# Jawn Lim - AI That Powers Experiences - TOKEN2049 Singapore 2025 - Channel: [TOKEN2049](https://streameth.org/token2049) - Date: 2025-11-07 - Duration: 16:49 - Topics: TOKEN2049, Web3 event, Singapore event, Byteplus, crypto and AI, AI experiences, generative AI - Watch: https://streameth.org/watch/yt-_c8aaFgZfFA - YouTube: https://www.youtube.com/watch?v=_c8aaFgZfFA ## Description Learn more about TOKEN2049: https://token2049.com Follow us on X: https://x.com/token2049 Keynote: AI That Powers Experiences Speaker: Jawn Lim, Regional Head of Solution Architect, GenAI @ BytePlus Stage: TON Stage #token2049 #byteplus #genai #artificialintelligence #web3 Recorded on 1 October 2025 ## Transcript Maybe most of you guys haven't heard about bite plus before. So what am I going to do here is that I just wanted to show you guys that bite blast essentially we are actually a company of bite dance right I think most of you guys here heard of tik tok you know capcotina or maybe things like mun mobile legend so in the past we have been doing a lot of things you know B2C tik toking right but because we actually know that customers they want to actually build some infrastructure so that's the reason five years ago biplus has been established And we do a lot of things not just you know live streaming in Tik Tok but we also do a lot in terms of AI which is one of the most core topic that I want to talk about today and in terms of the enterprise growth imperative you know most of the customers be a small scale B enterprise MNC's the key three things is always about the speed right the scalability and also the sustainability right so these are always the few things that they want to think of how to balance how to make sure that if you have any legacy items that you want to integrate to AI how would you be able to do that right so these are always of concern and let's say if you want to scale up you adopted the cloud journey probably from hybrid to multicloud journey how would you be able to continuous save the cost the total cost of ownership to bring it down with the best you know quality now this is the whole portfolio of biplas AI native cloud platform it's really a lot of modules over here but Today we're just going to focus on a few things right essentially it's going to be the one at the middle models and also the AI platform well over here you see that at the very bottom layer we have this thing that's called AI native infrastructure so what I meant by that is that you know if customers they want to build their own GPU whether is it a form of deep training or inferencing we do have the GPU service to provide to the customers right however you know most of the time 80% of customers is they wouldn't want to focus on building their own GPU servers because it is so you know you have to spend a lot of high capac so rather you get let us manage the GPU server for you and then you just call the API or tokens which is the models seed seat dream seat dance seat speech and also all the open source model let me bring you to the next slide to show you um bite dance seat model right so bite dance we have our own proprietary model is developed by our bike dance AI team. So this is called seat. So everything that you see with seat dream, seat dance, seat 1.6 is going to be the bike dance seat model. Right? In China we call it toba but then in international we actually kind of translated it to bike dance seat. So the very first thing is called seatream 4.0. So I'm not too sure anyone heard of this before. Sipdream 4.0 is is was released about three weeks ago. It is an image generation. So you can really do a lot of things. For instance, text to image, image to image and then later on you can actually use this image to put it into seat dance 1.0 which is a video generation. Right? So some customers they want to actually have some you know 5 seconds 10 seconds short clip of these particular images. So this is actually a combination with the seat dream 4.0 and seat dance 1.0. You can actually create a short drama as well. Now the third one is called Omih human. I think Omih human is really something that you know many of the customers always ask can I actually clone the Avada of me myself and then probably you can clone the CEO to do some opening message. So this is what we're going to actually see a short video later on on the omni and the last thing is the proprietary seat 1.6 model. So seat 1.6 ICS essentially this is a multimodel with reasoning capabilities and also VRM and OCR capabilities. I'll show you guys more of the things later on. Now before I start this um this short video uh just to give you some context right so Dream 4.0 is the image generation solutions and how we actually built this video is to combine SID dream and seed dance which is image and video. Let's uh sit back and then watch this for a quick one. [Music] Hello. [Music] Heat. What? Should I call you? I want to hear your voice. [Music] Right. Amazing. Right. I Wow. Feels as though I'm actually watching a Cinema M in theater. Right. So really very good the sound system the beat and the drum and bass is really good right so this is actually the sip dream 4.0 and we we're proud to say that you know we surpass most of the image geni providers in the market right now it was actually three rates ago I think 10 of September we actually released the newest version is 4.0 zero. Right? So, you can actually do this with a lot of things. You can generate like very uh studio grade uh visuals within 3 seconds and you can also go up to all the way to 4K resolutions. Before we were actually using SIP drinking 3.0, but now that we have 4.0, it can actually do way more than just the image generations. You can even do editing. For example, let's say if you actually have um if you are the property agent, you want to sell this house, the room, you know, showcase to your cl your clients. So, how would you be able to do that? So, you could put things like, you know, you can upload your own image and give a prompt saying that, hey, hey, John, I want to turn on the light to brighten up the light living room. However, I want to make sure that the the the room outside is evening, right? So, two key things. One is brighten up the room and second thing is keeping outside evening. So what it could do is that the model itself is able to understand clearly certainly the prompt is very important right so you can see that everything else including the sofa the tables the light the door is still the is still the same it doesn't really change much because the model is able to understand it so this is what you know seed dream 4.0 zero can do is able to understand is able to have this reasoning capabilities to just stick to what the prom is then obviously you know when I talk about you know the sip drain 4.0 zero because it is still a steel image image generation. So some of the media industries influencer you know KL what they did here is that they actually use SDR to generate image and then they convert this image to the video by using SIDS 1.0 pro. So you could actually generate 10 12 seconds video right a steel image moving around everything is going to be the prom and effectively you can actually stitch this video to make it a short drama right for example you know nowadays people use Tik Tok very often and just for for your information I think two months ago we have gotten some of the data the total DAU of Tik Tok user is about 1 billion he has crossed one billion DAU in Tik Tok itself whether is it live streaming or just you know you watching or scrolling in the in the in the live streaming or the Tik Tok. So, and one thing here is that you wouldn't be able to know whether this is real or not, right? In fact, it looks so real because of the resolutions. It go up to 4K high definitions. So, that's what we can do the combination of sipd dream and also seed dance. Now, the next thing that's more interesting is what we call the Omihuman. Uh two days ago, we released the newest version. is called Omihuman 1.5. Essentially, this is to really clone yourself as an Abara and then you can actually sing song, dance, you know, you can really do a lot of things. You speak multiple languages, right? For example, if I want to see but I think he can talk. So, you can see something like this. >> Omnihuman, >> a photo, a voice brought to life, smooth motion, vivid scenes, a magic that makes every story move effortlessly real. And then you can also let Labubu I'm not too sure who's Labu fan here but you can actually let Labubu sing and the lip sync is very accurate [Music] or even dancing. Okay, you might be wondering why the second pretty lady is not moving because we always leave the good thing at the last. So, anyone knows dance monkey? Anyone? Yeah. With Omi human effectively, you can really sing very well any song. For example, Take your handle. [Music] >> Now you can see that all of these are being generated based on the cloning of the AA. And what we do here is that we clone it and you give any text any you know uh song then you will be able to actually sing right. So this is what we actually been doing for omihuman and essentially you can also push this to any live streaming platform right. So this omi human now the next model I want to talk about is you know we have seen through seed dream which is image we have seen through the seat dance the video we have also seen through the omihuman the last one is called a seat 1.6 model. So bite dance we actually developed the seat 1.6 pretty much like you know deep seat R1 reasoning model it's a multimodel but it actually has this visual understanding right for instance if you want to do OCR on the invoices PDF or you want to do no VRM to be able to analyze a video. So what you could do here is that instead of you sitting through the whole watching the whole one or two minutes you know clip you could actually upload the whole video to the model and the model is able to understand and analyze and how do we do that is that we actually take every snapshot every frame from the video and we start to analyze it right so this is a multi model which encompasses the reasoning as well as a VRM model right something like this you can you can actually see there really a lot of use cases you can use when it comes to a VRM and reasoning model right can use it for e-commerce to see to analyze which of the brand is this products you can actually use it for online education and to be honest these days you know there all these the the students in the universities they're using the chat GPT DC to do their whole FIP and they're very concerned about knowing being a lecturer how are you be able to measure because output most likely is going to be very much the same thing when it when it comes to you know using the ROM to generate some of this content Right. Um the last two things that I want to share is really revolving between agentic tasks. I think everyone heard about agentic task before. So uh you know in the market there's really very good friendly competitors like n defi comfy ui right madness ai. So we have the corporate enterprise version of the agentic task workflow which is called high agent right. So by using high agent you could actually create multiple agents from many different use cases. So one of the use case that you know we help in Thailand is that they want to it's a healthcare industry they want to automate the whole booking process of the appointment. So they actually go through the chatbot the patient will ask hey I'm actually having a flu. So Gada will be actually going through the own the knowledge base to tell exactly where we should navigate you to which hospital which clinicates and then eventually we'll recommend the specialist accordingly for example the the ENT departments and then at the end of a day we will also you know check against the calendar of the doctor and if this time frame the doctor is available we will just notify via email or SMS and finally if the doctor approve it the whole booking process you know will be automated. And all of these is just one of one of the use cases that is applicable in healthcare. But certainly there will be more use cases that you can do especially with having the genai solutions image generations and video generations. I spoke about you know you can actually use your own product catalog to do some form of you know animations or even transforming to a cartoon because at the end of the day the users they want to see something that is very quality very nice very high resolutions and it best if is a is if the image is moving right so essentially you can combine the image and the video to make it very user interactive now the last two slides you know the key takeaways I want to really do a very quick recap right The first one is the AI infra and flexibility. Now we have all the GPU servers that's able to help you if you need to do some deep training inference but if not you can also use our foundation model which is a second one. You know I spoke about all the seat 1.6 seat dream seat dance omi human all of these is a it's a pass service. It's a start service for you. You don't have to use your own GPU. You can just call the API to start using it. focus on your main business and leave all the infrastructure to be managed by back plus and then the third one is the AI you know platform agentic AI so if you want to do you know when you talk about AI workflow sometimes the complexity really depends on what kind of use cases that you want to build right in the past when there's you know in the past most likely everyone heard about RPA robotic process automation so it's kind of like RPA but with the L with the gen AI this thing actually grows up so trendy everyone is talking about agentic task and the last thing here you know I think maybe um next year or something you probably will heard a lot in terms of AI security because I realized that there are so many customers they're using AI solution but they neglected or probably they compromised in terms of AI right because you need to actually use AI to secure AI so actually by plus we have been developing this thing that's called AI security it's a l firewall which able to help you to protect model abuse model, you know, modifications and so on and so forth. And this is just one of the last slides I want to show uh some of the customers that have been with us and uh you know, later on I'll still be around this area. If you want to know more about Pack Plus, GI solution testing you can feel free to let me know.