Transcript Auto-scroll ON 0:00 Hello, everyone, welcome to Innovator Cafe, a podcast that bridges the people and the 0:07 world of AI and innovation. 0:08 Please follow us for the applications, insights, and emerging trends. 0:12 Today, we're very honored to have three special guests. 0:16 I'd like to have them introduce themselves very quick. 0:19 Maybe we can start with Joe. 0:20 Hi, everybody. 0:21 I'm Joe Yu. 0:22 I'm a professor at Columbia in computer science. 0:26 I'm also a founder of Arclex AI. 0:30 Arclex is a company that delivers an AI agent framework that supports scale deployment, 0:37 frictionless automation, and improving over time. 0:41 Hello. 0:42 It's very honored to be on this podcast and also have the opportunity to talk with everyone 0:46 here. 0:47 I'm Chuanlai. 0:48 Previously, I worked at Uber and Robinhood on their machine learning platforms. 0:52 Two years ago, I started my own company, which is called Chatslide. 0:56 So initially, we help people to create slides, and now we kind of morphed into a personal 1:01 knowledge playground where everyone can upload a lot of their personal resources, and then 1:06 they can generate their own LLM and create all kinds of different content from our platform, 1:12 including personal avatar, including slides, videos, et cetera. 1:15 It really empowers a lot of knowledge workers. 1:17 Go ahead, Anson. 1:18 Folks, my name's Anson. 1:19 My name's Anson. 1:20 It's really honored to be here. 1:22 So previously, I was building Alexa at Amazon for five years, and after I spent three years 1:30 working for TikTok monetization platform. 1:33 So two years ago, I quit and started my own startup. 1:36 It's called Holiday. 1:38 To begin with, we started giving the AI knowledge base for general customers, and later we switched 1:46 to bring agents to small size and medium agents to cover the SOPs for our clients. 1:54 Yeah. 1:55 Yeah. 1:56 By the way, I'm Weiki. 1:57 I'm the podcast host, and Tom. 1:59 Many people say that 2025 is Agenda AI's year. 2:02 So we are very excited to see many companies doing Agenda AI. 2:06 The first question is, regarding your company, how do you each define next generation work, 2:11 and what problems are you most focusing on solving today with AI? 2:16 Maybe we can start with you. 2:17 Yes. 2:18 So we work on agents' workflow automation. 2:22 So I'll give you an example. 2:23 If you're a salesperson, a lot of your top funnel could be automated through agent workflow. 2:29 But this is what we do. 2:31 And we mostly help people to scale these workflows, and at the same time, use AI to help you to 2:37 maintain the workflow. 2:38 They don't have an agent, like a human being, to maintain it. 2:41 You can use AI to automatically maintain it for you. 2:45 Many people have already mentioned last year that 2024 is the year of agents, and 2025 2:52 is also the year of agents, but also the year of many other new technologies. 2:56 So I'm very excited about this. 2:58 So basically, I think the next generation of AI is about reducing the cognitive load. 3:03 A lot of people are talking about AI-native product. 3:06 What are the AI-native products? 3:07 So basically, it feels that you are not really learning a lot of things. 3:12 You have an AI co-pilot or AI body that helps you, or even an AI agent that helps you with 3:17 certain kinds of tasks. 3:18 So for example, if you are using a chat slide, it feels like you actually have someone who 3:24 is like a human being that helps you create a certain slide, read a certain kind of content, 3:29 and update a certain part of the document. 3:32 So I feel that's the next generation of workflow, and that's how we are going to solve it. 3:37 We'll probably build AI agents that can understand all the context of all these slides or any 3:44 other kind of workflow, and then help people with all those kinds of tasks. 3:48 Yeah, same as John. 3:49 I'm super excited about the so-called agent year of 2025. 3:55 So back to what we are doing here right now. 3:58 So we are basically trying to bring agents to the real users in enterprise. 4:04 So we are building agents for salespeople, as well as the design teams. 4:09 So our clients, including the e-commerce streaming companies, and also the publishing companies. 4:16 So at the very beginning, when we started this startup, we were building the AI knowledge 4:20 base for general consumers. 4:24 But later, we figured out there's still a huge gap between the actual users in the business 4:30 and the frontier AI models. 4:33 We got pretty good products from Google, from OpenNet, from Anthropix. 4:39 But there are more general data uses. 4:41 You get a question, you use them as the new search engine. 4:46 You try to get an answer, you try to get a piece of documents or PRD or things like that. 4:53 But in the real business, there are complex workflows, as well as SOPs. 5:01 For example, one of our clients is the live streaming in e-commerce. 5:08 They got hundreds and thousands of digital assets they need to create every day. 5:13 They want the abilities to add it on the top of generated contents from AI. 5:19 So we're building specific agents for them to accelerate those processes. 5:25 We're trying to bring up the efficiency. 5:27 So that's what we are doing here right now. 5:29 We're trying to bring the AI agents into the space to help them to bring up the efficiency. 5:34 And we call them AI interns. 5:37 That's much more friendly to the actual users. 5:41 So they just use the agents to speed up their work, and they can always keep themselves 5:46 in the loop. 5:47 That's what we do right now. 5:48 I have a follow up question for Joe, basically, because I want to learn more about your framework. 5:53 Actually, AI framework, there are many competitive components and people. 5:58 What's the uniqueness of your framework and what can really be called next generation 6:02 from your side? 6:04 So we are basically a tool to help engineers to build agents faster, safer, and more scalable. 6:11 So the unique things comes to two things. 6:14 One is about scalability. 6:16 So if you think about using other frameworks like LendChains, or you build it your own, 6:20 it's very fast to build a prototype, but you wanted to scale it and deploy it in real enterprise, 6:25 you hit a wall. 6:27 Because you run into concurrency issues, like provisioning issues, all this, right? 6:31 Because many of these AI engineers or data scientists will build the agent are not back 6:36 end people. 6:37 They don't have enough back end knowledge to scale it up. 6:41 And then like existing services for that wasn't enough. 6:44 So you hit a wall. 6:45 So what we provide is a seamless transition from fast prototyping into scalable deployment. 6:51 So just as simple as by our enterprise version, we already did all the auto scaling for you. 6:56 It's easier to do it that way. 6:58 And the other part is really on the, we call it, we use reinforcement learning to help 7:03 you to update your workflow. 7:05 So you start with a whatever is the best for you at this moment. 7:10 But when end users come in and interact with your agent, we can optimize the flow and orchestration 7:16 based on the end user's role. 7:19 For example, a sales is how much things you can sell, or automations, customer service 7:24 is how many questions you can resolve, right? 7:26 So these metrics could back propagate and change your orchestration for them. 7:33 It can largely reduce the workload and streamline the workflow, which is really nice. 7:37 Chenlai and Nathan, just a follow up question of Joe's description, right? 7:41 When you guys try to build up the agent, right, did you guys encounter those problems? 7:46 Maybe share some of yourselves. 7:48 Yeah. 7:49 So I can probably share a little bit about my experience. 7:53 So first, I think like when we are building the agents, the first question is that does 7:57 the agent learn know all the context? 8:00 Basically what we think about agent is that agent will be a kind of replacement of human 8:05 being. 8:06 Like if human being can make certain decisions, the agents should also be able to do that 8:10 and in a more efficient, faster way. 8:13 But when we are building the agent, what we figure out is that agent might not always 8:19 be able to understand what we are asking them to do, especially if the task is related to 8:26 something visual. 8:27 For example, if something is text, if I ask the agent to help me summarize a certain kind 8:31 of text and choose a better way to move forward, then agent is going to do that pretty well. 8:38 But if agent is going to work on something that includes visual, for example, if the 8:43 agent is going to look at my screen and say, you need to click this to cancel the subscription, 8:47 then that's a little bit hard. 8:49 But that's also the fun part of developing the AI agent. 8:52 There are a lot of new technologies around this and we are actively exploring this field. 8:58 Yeah. 8:59 Yeah. 9:00 I think for us, because we're basically building customized agents for different clients, right? 9:05 I think the more important question for us is about, I think what Joe was referring about 9:11 the matrix, how we analyze the data results, whether it's good for the, whether it's useful 9:16 for the client, how the clients like it. 9:20 So in our cases, we figure out when we talk to the clients, most of them understand like, 9:25 okay, AI is still a fossil growing and we're probably not going to use AI to replace humans. 9:32 So they are more focusing on like, you know, how many times you can save for each individuals 9:37 in the organizations. 9:39 For example, when we're creating these agents for the design teams, it's going to create 9:43 a digital assets. 9:45 So they basically estimate that the effects of our agents by like, you know, if we're 9:50 doing creating a 500 assets, it used to take them like, you know, four people and cost 9:57 them three days. 9:58 But for us, if we finish it like, you know, within three hours with only one people, they're 10:03 super heavy with that. 10:04 So for us, when we, you know, like a change to most important part for us. 10:10 So we need to define the right matrix and estimations and analysis. 10:13 Yeah. 10:14 Yeah. 10:15 So Joe, just a quick get back to you. 10:16 Right. 10:17 In your framework, it sounds like some of the pinpoints you identify and they encounter 10:21 already. 10:22 I don't know your framework. 10:23 Did you guys have some solutions on those? 10:26 Currently, we don't support like videos, we only support audios and texts. 10:31 But a multimodal is always very difficult because it makes the workflow even more complicated 10:37 to auto scale. 10:39 On the other side, I think a couple of other things I want to stress is also AI agents 10:44 is very diverse. 10:46 Their workflow requirements sometimes are different. 10:49 If you're from office facing, then, for example, long sessions, because people may chat with 10:55 you for a long time. 10:56 How do you maintain long sessions when doing auto scale? 11:00 And the other part is about real timeness. 11:03 If it's a customer facing real time agent, you want to make sure that your latency is 11:07 within a certain millisecond. 11:09 And if you have users from everywhere, they want to make sure how can you actually store 11:15 data in different regions in your cloud in order to be able to serve and reduce the latency. 11:21 So it's a very complex problem when the scale is up and then the requirements are very tight. 11:26 I see. 11:27 Well, this is a very good discussion. 11:29 Let's move to the next question, right? 11:31 AI is often seen as the final frontier of AI in knowledge work, right? 11:37 How do you ensure that your tools enhance teamwork rather than replace or complicate 11:42 it? 11:43 Our product itself is actually about helping people to communicate with each other. 11:47 For example, we help people create slides, we help people create posters, we help people 11:51 create videos. 11:52 All of them are like medias for communication, right? 11:56 So inherently, we help people to do more teamwork, to do more communication rather 12:01 than just replace them. 12:02 And at the same time, there's another problem. 12:06 So basically, we want to actually enable people to share their knowledge. 12:10 However, we don't want to just hijack all people's knowledge and then create something 12:14 that is not relevant to this person at all. 12:17 So first, it will revolve around understanding the intent context and also the person's knowledge. 12:24 And then at the same time, we help people to actually keep humans in the loop. 12:28 This is really important. 12:30 We still want the people to actually interact with us rather than us taking 100% of all 12:34 the work. 12:35 We are not a dishwasher, which is totally okay to take 100% of people's lives. 12:40 But what if users want to tweak the tone? 12:43 What if users want to adjust something or anything, like change the narration? 12:47 So now we enable people to actually build the AI agent to enable people to actually 12:54 change the content as they want. 12:56 And then we can actually create higher quality content. 13:00 And that is much easier. 13:02 And also that conveys the user's intention, knowledge, and also their brand ideas, etc. 13:09 And that will actually, for example, 10x users' productivity, at the same time, keep people 13:16 in the loop, not replacing human beings. 13:18 Yeah, sure. 13:20 So I have a different view, because when we talk to our users, I think there's a balance 13:26 between the collaboration as well as working individually. 13:30 I think with the AI agents, we basically become like a mad human, or we are more powerful 13:37 and capable of doing multiple types of things. 13:39 Some things we're not able to do, or we do it not as efficiently as today. 13:45 And when we're doing the investigations, trying to understand the SOP and build out 13:50 our know-hows with our users, I figured out there's a lot of inefficiency that was brought 13:58 up by the communication within the teams. 14:00 For example, one of our clients, their marketing team has to talk to their design team every 14:06 day, trying to gather posters and all the social assets they need to do the marketing. 14:12 And they don't have the skills. 14:13 They're not capable to build up all those posts or make the designs by themselves. 14:18 But when we deliver an agent, they don't need to depend on the design teams anymore. 14:23 So in terms of collaboration, they don't need to collaborate with the design team that much 14:29 anymore. 14:31 So we do bring up the communication, we do bring up the collaboration, but in the other 14:36 way, we do bring up the efficiencies. 14:39 There's a lot of back and forth communications within the organizations today. 14:44 And I think agents definitely are giving more capabilities and giving more powers to individuals. 14:51 And in terms of that, we do see less collaborations, but that brings more efficient communications. 15:01 So I would say it's probably a different type of collaboration with AI agents in between 15:08 of teams. 15:09 And in that case, we do bring up the efficiencies. 15:13 So that's my point of view. 15:15 Jill, how about you? 15:17 To me, I think the entire labor market will be reshaped because of the AI agents. 15:26 Some of the jobs will be repositioned for other things. 15:29 And this process could be very painful. 15:32 We always say like the technology is easy, the people are hard. 15:37 We're changing not only the technology, we're changing the process, we're changing the process 15:42 of how things are getting done. 15:45 That requires people to be rescaled and to be open to new changes. 15:50 So that's very difficult for transformation. 15:53 I think it's the same thing with electricity. 15:56 We have to adapt because there is no way of going back. 16:00 So this process is going to take some time. 16:03 I think the first thing for everybody is to be aware and educate themselves about what 16:08 AI can do and what AI cannot do. 16:10 And really think ahead of time, what could they be doing things and utilizing AI to empower 16:16 themselves, either rescaling themselves or either learn how to harness it to be able 16:23 to make themselves more efficient or expand their work scope, bring more value to the 16:28 company. 16:30 I'm just curious, right, from what you do so far, what do you think an agent can do 16:35 very well and what do you think an agent just cannot do at all? 16:39 So it's the same thing with all AI models. 16:42 If you have clear rules about what is correct and what is incorrect, that's the best way 16:47 for the AI to learn. 16:50 Whatever it's very clear in terms of pipelines and processes, it's easier for the AI. 16:55 Whenever it comes into data securities and other things, then it's always hard because 17:01 of the barriers of being able to see the real data and be able to develop upon that. 17:06 So there's always a couple of things we say, first is areas or verticals that have data 17:13 security, privacy constraints are going to be a little bit later in terms of adoption. 17:19 And then there's a couple of other things. 17:22 The criteria is more than clear. 17:24 For example, is this a good writing or is this a bad writing? 17:28 Everybody could be slightly different. 17:29 But did this agent sell the product or not? 17:33 That's clear. 17:34 So whatever things that are very clear, it's easier for the agent to learn. 17:38 That's probably the reason that the current large-language model is based on Transformer. 17:41 It means the transformation of the whole world. 17:45 Let's go to the next question. 17:46 Since the rise of the LLMs and we see many, we say, emerging AI capabilities, right? 17:52 You can mention a lot already. 17:53 What are those emerging AI capabilities that you're most excited about in your work or 17:57 in your agent development? 18:00 And also, what are those you think is still under-typed? 18:04 And how are you preparing your product roadmap to leverage it? 18:07 So I think the general trend is that we can see the large-language models, if they are 18:13 not getting smarter, at least they are getting faster, they are getting cheaper. 18:17 So this enables a lot of scenarios like we previously didn't see. 18:22 For example, some products like Minus. 18:25 I've tried Minus. 18:27 It's not as perfect as many people expect, but I think it unlocks a lot of potential. 18:34 For example, it might be able to actually create, for example, it might be able to actually 18:40 give you answers for a lot of complex queries. 18:45 For example, if I am going to Southeast Asia in seven days, please help me figure out the 18:52 cheapest way. 18:54 And also having these issues with probably Malaysia and Singapore, etc. 19:00 It can help you with a lot of very detailed requests and figure it out. 19:04 that out. So previously, people cannot do that automatically. For example, you can do research 19:10 online, for example, let's say at bookings.com, but bookings.com is not going to help you figure 19:16 out your visa issue in Malaysia, or it's not going to figure out a hotel issue in Cambodia. 19:24 But I think with large-language models, it can actually plan everything. Not 100% perfect, 19:31 but maybe 80% perfect for you. And I think that's something very exciting for us. 19:36 Yes, thank you. How about you, Joe? Because you also have a strong background in academia 19:40 together with industry. So from these two points, how do you see the emerging AI capabilities and 19:45 how you predict the future AI capabilities and role model? 19:48 Yeah, so definitely multimodal is one direction. Not only text, audio, and video, but also other 19:56 modalities. I'm also very interested about sequential models, because a lot of times, 20:02 it's not just words and images. It's about sequence of data points, like the stock market, 20:11 the watches that you wear that has the haptic sensors, or various other sense of data. You 20:17 can actually mine information and knowledge from that to empower other applications. 20:23 A couple of other things I'm always very excited about is about the embodiment of these AIs. 20:30 So that goes into robotics. So general robotics models that could potentially 20:36 do manipulation navigations all together. One other thing I'm really very interested, 20:44 and I'm working on a lot, is teaching large-language model with self-improvement and 20:48 self-correction about these. Because previously, the models were trained on static traces. 20:54 We're talking about static data on predicting next word, predicting next token, or is this a 21:00 dog, is it a cat? There's no traces that is interactive. So right now, we're doing a lot 21:06 of work in terms of simulating environments, creating traces that are interactive, so that 21:11 the model can actually learn how to do interactive predictions. So for example, how to self-correct 21:18 teaching them the ability of self-improvement as well. 21:21 Yeah, so I think there are two things I'm pretty excited about. One is the agent-to-agent 21:25 protocols released by Google. So right now, we're still using the MCPs a lot because we're 21:30 building individual agents for different departments. But I think in the near future, 21:36 very soon, especially in the enterprise, there's going to be a lot of agents talking to agents 21:43 to complete more complicated tasks. So that's definitely one thing we're looking into. 21:52 For example, a design agent talking to a sales agent or a marketing agent, things like that. 21:57 So that's one thing. And the second thing is about a paper released by Richard Sutton and 22:04 David Silver talking about the error of experience. So I really like last month as well. So it's more 22:11 about how the AI models, as well as agents, to learn from the experience, just like humans. 22:18 So I think we don't have any technologies around it yet, but a lot of people are working on that. 22:24 I think that's something I'm super excited about as well, because if you think about 22:28 in the enterprise, so when we build agents, we're basically trying to mimic the SOPs 22:33 and workflows of how humans work. There's a lot of experience inside. 22:38 So today we're defining the rules, we are defining the workflows, we are defining 22:45 different types of criteria to make sure agents behave as we want. But there are still a lot of 22:52 learnings that we have to loop human insight. People like us, when we build agents, 22:59 we're pulling the SOPs, we're pulling the learnings, we're pulling the know-hows. But later on, I think 23:06 they should be able to self-iterate themselves, the agents, learning from the experience, 23:11 learning from the interactions between the human and the agent. So that's something we're also 23:17 pretty excited about. I think a model context protocol, which is MCP2s, empty mountain 23:23 contextual memory across sessions and applications, it will transform how AI agents support long-term 23:31 workflows. What do you guys think? I'm a huge fan of MCP, and we started looking into MCPs 23:37 right after they were released last November. So basically, a lot of people say it's right 23:44 around two-use and function calling, but we think it's a huge thing. So first, 23:51 it's increasing the reusabilities. It gets a lot of things like the APIs, the tools, 23:59 the capabilities. They give those capabilities to agents pretty easily. When we build new agents, 24:05 we can reuse a lot of things. And in the protocols, they got a lot of other things, resource 24:10 prompts defined. So from the development perspective, it's lower the barriers to 24:17 build an agent, and it's increasing the capabilities of the agent can do. And especially when we do 24:23 this client, when we're working with our client, when we're building multiple different types of 24:30 agents for different departments and teams, it definitely speeds up our development process a lot. 24:36 And from a personal perspective, we're using cursors and things like that, some similar tools. 24:44 So it definitely speeds up and accelerates a lot on our development, designing, and a lot of 24:52 different types of process. We use different MCPs to accelerate. When we build up our PRD, 24:58 when we do our designs, when we convert our designs to our MIPIs, doing the code iterations, 25:05 it feels like it's pretty capable of doing a lot of things. But still, there's a long way to go 25:12 because currently, it requires you to have some backgrounds of engineering. You need to 25:19 know how to set up configuration MCPs to different hosts and clients. It's still 25:26 pretty hard for users without any engineering background. So I think that's something we are 25:33 trying to solve as well, because we want our clients to be able to self-serving them to 25:38 building agents by themselves without knowing any MCP setups and coding skills. 25:45 We are actually in the process of integrating with MCP. So currently, we use APIs to call 25:51 different agents. But having MCP just makes things a little bit easier for people to 25:56 plug in different services. Our product is not yet MCP, but we do use a lot of other products, 26:03 as mentioned before, like cursor. So I believe MCP model, it will be very powerful. Basically, 26:10 it will be the glue between the tools and also the agents. We actually are planning to develop 26:16 something that helps people to actually upload their content to YouTube and also to do SEO 26:22 monitoring and do a Reddit engagement in one flow. So I believe this will be actually powered by MCP 26:28 and it will be very powerful after we develop everything out. Interact with MCP or leverage 26:34 MCP. What do you think the pros and cons of this protocol? Do you feel there's some pinpoint 26:40 where you guys start to deploy it? Joe, do you want to go first? I don't really see a huge 26:47 pinpoint on that in terms of implementing. It's just extra work that you need. So I think it's 26:52 extra work in the front, but eventually it will be much easier for all the things to communicate 26:57 in the same protocol. Yeah, I see both pros and cons. So for ourselves, 27:06 the pros would definitely be the reusabilities, the scalabilities. So there's a huge open 27:12 community around MCP, right? There are a lot of people building open source MCP clients and servers 27:18 every day. And it's definitely helping us to speed up our development of different types of agents. 27:25 It's a great community and also Anthropix is building up the protocols proactively. 27:32 That's definitely something we love, but there are also pinpoints. But I think it's basically 27:37 at this point, I think they're going to stop it later. It's about authentication. So think about 27:43 some general tools like Google and Google APIs or some other like Slack or things like that 27:51 requires authentication. So they do support different types of authentication right now. 27:58 You can use API keys or you can use the OAuth. But still, to me, I think it's a complicated 28:07 process. It's also for the users. So in that case, I do hope it can be more user-friendly 28:15 in terms of authentication. When you try to leverage other products, 28:21 MCP to support yours, did you see any pros and cons or something difficult to use? 28:26 So first, I think everyone has already learned the benefit of doing that. And as Nissen said, 28:33 that the authentication is a problem. Basically, if you have a lot of contacts, 28:39 you are going to... That's actually owned by different applications. Then you are trying to 28:47 basically reduce the barrier between those applications. It will be pretty hard. For 28:53 example, if you have something in cursor and you have something in the browser, 28:57 and then you want to connect with each other, it's a little bit hard. So I do see that there 29:02 could potentially be some new product or a change towards this to allow for a better 29:10 communication between different contacts. I also have a follow-up question about MCP. 29:14 MCP is a two-use protocol, basically. And Nissen, you have already mentioned about the A2A as a 29:19 protocol, basically agent-to-agent. So what are the differences? Why is that two-using and 29:23 agent-to-agent need two different protocols? What are the differences between the two scenarios? 29:28 And also, do you think MCP can be further extended to agent-to-agent or... 29:32 Experience? We are not using A2A. We are actually leveraging MCP for agent-to-agent, 29:38 talking to agents. So we do write some AI calls as a SCC server. For example, 29:45 we're building this agent for design team or marketing teams to create posts. Basically, 29:51 there's going to be this agent making all the plans, how many layers we need, what kind of 29:56 image we need, and it's going to call MCP server, which is another image-creating agent. 30:02 It's going to create the image we use as a background or it's going to call a customized 30:08 model to create the figure we use in the post. So we do see MCP can be extended as A2A protocols. 30:15 You basically just write the agent as a MCP server. And we're also building something in this 30:22 way. But it's hard to talk to the pros and cons between MCP and A2A today as for now, especially 30:31 for myself. We haven't used A2A yet, but I do see a lot of benefits going forward. Because when we... 30:40 Back to my days at Amazon. So at that time, we don't have... We don't have all the things we 30:47 have today. But we're also using some similar architectures. So they're basically going to be 30:53 the master agent at that time. It's going to route user's query to different types of domains. 31:00 Like when a user has a shopping intent, we're going to route it to the shopping domains. 31:06 If it has a gaming intent, we're going to route to gaming domains. It's pretty similar to protocols, 31:11 which you have an agent and they talk to different types of agents directly. 31:17 It's like MOU, right? You route different intent to different agents to finish all the follow-up 31:23 tasks. I see A2A as a huge use case, because in terms of scalabilities, think about it. You can't 31:34 have an agent which has 100 MCP servers, right? Even though the AI models become very powerful, 31:41 but it's hard for them to understand all the context of which tools I should use when they 31:45 have hundreds of MCPs. And in that case, we probably have thousands of tools under those 31:51 MCP servers. So at that time, it's definitely you want to go with a new architecture. You just split 31:57 it to a different type of agents, a different type of experts. You have major agents and you want to 32:04 sub-agents or different agents to handle design works. Agents handle the research work. An agent 32:10 handles the image creation or video editing works. So at that time, I think A2A is definitely going 32:15 to play a more important role. But again, back to whether we can extend the MCP to cover the works 32:24 that A2A is going to do. Yeah, I do think so. I think you can basically write the agent as an MCP 32:33 server, but we still need to test it out to see how it works. I think, because fundamentally, 32:42 A2A is more focusing on agent talking to agents, and it has a lot of designs. They have a lot of 32:50 architecture design that's specific for this use cases. So I think it's probably going to work 32:54 better, I think, than MCP in terms of the agent talking to agent. Currently, when we design the 33:01 MCP client, we have to put some kind of MCP server into the configurations. So do you mean that A2A 33:07 can, in the future, help to discover the MCP servers, right? Yeah, I think both agent discovery 33:16 in A2A, as well as MCP server discovery, are going to be more dynamic. I think it's going to be 33:26 impacted by the frontier models, the fundamental models. When the model becomes more powerful, 33:31 it's going to have a better performance in terms of the agents or the MCP server discovery. 33:37 But in terms of scalability, I think A2A is probably going to have a more important role. 33:43 Because when you build an agent and it has tens or hundreds of MCP servers, 33:48 it's going to be super hard. That's going to affect the performance, for sure. So in that case, 33:54 I think a multi-agent architecture could be more helpful, because you can just basically have 34:00 rolling agents and route different queries or different intents to different agents. 34:06 Yeah, thank you. And Joe, how do you think about this question, that the MPC and A2A protocols 34:12 right now are rising, and also talking about scalabilities and also the discovery of the 34:18 MCP servers. So how do you see this problem? It's just two different protocols. Everybody wants to 34:23 be the control one. You never know which one is going to get more popular. But at least I see 34:30 MCP seems to have more adoption because they're earlier. But I guess that's where I would bet 34:38 my money on. But it's hard to say. Maybe Google will catch up. I don't really see a huge difference 34:44 right now. Yeah, so similar feeling. I would prefer MCP over A2A, but I haven't used much 34:51 A2A yet. And actually, I think it's probably too early to make any judgment. But I would be happy 34:58 to see how the technology develops. So our next question actually already mentioned about the 35:04 datas. Currently, as far as we know, still we don't support the authentication things. 35:10 And many users express concern about the data privacy, model hallucination, and over-dependence 35:15 on AI tools, as well as these authentication issues. So how do you or your companies mitigate 35:22 these risks while still pushing innovation forward? Can we start with Chennai this time first? 35:27 Yeah, sure. So I think the hallucination data privacy is a very big concern, especially if 35:32 our clients are professionals. They want their content to be very accurate. So what we do is 35:38 that we draw a line between the AI assistant content and also the user's intent. So our 35:45 users will actually review. We will ask them to review all the content before they publish 35:51 those content. This is the way how we keep the human in the loop when the judgment is needed. 35:58 And because a lot of our content actually relates to healthcare, we want our content to be as 36:03 accurate as possible. And because we control our own agents and the techs that avoid passing 36:10 sensitive data to the unpredicted third-party APIs. And in the long term, we don't want AI to 36:16 replace human being totally, but we want AI to execute the human's command efficiently. 36:25 And that's our goal. Very awesome. So I totally agree, hallucination is a huge problem in terms, 36:32 especially we are working with business and workflow SOPs. But I think agents 36:40 basically is helping a lot in that case. So if you are using TRGPT or using the model 36:47 directly to try to get an answer, it's definitely a higher possibility to get a hallucination. 36:54 But with the agents, we are setting up different tools. Some of the MCPs are going to be 37:00 the rules you basically have to follow. One of our agents built for our client is basically 37:05 a customer support. So we write it up and enforce all the responses to go through the judgment 37:13 tools to decide whether you have some hallucinations or some punishments 37:18 or things like that before it returns response to the customers and the clients. 37:23 So we do have some judging tools before we return the response. That's going to help. 37:30 And especially when we're talking about the enterprise agents, they got a pretty specific 37:37 SOPs and they got some specific rules we have to follow. So I think that also helps. 37:44 So we are more leveraging the models for the planning and the raising capabilities. We don't 37:51 theoretically get an answer from the model, but using the internal data source. 37:57 So I think that also helps. So in our case, we are basically following this practice. 38:04 So I think there are a couple of ways that you can improve. 38:08 Reduce hallucination, RAG is one type of way so you can do it. 38:15 It's always like if you have a confidence score, that helps you. 38:18 If you ground down the source information, that helps you too. 38:22 There are many companies specifically like SolidData Valuations, Gartrails, Laker, Confuse, 38:29 and there's also work on more security related like Virtual. 38:32 I think for us, we mostly leverage other people to help us on these things. 38:38 We plug into their offerings as well. 38:40 Now, how about the data privacy things and also 38:43 authentication things, do you encounter this also? 38:46 Of course, OAuth is always a problem. 38:48 Everybody finds it very painful, but it's also very useful for protecting the security, 38:55 like especially on transactions on money and things like that. 38:58 I think it's just a way that we need to put in these sort of 39:03 what I would usually call safety checks. 39:06 Try to distill it into the infrastructure instead of on the application layer. 39:11 So whenever you use the infrastructure to build things, it's by default compliant and secure. 39:16 It's the same thing if you think about databases, right? 39:19 Back in the days in databases, you have to write a lot of things yourself. 39:22 But now, whenever you use a certain standardized database, you don't have that problem. 39:27 So I think it has to be on the infrastructure level to build on the security layer. 39:32 Yes, definitely. 39:34 Actually, going back to NCP, client-server model, similar to HTTPS, 39:38 but we have HTTPS protocol to guarantee security. 39:42 Do you think that NCP will be evolving this way, 39:45 architecture way, or do you have other opinion for the security part? 39:49 I think that's interesting because I'm not sure if there's a question for me, 39:52 but I just have an answer. 39:55 So basically, HTTPS is like you add SSL to HTTP, right? 40:01 So if you add SSL to NCP, then certainly you can add a layer of production. 40:06 Basically, it means that who you are talking with, another person cannot spoof. 40:12 But I think their security is more than that. 40:16 Of course, maybe we will have MCPS. 40:19 That's a very interesting concept. 40:22 But beyond that, there is going to be a lot of other organization or other parties involved. 40:28 For example, you need to have a certain party. 40:30 For example, for HTTPS, you need to have a certificate, right? 40:34 So maybe you will need to have a party that issues all those certificates to AI agent, 40:39 and I think that's totally a possibility. 40:42 So it seems like currently, I think for this type of questions, 40:46 there are some startup working on those, especially for the data privacy, right? 40:51 I do see a lot of startups. 40:53 They are working at the time data, before the data get into the pipeline. 40:56 They do have some embedded web cart. 40:58 At that time, they started doing something there, which can solve part of this problem. 41:03 And also, for authentication, there are some companies working on that too. 41:08 But eventually, I think SRAPIC may try to do something and try to mitigate the risk. 41:14 If you think about this, right? 41:15 It worked on something on the data package. 41:17 So it should be the same as what we tried to do before. 41:21 So for the next one, if you were to design the perfect AI teammate, right? 41:27 One that complements human creativity and decision making, what would it look like? 41:32 And what are the current technical gaps that need to be bridged? 41:37 Who wants to be the first? 41:38 Maybe start with Jo. 41:40 So I want a chief of staff that is AI, that helps me do meeting booking, 41:48 proactively remind me I forgot certain sales, I need to follow up. 41:53 So it's just like everybody wants a chief of staff that is super smart, super proactive. 41:59 I think that would be everybody wants executive assistant or something like that, 42:03 or personal assistant. 42:05 I think everybody deserves to have one. 42:08 But right now, it's really about where do they get the data, right? 42:11 A plugin is always problematic. 42:13 You have so many different ways that you have your phones and you text people. 42:17 You have your WhatsApp, you have your WeChat, you have your email, 42:20 so you have a phone call, right? 42:22 All that information is scattered everywhere. 42:25 And then filtering through the information, 42:28 draw insights and doing proactive workflows is also difficult. 42:31 It's very hard to identify or just say, what does a good chief of staff do, right? 42:37 They do everything. 42:38 So that's the hard part in terms of how do you, through all these, 42:42 can you actually automatically find what is the best workflow from the traces you record, 42:47 instead of asking people to decoratively say what is the best practice. 42:53 So I think it's more about how do we learn these best practice automatically? 42:57 How do we go from that and improve over time when people are starting using it, 43:02 instead of asking people, oh, do you want to sum something up or something down, 43:05 which is like a lot of people don't like to provide feedback for. 43:09 Yes, I actually agree with what Joe said. 43:12 Personal chief of staff is really important. 43:15 This reminds me of what I saw on Mark Andreessen's Twitter. 43:20 So basically he mentioned that if some developer is only going to build a product for an investor, 43:26 then that product will be a very large human database, 43:31 which essentially has all his personal contacts and all the information related to that. 43:37 So I agree. 43:38 It's actually very hard to have everything, all your contacts, all your schedule, 43:43 et cetera, in one place, and then you can automatically interact with it. 43:48 But I think it is going to be a possibility. 43:52 So that's what Arvind Srinivas, the founder of Perplexity, replied. 43:58 He mentioned that actually there is going to be something 44:02 probably provided by Perplexity to help people with that. 44:06 I think it's a good idea and I will definitely sign up to try it. 44:10 So first of all, I'm a person hard to satisfy. 44:13 So I don't think there's going to be... 44:15 It's hard for me to say what is a good, perfect AI teammates. 44:20 So one thing is I think it's hard for AI to take a decision making, 44:24 especially if it performs as my teammate. 44:28 So I would say a pretty good AI teammates for me would have three. 44:35 One is pretty personal. 44:37 It knows me, like the agent knows me. 44:40 And it has contacts about like, okay, your schedule, 44:45 your relationships and your friendships and all that stuff. 44:48 Second, it should be proactive. 44:50 The proactive would be like the agent should remind me 44:54 about the things I need to do at time I need to be reminded. 44:58 If the agent just keeps sending me a message and notification, 45:03 it's going to be like another app on my iPhone right now. 45:06 So I don't like that. 45:08 I think I need to be smart and proactive. 45:11 And I think the third thing would be personalization. 45:16 It's more like it has to have the... 45:19 Sorry, personalization in terms of the experience. 45:21 It has to learn from the interactions between us. 45:26 So that's it. 45:27 If this teammates help me with the marketing or design stuff, 45:30 which is I'm not good at it. 45:32 It has to learn about my taste, about my preference, right? 45:36 Like, okay, what kind of designs I would like? 45:39 What types of... 45:40 Which artists I love most? 45:42 So it has to be learned from the interactions 45:46 and iterate from those experiences. 45:49 So I think those are the three aspects 45:51 I would love to see from pretty good AI teammates. 45:56 But again, I'm probably not a good teammate to this agent as well. 45:59 Next, we have some individual questions for actually your companies. 46:02 First one for chat slide. 46:04 What inspired you to personalize the PowerPoint slide using AI? 46:07 And what were some of the biggest challenges you faced so far 46:10 in aligning generated models in cooperate presenting standards? 46:14 Yeah, that's a pretty good question. 46:16 So first, for my own experience, 46:18 when I'm actually creating slide, 46:20 I did a lot of copy pasting and a lot of alignment 46:23 and a lot of searching. 46:24 So basically, all of those things are actually very mechanical. 46:29 And if something is mechanical, 46:31 this means that you can actually automate that through AI 46:34 or through a program. 46:36 So that's why we started this. 46:38 And we kind of imagine slides as a canvas 46:41 with a lot of moving parts. 46:44 It's not just one paragraph. 46:47 It's actually also visual. 46:48 So we want to increase how the AI 46:53 is going to understand this model. 46:56 That's actually something different 46:58 from a lot of other similar products approach. 47:02 The challenge we are facing right now with corporate standard 47:04 is that a lot of corporate actually has their design system. 47:07 For example, they will use the images from a certain library. 47:12 They will use the icons from a certain style 47:15 and they probably choose a certain font. 47:17 They probably choose certain colors. 47:19 So that's actually what we do. 47:21 So we help people to pick the colors, 47:23 pick the templates that they want to use 47:25 and then put all those things on the template. 47:29 This is very different from existing competitors 47:32 like Gamma or Tome or Beautiful AI. 47:35 They just give you a new fancy template 47:39 without actually fitting that into your corporate templates. 47:43 So that's what we did. 47:44 And a lot of corporate buyers love our product. 47:48 Second one, we want to talk to how the future is take 47:52 and deeply human-centered. 47:55 What does this look like in practice? 47:57 And how do you see it evolving over the next three to five years? 48:01 The idea came out because we want to be more capable. 48:05 Me and my co-founder, when we just started startups 48:09 back to 2003, we just feel like, okay, with AI, we can be more. 48:16 We can have extended memories. 48:19 We can let AI to memorize everything we know 48:23 and let them to handle a lot of things for us. 48:27 Writing emails, reading news, follow the trends and things like that. 48:31 And so that's why we started building the AI knowledge base. 48:36 And later on, when the agent terms and architectures 48:40 like the frameworks, like MCP came out, 48:43 we just feel like agents are probably a better way 48:46 to making people more capable by giving them an extra brain. 48:51 So we still believe in that. 48:53 And we're trying to use the agents to help people 48:55 to be more capable of doing a lot of things 48:57 they're not capable to do today. 49:00 Remember, memories, more information, 49:04 process information and data more efficiently, 49:07 and even do a lot of things they cannot do today. 49:11 For example, marketing people like to do the high-level designs 49:15 for designers to writing codes or even actual backends 49:19 on building some products directly. 49:22 So within three to five years, 49:25 I think that's a pretty long time period in the AI area. 49:30 I think within two years, there's a lot of things going to be changed. 49:34 Today we have agents, we have a lot of great support from MCP from A2A. 49:39 I think even within two to three years, there's going to be huge. 49:43 I think we talk about the one-man startup 49:46 after the LLM came out a lot on AXA and on social media. 49:51 I think that's going to be pretty popular in two to three years. 49:55 One people can start a startup doing a small business or something 49:59 because with the AI and agents and AXA brains, 50:03 they're capable of doing a lot of things and not just startups. 50:09 Even for education, for kids, 50:12 they're definitely going to be capable of doing a lot of things 50:14 that are not capable to do today. 50:18 That's what we see in two to three years time frame. 50:23 Thank you. 50:24 Next question is for Joe. 50:25 You're building a framework, right? 50:26 So definitely there are many people that adopt your framework. 50:29 So what are some real-world cases you see so far? 50:32 Where these agents could already be making a big impact? 50:35 And also, is there any specific industry or task 50:38 you see that people are doing agent using your framework a lot? 50:42 There are different people using it for different purposes. 50:46 But right now, we're mostly heavily on e-commerce and financial services. 50:50 So workflows-wise, sales, customer service, recruiting, 50:55 internal back office, due diligence are pretty popular. 50:59 What's your predictions for AI agent development for the next three to five years? 51:04 I think talking about three to five years, it's totally different. 51:06 AI is evolving really, really, really fast. 51:10 You know, like a lot of things are going to be changed 51:12 within six months or just a year. 51:14 So it's hard for me to think about three to five years, 51:19 but more about one year or two. 51:21 Yeah. 51:23 Yeah, that's my prediction. 51:25 Look at Google I.O. 51:27 I never imagined like Google can release that much consumer apps 51:31 and to integrate a lot of agents into their existing products. 51:34 But they did. 51:36 You mentioned once today about the agent 51:38 iterations that they involve, right? 51:40 Agent, you mentioned that agent talk to agent, 51:42 maybe they can learn from those agents that involve an iteration. 51:46 Then it will happen or do you see any possibilities? 51:48 Yeah, for sure. 51:49 I think so. 51:50 It's also like a paper released by Google. 51:53 It's Richard Sutton and David Silver. 51:57 I think it's a huge thing. 52:00 I think probably, I don't know that much, 52:02 but I assume a lot of research is being conducted 52:06 on that, the area or the era of experience. 52:10 So basically AI, like they're going to leverage 52:13 reinforcement learnings in test time 52:14 and let the models and the agents to upgrade 52:20 from the interaction with humans. 52:23 So that's definitely something like we want to see 52:25 and we want to learn. 52:27 And if anything comes out, 52:30 I think there should be something, you know, 52:32 in terms of the model level or in terms of framework level 52:36 within two to three years. 52:37 That's for sure, I think, 52:40 because we do believe in that direction 52:43 because I think that's something we need to do. 52:45 Because for agents, right, 52:47 they need to be able to evolve 52:50 and upgrade themselves from the interaction. 52:52 Otherwise, it's going to be really hard to scale. 52:56 Yeah, I also have additional question for Joe. 52:58 Since you're a professor from Columbia, 53:00 if you are going to open a course about agents, 53:04 what kind of important elements 53:05 or whatever topics that you think should be included? 53:08 So I'm actually thinking about doing a vertical course, 53:12 agents for financial services. 53:15 Not only all the use cases 53:16 and then what models are good 53:18 and what models are not good for certain things, 53:21 but also about backend systems. 53:23 How do you do autoscaling? 53:24 What database should you choose? 53:27 How do you reduce latencies and everything together? 53:30 We also see a lot of AI agent startups. 53:32 I'm curious about the go-to-market thing. 53:35 What are the challenges, 53:37 whatever things you experience 53:39 when you try to make your product go to market? 53:41 Can you share some of the experiences? 53:44 Yeah, so first, I think it's a very competitive market. 53:47 So the go-to-market strategy needs to be clever. 53:53 What we figured out is that 53:54 you need to find out the right channel. 53:56 For example, what we figured out is that 53:58 if you go to Reddit, it's a very good channel. 54:01 And there's a lot of people 54:04 with personal knowledge who are on Reddit. 54:07 So we actually did some posting on Reddit 54:10 and it seems that they generated very good traffic for us. 54:14 And that's what we are going to continue. 54:17 Three of you, if you try to reduce three things 54:21 for the AI agent you just developed, what are those? 54:25 It's very hard to do this kind of prediction. 54:29 I think there's too many moving parts. 54:31 Two years ago, I would do chat slide. 54:32 But now, if I'm going to start anything new, 54:35 I probably won't do chat slide. 54:36 I probably will do something that involves 54:39 a lot of more complexities. 54:43 But I mean, a lot of interesting decisions are made 54:47 when you don't have all those information. 54:50 There are just a lot of opportunities out there. 54:52 It's hard to say which one will win. 54:54 I really see there's an interesting topic 54:58 about vertical agents versus horizontal toolings. 55:01 I think both of them will be flourishing 55:04 right now at the moment. 55:05 How about you, Anessi? 55:07 It's hard to say. 55:08 As I said, everything is changing really fast every day. 55:13 But if I start from the beginning, 55:14 I would directly jump into the agent space. 55:17 I won't touch any knowledge base and things like that. 55:19 I think I like a lot of imagination two years ago. 55:23 And I think the paradigm has shifted. 55:28 So I would like also focusing on the interaction part 55:32 between AI agents and the human. 55:35 I think that matters a lot from our experience today. 55:40 Because we do see a lot of people and users 55:42 not using some agents just because of the interaction 55:44 or there are a lot of things they can't do 55:47 on the results AI generated. 55:50 So that would be definitely two parts. 55:53 Jump into the agency directly 55:54 and spend more time on the interaction 55:57 between human and agents. 55:59 Yeah. 55:59 And is there anything that you want to add to the audience? 56:03 Yeah. 56:03 So first, I want to do an advertisement to our product. 56:06 If you haven't tried and if you are a knowledge worker, 56:08 feel free to try chatfly.ai. 56:11 And second, as I mentioned, 56:13 we are trying to help people to... 56:16 What we experienced with our own platform 56:18 is that we figured out that Reddit 56:20 is a very good source for traffic. 56:23 So if you are also interested in that, 56:25 please feel free to reach out 56:27 and we can discuss the strategies 56:28 for the growth of your product. 56:30 Yeah. 56:32 Never under-hype about AI. 56:34 Everything can be changed overnight. 56:36 So if there's something you think AI is not doing good, 56:40 take another look tomorrow morning. 56:42 Thank you. 56:43 How about you, Joe? 56:44 Check out our colleagues open source. 56:47 Ark is NeurosArk. 56:48 Lex is Lexicon. 56:50 It's just a better way to start building our agents. 56:53 Don't think about using other things 56:55 and then very painfully have to strip the way 56:58 to do scalability. 56:59 It will save you a lot of work. 57:02 Thank you. 57:03 Thanks for your time today. 57:04 Thanks, everyone. 57:05 Thank you. 57:06 Nice to talk with everyone.
