Notes provided by Gemini
HSV AI – Deloitte Room: Hi. My name is Benjamin Robinson and I’m going to read out this monstrosity you for one more time because it’ll be my fifth time doing it and it sounds like fun. A mutually exclusive 3D convolutionbased time dilation framework of future frame prediction and video streams of natural events. Fifth time was a fun. Uh, this paper was done in research while I was at UA and I really have to thank doctors Mukerg and Bannery for all the help they gave me doing this.
HSV AI – Deloitte Room: So this is pretty simple when you get down to it. So going over it is going to be a lot of history because we need to really answer why are we even bothering to do ins. So all of you are going to get a quick history of next frame prediction from what I personally consider modern next frame predictors. So when we’re getting down to that, what is next frame prediction? AP terminology. When talking about a frame, we’re trying to take a video and splice it up into individual frames where a frame is any one still of the video. And these individual frames depend on the frame rate of a video. So if you’ve ever heard of 30 or 60 frames per second, that means in any one second of video, there are 30 or 60 of these still images. So we’re also going to be talking about general model things some of you may need some updates on.
HSV AI – Deloitte Room: We’re talk about parameters which when you’re talking about models is the internal weights and how many parameters you have is an important discussion for this paper because next frame predictors hugely weighty as you’re going to come to find out. uh when also talk about next frame predictors and why we wanted to do any of this at all. We’re talking about stability because when it comes down to learn the history of next train prediction, you’ll learn that part of the biggest issue is that the train process is disgustingly unstable and will just collapse at any time place for any reason. So what is frame? We talked about this a little bit. What is frame prediction? We’re going to take a sequence of these individual frames. We’re going to learn the features we can from them and we’re going to predict the next one. We’re going to predict the future.
HSV AI – Deloitte Room: Woohoo. So why would we want to do this? The typical reason is for brain reconstruction. So if you’ve lost video in a video stream, how can you recreate that video so that you could see what happened in between just lost time? Another application for next frame prediction is say you want a targeted system to have better predictability and where it’s going to go. So what do you do? You predict out what is going to be in the future so that your model can then say oh well we want to go in this direction for say flying a drone or whatnot. So why is this a hard problem? Well, the difficulty of this problem is that spatial features, which is what the image is made up of, and temporal features, which is the change in any one image to the next image, fundamentally disagree. Uh, structural features want everything to stay the same.
HSV AI – Deloitte Room: Temporal features want every change. And the combination of these two are necessary for next frame prediction but cause models to collapse because one usually outweighs the other and then the model can’t make accurate predictions after that. And finally, the issue with next frame prediction is that machines and humans have different opinions of what makes. A machine can make an image that to it is 98% perfect, but to us all the colors are wrong. If all the colors are wrong by just a hue of a few degrees, we notice it, but the computer, it doesn’t. So, we have to have something that looks good to us, not good to a machine. And that’s not very easy to make a machine do. So, I added a couple of slides for people who may be less informed of the history of next predictions. It really got started off with RNN, which are recurrent neural networks.
HSV AI – Deloitte Room: Our hands pos the question that what if all in all all data isn’t independent? What if we could have sequential data data that matches what happened before to what happens next? So they add in a hidden layer into the model so that we could say well this little bit of information it’s important and further into the model we’re going to need it. So we’re going to store it and update it. The problem with that is that it didn’t do a very good job of keeping up with what was important. So they came up with LSDM which introduced this idea of gates. Well, we can’t tell what’s important. We’re going to make these gateways. Input gates, output gates, and forget gates. Input gates determine, okay, on this bit of incoming data, what is valuable out of it? on the output gate, what value that we’ve stored do we need to give out and forgets means well we have all this data and if we hold all of it it’s going to collapse.
HSV AI – Deloitte Room: So what isn’t important anymore that we can just throw away and get rid of? Both of them however have a problem in that they are slow, they require a lot of data and they are still unstable even after all of that. So this talk is based off of my thesis. So I had a lot of previous works in it. I thought it’d be interesting to hear about some notable entries in that spring prediction. Uh some of these are not as interesting. This is one of the first called LSTDN models where they took an LSTM with those gates and added a convolutional neural network to it to try to increase its stability. And then they did it on YouTube character action videos which were like couple second videos of something happening a ball going across the screen. They would just try to predict what would happen and they were able to achieve an SSIM of 0.95.
HSV AI – Deloitte Room: An SSIM is a similar structure index measure meaning this number equates to a percent of how much we think these two images are. So that’s what that means. So what is a CNN? Well, this is really important because that’s what my model is based off on. So we’re going to go into this one a little bit more. A convolutional neural network makes use of a kernel. Need to pick up a kernel as a filter of certain weights and it slides this filter across data to transform in some way. The most common use of this is as a dot product. So instead of having to learn all of the features that you need necessary to extract features, you just train with built them and that filter would take care of that makes this extremely costefficient which matters for people your parameter count. That’s why I ended up choosing a CNN to work with because one of the biggest problems I’m going to work against is bloated parameter counts that require millions if not billions of parameters which take up gigabytes terabytes probably of data.
HSV AI – Deloitte Room: So CNN’s also work well because they have a receptive field which kind of mimics how the eye looks. So you’re not looking at all of the data. You’re looking at a small spot of the data then scrolling across the same way you would read a book. Okay. Also, this is a rambling mess. If any of you would at any point like to ask a question more information to know something else has explained better, please just ask. You’re not going to interrupt. You’re not. So, I do have So, you’re saying it sounds like you’re using really really tiny models. Uh, so what sort of parameter accounts are you generally targeting? I am gener I’m parting sub 100,000. Okay. Okay. So super we’ll get more to that a little later.
HSV AI – Deloitte Room: Go ahead. Okay. So one of the first interesting models to work with is deformable convolutional. So what does it mean by deformable? So if you go back one, this is a 3×3 filter and it will only ever be a 3×3 filter. It doesn’t care what’s neat managed. It’s going to be trained as a 3×3 filter work across this image. Well, a deformable neural network says what if I don’t want a 3×3 filter? What if for this one image a 2x two filter? Or if we just want to nudge off this one zero on the top left and we have a 233 filter instead. Deformable means that the size of the filter itself can change depending on the needs of the network. So they chopped up again YouTube character action videos. They used this new network where you could change the filter size as the network trains itself.
HSV AI – Deloitte Room: And they used about six million parameters. I don’t believe they in their paper gave an accuracy measure which in case you don’t know if they don’t give an accuracy measure in their paper that means it was hilariously bad. Oh free. So Ralph and Ballard had a paper a long time ago about how they thought the brain works. And how they thought the brain works is that neurons in our neural pathway are interconnected and always sending messages down and up. So one neuron will take in a bit of data, say what it thinks about it and pass that along to the next neuron. Well, that neuron then at its own point says what it thinks about this data, sends that down, but also then sends back up to the previous neuron, hey, I don’t think what you thought about this data was correct at all. So, reassess what you have. And they had this long paper about that and this precet is based upon that research.
HSV AI – Deloitte Room: So, they took car driving videos. Throughout this talk, you’re going to hear nothing but driving cars and taking videos. Uh they used two million frames. So that’s not parameters, that’s 2 million individual frames of video to train on. And they were able to get an accuracy of 93, 82, and 69. You’ll see t = 1, t= 2. That’s a time step. So the next frame is t to one. The second to next frame is t to two. The third next frame is t to three and you traditionally see massive fall off when you predict out future frames. So the very next frame has a good score. The third next frame has a pretty terrible touch. Yes. Uh millions of parameters again. So, if you go on to the next slide, they the way they let me uh yeah, put the focus back in. Actually, could probably just swipe it.
HSV AI – Deloitte Room: That would be interesting. I’m not going to try that. U So, how does it work for a model instead of a brain? Well, at every single layer in the model, they do a prediction and they send loss data back at every single layer. So, this is computationally extraordinarily expensive, but they thought that mimicking this pattern that was described would in some way give a better model. Overall, it didn’t really work. I think they were more so just trying out something fun and it is kind of fun to think about. So loss can be not only something you do at the end of a bottle but at any point in the bottle several times throughout the model. I found this to be a fun paper even though the results were accurate. So go on transformation predictor. If you know anything about current model work, which I imagine many of you do, you probably hear a transformer and some like sirens are going off in your head.
HSV AI – Deloitte Room: Not that type of transformer. So we’re talking about the difference between 2D and 3D animation transformers, not large language model transformers. So there’s a couple of ways to predict a frame. Typically what you do is you have weights on every single pixel. You say what of this image is of any value? How is that going to change? And we’re doing that for every single image. This work said let’s not do that. Let’s segment out images and treat them as whole units. So instead of however big the image is, say 200 by 300 or whatnot, these are cars, so they were pretty big. What they would do instead is you’d have just a handful, I believe it was like 12 or so, no 14. 14 parameters per object in the scene, which is Yeah. So real quick, you’re talking about video of cars.
HSV AI – Deloitte Room: Yeah. My first assumption was like dash cam. A dash cam is exactly what you’re talking about. someone either on a dash cam sitting still in their cars videoing other cars or driving through a street from their dash cam video. Did you try to isolate the two or just mix them? Oh, well this is their work. I So in the history section I didn’t do cars. So, you’ll see what I did because I if you’re gonna find out everyone cheats and I’ll tell you how I cheated a little later. Uh they they instead had these 14 things that said here’s where the car is, what the car is, and it’s just 14 parameters, which is much less than say two, three, 400 or whatever many parameters that made up the actual car. And they would say, “Okay, so the car is here as this one object.” And then they would transform
HSV AI – Deloitte Room: as in like they would literally say move 200 units forward. And that’s the same way if you know anything about 3D animation, they don’t rebuild the same they don’t re rebuild the same model every time they want to move it. They just have it pre-built and it’s filming it doing a motion. That’s the same way they’re doing it. They’re not describing the changes of the pixels. They’re describing the changes of the images of these objects and then just using transformative features like you would see off of photo editing software to move them within that picture. So are the so then are the weights on the on the objects in the images the weights of the transformation uh the weights describe certain attributes of the objects and it’s then used to reconstruct the object. So you don’t trail the change of the image. You trail you trail the change of the image as numbers not as how the pixels are actually changing themselves.
HSV AI – Deloitte Room: So, as you can imagine, this is a pretty great idea. It’s like, wow, we can still much easily more ch uh follow how something moves with them only using 14 parameters instead of several hundred. So, what’s wrong with it? Moving on. Uh what’s wrong with it is that they cheating. So, anything that’s even slightly occluded breaks the system. So, let’s just take everything that’s oluded and throw it away. And we’re not going to include that in our reports. And you know what else breaks the system? If you’re only pass if you pass by a car quickly and they’re only on screen for like a second or two, that also completely breaks the system. So, anything that’s less than I think 4 seconds, let’s get rid of that as well. So, let’s only train on things that are on screen on screen for a long time and very good to track.
HSV AI – Deloitte Room: Well, if you’re gonna you’re gonna do all that, well, anyone could do that. So, uh they also Yes. So, anything less than 10 frames is ditch that per object sequences. Did they advocate anything? Don’t think so. No, one last thing is that the final Oh yes, the final thing. Um, this relies on a very similar problem that I have, which is that this is a model that does not work for character action. And what character action is is say, I’m going to go this way as a human. I can stop at any point I want. At any point, any time what you’re doing can change. Models like this don’t like that. If something changes, it breaks. It wants things to consistently go the same way. I am curious what kind of sudden change could happen longer than I imagine slamming on the brakes is most common.
HSV AI – Deloitte Room: So move on to the next you can. Yes. So let’s talk about dance for a second. We just talked about character action. How do you deal with the fact that for recordable people at any moment we could stop start walking, stop walking, jump, fall, laugh, whatever we want to do. Well, you have to use something that has to kind of combat itself. So, GAN stands for general adversarial network. And what that means is that these networks make two predictions, a true prediction and a fake prediction. And as part of the model, they have a discriminator sense. I think this is what is real and what this is, what’s fake. And then it trains itself both on its ability to predict and its ability to discern what is and is real. The reason they do that is to as a regular officer they want to combat the extreme chaos of well if it can do anything then how do you possibly predict the future?
HSV AI – Deloitte Room: Uh the problem is anytime you introduce some type of judicial system into a model in my experience it breaks because the judicator either becomes too powerful and then it’s now learning how to decide what’s real and what isn’t real which isn’t the goal. The goal is to make good pictures. So it’s not making good pictures anymore. or it can loses all power just like what’s real and what isn’t real in which case then you’ve lost the whole point of it and again you’re just left to the chaos of well what’s the person going to do we don’t so for this one in particular uh the unique attribute for them is that they introduced a format where you train initially on a low res image at Google the models layers you upres the image because for this model and I believe the next one they came to this idea that building your initial parameter count on a smaller image and then updating it later is cheaper than first building your parameter count on a larger image don’t need the rest of that So they worked on movie emnest, cage action.
HSV AI – Deloitte Room: These are all and cityscapes. Movie emnest and cage action are action data sets like the Google action data sets. A couple second videos of someone throwing a ball or walking across the street. But Cityscapes is uh car video again where someone drives through the street and shows their car moving and other cars moving. So not great accuracies on either none given for citycapes what we say earlier as soon as you don’t I did it again I apologize I guess we are going to try to swipe so if they don’t give you accuracy it’s hilariously bad to there we go now we’re on to grip We’re already on to single events. So yes, single image. There is a there are types of models out there with the stated goal that we want to predict the future off. And I’m here to be the prophet to you all to tell you that this does not work.
HSV AI – Deloitte Room: It’s a terrible idea. You shouldn’t do it. when it does work it’ll be bad for a number of reasons I can talk about later afteration. Um what they did similar to the last one is a pyramid system which they call pyramid system the last one developed where you uh train on a low res version of the image and then build up to a higher res updating the parameters instead of building them. This also works by training both the spatial and temporal features at the same time which most of the time you do one or the other or you have parts of the model that train on one and parts of the model train on another. This one does it all at the same time and I’ll talk about the rest of those whole scheme. I’m gonna try this. Hey, I’m fired. You’re the weakest link. Uh, wow.
HSV AI – Deloitte Room: Seven frames away, they got 9 to3 accuracy with just a single frame. How amazing is that? How amazing indeed. Who thinks they cheated? Everyone should put their hands up because they absolutely cheated. So, between that first frame and the seventh frame, they did only use one frame. one full frame. They also took all of the changes in the data that happened between that first frame and the seventh frame and added that to the model as well. So it’s one full frame and six partial frames. But in recontextualizing this problem set, you take a frame, you take all the changes you have to that frame to then build a new frame, and you only get 93% accurate. That’s terrible. In addition, woohoo, this and many like it have a problem where large transitions create holes because like the transformer model before its, if you’re not actually keeping track of everything that happens and then you move large jumps all at once.
HSV AI – Deloitte Room: Well, the car moves, what’s behind the car doesn’t know. So, it just leaves a big hole in there that has to be filled with a separate model that they used that they didn’t talk about that I had to read the paper three times to find out about. So, they cheated in they cheated in a number of ways. So, why did we go through all this history? Why am I talking about some other people’s work and not mine? It’s definitely not because my work would only take about five minutes. Uh all of these models are huge. The very smallest of them are in the millions of parameters. The largest of them are in the billions if not trillions of parameters. Uh that’s not great. If you want to use it for any sort of real use, which is one of my goals we’re going to get to in a second.
HSV AI – Deloitte Room: Hello. uh that’s unusable. If you’re going to put it on an actual camera, an actual device, edge device out in the world, you can’t have billions of parameters. That just doesn’t work. They’re also unstable. You don’t get 50 different tries to make it work sitting there adjusting it and then putting it out there. That’s what happens with games, with LSTMs and all. You have to just sit there and fiddle with it till it magically works one time and then you just ignore all the failures that cost you a lot of money and time. And many of these, I didn’t even get into this for a lot of them, they use cues and labels. All of the Google action videos, all of them have cues on them. They tell you what’s going to happen in the video, and you can use that data to make better predictions.
HSV AI – Deloitte Room: What’s the point? If you already know what’s going to happen, you don’t need to predict what’s going to happen. like expecting somebody to use a blinker. Yeah. Never or when someone is using a blinker, actually expecting them to stop. Nope. They’re going to go straight on through and you were fool to pull out. So, what do I want to work with? Well, the only thing to work with if you want a reasonable parameter count is CNN’s. Nothing else is going to help. You can’t get there with transformer models. You can’t get in there with RNN models. you have to do pure CNN’s. Well, what’s the problem with CNN’s? They love to overfit. And let me tell you about how much they love to overfit.
HSV AI – Deloitte Room: They just refuse to learn any amounts of temporal information because they’re so good at learning spatial. So, I had some goals coming into here. I wanted something simple, lightweight, stable, modular, and active. Meaning I want a tool that works any which way you want to use it for the problem set it’s meant to use. It works every time and you can use it on any device. So, I wanted something that really worked well, which is not something you get for any of these other mods. So, here is my baby. You could coo all it if you wish. The things to take away from it is that there are phases, there are windows, and there’s this repeating pattern of convolutions, pooling, unpooling, and then the creation of a new window to move on in the model. Everything in this model is just convolutional neural network layers pooling and uncooling relu and one arrange function.
HSV AI – Deloitte Room: That’s it. So what are the phases? Well, let’s talk about how I cheat. I talked to you about how I want to cheat. It’s only fair I talk to you about how I cheat. We need a problem set and I believe in specific models because everything right now is general models. You want to go to chatgpt, it can do anything for you somewhat. It’s mediocre at everything. I want to build something that’s awesome at one thing and that’s my mindset. Well, I brought this mindset straight into my model. And what that means is my model runs on a basis of expertise and the phases are built out through this expertise. So each phase defines how the select group of windows learn. So in phase one you’re learning different colors. So if you have got three colors, you’ve got three windows and you define out I want these group of parameters to learn the reds, these group parameters learn greens, these group parameters blues.
HSV AI – Deloitte Room: On the second phase, we’re going to learn dilations. And we’ll get into what dilations are in just a moment. And at phase three, it’s just what prediction you want to do. Because I wanted to make this modular, you can have any number of predictions you want to do with this model. uh how many you actually get accurately through mileage area. The separation of these is through windows. So a window is just a collection of frames. So you take the previous three frames to predict the next frame or you take the previous nine frames to predict the next three. Those previous frames are collected into windows. So the colors have their own windows. Each dilation have their own windows. And that’s how I use to control what parameters are learning what. Yes.
HSV AI – Deloitte Room: Never mind. This slide’s probably going to answer it. I came up with a funny word that some of you might not know. What is that all about? So we I’m building off the idea that CNN’s overfit because frame one and frame two are nearly identical. The difference between them is almost nothing. So how do you combat that without adding in a huge regularizer or a whole bunch of other things to it? But what you can do with only the data you already have to hand is just introduce gaps in. So say you want to predict the next frame. You can take the previous three. Well, it’s a sequential model. So it’s third frame, the second frame, the first frame, previous next frame. Well, you just gap the previous frame. The n minus two frame, the n minus4 frame jumps with a gap of one to the next frame.
HSV AI – Deloitte Room: What you’ve done is now the difference between any two frames is more significant. And because you’re only using the frames themselves to do this, you are learning both a more general idea of the model the more you push out your dilations while also combating its ability to want to overfit. All just using the frames you already had to hand. Yeah. So you’ve got this idea same as a frame rate. So every third frame, every fourth frame, every fifth frames and you can add again with the modularity you could add any number of different dilations you want to add. It’s all just another you So you would just expand this out for how many dilations you want to have. Yay, it didn’t break. I was a little worried going back again. So as I just described here T8, T minus4, T minus 2, all to predict T1.
HSV AI – Deloitte Room: So you have this run of big gaps all coming together. Well, if you have all of these gaps at different time dilations predicting all the same frame, you’re learning different features from the same images of the same image you’re going to predict to which allows you to take kind of different scopes of the video that you’re looking at. So yes, did you naturally settle on one and two great skips or did you try I did try I tried a lot of different things and um from what I tested out u the most benefit at least at the low parameter counts I’m looking at is you’re going to work with two to three dilations and you can do a three gap four or four gap, but a one and a two and then a larger one is the best combination if you’re going to get through them. The lesson is constantly fine tuning the mult like captioning an image and then you mask a like object like the keyboard isn’t the caption.
HSV AI – Deloitte Room: You mask the word and the object in the image. It sounds similar to that in a way that you’re just kind of offiscating away something and then potentially feeling it at different time like a masking. Um what you’re talking about kind of sounds more like a diffusion model what they did with those. If you know anything about those we could talk those afterwards. didn’t add anything into those transformers or diffusers because uh transformer models the type you use for chat bots they’re not any different for next time prediction than they are for sentence structure or anything else they predict everything the exact same way so what part of this image relates to the rest of the image and how valuable that is so I I didn’t want to derail but I do I do want to talk about the fusion transformers and why you don’t think that this those could be you know when they’re distilled down small enough to like sub one billion parameters.
HSV AI – Deloitte Room: They couldn’t do this sort of thing. I don’t want to dr your thing. I would be happy to talk to you about that at the end. I didn’t add that because that was kind of going on the same time as this research was going on. Yeah. Yeah. That’s all 247. Yeah. No, you’re good. All right. Y’all got me. It’s time to admit my guilt. How did I cheat? Well, these are videos of natural phenomena. So, you can’t do jumps and dilations with character action. As we talked about before, you can’t have random things happening if you’re assuming, because that’s what a dilation does, that for any one dilation that the same things are going to happen at any other dilation.
HSV AI – Deloitte Room: Well, you might jump over a gap where someone stops walking or jumps or something and then it all breaks down. So, we toyed with refining it and making it a physicsbased predictor, but that’s a lot more. Maybe if I go for a PhD, we’ll get into that. But right now, it’s kind of a pseudo physics predictor because what it’s predicting is how flows work. how if something is going to go in the same way because a wave crashing is going to crash in the same way no matter how fast or how slow you go through it then you can use this type of model to much much more cheaply predict out how it’s going to do in the future. So that’s how I cheated. I defined a problem set and made something that works specifically for a single type of problem. So, uh, took, uh, people’s videos from their vacations from a copyright free website.
HSV AI – Deloitte Room: Some people have really expensive drones flying out there. They can get 4K images of waves crashing on the beach. Thank you to those people. Uh, turn them into frames using FFmpeg. Love that software. And this is my my happiest thing for this model. Simplicity. I didn’t touch the frames. No pre-processing has been done to them at all there. I might have cheated on the problem set, but nothing has been done to them to cheat otherwise. And that really matters because I only got 600 of them. Remember what we had before? several million frames had 600 of them and I had them at because I was working off my own machine I couldn’t train working images so this is off of 240p and 360p so fairly standard definition use AdamW a lot of standard stuff for this because a lot of it is in the model itself the framework model itself um my other very happy 79,000 parameters on a final model I worked off of.
HSV AI – Deloitte Room: You can put that literally on anything. I could put it on a camera at the very edge of an Alaska rainforest to predict the health missing footage and it would still work. So, yay. Stability. That’s a whole This is just to show you choose a CNN. They won’t betray you. They work fine every time. Dan will betray you every day. There’s a lot of hurt behind that statement. There is a lot of hurt behind that statement. If you’ve ever had something trained for several days, they just decide bye. Goodbye. LSDM and games will betray you. Yay. So, this is supposed to be a GIF, but it’s PDF, so it’s not done. Oh, yes. So this is gonna play showing how great it is but you know just assume prediction.
HSV AI – Deloitte Room: I can’t tell. So one thing from here is that what we saw from the previous model GAN feature model is that you had a drop of like 15 20% per next frame of accuracy. We saw almost across the board a drop of 3 to 5% from the first to the second and then like two to three from the second to the third with smaller and smaller diminishes as you go on up until a point and then it just lapses. So the SSIM is pretty decent on a lot of these. It would be better if they were higher definition videos, but I had what I had to work with. But from the eye, I don’t know about I’m nearly with blondes, so I’ll have to have you all tell me. Those look about the same. Yeah. What matters to me? Hope that’s what matters to you.
HSV AI – Deloitte Room: 95% accurac. So waves crashing, clouds floating, that sort of deal. So this is kind of like from my thesis. So I’ll order from the thesis just to say, hey, I did the thing and it worked well. Yipp. So uh more so it’s just now questions if y’all want to talk about this talk about next predictors talk about diffusion models and how they also cheat. Sure. Yeah. Well I mean yeah can we talk about the I think there’s two two areas. So when did you do this research? Uh I did this research over 2024 into 2025. Okay. into the just like whenever rectified mode all that stuff was coming out with the deal because I mean that seems like that is obviously you’re not going to get down to 70,000 parameters that’s nuts. Um so I I definitely buy that this might be the only way to get that but you know if we can get down to 600 million parameters there are a lot of things that could reasonably run that sort of thing.
HSV AI – Deloitte Room: uh and with diffusion transformers and you could do flow matching and hardcode it to you know or hard line it into this sort of uh regime pretty easily I think and so there’s like the diffusion transformer thing but also what about things like JA you know the world model predictive where they’re doing latent sort of thing and they can get those pretty small too so I mean aren’t those two reasonable directions to do this sort of thing I’m not as familiar with the second one uh joint embedding prediction model is the young lun where they’re basically learning not transformers but they’re learning from video just pure video prediction to kind of get a a world model essentially. So for more on diffusions more about that. Um from my understanding of diffusion is that you could do a lot of great with it but a lot of diffusion is a way of doing because a lot of diffusion models are backed by something else like most of the paper models on diffusion are backed on by like unit they have a unit architecture coming into it and they attach then a head right.
HSV AI – Deloitte Room: Yeah, they attach it’s for attention essentially. Yeah. But um so I think you could definitely do something with diffusion, but you’d have to effectively define what you’re attaching it to as well. Like you already had to have something good and then diffusion would come in. Sure. But there are a lot of good things out there. Yeah. So I mean that that I mean so if you’re you’re going like from scratch. So I I you know if you’re trying to go straight from nothing into a model I could see that being a concern. But if you’re willing to take something that is you know a reasonable unit and then distill it down with you know flow matchings and get down to one step you know I think like hey hello someone else say something on this. Christopher raised his hand. Ah, I’ve been I’ve been watching chat.
HSV AI – Deloitte Room: Um, uh, well, I’ll say like any of those other interesting models that we dealt with. Yeah, you could add diffusion to raise that accuracy. Something else like that as well. It’s I think the fusion is good, but I think it’s I think it’s waiting for someone to find what best and really down because there’s your PhD thesis will help you. Yeah, he’ll be happy. I love it. He’ll be happy no matter what I do. As long as I’m there as a PhD. Yeah. Christopher, do you want to talk or type?
Christopher Bales: Uh, hey, can you hear me real quick? Hey, Benjamin. Uh, awesome work.
HSV AI – Deloitte Room: Yes.
Christopher Bales: By the way, I had a question about understanding the time dilations. You said that you were kind of skipping frames based on a linear. Okay,
HSV AI – Deloitte Room: Yes.
Christopher Bales: we’re going to take the first, the third, the fifth, or potentially like the fourth or the eighth, whatever. I was confused about how you use those dilations in in forwarding the model. Can you explain that just a little bit more concisely? Um, you had you had one picture where you were going I minus one and then Oh yeah, right here you were saying h can you just explain this part just a little bit in
HSV AI – Deloitte Room: right here.
Christopher Bales: depth?
HSV AI – Deloitte Room: Sure, I’d be happy to. So, the idea is that each dilation gives a different idea of what’s happening even though it’s pretty much the same pictures. So because CNN’s love to overfitit, this gap helps to remediate that. But say if I get what you’re talking about from this point, are you talking about like how it further goes off into the other T plus2 T plus 3 or are you talking about how it flows in the
Christopher Bales: Yeah. Well, in in the real real time space dimension,
HSV AI – Deloitte Room: model?
Christopher Bales: it flows linearly like we have one second to the next.
HSV AI – Deloitte Room: Yes.
Christopher Bales: I guess what I’m asking is how do you make a a a a future from this model? How does the how does skipping time from frames help it ingest that information or or what was the theory behind that?
HSV AI – Deloitte Room: Um, it doesn’t help it ingest the information. It helps to regularize it because CNN’s really don’t need any help reading an image. Their problem is that they fail to stop reading an image. Effectively, they can’t let go of what they’ve seen. So, you’re adding, you could think of each image as what. So t minus2 to t minus5 might not be much of anything at all but what it is is a lot of noise compared to what it expects from say like t t minus one t minus 2. So it gives this kind of effect where it’s a lot of the same data but it’s kind of I don’t want to say like it’s a gossian noise effect but effectively because it’s ever so more slightly different it kind of forces the CNN to let go.
HSV AI – Deloitte Room: So the time deations really don’t help for um pure accuracy in the way of like say just the ones in the ones a CNN is perfect. You want to get something that’s going to predict the very next frame. The problem is that it can’t let go. So, and another way that it kind of helps is that say for t plus two, well, instead of just coming in from these three, you could say t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t minus 2 T minus 4 then jump to t plus two because you’re keeping a consistent pace. So what the other thing that time dilations do for say the feature frames like three you need to have like some amount of matching runup to the further predictions out that you want to do.
HSV AI – Deloitte Room: So t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t tus 5, tus 8, you’ve got a gap of three that flows to then predict the third. So for the first image where it’s just doing its own one, all of the additional dilations just help to regularize. But say for t plus two and t plus three these time dilations help because um what’s happening at t minus one at the pace of t minus one t minus 2 doesn’t really help tell you as much about t plus2 as it does to t+1 because the cnn’s want to say this is identical this is neo identical this is neo identical this also has to be neo identical the gap kind of keeps a flow So for the first prediction of any time dilation, it helps maintain a consistent flow to the video.
HSV AI – Deloitte Room: As you might think of it as um a lower frame rate of video where you’re seeing every third frame, but for every other dilation that’s added to those individual predictions, it just helps to regularize. I’m sorry that was a mad ramble. Hopefully that helps clear something up. So then the way I think about it is and and I know St. J’s that you’re doing wave predictions. So you’re doing ocean wave prediction or cloudbased prediction, right? So these are physical things in the world that are very predictable on their own. Yes. So if the wave is occurring on the evens for example, then choosing the even makes it really easy when you’re predicting hey the w it seems like um one one way to keep CNN from over fitting is basically a blank dropout. This seems like a forced drop out of certain things. Did you a system a systematic pattern force dropout right?
HSV AI – Deloitte Room: Try a randomized kind where one run through. Maybe it’s just a random selection of these eight pieces go to this layer and then roll the dice again for the next layer. Um it would keep you probably from having like the pink block from only being these three and this one being I don’t know if that would make it more generalizable. That would be interesting but be difficult because as I said before after a certain point a certain number of dilations adding on to one prediction stop adding much of a benefit at least at these very low parameter counts. Yeah. So I think you would have to control it in such a way that it’s semi- random but distributed across the frames to add enough. Right. Okay. And another issue is that say the easiest one to do is um ones and so you’re predicting out the next one for the next one but to say to predict out the third in this model uh it has to be say t t minus 3 t minus you know it has to be a run it can’t be just like these three don’t predict out two plus three So it’s all in the same flow.
HSV AI – Deloitte Room: Okay. So they can only be randomized so much, right? Thinking given the limited number of frames you have for training. Yes. So that was part of the thing is that for a real world problem, you’re not going to get millions of frames to work with. You’re not going to have weeks to work with. This trains in like a little more than an hour on your hardware. Well, uh, on my old hardware, yes, maybe 1070. Okay. Wow. DX. I was given the given the the number of samples and I was thinking an hour. What what are you working on there? Yeah, like an hour or two if I really just want to let it run for no reason because it’s it’s also highly efficient. So, it’s done within like 10 bucks.
HSV AI – Deloitte Room: So, more than ever overdo. You tried to run it on anything newer? Uh, I have now a 5070Ti and I thought about running it on the millions accounts. Uh, but then I got a job so I kind of tough time to just sit around and watch tell me about it. It be kind of neat to throw in a Google Collab or something. it would run there. You could easily Yeah, we we’ve done things like that before back when you could actually run like GPT2 and things. I mean, small thing, whatever. It was super good for like presentation, especially like college kids and other stuff as a they can run through it with you as you’re as you’re walking through it. So, you trained it on waves and clouds. The only two uh lots more things, some better, some worse. Did you do fire?
HSV AI – Deloitte Room: Yes. Yeah. You remember that from your poster? Yes. Yeah. Uh that was a kind of continued research sort of deal, right? Um but there’s most of the videos up there were of people at the beach or people pointing the camera at the center. So you kind of just I picked two way. I picked a wave. I picked a cloud. I went with that. But it should work on fire waves or any sort of water flow. It will work on people and animals as well as long as they maintain consistent paces and that kind of like I can never just to see how it does. You did try it out with as long as it’s consistent like it does not do great for well some character action.
HSV AI – Deloitte Room: Some it does great like ball of wind across the screen perfectly fine for that. Let’s say you want to have someone stop and jump, do a tuck and roll. Well, every big gap where it’s predicting it’s going to keep doing the same thing if has a little bit of uh Okay, we’re back to now. So, uh dumb question. If you had to do that, you know, you had to have a model that predicted inconsistent character action and you still wanted to use CNN because you love CNN. Yes. How would you uh do that? Um, well, I think part of the reason why I wanted to do something like this is that I don’t really need to because you could just take a GAN or a different model built for character action and say, okay, we’re detecting that for this part of the video, everything stays the same. Use my lightweight model.
HSV AI – Deloitte Room: Okay, I see. And now something now you have a sudden drop in loss. swap to something that changes to train on character action because it’s so lightweight. I mean, you can attach it to literally anything basically. Okay, that makes a lot of sense where you go the complete opposite direction. You feed in every possible movement and then you just detect if something breaks a possibility. And what it does actually work as a warning system because if the accuracy drops, say you’re doing waves, well suddenly the boat comes in, well the accuracy is going to drop. So hey, you’ve got a motion detector now. Yeah. Because you don’t know how to get this way. would tell you if something was steering or not steering. So, you’d have to find a threshold of loss of accuracy and then say under this threshold loss something.
HSV AI – Deloitte Room: So, because we’re talking about applications, did you have any application specific to this? Well, for this most uh the most specific one I had for this one was video restoration. So I worked with vacation photo vacation videos. What’s this best for? Well, it’s best for what you work with. You have a vacation video, you lost some frames out of it. Hey, plug this in an hour later. You’ve got a perfectly rebuilt video again. to speak that’s kind of like population. That’ll be interesting to try that. Say it’s not physics, but they are behaving. You know what I mean? They behave the same way. It’s morgism of Yeah. If you could increase the fidelity enough and have enough parameters to train on everything has even we can’t technically just do anything any time.
HSV AI – Deloitte Room: Your muscles will give away what you’re doing. Right. But to have the fidelity account to do that would be huge. That’s where the question comes in my mind true feed in all those. It’s easier take it from a cheater. It’s easier to cheat. Did you talk like that when you defended your I did not. It’s so I wasn’t there. So, it’s so free. This is everything you wanted to say. Yeah, this is everything I wanted to say. I’m getting it all off my chest. I’m admitting my guilt. I guess they can’t take it back, right? No, I’ve got my masters. I can do whatever they want. So, it’s 79,000. What can it be?
HSV AI – Deloitte Room: Or what can’t it be on? How about that? What can’t it be on? Yeah. Something pre-200 Yeah. So you probably building a whole system you probably couldn’t put it on like an original game. Nano running. Nano brought one of these. Yeah. Yeah. Is that what you’re thinking? Nano. Nanoact. I see all those ran it on a P zero. No problem. Put it on. You can put it on. The idea for it was that it could run on a edge camera. So any modern edge scammer would run it. It would do nothing. It wouldn’t even know something.
HSV AI – Deloitte Room: So So what are your thoughts on the single frame video? Yes, because it’s it’s really fun. Uh fun for me in my playground with different video generation models. And I know there’s a lot going on in the background. So imagine pretty much they’re using a combination description and then determining what’s in the image. When we feed that image in generators going forward, but feeding in a single image to some of the integration models without a problem just see what it does. It’s always interesting. Well, it’s it’s the dream of next frame prediction is to get to working with a single frame because whenever you’re working on problem set you have to work off assumptions. Well, the best assumption to be working off of next train prediction is all I’ve got is this one. What happens next? But I at least haven’t seen any that that shoot.
HSV AI – Deloitte Room: Almost all of them do something like the model I showed earlier which is one true frame and all the other. So you have differences between the true frames, you have uh pre-trained nets off of it beforehand. So almost all of them have some sort of training or additional information. Very few are lead here’s one frame, give me the next and nothing else. You know, I love just using that as I love feeding it just a still image of like here’s a person holding a cup and no problem seeing all the videos that it determines from going forward. Sometimes person makes it up to take a drink. Sometimes they just leave it on the table. Sometimes nothing happens here. The bird flies through. So it’s just it’s always really just an interesting subject. It is an interesting subject because almost all of those would be transformer models and the idea then becomes what parts of your positioning is getting is theoretically attached to what action.
HSV AI – Deloitte Room: So you’re sitting here holding a cup. Well, model just thinks if he’s holding a cup, he’s doing one of a couple of things with it. The number one thing any human doesn’t hold a cup like this, they’re going to knock off the back. Um, so almost it is interesting because it plays with association and kind of gets into the topic of large language models in general and the dangers of association and how what assumptions do we feed into models that then yield their own predictions because almost all videos on the internet fall into a certain number of categories of people inevitably doing stupid things. Yeah. Okay. So, like holding up a cup like this, I guarantee half of all the videos is some guy drinking. Yeah. I’d be interested to see if the camera is moving, but you have insight into So, so in other words, the camera is steering in a particular panning or something like that.
HSV AI – Deloitte Room: But if you knew the direction and you know velocity whatever of the camera moving if this would still hold um as you’re there there’s an application I’m thinking of but um that’d be kind of kind of interesting where everything else is still the waves are still crashing or something but I want to take this camera and move it up this way at you know 3 meters per second while maintaining whatever. So you have some type of like I don’t know what the right way to say it is. I would think of it as a hyperposition where it says for all this training that’s going on associated with a queue and they do that with the Google action they’re cued with this is what the video is. So this is what you can kind of expect. So you could have train the whole thing with an association with a Q and then when you give a queue in the future say this is a video of someone turning left.
HSV AI – Deloitte Room: Well most of the video is not going to be turned them turning left but it’s trained with that association. Well now in the future if you give it a video and cue it to turn left it’s going to preemptively try to do that for you. If you look at any of the diffusion uh transporter papers that are doing the spatio temporal stuff, yeah, they always I mean the cameras thing is very hard. It’s almost always an additional module. It’s just there’s a I can’t remember the actual name, but there’s a there’s a there’s a a coordinate system that basically does it in 3D space and it coordinates the movement of a camera relative to the world and that’s it whole like system. I was thinking, you know, predict the next frame given that my camera is moving. Yeah, it’s incredibly difficult to trade on multiple separate blocks of doing 2D. It’s doing height time width times time.
HSV AI – Deloitte Room: It doesn’t there’s no that third dimension is not you know drone footage moving. I didn’t do it with that was really some of the uh footage was of waves crashing on a rock, but the drone was like hovering around. It wasn’t flying over. So, not quickly, but it was it was it was moving. It wasn’t just like floating over it. It was moving in a pattern around it. And some of the drone footage was like it’s coming up from the ocean onto the beach. And uh when you add in accuracy, when you add in the second form of movement, the accuracy drops somewhere usually between 85 to 75% comparatively to on average most of my videos got about 95% accurate for the next frame. About 90% for the second frame. Okay. Um were you able to try this on anything with stereoscopic?
HSV AI – Deloitte Room: This is all just people’s education videos. So you don’t think uh it was only later that I got access to any actual real data. So I was just stealing people’s vacation. You could you could solve the moving camera problem from an aircraft with a known altitude focal length and all that. You got your own stereoscopic. Well, I was think a Z buffer for what you’re trying to No. No amount of metadata was used in this at all. So, it didn’t know anything aside from the picture. At any point did you let it just run and then see predictions like 15 crates away and just look at it. Did it look crazy or did it look good? It look terrible. What is What does that mean? Um, most of it is. So, it starts breaking down at some point.
HSV AI – Deloitte Room: depends on how far out and how big the dilation tube do. So at first how it breaks down is just there will be patches of like noise and for at least what I’ve seen for the ocean it’s usually purple. So it’ll just have a purple blob in the middle of the ocean where it couldn’t make a good prediction and then the purple blob consumes all around it with hunger unmatched. So first just goes the color that falls out then the structure after that just becomes a visual saw the one where I created the diffusion. Yeah. the nebulas and like what would happen to the nebula is that it would it would do it fairly well and then by a few seconds in which I’m sure is sitting here it would it would like explode again I imagin like sitting in a petri dish it looks all solid and fine and it just evaporates into the That’s what happens to the videos after you get too far out.
HSV AI – Deloitte Room: It evaporates enough. Interesting. So regardless if it’s like a diffusion or it has a nice My mind blinked the term where you focus on one object in the video, keep it stable and everything moves around it. So it look instead of my hand moving, my hand will be focused in the middle while the whole background was moving. Um, have you seen the videos like that? I have, but I don’t know what’s wrong. So, I’m curious how these models to that type of people are experiencing. I think it would uh strongly because a lack of focus usually is attached. I you get a lack of focus just by adding noise to an image. So I think that would just define itself as the noise that be harder to predict, which then a diffusion network would come in and be perfect for.
HSV AI – Deloitte Room: Anyone else have any questions? Any thoughts? So how many forms of predict? You said there was like predictable like waves and then unpredictable was there levels like semi-predictable or uh semi-predictable I would call what I mentioned briefly as like a person’s waffle is semi-predictable as long as they maintain a pace and maintain stride and don’t do anything too. So it doesn’t do well for random actions, but not everything chooses to be random because you still are subject to gravity, are subject to the laws of your own motion. And if you follow those laws without intentionally breaking them, I’d say it’s semi works well for them, but uh it breaks down full determinant action. That would be so weird from a uh so If you’ve got videos of like a train station and crowds moving through and stuff, those generally operate in a particular kind of a way. And if you were doing this looking for error, you could note if somebody was moving in a different way than the typical crowd moved through this place.
HSV AI – Deloitte Room: So if all of a sudden you saw somebody like running through a ballroom the wrong way or whatever, it’d be a pretty interesting flag of a weird maybe somebody should go see people who are speeding on the highway because people all the cars typically are flowing the same speed even if it’s above or below the speed limit as long as they’re in a consistent flow with each other but someone races through and breaks the flow of the situation, Right. So to your point, do you have to be trained on a scenario? Like does the camera have to be in the ballroom pull the data trained somewhere else? Obviously back edge camera like is that the use case? Um or you to pick because like I said 70,000 parameters there. All you need is like the worst pie you’ve ever seen in your entire life and you can do it there. Train there. Yeah.
HSV AI – Deloitte Room: Why? Just need somebody to know when to this is not a model that’s like you train on a bunch of videos and then then you deploy. This is individually trained. You can take whatever you got, tramp it in there, and spit out some especially when it’s for this purpose at this location. And if you want another one, we’ll share another one. Yes. I don’t know if the location matters. Can’t you people or the roads matter? Maybe layout is agnostic. Yes. Just full of people. I don’t care if it’s train station, ballroom. Correct. You could just have it run people diagnostic. There’s probably some nuances to certain structure whatever and stuff. But in general interesting for the uh we did a talk a couple of weeks ago at it was a group of architects, city planners, things like that.
HSV AI – Deloitte Room: Um and they’re always worried about how people move through their spaces and things like that. That’d be an interesting thing to drop on for them. Well, it started off as water. So uh as he said it though we don’t want to do water flow because the me moment you mentioned water flow you get a lot of associations with that and so yes we kind of refocused on more general videos because yes it’s agnostic as long as it’s a flow of semi semifysics it’s not true for this video like I wish it Okay. One thing that sounds kind of fun too because we’re talking about you’re maybe doing an escalation of like you’re using this for your low power mode or whatever it is. So you have this high power model on it. So if you have that and they’re both connected, one of the things you could do is your low power mode, whatever that’s going using that high power capacity to do that training, you know, on board.
HSV AI – Deloitte Room: But then, you know, kind of using that so I’m on, but when I’m not on, I’m using the rest of my resources for whatever it is. And then uh this super low parameter count becomes very useful even as part you don’t even need to switch you only need to switch all the high the higher pressure they’re more extensive you don’t ever need to turn off the Yeah that’s what I’m saying you can always have that on but it’s just it’s either I’m using the the executive mode or training you know or whatever it mention just 10 that uh I believe this trains in two or three. So it’s highly efficient. un I got for so yeah I went for 100 because you’re supposed to but um by 20 by 20 nothing happened so it’s all done by about 10 or so is that just are you still after about 15 or so nothing’s happened did you try bring topic. What?
HSV AI – Deloitte Room: Have you heard of rock candy? Well, I’ve heard of something that’s similar for rocks and sedimentation, but I don’t think it’s the same thing. K they started finding where sometimes if you passed way beyond any reasonable that you would start to see performance. So we’re talking about where somebody might get 10,000 ethics like okay we’re going to but then somewhere along the line they would actually get like little super I’ve actually read a lot of those papers and it’s um it is actually a side effect of memorization basically so they would learn the semi-correct answer pretty quickly but then to actually understand the underlying uh system long enough it would looked like nothing was happening but it was learning the entire way the performance isn’t increasing And suddenly all of a sudden it knew enough to get like the true right answer and then suddenly performs up like it’s magic and then it’s like no interesting things with if you go to frequency analysis maybe if you let you get an answer right away but then you’re just moving on with your two weeks later like there it is sort of thing you know like it’s pretty different I guess um I suppose that how the loss function is like how is it combined
HSV AI – Deloitte Room: So, uh I use a custom loss function because MSC is great for color and structural similarity index is great for structure. Uh neither are good for the eye. So waves are mostly a form of structure. So I weighted it more towards SSIM, but I had a 7030 splits of them, combined the lawsuits, and pushed them back for trade. So that’s what I did for it. And we used Adam W because Adam W is the best. Bite me on that. I Nobody’s gonna just turn it on. s***. and a slightly unique relu where it’s effectively um a positively shifted reload. So you drop out every positive value is doubled effectively where every negative value is taken away same. So instead of losing all those values, you’re just promoting the same. So that’s more of the in the model kind of way.
HSV AI – Deloitte Room: I don’t know if I got too deep into it, but um the model itself, this is just a repeated unit. So yeah, this same little core unit is repeated twice here. And you would repeat it however many dilations you want. and you repeat it in the next phase, however many predictions you wanted. You’d always start with just one because you only were working off of ever how many colors. But um so everything is just convolutional layers pooling, unpooling some reloop. I actually on my newest version of this model cut out all the reloop. So it doesn’t even have that anymore. All of the nonlinearity is through uh pooling and unpooling. and T is desperate for a paper about all about pooling and unpooling. So you gonna do it? We’ve been working on it for a year or so.
HSV AI – Deloitte Room: We’ll see how that goes. Um so create a window one convolution. Split the convolution into make a copy of the information. Use the information once to min to max pool. a min poolool. It’s just a max pool, but you put a negative in front of it. Uh two convolutions for the midpool and the max pool individually. Unpool them respectively. put it all back together, including just a tiny little taste of an RNN, and that you take the original convolutionals information, put it back on the end before you create the new windows. So you have the poolled unpooled the max pool and unpoolled the min pool and unpool and the original information put that into the final convolutional layer then send that into the next phase which is to create new windows. When you say a tiny bit what does that mean?
HSV AI – Deloitte Room: Well RNN’s have an entire module that go throughout the whole model. Yeah, I just take from here from one phase. So only within the phase do I keep the information flowing from a previous point. So I find that LSTMs are too obsessed with keeping information for the whole model. breaking it down like this, you could think of these as effectively three individual models in a sort of way because these other phases don’t really care about what happens to the other phases. They’re all independent of each other. And the way I rearrange how the windows are done u the parameters are not mixed up. So the reds and the greens and blues are always kept separate. The dilations are always kept separate. So everything is always agnostic of everything else going on within the model creating this what I called earlier expertise everything in the model is okay we’re the reds we’re the reds for dilation one okay we’re the reds for dilation one for prediction one we do this prediction and we do this prediction because that’s how I think models all model should be done to a single purpose for because I don’t really believe in drilling.
HSV AI – Deloitte Room: I believe in specific models, a specific model for a specific task. You don’t fix a whole house just using a hammer. You have a saw, you have screwdriver, you have everything, a whole bucket of tools, right tool for the right job. And I put that into my muscle. So, this is great. Thank you. agree with your philosophy and thank you for that because that reminded me to speak more about that. It just repeats. Um I don’t necessarily know how necessary this bit is. I added it because effectively because I like it. But like with the real supports, I think the next part of this model, even more so than making a baseline, would be seeing how much I can just rip out of it because I find that I think just the core idea of the dilations themselves are uh what’s really I really want to see what all can take away still.
HSV AI – Deloitte Room: Can I ask if you’ve done any experiment condensing the color channels into one. No, just part of the T and Dr. Facilia like their stories and the way they like the way they’re done. So the professor I worked with has a way he likes to tell stories and he likes to have his own narrative. He keeps looking at me because T’s my advisor for PhD. But he’s there. Uh I’ve worked with a researcher and my professor for a lot of my research and they like things in their own stories in their own way. So uh it has to be interesting. It can’t just be what it is. So to actually make a good product afterwards, I have to go back in and take all the interest. I have to take the pencil down. Take the strip the paint off. Let’s thank Ben for super interesting.
HSV AI – Deloitte Room: Uh Christopher, do you have any other questions online? I think you are it.
Christopher Bales: Uh, no. I mean, it’s it’s crazy. I don’t understand half of it. And um, it’s really interesting nonetheless, and I I appreciate everything. So thanks Benjamin.
HSV AI – Deloitte Room: No problem. If you have any questions, be happy to answer. And we have another group or worry faces. Thank you everyone for enduring my mad. Thanks for the cheating and go recommending that I come.