Agnes AI: Southeast Asia's Answer to ChatGPT  ## Document Notes ## Highlights - **PhD Sparked The Startup** - Bruce Yang started a PhD at NUS during COVID and used that time to research agents and memory for Agnes AI. - That academic period became the foundation and turning point for founding Agnes AI. Transcript: Bruce Yang Good. So I've been studying in Ralfour's institution for my high school. And after that, I've been studying in the U.S. At UC Berkeley for undergrad. And right after that, I've been working in the Valley for over four years in Microsoft and LinkedIn. And after that, I've been doing startups in the Valley and decided to come back to Singapore, actually right before the COVID. So during the COVID, there's actually nothing much I could do. So one thing which I think is a turning point for my career is I started my PhD during the COVID at National University of Singapore. And I think this is a very good timing for me because, you know, it's a good time to think about the second curve of my career. And it's right before the start of the CHIGBT. So during my study and my research as a PhD researcher at NUS, we've been very well exposed to all the brand new stuff like the agents, like large language model and like the memory, agentic Memory. And there are these other foundations for the starting point of Agnes AI, which I think is one of the most exciting moments of my life. So I've been talking to a lot of people, seeing the opportunity of Agnes and any exciting consumer app of AI for recent years. If you miss this for once, you miss it for your life. So I'm definitely getting very much excited about, you know, finding this company. And I think that my PhD is a very important turning point for me. ([Time 0:02:58](https://share.snipd.com/snip/3d1cb774-11c9-492f-a823-0ad85f11e149)) - **Product Data As Model Fuel** - Agnes built its own models to close the gap between closed-source and open-source LLMs using product data. - They leverage user prompts and reinforced distillation from SOTA models to improve performance. Transcript: Bruce Yang Lot of models and this is actually the reason we want to start our own one. The models have a division of two kinds. One is like the Soda, the closed source models provided by ChaiGPT, Anthropic, and maybe Gemini. And another set is defined by the open source ones. You know, one of the very famous ones, DeepSeek, and Queen, and Lama, and a few more. We've been through a time that, you know, the SA models are mostly from the closed source ones and open source ones are actually lagging very far behind. Right now the gap is very very small. So we think that this is the best time for us to find our way to close the gap between the open source ones and closed source ones. We have all the requirements which we need, all the ingredients for us to do this job because we have a very strong research team. My advisor joined us as part of the funding team of Agnes. I have another friend, a Brooklyn alumni who went to MIT for his PhD. He's part of my research team. And we have a lot of users using our product out there, especially in the Southeast Asia. We have over close to 4 million registered users with something like 300k DAU by end of this month. So we have a lot of people using our product every day. And that generates a lot of data, a lot of prompts. Using these prompts as the seeds to train our post-train our model through a reinforced mirroring, we could potentially also distill our model from deploying the SOTA models in our Production. ([Time 0:06:40](https://share.snipd.com/snip/5372122a-f077-40c5-ad7c-17ecc02360fc)) - **Specialize To Cut Cost And Latency** - Agnes specializes models for narrow daily tasks like research, slides, images and video instead of general chat. - Narrow specialization lets them use smaller models with routing to be cheaper and faster. Transcript: Bruce Yang If you're talking about the model itself, we are very much specialized for the work cuts like doing research, writing a PowerPoint, or maybe something like designing for a picture, Generating a video. We have both the large-duncture model to generate the props and the image and video generating models, which is the DIT diffusion transformer model to generate the content. And we also have a model to support our group chat. So you can think that all the models which we are trained are specialized for your daily tasks instead of your daily conversations. So this is one thing which we differentiated from TriGPT, Gemini, or Anthropic. Another thing which differentiates us a lot is we are cheaper and way faster than most of the big models because we are specialized for more narrow tasks. We don't need to use very big MOE models. Smaller models will suffice with some proper routing. By routing, we mean that if you know what the intention of the user, we would not need to direct the request to the big model all the time. So for example, if you need to work on your PowerPoint, the model specialized for the PowerPoint by writing the HTML content for the PowerPoint is specially trained which is unbarred With the SOTA models like the Gemini or Ch-LGBT, but it's much faster and much cheaper. So this is how we differentiate ourselves from the other models. ([Time 0:10:05](https://share.snipd.com/snip/e2f09c49-f9d2-44ea-86ed-1fd48a461f52)) - **Lower Cost Enables Inclusion** - Emerging markets rarely pay subscriptions, so cost-per-task must be far lower to reach mass users. - Agnes reports costs per task at one-tenth to one-twentieth of major competitors to serve these markets. Transcript: Bruce Yang So one of the observations which we had is, you know, only like 0.5% of users are paying for generative AI products. So this is the data from Manor Ventures. There's about 500 million to 700 million users using Gen.AI product every day. Compared to the netizens in the world, which is about 6 billion, it's about 10%. And out of this 10%, if you look at the products like Gemini or ChaiGBT, that's only about 5% of people are spending money for subscription. And this is centered in the high-paying markets like the us the europe korea and japan and for emerging countries like the southeast asia like latin america very few people spend any Money on these products and if you can really reduce the cost for them reduce the cost for ourselves, we're, you know, kind of block ourselves from serving more people. Our goal is to create a product which will be everyday AI for everyone. Serving the emerging countries, be very inclusive. We want to keep our costs really low. Much cheaper than that of ChaiGPT, of Publasty, or even that of DeepSeq or Queen. So the way we solve the problem is reduce the size of the model while achieving the same kind of standard of the results. So we've been doing extremely well for this. Our cost for each of the tasks is like one-tenth or one-twentieth of that of all the major products out there. ([Time 0:12:11](https://share.snipd.com/snip/e35b9a1e-8e10-4f12-81d0-91777df44364)) - **Local Data Improves Minority Languages** - Agnes prioritizes Southeast Asian languages and local conversational data to improve minority-language quality. - They post-train models with platform-collected informal language to better serve local users. Transcript: Bruce Yang So this is a very important thesis for us, because if you look at the population in the Southeast Asia, there are a lot of people speaking minority languages like Bahasa, like Thai, Malay. These languages are not very well served by most of the major national language models because of the lack of corpus of data from the literature. So what we can do from here is instead of trying to make a product which is doing well everywhere, we want to make a product which is specially aligned for the local interest, especially Serving for the regional languages. And the way we do this is post-train from open source models, but with more data, more language, more literature, and more data from our own platforms, from all the compositions, from Agnes. That way we can bring more interesting content like the slams, the spoken languages to our platforms for people to understand about the culture, understand about the ethnics. ([Time 0:14:19](https://share.snipd.com/snip/a48cf9c2-e9aa-474b-8ec1-8668b07a2566)) - **Use Multi-Agent Checks To Reduce Hallucinations** - Reduce hallucinations by combining citations with multi-agent pipelines for generation, evaluation and correction. - Use agents with distinct roles to check and fix outputs before returning results. Transcript: Bruce Yang We do a lot of post-training from the pre-trained models like Queen and DeepSeek. For the newer versions, they are already performing much better, so we take some free ride. Besides that, we also use a lot of citations. We're sourcing our support from all the sources for all the claims which we provide in our research and search results. That gives us better fact for us. Besides that, another way to solve the problem is using multi-agent systems. So sometimes one large language model can be wrong and the model itself as one agent do not really know where the problem is. But it's very easy to be checked by another agent. So if you have multiple agents, especially multiple agents from different models, looking at the result at the same time, leading different roles. One is working on the generation. Another one is working on the evaluation and verification. The third one is working on providing the feedback for some correction. All three of them working together solve a lot of hallucination problems. So we've been dealing in this kind of situation a lot, and we enable this trio situation of generation, devaluation, and correction. ([Time 0:18:24](https://share.snipd.com/snip/4fed843b-109e-44d9-929c-caf68b826fb4)) - **Contextual Memory For Team Handovers** - Agnes offers PC productivity (research, slides, Excel) with persistent memory and handoff via shared context. - The product remembers your last task and lets colleagues pick up work with context preserved. Transcript: Bruce Yang So Agnes is both the name of the model and the name of the product. In terms of product, we think we are doing pretty well. The number speaks for itself. And we have both two different platforms. One is on the PC and one is on the mobile. The positioning of the two platforms are slightly different. If you're working on the PC, a lot to do with the productivity. So you can do your deep research on Agnes with the same context. You can start a PowerPoint, or you can generate images or videos, and you can inject the video images into the PowerPoint. It can also help you to work on the Excel. It reads the Excel sheet very well um a thousand you know lines of record can be read within half a minute and if we understand the content the relationship and you can draw the draw the Graphs for you within half a minute too so it's some of a lot of problems of having lifting we would call it dirty work if you don't't have AI supporting you, you might spend hours. I would spend hours, five hours, 10 hours working on PowerPoint, and you have to start with the research. It's like a pain in the ass. With AI, it solves the problem very well, and the good thing about Agonist is it has a genetic memory. That means the first day after you use it, it knows where you stopped from yesterday. And you will start the task with a contest with you, which you left off. And you can also pass a contest to your colleague through a group chat or from a shared collaboration. That means you don't have to just drop an artifact like a PowerPoint to him and he doesn't know how you come up with this result. ([Time 0:20:24](https://share.snipd.com/snip/b22231b3-561b-4264-a88e-e2c8bbcc6e12)) - **Blend Social Features To Boost Retention** - Mobile is positioned for social and entertainment to increase daily usage and retention beyond productivity. - Agnes blends social features with AI to attract an AI-native generation used to long daily sessions. Transcript: Bruce Yang On the mobile side, we are slightly more on social, entertainment, slightly more, you know, less formal. We really want our mobile experience to be more inclusive compared to that of Chagibity and not only limited to productivity. If you look at the retention numbers, usage hours for different apps, you know, the social apps like TikTok are way better than that of the productivity apps like the chat GPT. GPT, their daily usage time is like by average like 14 minutes. If you start with TikTok or WhatsApp, how many hours would be spent? So I definitely think that when we talk about AI right now, people are still limiting themselves into the productivity category. And if you look at the entire internet, entire usage of the internet, right from your wake up until before you get to sleep. You will spend most of your time working on communication, like how we are doing communication right now on the WhatsApp. And also you spend a lot of time and your free time to look at, you know, entertainment or contents and things like from Instagram, TikTok, maybe Facebook. So we think that all this content can be very well supported, assisted by AI. It's just that nobody has done that just yet. And it's very difficult for QBT or for Gemini, for any other products to do the same thing because they have already have a very strong use base and if they want to make any change they have To start a new app so we are able to integrate all these social features into the productivity category because we're still early to the market and we want to really cater to the young Audience um the ones which we called the AI native population, the ones which, you know, had their high schools or college while during the COVID, stuck in their own room and, you know, Have to jump on the virtual world right after they get, you know, unblocked. It's the year of the birth of JITBT. So they're very native to AI. Any product without AI seems odd, seems a bonnet for them. So I think this is something which our product will help, will get a chance if we are able to serve this AI native generation. If we get them used to using our product instead of that of WhatsApp or Snapchat, we get a chance to become the next meta. ([Time 0:22:26](https://share.snipd.com/snip/9f74a3b9-8a8c-4070-9c14-62b7fb84f44b)) - **Put AI In Group Chats** - Embed AI inside group chats to act as shared memory, summarizer and mediator for misunderstandings. - Use group-AI features to create social games and increase engagement via collaborative experiences. Transcript: Bruce Yang So, you know, we are now the first players to include group chat in AI. ChatGPT has just launched their group chat recently and Snapchat as an incumbent player, they are working with Populacity. I think they are going to spend something like $500 million to gather services from Populacity to support search in their chats. So I definitely see the trend of AI and chats working together and the chat, which we're meeting a real chat, not a chat with AI, I mean, like the chat LGBT. So AI in the chat can be very disruptive, I mean, very innovative because during your normal chat with your friends, there's a lot of loss in memory and loss in context. If you miss the communication in the morning, you would spend a lot of time to look at all the communication in the morning, but it may not really contain a lot of information. And AI can help you summarize the content and let you jumpstart with a topic. First of all, AI serves as a common memory for all the people in the group. Second, you know, sometimes people won't really communicate very well. There's a lot of misunderstanding. AI sometimes can understand people better. One sentence may mean two different things to different people, and some of the people are very inadvertent, and other people would be more expressive. We would hope that AI would fill the gap to make all people very expressive. If the person can't really express for himself, AI will fill the gap. There are a lot of fun things. ([Time 0:26:00](https://share.snipd.com/snip/f48ab7c7-e959-4930-bfb4-c699863a94dc)) - **Offer On-Prem Hosting For Sensitive Data** - Protect privacy by hosting models and data on your own servers and following regional data regulations. - Offer enterprise on-prem deployments to reduce leakage risk for sensitive data. Transcript: Bruce Yang Well, so one thing which we can support compared to all the other, you know, application apps is we have our own models and we have our own, you know, database. So that means we are able to solve the problem from our own end, instead of playing, you know, switching between different apps, different models, and that could introduce a lot of complications. So privacy is a huge problem for us. I mean, a very important topic for us. We're working with all the GDPR, all the guiding rules for different regions. If we talk about enterprise solutions, we're able to deploy our service on our clients, their own servers. They also reduce any risk of information leaking. So I would say, besides all the important measures, we can handle this slightly better because we have our own models and we have our own database. And that means we don't have to deal with data exchange problems. ([Time 0:29:54](https://share.snipd.com/snip/e1bbf85d-c363-44b6-9517-b34a1291da80)) - **Keep Humans In The Loop** - Keep humans in the loop at major milestones to retain critical thinking and originality in outputs. - Require iterative prompts and manual edits to ensure AI output aligns with human insight. Transcript: Bruce Yang Yeah, I think Jason Huang has mentioned that in the, you know, symbiosis of human and AI, humans will need to give a lot of inspiration while AI doing the heavy lifting work. I don't really think that AI would serve the purpose of introducing all the inspiration and entire end-to work for human. Because if you look at Sora, it's a huge success at the beginning, but nobody is using it right now because everything is AI. There's no real human content in it. And the density of human UGC user-generated content will be a very important success for the community because it's like a mutual respect. If this is content generated by AI, why do I need to read it? If it's content generated by humans, there's a lot of in reading the content. So I would say that Agonist serving as a product of AI native social app, you know, integrated with productivity, we will value a lot into human inspiration, human content, and we put Human in the loop at all the major milestones. So for example, if you really want to produce a very good PowerPoint, you would do a very thorough research yourself, and you would go for multiple rounds of prompting to get the content You needed, and you would really ask Agonist to help you look into the content, verify the content, and put up something like page-by layout. And we enable people to have multiple rounds of communication to edit each of the page content before starting generating of the HTML, which is the end part of the slides. And even the first round of slides are generated, you are able to edit very well. We support an advanced editing function, which is almost like manual editing on PowerPoint, but it's much more smoother. And you can also have agonists help you change the page, like change the logos, change the image, change the background. By the end of the process, we would hope that there's very little signals of AI. Very little, you know, uh, uh, you, you won't really see the, the, the AI content, um, within these slides because AI is doing everything which human wants it to do, instruct AI to do. ([Time 0:32:17](https://share.snipd.com/snip/635f7189-87ad-4d5d-8654-0a454c071765)) - **Using AI Helped Him Top Class** - Bruce used AI extensively during his PhD and still topped a robotics class by combining AI with his own work. - His professor encouraged using AI tools and sharing AI-assisted results with peers. Transcript: Bruce Yang While I was doing my PhD, one of the classes which I took, the robotic class, I was an A-plus student, A-plus, like, top student in the class. And the reason I was able to do that is, well, I started my PhD not right after my honor of red. I'm like 10 years older than all my peer classmates, but I'm still like the number one in class, which is very surprising. And the secret which I share to all my friends when they ask about the reason is I use a lot of AI. It's a lot of, you know, tools like the Noble AIM, Chagibity to help with, you know, curating the content of my paper and doing the research, doing a survey. And this is all, you know, very well supported by our professor. He studied from the first day. He's very open-minded. He said that you should use AI tools. Just let us know. If you can generate a very good results with AI tools, you should share that in the class because everybody should know about using AI tools to boost that creativity, boost that productivity. But one of the things which we don't really want to see with this kind of AI facilitation is if AI is diluting the entire community, that could be a huge problem. ([Time 0:35:33](https://share.snipd.com/snip/247a6b4a-d0c0-444b-aa67-faba67d62a54)) - **Limit Early AI Use In Education** - Teach foundational critical thinking and manual problem-solving before heavy AI use in education. - Reserve early-stage non-AI practice so students learn to evaluate and prompt effectively later. Transcript: Bruce Yang Is exposed to AI because I'm an AI entrepreneur. But I also give her a lot of heavy lifting schoolwork, like the Mass Olympia. He just got in the gifted program in Singapore, which is like top 1% of the batch. And at age of nine, she has a vocabulary of 10,000, which is quite a lot. She does all the math Olympiad. She was like top 40 national level. So I definitely see the advantage and the benefit of training the intellectual without the assistant of AI at early stage. Because if you start to rely on AI, a lot of things can be very easy, but you will just lose the kind of habit or lose critical thinking. You might not even know how to prompt. That's a problem. If you know the end results, you know, human analysts have to be the evaluator at the end. You will know what kind of result would be a high quality result. And the only way to appreciate, to have the kind of taste is going through the process manually without dependence on AI. Even though I know the benefit, I know the disruption of AI in education, I still instruct my kids to work really hard, to be a critical thinker, to understand about rudimentary logic And high-quality critical thinking, because this will be very important prerequisite for you to be an AI adopter, for you to really highly utilize AI in your future work. ([Time 0:38:10](https://share.snipd.com/snip/ec3bb9ed-586c-42b4-b80c-9ec778810d6f)) - **Products, Not Models, Win The Market** - Most current models will consolidate into a few dominant products; product experience, not model alone, will decide winners. - Users will choose apps for features and traffic rather than knowing the underlying model. Transcript: Bruce Yang Don't know about that, but, you know, most of the models won't really exist after like three to five years, because the only reason they are existing right now is there's not a huge dominance And we are sitting on very shaky land that everything is disrupting just like how the chat gpt is is focusing a lot on all different kind of things and we just see that the model itself is Not a huge barrier for a product to grow down the road because, you know, one year ago, ChaiGPT is still like number one in the world. Within the last one year, there are quite a few new players which is competing with ChaiGPT like Anthropik, Gemini, and XAI. Within the, you know, a year from now, GPT might not even be the best model in the world. Right now, after the release of Gemini 3, everybody's saying that Gemini is doing better. You know, a merge of both a good model and a good traffic, huge used product. And that means that it's not about the model which make people, you know, doing the choice, I'm making the choice. It's the product which providing all the features, utilizing all the models, which get people stick on it. So that's why I think, you know, Chai GPT is definitely doing well. There are quite a few other players like maybe Agnes, maybe us, maybe a few more, But most people won't be really knowing what's the model behind it. Because the product, the features, the use cases will definitely be more prominent for people to understand, to resonate. So in terms of whether people would have a product built by their own model down the road in the future, it's a potential, I don't know, I think it's pretty wild imagination. It's something very different from what people are using the apps right now, but it has brought a lot of new things. Won't say it definitely won't happen. But as I mentioned, I don't think there are going to be a lot of models down the road. There are very few models supporting very few products which will dominate in the world. And all the other smaller models will be on the long tail. Very few people use it. ([Time 0:43:57](https://share.snipd.com/snip/1e9e84e0-4146-4b25-b473-dd128e26ea9c)) - **Prioritize DAU Then Monetize** - Focus on user acquisition and DAU early; prioritize one metric to build momentum. - Monetize later via subscriptions in rich markets and low-cost one-time payments in emerging markets. Transcript: Bruce Yang Yeah, that's a very good question. So we definitely focus a lot of user acquisition at this stage because we're a newcomer. Well, not like one year ago or two years ago or three years ago when Chaggbt came up. So we want to catch up with their traffic. That's why we do not focus a lot on the AR, on the payment right now. But the reason that we're able to do that is we're able to support the high usage with our own models with very low cost. We're like 1 20th of that of our competitors, other players. And how we want to make money? Two-folds. One is still on subscription, but this is on the high-paying markets like US, Japan, Korea, Europe. For these regions, we want to enter, but we will give a very big discount in the subscription, provide equal or more features, better quota, but the cost will be one third of that of the Other place. The other fault is on the emerging countries. We understand that most people are not paying, not only limited to all products, but also TriGPT, Gemini, nobody wants to pay, especially if it's a monthly subscription. One of the reasons is everybody is afraid that they're missing the time of cancelling or unsubscribe, but the product is not used anymore. So we try to solve the problem by giving a lot of one-time payment, which is very low cost, like $1 or half a dollar for people to complete a task. If you don't really want to spend any money, any dollar, you can also get people to our platform by inviting more people, be very active on the group chat because the user acquisition Is also cost from outside. By saving cost, it's making money for us. And long-term speaking, we're definitely believing in a thesis of traffic, just like how Google, Meta, TikTok become so valuable because everybody is talking about it. Everybody believes about the brand. ([Time 0:57:53](https://share.snipd.com/snip/9b1a109e-8523-45a8-980b-877977b1966f))