Evaling Video Slop — Maor Bril, Character.ai

>> So, hi. I'm Mayur. I've been with >> So, hi. I'm Mayur. I've been with Character for a bit over 2 years and Character for a bit over 2 years and Character for a bit over 2 years and we'll talk about we'll talk about we'll talk about AI slop, right? AI slop, right? AI slop, right? I think that, you know, I think that, you know, I think that, you know, when we look at video generations as a when we look at video generations as a when we look at video generations as a whole, right? We have like two kind of whole, right? We have like two kind of whole, right? We have like two kind of uh parallel tracks. One is the the video uh parallel tracks. One is the the video uh parallel tracks. One is the the video generation, which generation, which generation, which became insanely good from from models became insanely good from from models became insanely good from from models like Kling and SeaDance and VEO and like Kling and SeaDance and VEO and like Kling and SeaDance and VEO and Sora. We still Sora. We still Sora. We still remember Sora. remember Sora. remember Sora. But but but the part that got left But but but the part that got left But but but the part that got left behind is how we evaluate the quality of behind is how we evaluate the quality of behind is how we evaluate the quality of the video that was the video that was the video that was generated, right? So, on on the one generated, right? So, on on the one generated, right? So, on on the one hand, um we we we still kind of squint hand, um we we we still kind of squint hand, um we we we still kind of squint at it and decide whether or not it's at it and decide whether or not it's at it and decide whether or not it's it's good, but on the other hand, we it's good, but on the other hand, we it's good, but on the other hand, we know that the generation has gotten a know that the generation has gotten a know that the generation has gotten a lot better. And when when we when we we lot better. And when when we when we we lot better. And when when we when we we look at at X or look at at X or look at at X or whatever whatever whatever social you're you're you're consuming social you're you're you're consuming social you're you're you're consuming your content on, there are a lot of your content on, there are a lot of your content on, there are a lot of guides on how to create amazing videos guides on how to create amazing videos guides on how to create amazing videos with um this model or or that. So, the with um this model or or that. So, the with um this model or or that. So, the hard part was never how to make video.

hard part was never how to make video. hard part was never how to make video. The hard part was how do we generate The hard part was how do we generate The hard part was how do we generate um um um good enough video and how do we judge if good enough video and how do we judge if good enough video and how do we judge if the video is good enough? So, so now the video is good enough? So, so now the video is good enough? So, so now we've gotten to a world where we've gotten to a world where we've gotten to a world where the the generation of video is basically the the generation of video is basically the the generation of video is basically free, right? Free as especially when you free, right? Free as especially when you free, right? Free as especially when you compare it to how much studios would compare it to how much studios would compare it to how much studios would charge. Um and but but the problem is charge. Um and but but the problem is charge. Um and but but the problem is the most the grand majority of videos the most the grand majority of videos the most the grand majority of videos that is generated is not that good, that is generated is not that good, that is generated is not that good, right? We have like right? We have like right? We have like a lot of hallucinations like a third a lot of hallucinations like a third a lot of hallucinations like a third limb, uh opening and closing the the limb, uh opening and closing the the limb, uh opening and closing the the door at the same time, hovering, door at the same time, hovering, door at the same time, hovering, physics, um etc. So, unfortunately, in physics, um etc. So, unfortunately, in physics, um etc. So, unfortunately, in order to get high high high-quality order to get high high high-quality order to get high high high-quality content, we need a human to judge. And I content, we need a human to judge. And I content, we need a human to judge. And I I know when was the last time you've I know when was the last time you've I know when was the last time you've seen how someone is creating these seen how someone is creating these seen how someone is creating these long-form long-form long-form um um um generated video. It's usually a lot of generated video. It's usually a lot of generated video. It's usually a lot of shorter generations and a lot of shorter generations and a lot of shorter generations and a lot of editing. editing. editing. The problem is because we're using a lot The problem is because we're using a lot The problem is because we're using a lot of the tools that we built for the ticks of the tools that we built for the ticks of the tools that we built for the ticks for the text era, for the image era, for for the text era, for the image era, for for the text era, for the image era, for videos, right? We're using things like videos, right? We're using things like videos, right? We're using things like clip score, which is is great to to to clip score, which is is great to to to clip score, which is is great to to to judge a single frame.

judge a single frame. judge a single frame. Things like LP LP IPS will help us kind of detect the IPS will help us kind of detect the IPS will help us kind of detect the drift between frames, but we don't have drift between frames, but we don't have drift between frames, but we don't have I mean but the problem is when you kind I mean but the problem is when you kind I mean but the problem is when you kind of combine all these together, all these of combine all these together, all these of combine all these together, all these tools are good at watching the tools are good at watching the tools are good at watching the individual frames. They're good at individual frames. They're good at individual frames. They're good at checking this this this this does this checking this this this this does this checking this this this this does this specific frame does it match the the specific frame does it match the the specific frame does it match the the prompt that generated it, right? It will prompt that generated it, right? It will prompt that generated it, right? It will check consistency between frames and it check consistency between frames and it check consistency between frames and it will check whether or not it match the will check whether or not it match the will check whether or not it match the prompt that the drove it. But but what prompt that the drove it. But but what prompt that the drove it. But but what it won't do it doesn't tell you if it won't do it doesn't tell you if it won't do it doesn't tell you if if it's if you told if you told the if it's if you told if you told the if it's if you told if you told the story that you you meant to tell, right? story that you you meant to tell, right? story that you you meant to tell, right? If you think about what is video, video If you think about what is video, video If you think about what is video, video is a storytelling medium. is a storytelling medium. is a storytelling medium. Video is just another form on how we Video is just another form on how we Video is just another form on how we tell a story, right? From for tell a story, right? From for tell a story, right? From for any type of story. So so one of one of any type of story. So so one of one of any type of story. So so one of one of the things we have to look at, does it the things we have to look at, does it the things we have to look at, does it tell the actual story? Does the physics tell the actual story? Does the physics tell the actual story? Does the physics make sense? Like for example, if you make sense? Like for example, if you make sense? Like for example, if you want a video of a character walking want a video of a character walking want a video of a character walking downstairs, does it actually walk or or downstairs, does it actually walk or or downstairs, does it actually walk or or hover?

hover? hover? Does the character stay the same Does the character stay the same Does the character stay the same character across multiple shots? character across multiple shots? character across multiple shots? Does the pacing make sense? Like you Does the pacing make sense? Like you Does the pacing make sense? Like you know, for example, people take time know, for example, people take time know, for example, people take time going from one place to another. We need going from one place to another. We need going from one place to another. We need to make sure that the pacing makes sense to make sure that the pacing makes sense to make sure that the pacing makes sense as well. And especially when we add as well. And especially when we add as well. And especially when we add audio, we want to make sure that the audio, we want to make sure that the audio, we want to make sure that the audio is kind of synced with the audio is kind of synced with the audio is kind of synced with the imagery. Like for example, if someone is imagery. Like for example, if someone is imagery. Like for example, if someone is slamming a door, we want that that that slamming a door, we want that that that slamming a door, we want that that that that that that sound of the door being that that that sound of the door being that that that sound of the door being slammed to be exactly when the door is slammed to be exactly when the door is slammed to be exactly when the door is actually being slammed. actually being slammed. actually being slammed. Now, the the next iteration we all went Now, the the next iteration we all went Now, the the next iteration we all went to a while ago. We started using LLM as to a while ago. We started using LLM as to a while ago. We started using LLM as a judge for everything and we have a judge for everything and we have a judge for everything and we have amazing foundational models that we just amazing foundational models that we just amazing foundational models that we just throw throw throw videos at them. The problem with them is videos at them. The problem with them is videos at them. The problem with them is that A, they're slow. B, they're only as that A, they're slow. B, they're only as that A, they're slow. B, they're only as good as your prompt and multiple people good as your prompt and multiple people good as your prompt and multiple people will prompt multiple ways and the same will prompt multiple ways and the same will prompt multiple ways and the same model may respond in a very very model may respond in a very very model may respond in a very very different way. And sometimes the prompt different way. And sometimes the prompt different way. And sometimes the prompt we use like is it consistent? we use like is it consistent? we use like is it consistent? Does this match the the prompt? But but Does this match the the prompt? But but Does this match the the prompt? But but then the question we really care about then the question we really care about then the question we really care about is it good? And the answer varies.

is it good? And the answer varies. is it good? And the answer varies. So, So, So, oops, sorry about that. So, our first oops, sorry about that. So, our first oops, sorry about that. So, our first iteration is like let's take all these iteration is like let's take all these iteration is like let's take all these things and build a repeatable things and build a repeatable things and build a repeatable benchmark on how we test video that we benchmark on how we test video that we benchmark on how we test video that we can rerun over and over and over again. can rerun over and over and over again. can rerun over and over and over again. So, so that combines both metrics as I So, so that combines both metrics as I So, so that combines both metrics as I said earlier that that knows how to view said earlier that that knows how to view said earlier that that knows how to view individual frames, but also individual frames, but also individual frames, but also consistent LLM as a judge, right? Where consistent LLM as a judge, right? Where consistent LLM as a judge, right? Where we also use human annotation to to we also use human annotation to to we also use human annotation to to calibrate the the LLM as a judge. So, calibrate the the LLM as a judge. So, calibrate the the LLM as a judge. So, for every report that we generate with for every report that we generate with for every report that we generate with that harness, that harness, that harness, we're able to have humans annotate it we're able to have humans annotate it we're able to have humans annotate it and basically feed that feedback back in and basically feed that feedback back in and basically feed that feedback back in into the the into the the into the the the LLM as a judge prompt to make sure the LLM as a judge prompt to make sure the LLM as a judge prompt to make sure that it's it's aligned with what I think that it's it's aligned with what I think that it's it's aligned with what I think or what the annotator thought is good. or what the annotator thought is good. or what the annotator thought is good. And and we use it to to score the And and we use it to to score the And and we use it to to score the videos. The problem with videos. The problem with videos. The problem with this approach, it's very slow, it's very this approach, it's very slow, it's very this approach, it's very slow, it's very expensive and especially when we we want expensive and especially when we we want expensive and especially when we we want to bring it in for our users to be able to bring it in for our users to be able to bring it in for our users to be able to to to generate a lot of video because creation generate a lot of video because creation generate a lot of video because creation is a very hard process. And is a very hard process. And is a very hard process. And so the problem as I said, the problem is so the problem as I said, the problem is so the problem as I said, the problem is when when So, this is a slow process and when when So, this is a slow process and when when So, this is a slow process and we need to bring it as close to the we need to bring it as close to the we need to bring it as close to the users as possible and also earlier into users as possible and also earlier into users as possible and also earlier into the process. The reason for that is the process. The reason for that is the process. The reason for that is if we take a look at all the metrics and if we take a look at all the metrics and if we take a look at all the metrics and there's there mistakes that we can find there's there mistakes that we can find there's there mistakes that we can find earlier than earlier than earlier than than later than it's a lot cheaper to than later than it's a lot cheaper to than later than it's a lot cheaper to correct that that particular mistake.

correct that that particular mistake. correct that that particular mistake. So, for So, for So, for example, right? On example, right? On example, right? On on the left we have two starting frames of different two starting frames of different different shots, right? But it's easy to different shots, right? But it's easy to different shots, right? But it's easy to to correct to to view it at this point to correct to to view it at this point to correct to to view it at this point and see did the character drift between and see did the character drift between and see did the character drift between frame one and and frame two because frame one and and frame two because frame one and and frame two because those frames will be used as starting those frames will be used as starting those frames will be used as starting frames to generate videos. So, if you frames to generate videos. So, if you frames to generate videos. So, if you can correct catch the drift at this can correct catch the drift at this can correct catch the drift at this point and correct it, then point and correct it, then point and correct it, then then it's much cheaper to generate the then it's much cheaper to generate the then it's much cheaper to generate the video as a whole because we we can video as a whole because we we can video as a whole because we we can correct it at a much correct it at a much correct it at a much cheaper cost. And the same thing applies cheaper cost. And the same thing applies cheaper cost. And the same thing applies when we look at longer form video, when we look at longer form video, when we look at longer form video, right? When we see all these three, right? When we see all these three, right? When we see all these three, four, five-minute long videos, they're four, five-minute long videos, they're four, five-minute long videos, they're usually a collection of a lot of shorter usually a collection of a lot of shorter usually a collection of a lot of shorter videos. And being able to catch a videos. And being able to catch a videos. And being able to catch a six-second generation that drifted and six-second generation that drifted and six-second generation that drifted and and regenerated that regenerate that and regenerated that regenerate that and regenerated that regenerate that before we combine the whole video will before we combine the whole video will before we combine the whole video will end up being a better result as a whole. Um Um >> [clears throat] >> [clears throat] >> [clears throat] >> And >> And >> And now the the other problem what we're now the the other problem what we're now the the other problem what we're trying to solve is some of these axes, trying to solve is some of these axes, trying to solve is some of these axes, right? Only exist across time, right?

right? Only exist across time, right? right? Only exist across time, right? So, for example, when we look at the So, for example, when we look at the So, for example, when we look at the right? We we mentioned the story, right? right? We we mentioned the story, right? right? We we mentioned the story, right? So, does does the story that we're So, does does the story that we're So, does does the story that we're trying to tell with that video, does it trying to tell with that video, does it trying to tell with that video, does it hold in that video? Does the video tell hold in that video? Does the video tell hold in that video? Does the video tell the exact story? Does the pacing make the exact story? Does the pacing make the exact story? Does the pacing make sense, right? And we mentioned the the sense, right? And we mentioned the the sense, right? And we mentioned the the the the sound. So, the the sound. So, the the sound. So, as I said, right? The the you know, the as I said, right? The the you know, the as I said, right? The the you know, the underlying goal is to bring that evals underlying goal is to bring that evals underlying goal is to bring that evals closer to to the to the the online closer to to the to the the online closer to to the to the the online generation because the sooner we're able generation because the sooner we're able generation because the sooner we're able to catch those mistakes, we were to to catch those mistakes, we were to to catch those mistakes, we were to the sooner we're able to catch that the sooner we're able to catch that the sooner we're able to catch that drift, right? Then it's um drift, right? Then it's um drift, right? Then it's um it's much easier, much much cheaper to it's much easier, much much cheaper to it's much easier, much much cheaper to fix. fix. fix. Now, Now, Now, so so that now now the problem is that so so that now now the problem is that so so that now now the problem is that as I said, this is a very slow process. as I said, this is a very slow process. as I said, this is a very slow process. So So So the the solution is actually actually to the the solution is actually actually to the the solution is actually actually to take all these committee of experts and take all these committee of experts and take all these committee of experts and distill it into one small model that is distill it into one small model that is distill it into one small model that is also very very fast, but also very very fast, but also very very fast, but it is able to give us it is able to give us it is able to give us a response that is not whether or not a response that is not whether or not a response that is not whether or not this video is slap or not, but why is it this video is slap or not, but why is it this video is slap or not, but why is it slap? Right? Why Why is that video slap? Right? Why Why is that video slap? Right? Why Why is that video scored low versus the other? Because for scored low versus the other? Because for scored low versus the other? Because for example, it added an extra limb, because example, it added an extra limb, because example, it added an extra limb, because it didn't obey physics, because it didn't obey physics, because it didn't obey physics, because the the the audio was was out of sync. So so the the audio was was out of sync. So so the the audio was was out of sync. So so the the goal was a build it on top of a small goal was a build it on top of a small goal was a build it on top of a small VLM and why is it a VLM? VLM because we VLM and why is it a VLM? VLM because we VLM and why is it a VLM? VLM because we needed the video needed the video needed the video the model to be able to see the image, the model to be able to see the image, the model to be able to see the image, but also we needed to work fast, right?

but also we needed to work fast, right? but also we needed to work fast, right? Because we bought it closer to the Because we bought it closer to the Because we bought it closer to the generation, where in fact it takes about generation, where in fact it takes about generation, where in fact it takes about the model we have trained, it takes the model we have trained, it takes the model we have trained, it takes about 3 seconds to score a three about 3 seconds to score a three about 3 seconds to score a three 15 second video. Now, the Now, the we we also we we also we we also tested a bigger model and the results tested a bigger model and the results tested a bigger model and the results were better, but it was significantly were better, but it was significantly were better, but it was significantly slower. And and the decision was to go slower. And and the decision was to go slower. And and the decision was to go with the the with the the with the the the smaller model because the the added the smaller model because the the added the smaller model because the the added value from the bigger model didn't value from the bigger model didn't value from the bigger model didn't justify the the the justify the the the justify the the the the slowness. the slowness. the slowness. The other very The other very The other very interesting realization we came to is interesting realization we came to is interesting realization we came to is don't score compare. What does that don't score compare. What does that don't score compare. What does that mean? For example, if I'll ask any mean? For example, if I'll ask any mean? For example, if I'll ask any person in this room to look at a person in this room to look at a person in this room to look at a particular video and and rank it from 1 particular video and and rank it from 1 particular video and and rank it from 1 to 10 on storytelling, right? I'm pretty to 10 on storytelling, right? I'm pretty to 10 on storytelling, right? I'm pretty sure that you know, what will be a six sure that you know, what will be a six sure that you know, what will be a six for you will be a five for you, will be for you will be a five for you, will be for you will be a five for you, will be a four for you and and and an and an a four for you and and and an and an a four for you and and and an and an eight for you. Right? But if I show you eight for you. Right? But if I show you eight for you. Right? But if I show you two videos and I'll ask you which one of two videos and I'll ask you which one of two videos and I'll ask you which one of them is telling a better story, the them is telling a better story, the them is telling a better story, the grand majority will probably agree that grand majority will probably agree that grand majority will probably agree that B is telling a better story than A, B is telling a better story than A, B is telling a better story than A, right? And if you do it enough times, right? And if you do it enough times, right? And if you do it enough times, then it it uh it's easy to um uh then it it uh it's easy to um uh then it it uh it's easy to um uh generalize the model at um towards generalize the model at um towards generalize the model at um towards detecting what's detecting what's detecting what's better uh uh versus better uh uh versus better uh uh versus uh not.

uh not. uh not. So, So, So, >> [clears throat] >> [clears throat] >> [clears throat] >> we trained on pairs, right? Um uh A >> we trained on pairs, right? Um uh A >> we trained on pairs, right? Um uh A versus B as opposed to 1 through 10. versus B as opposed to 1 through 10. versus B as opposed to 1 through 10. Now, we manufactured badness. So, Now, we manufactured badness. So, Now, we manufactured badness. So, luckily, the internet is full of very luckily, the internet is full of very luckily, the internet is full of very high-quality videos, and it's very very high-quality videos, and it's very very high-quality videos, and it's very very easy to get good videos, and it was very easy to get good videos, and it was very easy to get good videos, and it was very fun to create bad videos A by either fun to create bad videos A by either fun to create bad videos A by either corrupting corrupting corrupting good videos or by, you know, just good videos or by, you know, just good videos or by, you know, just generating random slop. generating random slop. generating random slop. >> [clears throat] >> [clears throat] >> [clears throat] >> Um >> Um >> Um Now, Now, Now, we shipped V1, and it was so wrong. It we shipped V1, and it was so wrong. It we shipped V1, and it was so wrong. It It was wrong, but it was wrong in a very It was wrong, but it was wrong in a very It was wrong, but it was wrong in a very confident way. So, uh uh for for for confident way. So, uh uh for for for confident way. So, uh uh for for for example, the example, the example, the the frame you see here is from from a the frame you see here is from from a the frame you see here is from from a video that the model scored 9.2 on the video that the model scored 9.2 on the video that the model scored 9.2 on the camera work, and the camera didn't move. camera work, and the camera didn't move. camera work, and the camera didn't move. For for 4 seconds, it was like a still For for 4 seconds, it was like a still For for 4 seconds, it was like a still image of the same character, but but but image of the same character, but but but image of the same character, but but but the model was was very very happy with the model was was very very happy with the model was was very very happy with with, you know, the with, you know, the with, you know, the cinematography. cinematography. cinematography. Um Um Um so, the physics in in in in some so, the physics in in in in some so, the physics in in in in some other videos, which I'm not showing the other videos, which I'm not showing the other videos, which I'm not showing the cuz of uh time limitations, it says that cuz of uh time limitations, it says that cuz of uh time limitations, it says that the physics look great, but it said it the physics look great, but it said it the physics look great, but it said it on on ghosts hovering and people flying, on on ghosts hovering and people flying, on on ghosts hovering and people flying, etc.

etc. etc. So, So, So, the I mean, so the so then the question the I mean, so the so then the question the I mean, so the so then the question is like, why was it wrong? The The is like, why was it wrong? The The is like, why was it wrong? The The reason it was wrong is because how we reason it was wrong is because how we reason it was wrong is because how we generated that data, right? It It um it generated that data, right? It It um it generated that data, right? It It um it scored the vibe as opposed to the the scored the vibe as opposed to the the scored the vibe as opposed to the the the axes. So, it it learned how to how the axes. So, it it learned how to how the axes. So, it it learned how to how to detect um cohe- coherent videos and to detect um cohe- coherent videos and to detect um cohe- coherent videos and it learned how to detect the the the the it learned how to detect the the the the it learned how to detect the the the the the the artificial artifacts. Basically, the the artificial artifacts. Basically, the the artificial artifacts. Basically, the gloss of the video as opposed to uh the gloss of the video as opposed to uh the gloss of the video as opposed to uh whether or not the video actually um whether or not the video actually um whether or not the video actually um told the the the the the Sorry. told the the the the the Sorry. told the the the the the Sorry. The videos actually told the the the The videos actually told the the the The videos actually told the the the story. And story. And story. And um um um And so so so the solution was to fix the And so so so the solution was to fix the And so so so the solution was to fix the data set. And so the way we we fixed the data set. And so the way we we fixed the data set. And so the way we we fixed the the the the the the data set, we the the the the the data set, we the the the the the data set, we actually I actually I actually I uh um I started pairing real footage uh um I started pairing real footage uh um I started pairing real footage versus versus versus um AI footage. Now, the risk with that um AI footage. Now, the risk with that um AI footage. Now, the risk with that and that that's the reason why I voted I and that that's the reason why I voted I and that that's the reason why I voted I voted I avoided doing it at first is voted I avoided doing it at first is voted I avoided doing it at first is because I didn't didn't want to create because I didn't didn't want to create because I didn't didn't want to create an AI detector, right? Because if you an AI detector, right? Because if you an AI detector, right? Because if you start creating pairs of good is is is start creating pairs of good is is is start creating pairs of good is is is human-generated video and bad is AI uh human-generated video and bad is AI uh human-generated video and bad is AI uh uh video, then then there's a very big uh video, then then there's a very big uh video, then then there's a very big chance of of the model overfitting and chance of of the model overfitting and chance of of the model overfitting and becoming an AI detector as opposed to a becoming an AI detector as opposed to a becoming an AI detector as opposed to a um uh uh um uh uh um uh uh video quality detector. So, the video quality detector. So, the video quality detector. So, the There are two things I did in order to There are two things I did in order to There are two things I did in order to avoid that. A, I made sure that the avoid that. A, I made sure that the avoid that. A, I made sure that the encoding is is consistent across both encoding is is consistent across both encoding is is consistent across both sides of of the equation. So, uh there

sides of of the equation. So, uh there sides of of the equation. So, uh there there there there's no there there there's no there there there's no um artificial artifacts for for video A um artificial artifacts for for video A um artificial artifacts for for video A versus video B. versus video B. versus video B. Uh and I I used the exact same method of Uh and I I used the exact same method of Uh and I I used the exact same method of annotating both videos. So, so both the annotating both videos. So, so both the annotating both videos. So, so both the axes So, all the axes in those videos axes So, all the axes in those videos axes So, all the axes in those videos were annotated in the same way. were annotated in the same way. were annotated in the same way. Uh and surprise, it it turned out pretty Uh and surprise, it it turned out pretty Uh and surprise, it it turned out pretty awesome. Um and so so so now what what awesome. Um and so so so now what what awesome. Um and so so so now what what what we're able to do, especially when what we're able to do, especially when what we're able to do, especially when we you looking at at at videos, we you looking at at at videos, we you looking at at at videos, A, we we changed from a very complex A, we we changed from a very complex A, we we changed from a very complex pipeline, right? To an pipeline, right? To an pipeline, right? To an agentic workflow. The reason behind this agentic workflow. The reason behind this agentic workflow. The reason behind this is, a, the pipelines work great if you is, a, the pipelines work great if you is, a, the pipelines work great if you have a very very unique use case. But have a very very unique use case. But have a very very unique use case. But one once you put put one once you put put one once you put put put it in front of users, they'll have a put it in front of users, they'll have a put it in front of users, they'll have a very very distinct story that they want very very distinct story that they want very very distinct story that they want to tell with their own characters, with to tell with their own characters, with to tell with their own characters, with with their own images, and their own with their own images, and their own with their own images, and their own voice. So that's where it starts to voice. So that's where it starts to voice. So that's where it starts to drift. drift. drift. But by providing the agents with tools But by providing the agents with tools But by providing the agents with tools to validate the quality of of the to validate the quality of of the to validate the quality of of the outputs it's creating, it's able to outputs it's creating, it's able to outputs it's creating, it's able to adapt to changes better. It's also able adapt to changes better. It's also able adapt to changes better. It's also able to verify its own work and and and fix to verify its own work and and and fix to verify its own work and and and fix things as they go along. So if you're things as they go along. So if you're things as they go along. So if you're going to steal from from these going to steal from from these going to steal from from these from the stock of a few things, one, go from the stock of a few things, one, go from the stock of a few things, one, go relative, not absolute, right? As I relative, not absolute, right? As I relative, not absolute, right? As I explained earlier, explained earlier, explained earlier, the the value of the the value of the the value of comparing video A versus comparing video A versus comparing video A versus video B will always give you a better video B will always give you a better video B will always give you a better result going forward.

result going forward. result going forward. Be score the real axis that you care Be score the real axis that you care Be score the real axis that you care about. So if you care about about. So if you care about about. So if you care about storytelling, if you care about pacing, storytelling, if you care about pacing, storytelling, if you care about pacing, if you care about if you care about if you care about physics, physics, physics, score those axes. score those axes. score those axes. Don't expect them to Don't expect them to Don't expect them to miraculously appear. And put eval inside miraculously appear. And put eval inside miraculously appear. And put eval inside the the the generation loop, right? Especially if generation loop, right? Especially if generation loop, right? Especially if your goal is is is is to have a higher your goal is is is is to have a higher your goal is is is is to have a higher quality of generation, get the the eval quality of generation, get the the eval quality of generation, get the the eval as close to to the generation loop loop as close to to the generation loop loop as close to to the generation loop loop as possible. as possible. as possible. Eventually evaluate it as a story. Eventually evaluate it as a story. Eventually evaluate it as a story. Videos are stories. Videos are just Videos are stories. Videos are just Videos are stories. Videos are just another way for us to tell stories to another way for us to tell stories to another way for us to tell stories to others. And others. And others. And Thank thank you very much. Thank thank you very much. Thank thank you very much. >> [applause] >> All right. Any questions? Okay, down >> All right. Any questions? Okay, down here. Awesome. here. Awesome. here. Awesome. All right, I got two down here. Here you All right, I got two down here. Here you All right, I got two down here. Here you go.

>> Hi. How do you eval sound? >> Hi. How do you eval sound? Sound and video matching. Sound and video matching. Sound and video matching. >> I'm I'm I'm Can you you repeat? >> I'm I'm I'm Can you you repeat? >> I'm I'm I'm Can you you repeat? >> How do you evaluate sound? Sound effects >> How do you evaluate sound? Sound effects >> How do you evaluate sound? Sound effects and matching with the video. and matching with the video. and matching with the video. >> oh yeah. >> oh yeah. >> oh yeah. That's a fantastic question. So, That's a fantastic question. So, That's a fantastic question. So, um um um so so sound is actually a a combination so so sound is actually a a combination so so sound is actually a a combination of few things. One, I'm using of few things. One, I'm using of few things. One, I'm using Atmos to Atmos to Atmos to to to make sure that that that the the to to make sure that that that the the to to make sure that that that the the the sound quality the sound quality the sound quality is high enough and is is high enough and is is high enough and is understandable. B, the model will will understandable. B, the model will will understandable. B, the model will will will will learn to learn to identify key will will learn to learn to identify key will will learn to learn to identify key frames. frames. frames. Right? And and and especially because Right? And and and especially because Right? And and and especially because when I feed something into the model, when I feed something into the model, when I feed something into the model, it's it can be just a video or it can be it's it can be just a video or it can be it's it can be just a video or it can be the video plus the prompt that generated the video plus the prompt that generated the video plus the prompt that generated that video. So, for example, if the that video. So, for example, if the that video. So, for example, if the prompt will say prompt will say prompt will say the door slammed, right? It will look the door slammed, right? It will look the door slammed, right? It will look for a door being slammed and for a door being slammed and for a door being slammed and and and will will will match the sound and and will will will match the sound and and will will will match the sound at that same frame. I I Did I answer your question?

Did I answer your question? Did I answer your question? >> How does the model recognize sound? >> How does the model recognize sound? >> How does the model recognize sound? >> Uh so, it's both by using um um >> Uh so, it's both by using um um >> Uh so, it's both by using um um uh uh uh Atmos, but Atmos, but Atmos, but and also to to correlate the and also to to correlate the and also to to correlate the So, for example, when it's looking at So, for example, when it's looking at So, for example, when it's looking at the frames, right? It's making sure that the frames, right? It's making sure that the frames, right? It's making sure that that that that for example, the door being slammed at for example, the door being slammed at for example, the door being slammed at frame six, frame six has a a specific frame six, frame six has a a specific frame six, frame six has a a specific timestamp. So, it's looking for that timestamp. So, it's looking for that timestamp. So, it's looking for that that that that spike in the sound at that timestamp. It spike in the sound at that timestamp. It spike in the sound at that timestamp. It doesn't know that it is that sound, but doesn't know that it is that sound, but doesn't know that it is that sound, but it's looking for a specific spike of it's looking for a specific spike of it's looking for a specific spike of sound at at that timestamp. sound at at that timestamp. sound at at that timestamp. >> What about lip syncing? >> What about lip syncing? >> What about lip syncing? >> Lip syncing is an unsolved problem yet. >> Lip syncing is an unsolved problem yet. >> Lip syncing is an unsolved problem yet. >> [laughter] >> [laughter] >> [laughter] >> One question. >> One question. >> One question. >> We're trying though. >> We're trying though. >> We're trying though. Yeah. Yeah, I said the Yeah, I said the the question was what about lip syncing? the question was what about lip syncing? the question was what about lip syncing? >> Oh, Well, wasn't me, but >> Oh, Well, wasn't me, but >> Oh, Well, wasn't me, but I guess the lip syncing answer would be I guess the lip syncing answer would be I guess the lip syncing answer would be interesting before I ask my question.

interesting before I ask my question. interesting before I ask my question. >> Uh yeah, um as I said, it is an unsolved >> Uh yeah, um as I said, it is an unsolved >> Uh yeah, um as I said, it is an unsolved problem still. We're still problem still. We're still problem still. We're still working through it. working through it. working through it. Especially for us, Especially for us, Especially for us, some some of the characters that we're some some of the characters that we're some some of the characters that we're trying to do a talking head are humans, trying to do a talking head are humans, trying to do a talking head are humans, right? Which, you know, right? Which, you know, right? Which, you know, we can look at the different techniques we can look at the different techniques we can look at the different techniques to trying to identify the the lips, but to trying to identify the the lips, but to trying to identify the the lips, but some of them are just talking, you know, some of them are just talking, you know, some of them are just talking, you know, talking animations talking animations talking animations that have no real correlation between, that have no real correlation between, that have no real correlation between, you know, the the movement of of of the you know, the the movement of of of the you know, the the movement of of of the mouth and and and speech. So, mouth and and and speech. So, mouth and and and speech. So, unfortunately, I don't have a solution unfortunately, I don't have a solution unfortunately, I don't have a solution for that yet. for that yet. for that yet. >> So, I'm >> So, I'm >> So, I'm So, I'm curious about, for example, if So, I'm curious about, for example, if So, I'm curious about, for example, if you wanted to further enrich the data you wanted to further enrich the data you wanted to further enrich the data set with human evaluation. set with human evaluation. set with human evaluation. >> Yes. >> Yes. >> Yes. >> Um >> Um >> Um the question of of taste in what is the question of of taste in what is the question of of taste in what is good, because I think there is a big good, because I think there is a big good, because I think there is a big question mark about is that is that question mark about is that is that question mark about is that is that going to remain the domain of humans? going to remain the domain of humans? going to remain the domain of humans? But, I've also seen people say that, But, I've also seen people say that, But, I've also seen people say that, well, most humans well, most humans well, most humans they have terrible taste anyway in in they have terrible taste anyway in in they have terrible taste anyway in in videos and games and books. videos and games and books. videos and games and books. >> Fair. >> Fair. >> Fair. >> Um so, how would you construct and align >> Um so, how would you construct and align >> Um so, how would you construct and align sort of like any human sort of like any human sort of like any human judges?

judges? judges? >> Yeah. So, >> Yeah. So, >> Yeah. So, So, So, So, this is actually solved at So, So, So, this is actually solved at So, So, So, this is actually solved at first at the the the Judge Judy part, first at the the the Judge Judy part, first at the the the Judge Judy part, where where where every report it will generate a human every report it will generate a human every report it will generate a human can go and annotate it. And we actually can go and annotate it. And we actually can go and annotate it. And we actually we we we we do that. We we will periodically have we do that. We we will periodically have we do that. We we will periodically have sessions where everyone spends 10 to 15 sessions where everyone spends 10 to 15 sessions where everyone spends 10 to 15 minutes just minutes just minutes just annotating videos. And and that usually annotating videos. And and that usually annotating videos. And and that usually happens on happens on happens on multiple axes. I won't ask everyone to multiple axes. I won't ask everyone to multiple axes. I won't ask everyone to annotate the same video on on 10 annotate the same video on on 10 annotate the same video on on 10 different things. It'll It'll It'll be different things. It'll It'll It'll be different things. It'll It'll It'll be random. And I use random. And I use random. And I use the data to to calibrate the the AI the data to to calibrate the the AI the data to to calibrate the the AI judges. And and the results from that is judges. And and the results from that is judges. And and the results from that is actually being served as as a data set actually being served as as a data set actually being served as as a data set for for training for the next version of for for training for the next version of for for training for the next version of of that model. So, it's a process that of that model. So, it's a process that of that model. So, it's a process that does take a little bit of time and and does take a little bit of time and and does take a little bit of time and and hopefully and and it does evolve over hopefully and and it does evolve over hopefully and and it does evolve over time. Uh but it's not immediate you time. Uh but it's not immediate you time. Uh but it's not immediate you know, because also know, because also know, because also taste is very subjective and and things taste is very subjective and and things taste is very subjective and and things that are great for me, you know, some that are great for me, you know, some that are great for me, you know, some that I think are fantastics some people that I think are fantastics some people that I think are fantastics some people that come and say that come and say that come and say are you sure they're great because you are you sure they're great because you are you sure they're great because you know know know So, yeah, it's it's it's a process and So, yeah, it's it's it's a process and So, yeah, it's it's it's a process and and and and I use the human and and and I use the human and and and I use the human feedback to calibrate the models all the feedback to calibrate the models all the feedback to calibrate the models all the time.

time. time. >> Um how did you land on the Quan small >> Um how did you land on the Quan small >> Um how did you land on the Quan small VLM? Did you try any others? VLM? Did you try any others? VLM? Did you try any others? >> Uh I did. So, so the intent I had was to >> Uh I did. So, so the intent I had was to >> Uh I did. So, so the intent I had was to A you know, find a small enough model. A you know, find a small enough model. A you know, find a small enough model. The reason I went with Quan is because The reason I went with Quan is because The reason I went with Quan is because we also had a very good experience we also had a very good experience we also had a very good experience with post training Quan with post training Quan with post training Quan on other use cases. So, it on other use cases. So, it on other use cases. So, it I mean yes, I could have I did try a few I mean yes, I could have I did try a few I mean yes, I could have I did try a few others, but it just you know, everything others, but it just you know, everything others, but it just you know, everything was just there and it was good enough. >> So, my question is about scale. So, >> So, my question is about scale. So, obviously Character AI produces obviously Character AI produces obviously Character AI produces thousands, millions, bajillion videos. thousands, millions, bajillion videos. thousands, millions, bajillion videos. What scale does this become reasonable What scale does this become reasonable What scale does this become reasonable for my domain that is not Character AI? for my domain that is not Character AI? for my domain that is not Character AI? So, my domain has hundreds, maybe a So, my domain has hundreds, maybe a So, my domain has hundreds, maybe a thousand videos. thousand videos. thousand videos. >> Mhm. >> Mhm. >> Mhm. Sure. Sure. Sure. So, So, So, if you're happy with with the cohort of if you're happy with with the cohort of if you're happy with with the cohort of experts and and and and it and you don't experts and and and and it and you don't experts and and and and it and you don't need right so need right so need right so I'll I'll I'll rephrase that. rephrase that. rephrase that. The scale is both for speed, right? As The scale is both for speed, right? As The scale is both for speed, right? As well as capacity because I can serve well as capacity because I can serve well as capacity because I can serve this model as one instance on one GPU or this model as one instance on one GPU or this model as one instance on one GPU or I can serve it as you know, I can serve it as you know, I can serve it as you know, 100 instances, right? So, so that's that 100 instances, right? So, so that's that 100 instances, right? So, so that's that determines my scale. The reason I chose determines my scale. The reason I chose determines my scale. The reason I chose to go towards the model is because I to go towards the model is because I to go towards the model is because I wanted to to speed up the creation wanted to to speed up the creation wanted to to speed up the creation process, right?

process, right? process, right? It would work would have worked just as It would work would have worked just as It would work would have worked just as well if I didn't have this particular well if I didn't have this particular well if I didn't have this particular model, I would have used like the cohort model, I would have used like the cohort model, I would have used like the cohort of experts, right? From of experts, right? From of experts, right? From metrics metrics metrics that are available both on CPU and GPU that are available both on CPU and GPU that are available both on CPU and GPU as well as the as well as the as well as the frontier models. Right? So it's it was a frontier models. Right? So it's it was a frontier models. Right? So it's it was a balance balance balance as as as you know, A, how long did it take me to you know, A, how long did it take me to you know, A, how long did it take me to to to train this model and to curate the to to train this model and to curate the to to train this model and to curate the the data set and and get it to a working the data set and and get it to a working the data set and and get it to a working set, right? set, right? set, right? And and how much does it cost to serve And and how much does it cost to serve And and how much does it cost to serve it it it versus how much it would have cost me to versus how much it would have cost me to versus how much it would have cost me to do this A slower. Now potentially it is do this A slower. Now potentially it is do this A slower. Now potentially it is better, right? I mean like I assume that better, right? I mean like I assume that better, right? I mean like I assume that if you're going to use if you're going to use if you're going to use Fable which came back today, right? It Fable which came back today, right? It Fable which came back today, right? It will probably give you a a better will probably give you a a better will probably give you a a better result, but at what cost, right? result, but at what cost, right? result, but at what cost, right? If you do it for one or two, that's If you do it for one or two, that's If you do it for one or two, that's probably fine. If if you do it for probably fine. If if you do it for probably fine. If if you do it for thousands or tens of thousands per day, thousands or tens of thousands per day, thousands or tens of thousands per day, it adds up. it adds up. it adds up. So it's it's it's a matter of your So it's it's it's a matter of your So it's it's it's a matter of your your your unit economics. your your unit economics. your your unit economics. >> Cool. Over here.

>> Cool. Over here. >> Cool. Over here. On your right, there you go. On your right, there you go. On your right, there you go. Last question. Last question. Last question. >> It's very bright. I'm sorry. >> It's very bright. I'm sorry. >> It's very bright. I'm sorry. >> No worries. No worries. Um >> No worries. No worries. Um >> No worries. No worries. Um My question is I looked a bit at the My question is I looked a bit at the My question is I looked a bit at the repo. You guys don't export all tail repo. You guys don't export all tail repo. You guys don't export all tail traces of the LMS judges yet. traces of the LMS judges yet. traces of the LMS judges yet. >> Correct. >> Correct. >> Correct. >> Is that something are you open to that >> Is that something are you open to that >> Is that something are you open to that so you can connect to other so you can connect to other so you can connect to other platforms? platforms? platforms? >> Sure. So so the >> Sure. So so the >> Sure. So so the the the repo itself the the repo itself the the repo itself it's a harness and you can connect any it's a harness and you can connect any it's a harness and you can connect any any any any agents or any LLMs agents or any LLMs agents or any LLMs you want. We actually have an internal you want. We actually have an internal you want. We actually have an internal version of this which is running it as a version of this which is running it as a version of this which is running it as a service, right? We we have an agentic service, right? We we have an agentic service, right? We we have an agentic harness on top of it that has all the harness on top of it that has all the harness on top of it that has all the the the the the the the the metrics the metrics the metrics we we care about, but I I I I do accept we we care about, but I I I I do accept we we care about, but I I I I do accept your feature request and I'll be adding your feature request and I'll be adding your feature request and I'll be adding hotel hotel hotel telemetry to the the harness. telemetry to the the harness. telemetry to the the harness. >> Awesome. Thank you very much. A warm >> Awesome. Thank you very much. A warm >> Awesome. Thank you very much. A warm welcome or round of applause for Mayor. welcome or round of applause for Mayor. welcome or round of applause for Mayor. Thank you. Thank you. Thank you. >> Thank you all. >> Thank you all. >> Thank you all. >> Thanks.