improving and increasing these AI models one of the potential Solutions was using synthetic data or data that was created not by humans like a human
00:08:15
written book butd rather buy these AI models or in this case you know Unreal Engine 5 and up until very recently we didn't really know if it was going to work or not a lot of people questioned it they said the minor problems with the
00:08:28
synthetic data will lead to Corruption of the models little mistakes here and there will compound and kind of break the model so the argument was you know if you create a bunch of these images and you feed it into a model right and something's wrong with how the legs are
00:08:41
generated or how the fingers are generated you can kind of tell that she's kind of sliding on the sidewalk a little bit all those little errors will kind of compound and if we just use the AI generated data to train the next
00:08:54
series of models eventually those little errors will compound and just everything will kind of Fall Apart it'll corrupt and it'll actually start getting worse and worse but a lot of the more recent research seems to suggest that no that's
00:09:06
not the case Orca 2 was the Microsoft open source model that was built on synthetic data data generated by GPT 4 here we're seeing potentially the amazing Sora being built on synthetic
00:09:20
data as well and looking incredibly incredibly good and rumors are is that openi kind of understood this kind of early on and they're just running with it like they might have discovered it early and just have been building and
00:09:33
building with that idea in mind and certainly Sora if this is true kind of confirms that now I think this was uh very confusing to a lot of people because I guess according to what Dr Jim fan said here maybe some people in the
00:09:45
comments thought oh so Sora just uses the Unreal Engine to generate this stuff and just spits it out is that what's happening here cuz then is all of a sudden not that impressive right but that's not the case and let me show you
00:09:59
an interesting study that I think might help not only understand how this is happening but it's also kind of wild in and of itself so this is uh Beyond surface statistics so it was in November
00:10:11
2023 out of hardvard and the question that I was trying to ask is how do these Laden diffusion models an image generating model you know the inner workings of these they remain mysterious
00:10:23
when we train these models on images without explicit depth information they typically output coher pictures of 3D scenes how's that possible so really fast a lot of these image models this is kind of how you can think of how they
00:10:36
generate these images so the idea of taking this image that we can all recognize as a dog and kind of adding more and more what they call noise to it right until it turns into static where we can't tell what it is this is how
00:10:49
those diffusion models are trained and then to produce images they do the sort of reverse of that a d noising process so they start an image out as the this sort of static thing that you can't tell
00:11:01
what it is and slowly Den noise it until it becomes an image that you requested for it so for example here you can see this process take shape here step one the first part you can't even tell what that is it's just noise right and slowly
00:11:14
over time as the D noising process happens all of a sudden you see certain shapes take form and finally this is the final product you know as you can obviously tell this is a red car one of those antique cars on a Green Lawn and
00:11:27
so for this experiment they basically train their own kind of neural network their own AI model they produce these images they used a synthetic data set so basically they used stable diffusion the open source model you might have heard
00:11:38
of it so they basically created their own synthetic data set of images and I believe they just made images of like cars and animals and people Etc but the important thing to understand is they just FedEd 2D images here's a picture of
00:11:50
a car here's a picture of a boat here's a picture of a person so they had those kind of pairs of text the description of the image and the 2D image the model had no concept of what 3D meant of what
00:12:03
depth of field meant nothing like that and at the end of this it was able to you know if you said make a picture of a red car it would do something like this it would create a picture of a red car which is great this is what we expected
00:12:14
to do here's where it gets weird they kind of sliced into it at various parts of the generation process to see how it was sort of thinking about how to create this car and what seemed to be happening
00:12:27
is that very early in the process long before you could tell that it was a car the AI model learned to separate like objects in the foreground sort of near the camera like this wheel with objects
00:12:39
that were further in the background like the you know the background the trees in the background Etc so before it even made the image it had kind of an idea in its head about where things would be
00:12:51
placed within that image in that 3D space even though it was never explicitly taught what 3D meant or or how to construct a 3D scene in its head in its neural network whatever but this
00:13:04
was something that's sometimes referred to as an emergent ability in order to figure out how to build realistic looking 2D images it had to create a mental model of the 3D World in its head
00:13:15
so in other words even a very simple neural net AI model like this when we feed it 2D images it learns to have a 3D sort of representation of space in order
00:13:27
to create the those images we don't teach it to do that we don't push it in any way to do that it kind of happens and so when Dr Jim fan is saying Sora learns a physics engine implicitly in
00:13:40
the neural parameters by gradient descent through massive amounts of videos I think that's kind of similar to what this paper is saying so Sora learned a physics engine implicitly
00:13:52
meaning it wasn't taught we didn't tell it how physics worked we just showed it a massive amount of videos and its brain and its neural parameters it was like okay I kind of get how physics works now
00:14:05
by the way if you want to know what the really big deal why everybody's freaking out about Ai and why everybody's so obsessed with it I think this sentence if you understand what this sentence means this is kind of the big deal we've
00:14:18
figured out how to make computers think and learn and create certain mental models similar to I would say how humans do it so Dr Jim fan continues so there's some vocal objections apparently in the
00:14:30
comments so people are saying Sora is not learning physics it's just manipulating pixels into to deep and as somebody that's been posting videos about this for quite a while now I see this in the comments it's a small
00:14:44
minority of people but they really argue hard against this they say you know AI it doesn't reason it doesn't think it doesn't understand it's not learning physics and truth be told we probably
00:14:56
don't fully understand everything yet so I'm not really saying that we know exactly what it's doing but things like this right as this this paper here States you know how do these neural networks transform say the phrase car in
00:15:09
the street into a picture of an automobile on the road do they simply memorize superficial correlations between pixel values and words and this is what a lot of these people they want us to believe the people that are
00:15:21
arguing against it is just pixels in 2D this is just is just manipulating little pixels it has no understanding of how how physics Works how water works well this is just pixels on a screen that's
00:15:34
all it's doing right it has no concept of how fluids move or how coffee you know the the little foam buildup is different than water and at this point I got to say you know probably not it
00:15:46
seems like they are learning something deeper such as an underlying model of objects such as cars roads and how they are typically positioned and by the way there are other researchers that have found this idea that a neural Nets
00:15:59
create certain world models certain understanding of how their world Works in this research you can find by searching for aell GPT a GPT model similar to opening eyes Chad GPT it
00:16:14
started out as a blank slate so no words no images no pictures nothing and it was fed nothing but aell moves this game that you play on a board looks kind of like this so fed you know millions of
00:16:26
these until it was able to predict a legal next move it could make which is what we would kind of expect it to do but when they dug deeper it seemed like this neural net developed a
00:16:38
representation of the board state of where the pieces were of whether they the opponent's colors or my color now again it had no idea about any of these it didn't have any idea that there was a
00:16:50
board or pieces or rules of the game and yet it created a sort of representation in his brain of how that game is played based on nothing but moves of these games being given to it to train on and
00:17:03
in all these cases people are saying no is just predicting the next move nothing else is happening there's no understanding or anything like that it's just a stochastic pit repeating the data it's been fed and same thing here with
00:17:15
Dr Jim fan so people are saying no it's not learning physics you silly AI senior researcher at Nvidia is just manipulating pixels in 2D right cuz obviously random Twitter comments know better than the people that are studying
00:17:28
this their entire life and he's saying I respectfully disagree with this reductionist view it's similar to saying GPT 4 doesn't learn coding it's just sampling strings well what Transformers do is just manipulating a sequence of
00:17:41
integers token IDs what neural networks do is just manipulating floating numbers that's not the right argument Sour's soft physics simulation is an emergent property as you scale up text to video
00:17:52
training massively this is the controversial thing this is the thing that I think people are kind of a little bit scared of that they're arguing against anytime you suggest that these things understand something some portion
00:18:05
of people some percentage get really kind of mad and they call you names and they tell you you're crazy but again at this point so this is Dr Jim fan Andrew a has stated that he believes that
00:18:17
neural networks on some level understand or at least in terms of the fact that they build these mental models that that shows some level understanding Jeffrey Hinton was called the Godfather of AI
00:18:30
also said something very similar along those lines this isn't just a thing that spits out some something that's like statistically likely there's something deeper Happening Here There are these emerging properties and so he continues
00:18:42
GPT 4 must learn some form of syntax semantics and data structures internally in order to generate executable python Code gbt 4 does not store python syntax trees explicitly that's not the point
00:18:54
GPT 4 and other AI models they don't store the information they see there isn't sentences of text that are stored in it very similarly Sora must learned some implicit forms of text to 3D 3D
00:19:05
Transformations Ray traced rendering and physical rules in order to model the video pixels as accurately as possible it has to learn concepts of a game engine to satisfy the objective if we don't consider interactions the Unreal
00:19:18
Engine 5 is a very sophisticated process that generates video pixels sore is also a process that generates video pixels but based on end to-end Transformers they are on the same level of abstraction the difference is that ue5
00:19:32
Unreal Engine 5 is handcrafted and precise but Sora is purely learned through data and it's intuitive and this is the big difference between computers of old computer programs that were coded most of the things that we see computers
00:19:43
do they do because a smart software engineer was able to figure out how to do that these neural Nets they do stuff that we don't teach them they do stuff that we can't even imagine if you think I'm kidding take a look at what Alpha
00:19:56
fold 2 does and so he's asking will s replace game engine devs absolutely not its emergent physics understanding is fragile and far from perfect it still heavily hallucinates things that are incompatible with our physical common
00:20:08
sense it does not yet have a good grasp of object interactions see The Uncanny mistake in the video build below so basically it doesn't break the glass but notice that but while the glass doesn't shatter like everything else is looking
00:20:21
great the physics the ice cubes the water the fluid Sora is the gpt3 moment back in 20 gpt3 was a pretty bad model that required heavy prompt engineering and babysitting but it was the first
00:20:34
compelling demonstration of in context learning as an emergent property don't fixate on the imperfections of gpt3 think about extrapolations to GPT in the near future all right that said let's
00:20:47
quickly go over the technical report that opening I posted for Sora so we explore large scale training of generative models on video data specifically we train text conditional diffusion models so those are the D
00:20:59
noising models that create images jointly on videos and images of variable durations resolutions and aspect ratios we leverage a Transformer architecture so this is the big thing that probably
00:21:11
drives a lot of the AI progress today so that was a made in 2017 by Google a lot of smart researchers that created that and it's really pushed the whole field forward you could say and it operates on
00:21:23
space-time patches of video and image latent codes our largest model Sor is capable generating a minute of High Fidelity video so this is as far as I can tell I mean before this I personally
00:21:34
have not seen any AI model capable of doing a coherent one minute long scene the lady walking in Tokyo I mean this to me was I haven't seen anything like this
00:21:46
the fact that it's for one whole minute stayed coherent nobody like floated up into the sky this building didn't morph into five different buildings over time like you can tell she's on the same street it's it's the same people it's
00:21:58
the same building even the text doesn't morph I'm kind of gushing about it but it's it's a big deal and they're saying our results suggest that scaling video generation models is a promising path towards building general purpose simulators of the physical world so they
00:22:12
start by saying turning visual data into patches patches is kind of a new term here that they explain we take inspiration from large language models which acquire journalist capabilities by training on internet scale data so like Chad GPT GPT 4 just sucks up all the
00:22:26
data on the internet various books Etc and then it acquires generalist abilities it's it's able to do a lot of different things well and the success of this large language model Paradigm so GPT 4 all the breakthroughs that it did
00:22:39
this is enabled in part by the use of tokens that elegantly unify diverse modalities of text code math and various natural languages so you can think of tokens I guess kind of like letters it's it's a unit just like we break words up
00:22:52
into letters they're they're kind of like the a unit of a word they're similar and that they break up all the text into these tokens into units that the llms are able to understand and so here they're connecting the idea of
00:23:05
tokens to models of visual data how can they inherit the benefits that we've had with these GPT 4 Etc and so where llms have text tokens Sor has visual patches
00:23:17
so tokens basically are patches in the visual models patches have previously been shown to Be an Effective representation for models of visual data we find that patches are highly scalable an effective representation for training
00:23:29
generative models on diverse types of videos and images so they take a bunch of images encoded into these patches they're able to produce outputs video compression Network we train a network that reduces the dimensionality of
00:23:41
visual data this network takes raw video as input and outputs a latent representation that is compressed both temporarily and spatially so when you hear this idea of like a latent space you can think of it as sort of a a space where the model does its thinking and
00:23:55
processing that is not directly observable so if you have a library full of books and then you have like a little database that has all the books and the titles and what they're about you can think of that as like the latent space
00:24:06
that represents the library and the library has actually all the books and all the text and the database is kind of just a representation of it and compressed temporarily and spatially temporarily just means time so an hour
00:24:18
long video is kind of compressed to be shorter and so Sor is trained on and subsequently generates videos within this compressed latent space I think it's fair to think of lat space as similar to what humans have with like
00:24:30
abstract ideas or if you ever have some intuition about something that you can't fully put into words that ID exists somewhere but it might take you a while to like fully extract it into something that another human being might
00:24:42
understand and so these patches that they're talking about it makes the output much more flexible so looks like they can control the size of the generated videos it's not limited by an aspect ratio or anything like that and they're pointing out here that Sora is a
00:24:54
diffusion Transformer and Transformers have demonstrated Remar remarkable scaling properties across a variety of domains including of course language modeling computer vision image generation that's why this picture is kind of interesting because as we
00:25:06
increase the compute how much sort of resources we give this thing it gets better and better and better with no additional changes this basically means that a lot of improvement can be had just with kind of better Hardware with
00:25:19
more Hardware I think that's part of the reason why samin's talking about you know 7 trillion funding for his AI chip company cuz sometimes that's pretty much just all you need and they're able to rapidly create quick prototype content
00:25:31
and lower resolutions and generate full resolution all within the same model instead of having one model generate the low resin and another that kind of like turns into high resolution it's all within the same model and they notice
00:25:43
the improved Framing and composition in these models which again I'm curious how much of that you know if the idea that they've used the Unreal Engine in the Unreal Engine like a video game you can slightly shift the angle of the cameras
00:25:56
just a little bit left and right up and down to maybe give the model just a better understanding of how zooming in and out how framing Works it'll be curious to see if that's how they solved it and again so language understanding
00:26:07
so they're saying that they use the same recapturing technique that they use in Dolly 3 because these mods require a large amount of videos of corresponding text captions now notice what they say here we first train a highly descriptive
00:26:20
captioner captioner model and then use it to produce text captions for all the videos in our training set we find that training on highly descriptive video captions improves text Fidelity as well as the overall quality of the videos so
00:26:34
I don't know I'm reading this as you know they use GPT for vision maybe slightly fine-tuned to caption a boatload of these unreal five Unreal Engine videos right again we don't know
00:26:48
if this is the case or not obviously nowhere in there do do they say Unreal Engine but assuming it's true they use something like GPT for RI Vision to caption every single video and then feed that into this model and the whole point
00:27:00
of that is you just you just have unlimited high quality data for train the model you can create images of anything people spaceship cities whatever you can do realistic animated
00:27:12
firstperson shooter whatever the model can probably caption you know millions of hours of this footage like there's just no bottleneck to producing a staggering amount of very high quality
00:27:25
video of data data that is paired with text which is the pairs that you need to train these models and Sora can also be prompted with other inputs such as pre-existing images or videos create perfectly looping videos animated static
00:27:38
images Etc extending videos forwards or backwards in time Etc so you can take a still image and turn it into this or that or pretty much anything you want this one's pretty cool I got to say wow
00:27:50
yeah to me notice like the lighting and the how it handles the 3D space up above in the in the C Cathedral like I don't know I'd be hardpressed to say I mean that looks like a Unreal Engine sort of
00:28:04
how it would render the 3D space of a building it does not look like any of the other AI video generation things that I've seen and the lighting too seems like actual lighting it doesn't
00:28:16
seem like just texture it looks like it's a light source now take a look at this video I had half a mind to just have it Loop infinitely and uh not say anything but I won't do that so apparently you can also use Sora to
00:28:28
produce an infinite Loop a seamless infinite Loop by just looping the entire few seconds together you can extend the video forwards and backwards and produce an infinite Loop you can take two different sort of sets of images and
00:28:40
kind of combine them so you see a butterfly underwater or at one point it's going to be a drone or it's going to be a butterfly flying like a drone through the Coliseum there it is now here you're combining was that a
00:28:53
chameleon a gecko with a bird and you got this uh birdlike chameleon chameleon bird a very unique look I got to say I mean it really captures kind of the essence of both you can generate images
00:29:05
and here's where it gets interesting emerging simulation capabilities we find that video models exhibit a number of interesting emerging capabilities when trained at scale these capabilities enable Sora to simulate some aspects of
00:29:17
people animals and environments from the physical world these properties emerge without any explicit inductive biases for 3D objects Etc they are purely phenomena of scale and as I mentioned in the previous video and this is kind of
00:29:30
what we're talking about here with the emerging properties this is really where I would love to see much much more research like what is happening in those neural Nets that is making this real and
00:29:42
how far can we take this and a continue 3D consistency Sora can generate videos of dynamic camera motion as the camera shifts and rotates people and seene elements move consistently through threedimensional space long range
00:29:54
coherence and object permanence this is huge this is one one of the most striking things about how well they're able to keep the coherence in these images over long periods of time or as the camera shifts around this is unlike
00:30:07
anything I've seen and interacting the world so it can simulate actions that affect the state of the world in simple ways so like watching somebody draw that's pretty incredible and it's going to be able to simulate artificial
00:30:20
processes like in video games it's able to simulate digital worlds so it's that idea of simulating Minecraft as we've mentioned earlier I mean it looks it looks so close I mean everything 3D is like simulated perfectly As you move
00:30:32
around it's very very similar and these capabilities suggest that continued scaling of these video models is a promising path towards the development of Highly capable assimilators of the physical and digital world and the
00:30:44
objects animals and people that live within them and I'll leave you with this final Poe by Dr Jim fat none of this is meant to be a religious or spiritual or anything like that is just sort of a thought experiment but but think about
00:30:58
this if there's a higher being who writes the simulation code for our reality we can estimate the file size of the compiled binary so basically when you write code when you write out it's can be understood by humans it's long it
00:31:10
kind of spells everything out then when you compile it it gets turned into a smaller file that the computer can just execute and run directly it's smaller faster it just everything that it needs to run to do what it's supposed to do so
00:31:23
he's saying if you know meta ai's videos this many parameters and sore as this then the Creator's binary the sort of the person that built the simulation which is our lives might be no larger
00:31:35
than 111 gbt which is interesting to think that our world could be reduced to some tech spec like that some technological specification and he's saying Sora is not just compressing our world but all possible worlds our
00:31:48
reality is the only one of these simulations that Sor is able to compute it's possible that some parts of the physical world doesn't exist until you look at it much like you don't need to render every atom in the unreal five and
00:32:02
the Unreal Engine five to make it a realistic scene by the way this idea that it's possible that some parts of our universe our physical world that they don't exist until you look at it does that sound like nonsense to you cuz
00:32:14
you know it is true this is literally true you know how some people act differently if they're aware that somebody's watching them well so does light why I have no idea but I've said
00:32:27
this before I feel like AI is going to unravel some of the strangest mysteries in our universe my name is Wes rth and thank you for watching
End of transcript