WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=mOf-PP4mVjA NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/mOf-PP4mVjA.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:01.309 --> 00:00:03.309 [music] s2 00:00:12.480 --> 00:00:16.840 Thanks so much for having me and inviting me to to be a speaker at the the welfare. s3 00:00:16.840 --> 00:00:20.040 You know, I I attended last year and was so impressed by the quality of presenters. s4 00:00:20.040 --> 00:00:23.200 So so glad to to have a chance to be here and to present. s5 00:00:23.200 --> 00:00:26.720 So the title of my talk is, you know, video has no memory, right? s6 00:00:26.720 --> 00:00:32.480 And you know, this might sound strange because video is already in like a preservation of the past, right? s7 00:00:32.480 --> 00:00:44.560 If you think about like you have footage, you preserve recording training data, incident, creative work, history, etc. But actually most of the video AI systems these days do not have memory in the system sense. s8 00:00:44.560 --> 00:00:51.800 So actually for this talk, I will try to answer the question like what could it take to build a memory layer for video intelligence? s9 00:00:51.800 --> 00:00:55.920 To start, I want to be clear about what makes video different from other data type, right? s10 00:00:55.920 --> 00:01:01.040 So this is the first mental model that I want to highlight, which is that video is not a stack of frames. s11 00:01:01.040 --> 00:01:09.280 Um so in in a many of my conversations with like developers, you know, who are using our product, a lot of them still treat video as like a stack of images, s12 00:01:09.280 --> 00:01:16.800 maybe a transcript being attached or you know, like you know, but essentially like like a frame level, right? s13 00:01:16.800 --> 00:01:23.920 And that is a useful approximation for some tasks, but it throw away the thing that makes video very unique, which is continuity, right? s14 00:01:23.920 --> 00:01:29.160 So meaning in video derives from space, time, modalities, and sequence. s15 00:01:29.160 --> 00:01:33.320 So a better mental model for video is a special temporal volume. s16 00:01:33.320 --> 00:01:43.560 So what I mean that inside that volume, you have visual information, speech, sound, motion, OCR, camera changes, scene transition, metadata, and time, right? s17 00:01:43.560 --> 00:01:51.080 So the hard part here is really well, how can you preserve any relationship across this volume so that later an application can traverse it? s18 00:01:51.080 --> 00:02:01.880 And then you know, especially at the enterprise scale, you know, across like industry like entertainment, sport, you know, um, short-form content, then you're sitting on petabytes of of footage, right? s19 00:02:01.880 --> 00:02:03.560 So, finding moment is already hard. s20 00:02:03.560 --> 00:02:07.400 So, how can you preset meaning across millions of moments in the deeper platform? s21 00:02:07.400 --> 00:02:10.720 So, I work at Twelve Labs, which is a uh series B uh startup. s22 00:02:10.720 --> 00:02:15.800 Um, we build foundation models that understand, you know, video the way that human do. s23 00:02:15.800 --> 00:02:24.600 Um, and the way we talk about our positioning is like the existing uh stack of of dealing with video is not equipped to do that, right? s24 00:02:24.600 --> 00:02:27.400 Obviously, language model are very powerful. s25 00:02:27.400 --> 00:02:29.480 They are good reasoning interfaces. s26 00:02:29.480 --> 00:02:32.120 They are increasingly multimodal, as well. s27 00:02:32.120 --> 00:02:38.480 But, the supporting stack, right, around that is very um I'd say limited, and that create three problem. s28 00:02:38.480 --> 00:02:40.880 Number one, is wrong context, right? s29 00:02:40.880 --> 00:02:44.120 So, video is not naturally a sequence of text token. s30 00:02:44.120 --> 00:02:55.440 If we force it into that sequence by sampling frames, by extracting a transcript, uh by dumping everything into a prompt, you lose the spatiotemporal relationships, right, that actually define the event. s31 00:02:55.440 --> 00:02:57.000 Second, uh wrong memory. s32 00:02:57.000 --> 00:03:04.320 So, if you think about text system memory here is often mean retrieval and generation, vector search, or probably like larger context window. s33 00:03:04.320 --> 00:03:07.239 Uh those are very useful, but video memory has a different requirement. s34 00:03:07.239 --> 00:03:14.080 It needs to link today's scene for something that happened in another file, another episode, another camera angle, another season, another year. s35 00:03:14.080 --> 00:03:18.640 So, it actually need durable continuity, And the last part here is wrong reasoning. s36 00:03:18.640 --> 00:03:25.040 As I said, you know, text-first system cannot reason over, you know, um, natively over motion, causality, all that. s37 00:03:25.040 --> 00:03:41.600 So, uh you know, they do not automatically build like a persistent structure uh on, you know, who appear, what happened, what changes, etc. And so, my argument is that video intelligence need a memory layer that decide what to preserve, how to connect it, and how to retrieve later. s38 00:03:41.600 --> 00:03:48.120 So, um, I want to kind of ground it into the properties of video, right, to to make it even clearer. s39 00:03:48.120 --> 00:03:50.480 There's five challenges in dealing with video. s40 00:03:50.480 --> 00:03:52.280 Number one is temporal, right? s41 00:03:52.280 --> 00:03:55.360 So, depends on before and after. s42 00:03:55.360 --> 00:03:58.160 So, frame by itself can can be misleading, right? s43 00:03:58.160 --> 00:04:03.800 The same expression, product shot, physical action can mean different things depending on the sequence around it, right? s44 00:04:03.800 --> 00:04:07.720 Second is that video is obviously multimodal as I I explained already. s45 00:04:07.720 --> 00:04:13.040 Um you know, a transcript alone may miss, you know, the logo, a frame alone may miss the spoken claim. s46 00:04:13.040 --> 00:04:15.360 Video Video is also very dense, right? s47 00:04:15.360 --> 00:04:21.320 So, a few minutes can contain dozens of shots, people, objects, action, location claims. s48 00:04:21.320 --> 00:04:25.000 The useful signal is uneven across the distribution on on the frame. s49 00:04:25.000 --> 00:04:28.160 Some seconds are decisive, others are noisy. s50 00:04:28.160 --> 00:04:31.680 Fourth is that uh video is also ambiguous, right? s51 00:04:31.680 --> 00:04:41.280 Um people reappear under different lighting and angles, brands are partially visible, location are implied, concepts emerge over time rather than being named in a single moment. s52 00:04:41.280 --> 00:04:52.560 And lastly, it is uh expensive uh because in a lot of uh big enterprise and in complex workflow, you need to, you know, point back to the source moment, like where it come from, right? s53 00:04:52.560 --> 00:04:56.400 So, these are the five properties explaining why video memory um is is very complex. s54 00:04:56.400 --> 00:05:00.880 You need to preserve temporal span, multimodal evidence, continuity, all of that. s55 00:05:00.880 --> 00:05:04.919 Uh this is a very simple uh stack of how we build things at Twelve Labs. s56 00:05:04.919 --> 00:05:10.640 Um at the bottom, we have these semantic chunks that capture, you know, meaningful temporal units. s57 00:05:10.640 --> 00:05:19.240 Above that is our uh multimodal embedding coder called Marengo, which essentially turn those spans into spatial-temporal relations. s58 00:05:19.240 --> 00:05:22.480 Um basically, vector embeddings that represent video content. s59 00:05:22.480 --> 00:05:32.040 And then we have a special special-temporal context store, which is where it preserve pre-usable structure, like moment, entities, metadata, all of that. s60 00:05:32.040 --> 00:05:39.080 Uh we also build our own uh VLM, video context-aware language model called Pegasus, that essentially serve as the the reasoning layer, right? s61 00:05:39.080 --> 00:05:45.000 That can preserve uh prepare over video content, so think about summaries, like um metadata, synthesis, comparison. s62 00:05:45.000 --> 00:05:51.480 And we expose our our models as API because, you know, we want to get developers to use them as infrastructure. s63 00:05:51.480 --> 00:05:56.480 Now, moving beyond like kind of the stack right here, I want to talk about the difference between search and memory, right? s64 00:05:56.480 --> 00:05:59.480 Very quickly speaking, uh search is obviously super important. s65 00:05:59.480 --> 00:06:03.080 It's how you recover relevant moments from large video library. s66 00:06:03.080 --> 00:06:04.400 But then it gives candidate. s67 00:06:04.400 --> 00:06:06.480 It actually not give you like any continuity. s68 00:06:06.480 --> 00:06:16.360 So, memory on the other hand is is all the um you know, uh the things that enable the system to answer different class of question as you see here on the right side of of my screen. s69 00:06:16.360 --> 00:06:19.280 So, these are not the single retrieval call, right? s70 00:06:19.280 --> 00:06:24.840 They require the system to reason entities, timeline, evidence across an entire corpus. s71 00:06:24.840 --> 00:06:33.080 Um and so like you can actually build product moving beyond from like show me something like this to, you know, tell me what this collection knows, right? s72 00:06:33.080 --> 00:06:38.440 And so that that might sound, you know, simple and subtle, but uh the the the the output is completely different. s73 00:06:38.440 --> 00:06:47.960 Like with search, you you get like an output like a time-bounded moment, but with memory, you actually return like structured knowledge, timeline, uh explanation, composable output. s74 00:06:47.960 --> 00:06:56.600 And that like, you know, is very important because we can now move um the the unit output from clip retrieval to corpus memory, right? s75 00:06:56.600 --> 00:06:59.720 Um there are two scaling dimensions shown here in the slide. s76 00:06:59.720 --> 00:07:01.080 The first is time scaling. s77 00:07:01.080 --> 00:07:07.040 So, a real video system should be able to reason over years of footage without reprocessing the whole archive every time, right? s78 00:07:07.040 --> 00:07:16.600 That means memory first retrieval, uh be reusable representation once and then support multi-hop timeline, episodic recall, follow-up question at lower latency and cost. s79 00:07:16.600 --> 00:07:19.480 And then the second uh dimension is in space, right? s80 00:07:19.480 --> 00:07:29.960 So, many real workflow actually um involve multiple perspective like different camera angles, uh you know, live stream, creator broadcasting content, body cams, stock cameras, uh event feed, right? s81 00:07:29.960 --> 00:07:35.640 So, how can you build a system that can fuse evidence across all the sources and then maintain current understanding, right? s82 00:07:35.640 --> 00:07:37.320 And so, that is the challenges here. s83 00:07:37.320 --> 00:07:43.120 How can you build a representation that let application traverse video across time and across sources? s84 00:07:43.120 --> 00:07:46.320 Um since this is um you know a track on on graph, right? s85 00:07:46.320 --> 00:07:52.560 So the the best mental model that I can come up with is to represent you know, video collection as a context graph. s86 00:07:52.560 --> 00:08:02.080 So a context graph is a durable, queryable representation that connects video moment, entities, appearances, relationship, time span, metadata, and corpus level context, right? s87 00:08:02.080 --> 00:08:05.840 So if you take a look here on on the screen all the way in the bottom, you got time bounded moment. s88 00:08:05.840 --> 00:08:08.600 These are like the the scene, the shot, right? s89 00:08:08.600 --> 00:08:10.200 Uh these are evidence unit. s90 00:08:10.200 --> 00:08:14.520 One one level up are the appearances, where and when each entity show up. s91 00:08:14.520 --> 00:08:16.040 And then you got the actual entity itself. s92 00:08:16.040 --> 00:08:20.280 So think about the people on the video, the brand, the places, the concept. s93 00:08:20.280 --> 00:08:28.000 Next you have relationship, uh co-occurrences to the same brand, sequences between different places, the causality, and timeline. s94 00:08:28.000 --> 00:08:30.640 And finally at the top you have corpus level context. s95 00:08:30.640 --> 00:08:35.320 What are the main themes, the patterns, the gap, the coverage that this video collection cover, right? s96 00:08:35.320 --> 00:08:39.039 Uh this matter because different question travels different part of the graph. s97 00:08:39.039 --> 00:08:43.360 If you ask a simple search question, then that might go directly into the moment. s98 00:08:43.360 --> 00:08:47.960 But like an entity workflow might start with a person and then it expand into appearances, right? s99 00:08:47.960 --> 00:08:55.800 And if you ask question like a storyline, like narrative storytelling of certain you know you know person, then it may follow relationship across time, right? s100 00:08:55.800 --> 00:09:03.600 Uh so the key idea here is that memory in the context of video understanding is a navigable structure over the entire video volume. s101 00:09:03.600 --> 00:09:08.839 From that concept, I come up with these five principles as a building you know a memory layer for video intelligence. s102 00:09:08.839 --> 00:09:11.920 Number one is to ingest once and reason many times. s103 00:09:11.920 --> 00:09:16.200 So um you don't want to like do every single query from scratch, right? s104 00:09:16.200 --> 00:09:25.480 You want to pay the cost up front, do one interpretation from the video content up up front, pay the cost, and then you move expensive understanding into injection. s105 00:09:25.480 --> 00:09:28.200 So this is same mental model uh database, right? s106 00:09:28.200 --> 00:09:34.880 You you're not repeatedly have to parse your entire source of data um for you know every application request. s107 00:09:34.880 --> 00:09:38.000 Um second principle is to store primitive, not just answer. s108 00:09:38.000 --> 00:09:42.920 So, uh, you know, moments, entities, appearances, as we talked about that, those are the the primitives, right? s109 00:09:42.920 --> 00:09:49.720 That allows you to, uh, do downstream workflow, like search, editing, um, you know, analytics, all of that. s110 00:09:49.720 --> 00:09:51.520 Third is to ground every claim. s111 00:09:51.520 --> 00:09:56.920 They like basically, if you ask a question, you need to cite back into where that scene happening in the video. s112 00:09:56.920 --> 00:10:02.800 So, uh, evidence, like, you know, should be grounded to a specific timestamp within the video, right? s113 00:10:02.800 --> 00:10:05.160 Uh, fourth is to let intent shape memory. s114 00:10:05.160 --> 00:10:09.000 Um, this is important because the same footage mean different thing in different workflow. s115 00:10:09.000 --> 00:10:13.520 We work across sports, uh, application, brand safety, compliance review, clear analytics. s116 00:10:13.520 --> 00:10:15.880 All of them require different primitives from the same video. s117 00:10:15.880 --> 00:10:18.280 So, the memory layer should be configurable, right? s118 00:10:18.280 --> 00:10:21.600 Developers should, uh, should be able to tell the system what matters. s119 00:10:21.600 --> 00:10:24.160 And lastly, uh, keep the layer composable. s120 00:10:24.160 --> 00:10:33.280 Um, so, basically, being API first, you know, um, it should provide, uh, the layers that allows those application, uh, on top of that to to serve it, uh, structured grounded metadata s121 00:10:33.280 --> 00:10:36.320 that can be plugged into any sort of application. s122 00:10:36.320 --> 00:10:45.120 Um, so, moving beyond these five principles, I want to talk about like, kind of the the harnesses around building a memory layer, right? s123 00:10:45.120 --> 00:10:49.680 There's a lot of talk these days about, um, you know, building the right harnesses for the context of language model. s124 00:10:49.680 --> 00:10:53.240 So, what does it look like for for video understanding model, right? s125 00:10:53.240 --> 00:10:56.000 Um, a model could produce a single answer. s126 00:10:56.000 --> 00:10:56.880 It is stateless. s127 00:10:56.880 --> 00:11:01.520 It start fresh each time, start fresh each time, and doesn't have any constraint. s128 00:11:01.520 --> 00:11:04.440 So, the output is largely based on what the model is set to produce. s129 00:11:04.440 --> 00:11:09.240 A video worker, on the other hand, operate inside a a very deterministic system. s130 00:11:09.240 --> 00:11:21.360 Understand what is available, uh, it can plan the task, could ship evidence, inspect, uh, the relevant moments, synthesize, validate, return output, and then the entire workflow can be evaluated, right? s131 00:11:21.360 --> 00:11:30.920 Um, so, so for video understanding, this is very important because the worker need to know what memory is is available, what evidence matters, and also like how deep to inspect, because that will depend, s132 00:11:30.920 --> 00:11:35.040 uh, determine how much cost to spend, what output constraints to satisfy, right? s133 00:11:35.040 --> 00:11:43.920 Um talking about harness engineering for for video understanding, um I come up with these like different capabilities for for like a video worker, right? s134 00:11:43.920 --> 00:11:47.200 Um number one is memory, I talked about that already. s135 00:11:47.200 --> 00:11:48.400 Number two is task planning. s136 00:11:48.400 --> 00:11:53.920 So, given given a query from from an user, uh you have to decide like what task to execute. s137 00:11:53.920 --> 00:11:59.800 Is it like search or is it like summarization or like, you know, multi multi-step reasoning? s138 00:11:59.800 --> 00:12:07.080 Uh third is retrieval, like every single system should be able to like select the right evidence from from your video corpus to read for a specific task. s139 00:12:07.080 --> 00:12:09.280 Uh expert tools, right? s140 00:12:09.280 --> 00:12:18.840 So, we work with customer where they require like, you know, zoom in, zoom out, uh comparing different uh uh you know, frames, enriching uh content with like additional metadata. s141 00:12:18.840 --> 00:12:23.720 So, building expert tools inside uh like a like a video worker uh is very important. s142 00:12:23.720 --> 00:12:29.200 Operating envelope, so these are like explicit limit on time, cost, depth, scope, autonomy. s143 00:12:29.200 --> 00:12:30.400 Uh and output contract. s144 00:12:30.400 --> 00:12:35.400 So, sometimes natural language is not, sometimes the patient needs structured data with references and time stamp. s145 00:12:35.400 --> 00:12:38.120 And of course, finally you have evaluation, right? s146 00:12:38.120 --> 00:12:41.360 Uh like, you know, did the retrieval find the right evidence? s147 00:12:41.360 --> 00:12:43.160 Did the synthesis preserve it, right? s148 00:12:43.160 --> 00:12:45.280 Did the worker stay within the budget? s149 00:12:45.280 --> 00:12:48.480 All right, so so that's a lot of like, you know, uh slide and and and talk. s150 00:12:48.480 --> 00:12:55.960 I want to quickly jump into some demos uh that I actually built using Two Labs uh you know, video agent product. s151 00:12:55.960 --> 00:12:57.960 So, the uh there'll be three demos. s152 00:12:57.960 --> 00:13:01.880 Um the the video agent product that we've been building is called Jockey. s153 00:13:01.880 --> 00:13:05.240 So, this first example here is for sport understanding. s154 00:13:05.240 --> 00:13:09.160 Uh you know, obviously everyone is super excited about the World Cup that happening right now. s155 00:13:09.160 --> 00:13:16.200 So, what I did is I ingest um 67 videos from the 2022 uh World Cup in Qatar. s156 00:13:16.200 --> 00:13:21.760 And you can see here I asked it to find the near misses, uh the shot that almost become goal but did not. s157 00:13:21.760 --> 00:13:27.480 For each, explain why it was not a goal, uh but do not include actual goals, right? s158 00:13:27.480 --> 00:13:31.880 So, these are the the the top output that it returned. s159 00:13:35.200 --> 00:13:37.840 So, that that's hitting the woodwork. s160 00:13:37.840 --> 00:13:40.080 This is saved from the goalkeeper. s161 00:13:40.080 --> 00:13:44.880 I don't know if the sound is up, but like I'm playing the the video, by the way. s162 00:13:45.120 --> 00:13:52.160 Um Right, this is another save from the goalkeeper. s163 00:13:52.600 --> 00:13:57.400 It even catch like, you know, the upside uh from one of the goals. s164 00:14:01.120 --> 00:14:03.560 And then I ask question, okay, when is a goal? s165 00:14:03.560 --> 00:14:05.920 Uh Find the most dramatic actual goals. s166 00:14:05.920 --> 00:14:08.120 Show the build-up play and the finish. s167 00:14:08.120 --> 00:14:11.720 For each goal, describe the sequence, right? s168 00:14:11.880 --> 00:14:19.680 So, if you know this one, this is the um the first goal of the World Cup final like 4 years ago. s169 00:14:19.680 --> 00:14:30.120 And it it actually like returned like, you know, you you don't understand who who are the passer like it was Alvarez passing to Mac Allister and passing to Di Maria to score the goal. s170 00:14:30.120 --> 00:14:32.560 Um Take a look at this one from Richarlison. s171 00:14:32.560 --> 00:14:39.320 This is goal of the tournament uh from Brazil again uh South Korea, I believe, right? s172 00:14:40.560 --> 00:14:44.360 And it returned like, you know, um an outrageous skill in a build-up, right? s173 00:14:44.360 --> 00:14:47.839 The name the player who did the return pass. s174 00:14:48.960 --> 00:14:50.320 You can even do player tracking. s175 00:14:50.320 --> 00:14:56.600 So, I asked it to track Lionel Messi across this entire corpus including the shot where he's won the man of the match figure on the screen. s176 00:14:56.600 --> 00:14:58.920 Describe the camera framing, right? s177 00:14:58.920 --> 00:15:06.480 Uh so, this is uh hilarious of all the important moment in the game and this is a scene where Messi dribble past a sliding defender. s178 00:15:06.480 --> 00:15:08.800 You can see here. s179 00:15:09.080 --> 00:15:15.200 It pick up the scene where he scored the first goal against Australia in the round of 16, I believe. s180 00:15:15.640 --> 00:15:16.839 Right? s181 00:15:16.839 --> 00:15:21.320 This is another scene where he scored the third goal in the final. s182 00:15:25.440 --> 00:15:29.680 Yeah, so that one example on spot spot understanding. s183 00:15:29.680 --> 00:15:37.240 But then you can obviously build more interesting and more like real practical application of which security is one that we encounter a lot. s184 00:15:37.240 --> 00:15:42.720 So on the on this example I ingest it you know publicly available camera footage. s185 00:15:42.720 --> 00:15:55.080 For context this are the clip you have traffic jams suburban you know urban area and given this footage where I asked Jockey to count and classify every vehicle in the intersection s186 00:15:55.080 --> 00:15:57.880 break it down by type plus pedestrian. s187 00:15:57.880 --> 00:16:02.640 And it returned the number of vehicles and the big foot traffic as well. s188 00:16:02.640 --> 00:16:06.600 You can detect safety events right? s189 00:16:07.880 --> 00:16:12.840 So you see that a red SUV turn and almost gets struck. s190 00:16:13.400 --> 00:16:15.920 Another scene here. s191 00:16:16.080 --> 00:16:19.120 Turn left into an upcoming car. s192 00:16:22.800 --> 00:16:25.880 Yeah, so that a clear straight line entry. s193 00:16:25.880 --> 00:16:29.480 It works well in you know different scenario. s194 00:16:29.480 --> 00:16:34.640 This scene is very crowded area in Bangkok. s195 00:16:36.040 --> 00:16:41.120 It asked it I also asked it to work on the you know the rain right? s196 00:16:45.400 --> 00:16:51.320 So this is another scene where it understanding the um rainy condition. s197 00:16:55.440 --> 00:17:00.480 It identify the basis intersection vehicle window. s198 00:17:03.480 --> 00:17:07.760 So yeah, those are some example for for camera security surveillance footage. s199 00:17:07.760 --> 00:17:16.400 Finally advertising so um you probably seen this Adidas clip in all commercial leading up to the World Cup. s200 00:17:16.400 --> 00:17:22.800 Recently is a 5 minutes Adidas footage and I asked it to classify all the point where you can put an ad on. s201 00:17:22.800 --> 00:17:27.160 So, it find the review, the hard cut, the impact, energy big rate. s202 00:17:27.160 --> 00:17:31.240 It find the scene where a certain player appear on the screen. s203 00:17:32.520 --> 00:17:37.120 It identify like, you know, high impact action like this. s204 00:17:37.720 --> 00:17:41.920 Condition, um, the hard cut to Nike football underlies. s205 00:17:41.920 --> 00:17:44.520 And of course, it it point into the logo. s206 00:17:44.520 --> 00:17:48.840 Um, of Adidas. s207 00:17:48.840 --> 00:17:59.040 So, you know, uh, from from perspective of an advertiser, these are very important moments because they can, you know, find the scene with the slow motion hero. s208 00:18:01.800 --> 00:18:04.360 Hard cut on a beat. s209 00:18:06.440 --> 00:18:08.760 Or pick action. s210 00:18:10.280 --> 00:18:15.240 In which they can advertise their brand content against this footage, right? s211 00:18:15.240 --> 00:18:22.200 Um, yeah, so those are three sample demo application, um, that I want to highlight, uh, I've using two labs. s212 00:18:22.200 --> 00:18:29.280 Um, and again, um, you know, what can you build with with this sort of video memory layer based on the case example? s213 00:18:29.280 --> 00:18:33.040 These are the categories of of application that I believe developers can build. s214 00:18:33.040 --> 00:18:34.800 You can discover things. s215 00:18:34.800 --> 00:18:36.880 You can be reasoning experience. s216 00:18:36.880 --> 00:18:50.880 You can organize your content across different video library and it can be action workflow, assemble uh, different scene together, do compliance review, data operation, etc. The same framework apply for different verticals s217 00:18:50.880 --> 00:18:58.240 in media and entertainment spot, segmentation, highlight generation, in commercial security, evidence review, contextual analysis. s218 00:18:58.240 --> 00:19:03.720 In advertising, uh, brand safety, uh, creative intelligence, right? s219 00:19:03.960 --> 00:19:07.520 And, uh, yeah, so this is our product, uh, that I coming up. s220 00:19:07.520 --> 00:19:12.840 Um, one quick highlight is that we we try to as a video cognition infrastructure. s221 00:19:12.840 --> 00:19:16.400 So we have a knowledge style that basically become the video memory layer. s222 00:19:16.400 --> 00:19:20.680 Web configurable injection that let video shape what the system can can extract. s223 00:19:20.680 --> 00:19:26.880 Corpus digest so that you can understand in what is in the library and the resolution as you search responses API. s224 00:19:26.880 --> 00:19:29.840 So the thing we want to highlight here is it's not an application layer. s225 00:19:29.840 --> 00:19:32.720 It's not an editing platform, not a compliance product. s226 00:19:32.720 --> 00:19:39.520 It's the cognition infrastructure with the layer and the harnesses that enable like those product being to become available. s227 00:19:39.520 --> 00:19:46.600 Um and if you found the content of this talk interesting, definitely recommend you to scan this QR code. s228 00:19:46.600 --> 00:19:48.920 The product is currently in private beta right now. s229 00:19:48.920 --> 00:20:01.760 If you bring any sort of workflow that touch video content especially around content assembly, content organization, you know, think about media archive, content creator, YouTube, TikTok, sport analysis, media workflow, s230 00:20:01.760 --> 00:20:06.720 definitely add the scan this QR code and register for the interest or come talk to me after the talk. s231 00:20:06.720 --> 00:20:07.960 So that should be my time. s232 00:20:07.960 --> 00:20:10.160 Thanks a lot.