 Hi everyone, it's been a while since I came back to Sunset Air actually. The last time I was here I was like, I don't know. I was paid to exhibit and stuff like that. Are you school? Probably. They do all these math excursions. Sunset Air, hi everyone. Good afternoon. So I'm one of the co-founders of a company called Feud. So we actually focus on the mission for enterprises and governments. And today I'm looking to share with you guys a little bit more about building a micro-AI company in 2017. So, by micro-AI I think, what is micro-AI? Why is it not AI company? So I think the way that I personally define that is essentially a smaller subset. Essentially an applied AI company focused on solving a more niche problem within the entire sphere of whatever you can do, from image recognition to creating whatever it is. So I think, let me start by introducing ourselves for me and my company people. So we started about two years ago in Singapore, out of college. I was working at Twitter back then. So I was spending the day in the office and then after dinner we would spend block 71 working on solutions, working on products. And I think our first product back then was actually a product called Graphite. So it's a collaboration tool. Since then, about more than a year ago, we started moving into chat automation. We went out to San Francisco. We get a set of angel pads based in New York. And we started working with a lot of other companies, customers and so on. So right now we work with the government of Singapore on chat automation. We also work with other enterprises and e-commerce and financial services as well as pharmaceuticals and healthcare. So really trying to broaden the applications and look at how can we automate chat conversations. So what I'd like to do today is just in my short sharing to share more about what exactly is happening and I'm sure you guys have read the headlines around the world about AI and so on. But really what can we do with individuals? Whether you're looking for a company today, or whether you are graduated, or whether you are already running a company. Like what are some of the things that you can possibly do to introduce AI solutions into your product or to build an entirely AI focused company? And then of course how to get started with the micro part, like for example product, data and so on. So let's jump in. So I think if you look at, you know, it's really hard nowadays when you read TechCrunch or when you read any of these like news blogs not to transform the world of AI, right? It's like everyone kind of adds AI to anything, like AI for transportation, AI for classes, AI for whatever, right? So I think if you Google all these different trends and you keep up with it, it's really really difficult. But I think for us, you know, what we like to think about is what are some of the interesting things that really happen? And I think, you know, the AI development has been pretty exciting. I think if you look at systems like Liberatus, they actually managed to beat a couple of like really professional, like no limit heads up poker players, right? And the challenge with poker is that it's actually an imperfect information game. So it's actually pretty amazing. As compared to Go or chess where it's actually a perfect information game. So I think the great thing is that they actually chose this format of game which is no limit holden and of course heads up which is a little bit more confined. I think the next challenge for them could be to use the same system or a different system to solve multi-player poker. It'll be more exciting. So of course 1.7 million is a very significant amount after playing for 120,000 hands. I'm not sure how these people actually survived but you know, there's up to anyone to guess. And of course in AI space right now, the big companies are acquiring a lot of smaller companies. So I think one of the reasons why that definitely makes a lot of sense is because some of these smaller teams, they have very strong execution capabilities or they have very strong research capabilities. And then if you look at, you know, can they bring a massive solution to market with massive amount of users? Probably not. But if they work with, for example, I say Amazon or Microsoft they can potentially sort of allow their applications or their systems to be able to be used by tons of companies, users and so on. And here itself, we've definitely seen a lot of more developments as well. So some of the logos that has already been recently acquired is not on the list. But yeah, just a sense of how things are. And of course I think, you know, in our space of chat automation and so on, we see that in the last one year there's been a lot of developments, primarily really due to a couple of the messaging platforms and kind of this whole chat bot. So I think the first thing that happened was, you know, Facebook launched the Messenger platform last year in April at date. And since then, we've seen about, I think it's probably now 40,000-ish bots in the market from bots by big brands to smaller companies and so on. And I think in general, we've seen some really good experiences. For me, I spend one weekend, three days, talking to about 250 chat bots on Messenger. And after that, I just don't want to talk to anymore bots for like a week. Yeah, really interesting experiences from like dating bots to like, you know, news bots and stuff. So it's just like, yeah, it's interesting. And of course, in the self-driving car side, you know, you see a lot of great advancements and really nice that, you know, we do have a company in Singapore who's at least located in Singapore doing interesting things as well in the economy. And of course, for self-driving car side, you know, there's problems like immigration to, you know, the street rules and regulations and all these kind of things that needs to be considered. And of course, there's a lot of business problems as well like regulators and so on. So of course, you know, as people in the audience as well as myself, I mean, do we just sit back and wait for the companies to just like introduce tons of great solutions, Alexa, whatever that is, or can we do something about it, right? So I think I like to think that the latter is possible, meaning we can actually build some meaningful products in the space to help some people. And just to cite an example of how some people took it to a really strong extreme. So this group of guys, I think three or four people at this NRPS conference actually created sort of a concept called Rocket AI. And they dubbed it as, you know, this next generation technology temporal recurring optimal learning. No one knows what it means, right? And some people here in the conference were like putting it out and everything, you know, they're like, made a huge hoot about it. And guess what? It's all hoots, right? As you guys probably already know. But the numbers speak for themselves, right? So what happened was after, like, I think during the conference, they were able to get a, you know, party going on with like 300 people attending. They even had a police attendee event and, you know, of course, the amount of value, the amount of people that, I mean, it's really close to the thing that investors actually reached out to invest in a company, like five of them. And, you know, of course, like a lot of people were like sending resumes and all that kind of things, right? And you think about the typical sponsorship at an event like that. It's like $10,000. And these guys spent like 500 bucks and you know, like a huge amount of interest, right? And whatever they're doing. So I don't know if this is the best way to put a company. Probably not, right? Because, you know, more of these kind of things go on that I think people are just going to call AI like a bluff or something, right? It's not a good idea. So I think maybe just to default back to someone who actually knows what he's doing after incubating so many companies at Y Combinator. I think Paul Graham stated a pretty good point, which is, you know, you need to start, instead of thinking about problems, like thinking about problems out of thin air, you need to start noticing problems around you, right? So one of the things, you know, just to give an example. If that's it, for example, you have a friend, a farmer, who is doing a lot of agriculture and every single, every single day he has to go and check the weather forecast and if he has to, you know, do his own manual on the soil and stuff just to make sure that he finds it healthy. You know, if you notice that problem, what you can do is potentially use a solution to essentially maybe like problem APIs and you'll call it and automatically predict whether, hey, what is going on? So I think those are some of the things that, you know, as individuals, you know, when we see problems around us, we can potentially find solutions to help these people with the problems. And of course, I think this teenager from England did a pretty good job. So he, of course, received a lot of parking tickets for some reason and he built a chat bot where, you know, essentially he helped people to automatically appeal for those parking tickets. So I think the interesting thing about those kind of use cases is that he really took it to like, hey, this is the problem that I have and then he figured out all the automatable parts of like, for example, the website and the responses and so on and he helped people to automatically do that, right? And after he achieved the standard level results, he's not stopping that, right? He's taking that on the next step by looking at applying the same methodology to train the concept to helping, you know, refugees who are seeking asylum, you know, escape from the home country, which is amazing. So I think if you think about, you know, all these kind of solutions, like individuals like that from Stanford, for example, can, if they can do something like that, I don't see why we can't, right? So that's very how... So right now, if you guys are kind of interested, like, for example, building something now or something, usually what are some of the things that you guys would think about? You would be like, hey, how do I get customers? Or how do I actually do the actual product? How do I go to the market after the product is built, right? Is it going to be tasteful ads, is it going to be viral marketing, whatever it is? And also how do I assemble a team, right? Am I going to find my college buddies and am I going to find, you know, my friends from school or what would that be? What's the launch, right? Are you going to be... Is it going to be a local solution? Is it going to be regional? Or is it going to be... So all these questions kind of come in once you start thinking about all these kinds of things. Once you identify the problem you're going to solve, once you realize that, hey, you're really passionate about it, these are some of the things. And I think maybe just for me to highlight a little bit more about the product side of things for a micro-AI company. Really, I think for product is one of the most important things. Like, once you've decided all these other things, which are of course all the start-up books or whatever teaches you, then the product side is really, I think, a little bit different from an AI company's perspective. So I think to quote Mr. Andrew Lone, he's really good at consolidating that to simplifying the product perspective from the two things, right? So I think to share what he said was essentially if someone asks, what are AI products good for, there are a couple of things that are really good for. And I think one of them is he's able to help us to be able to solve problems that humans can potentially do in less than a second. So it could be, for example, showing a person an image of a cat and a person instantly tells, hey, this is a cat, right? Or a dog, for example. But let's say, for example, it's a task that even a person with a medical degree, after looking at a problem for 100 for one hour with multiple data sources and so on, is still not able to figure out, then yes, you can augment that, but it may not be good at actually making the correct decision. So I think we are seeing a lot of applications from various degrees of complexity and I think the easiest ones are those that can be solved by doing it in less than a second. And then, of course, predictive tasks. So we see these applications in many areas. We see, for example, Facebook surfacing the most relevant content or whatever you call it, that you would most like to click on and they tend to be bossy articles for some reason. And then, of course, some of the skills for us are applying potentially to lead scores, using all this data that you have to be able to figure out, hey, is the person going to convert? Is the company going to convert if you want to pay customer? So I think, very in this in mind, these are some of the good friends that I think about if you want to think about problems that you want to apply for today in whatever you're doing. And then it's really about how to get started. So I think the idea is that you want to get more users. But once you have more users, these users are going to generate a ton of data on your application of service. They're going to click a lot, they're going to share a lot, they're going to do a lot of action. And then, of course, with this data that you have, you can essentially design experiments, you can design algorithms, you can design systems to be able to see how you can potentially do predictive tasks or all these kind of interesting things you can do. And then, by servicing better than content, by servicing more relevant things, then, of course, you might be able to get a better product. And a better product, hopefully you'll get more users. And it's really what everyone is trying to achieve. So I think this loop, that's how we think about chat automation or any of these products as well. So let's dive into the data part. So for data, I think every company and every company, for example, SOTA or a big company has a ton of data everywhere. But it's really about how do you kind of sort of prepare and understand the data sharing is very important. So the first part is, potentially understanding where the data is going to come from. Is it going to be from an application? Is it going to be from your offline services, offline machines, and so on? And then it's about injecting the data into your database, for example, and then preparing the data. So preparing data, you can also label the data. You can figure out how to inform the data as well as the data. And then it's about how do you then publish the data to different kind of sources. And of course, a lot of services and stuff and people to consume the data. So now that you have kind of formed a data strategy, then you want to kind of collect more data, right? Bring more data into the system. And I think a lot of people that we have seen just go to companies and say, hey, give me another data. I'm going to do a lot of magic with it. And you're going to get this X amount of improvements in sales, revenues, or cost savings, whatever it is. I don't think that's a good way to actually show results, even though I think some of those sales pictures do get through. But I think I know the data. There's many things that as a company you can do if you just don't have enough data, right? So I think this one type of drop scope is recurring and coming up right now. It's called AI trainers. I don't think many people quite AI trainers, but that we do see some companies do that. So that X.AI, they are kind of the scheduling app. I think they have a very related job posting on there into this profile. Let's say it's AI trainers. Essentially, if you look at a drop scope, it's people who go in and label data. It's a mechanical process. So if you think about it, for these kind of job requirements, yes, you can have people do it in the house, or you can do it via an outsourcing agency, or you can have people doing that, you know, either through mechanical process or so on. By the end of the day, the purpose of having people label data is so that you can essentially generate more forces for your prediction agency to be better, right? And after you have done that, generally, I think the output would be more significant. Another way that some companies actually do it is, say, for example, for clarify, they are an image recognition company or as an edge user. So what they do is that, of course, they sell to enterprises, but, you know, when they do that, sometimes, they don't even have data. So they actually spun out an application of forgery. What it does is that it allows people to upload photos, tag them, and so on. And of course, they try to make it a nice user experience, so a lot of people started doing that. And then they started doing that. They essentially were able to capture all these tags photos, and these helped them improve their file from the back end so that they can generate more revenue from enterprises. So they started capturing a lot of different systems. And of course, sometimes, after you answer a question or after you do something, someone would ask, is this relevant? Yes or no? So all these kind of settled things. I would say data traps, but maybe like just traps, just like systems where they help with the type of user input, it's actually very useful for a data generation. And of course, I think right now we're seeing a huge movement for public data sense, right? Like, you know, if you look at companies, you know, corporate sources, some of the data sources and stuff. I think the great thing about that is that it means that you're going to have to start from zero when you are thinking about designing a system. And you can get a lot of interesting corporate system data that you can play with. For example, if you're designing an email automation of a company, you can potentially use an N-ROM email of corporate. I'm not sure if that would make your system good. But at the end of the day, and say for example if you're building a chat bot, it tells you the bus timings at bus angle. The good thing is that you can actually now use the APIs from the LTA and agencies for example to be able to get the bus timings. And by doing so you are able to then provide a good service for the user by focusing on the personality and focusing on the other things. So I think that's pretty interesting and hopefully more and more independent companies can do more of that. And I think we've seen companies, let's say for example if you can get all the data in the world, great. But in many cases like for example if you are trying to solve problems in like medical imaging or like cancer research or something like that, you may be in a very specific domain field that is very difficult to get access to open data sources or you know, or generate data. So what you can do is you can pick an error domain. So let's say for example in emails, you can't solve about every single problem in the world but you can solve every single problem from like proposal sending to whatever. So I think what XRAI did was pretty good with that. They actually focused on an error domain which is email schedule. So instead of burning the forest on every single thing they just pick one set of trees and it just probably exhausted them. So I was recently at a talk where Dennis, the CEO of XRAI was at and he shared that at this point right now that someone can ask to schedule a calendar and write it. So I think by being comprehensive in one narrow domain it actually provides particularly to have more data around it which is really exciting. So we'll see where this goes right now there's a battle in this email box space with the XRAI you know, calendar help in Singapore and so on so we'll see. And of course I think we've seen some companies being very innovative and generating data automatically so say for example if you're a you're trying to build a self-driving car system and you don't have enough like night time driving data what you can potentially do is to log on to GTA and not play the game but generating or essentially have the car just drive around automatically and generate a lot of data based on the game environment and so on I think of course if you want to generalize that from real world driving you do have to factor in other things but it's a really good way to at least experiment with very important but of course now gaming is again the point where you can actually limit the real world so it's getting pretty interesting and of course we've seen companies like Facebook they also use this data generation strategy to essentially take a paragraph of text remove one word and then use that as a training data for systems essentially to find if hey are you able to comprehend the paragraph that can potentially generate data automatically from for example existing data that you already have and I think another day for once you have all this data you have picked the domain you have done all these great things it's all about fast iterations it's not just about being benchmarks but it's also about hey how can you provide a service to people and how do you actually experiment and deploy and publish so I mean it's a solution and they will deploy faster so I think that's something that of course you'll see more and more such frameworks coming up so that's very exciting yeah so at the end of the day I think the way that we think about it for us it's really that hey AI has various applications for AI assistance right and I think if you break down this sort of AI in so many different components it's really meaningful and I think it can provide a lot of or a lot of people in different industries and then of course you know I think the available tools that we have right now it's pretty mature and you know if you are interested in solving any of these problems you can actually solve today almost and of course you know building better products in the AI space is really about getting more data getting more data or picking a narrow domain are these kind of interesting things that you can do and I think coming back to what I was talking about earlier because you know building a just like two anecdotes about you know what we have a personal experience in terms of like how that has been so from a team perspective we realize that you know we needed to have a team that was not just technical but also able to handle the different components like design, marketing and all these other different sales perspectives right because any other day we just wasn't doing product but we're not able to sell it then we have a problem right and of course you know AI products the way that we sell it the way we price it the way we lock it is almost exactly the same but the understanding on the business level I realized on this product it's actually pretty new still because of how nascent it still is so there's really a lot of adaptations required for us to sort of edit the client and also tell them hey it's not about the buses it's not about what is like the type of people telling you it's really about hey what is the result that you can drive and you need to have the right team to be able to structure to do that and whether it's in Singapore whether it's in the US whether it's in Europe whether it's in the net right as long as the team is always working together on these problems and able to grow in that kind of team so that's kind of been a pretty interesting problem journey and then I think about the global part so we realize that after going to the US the reason why automation in the US is very meaningful and very very looked upon is probably also because the cost of labour in the US is pretty high so it costs a lot to actually hire people to be able to for example do customers support do sales and so on so as a result they either outsource that or they try to look at automation opportunities but in terms of in Asia I think we do see that that is still not at the same level it's as neat as as in the US so it's interesting how the contrast in the markets work as well and also in terms of the environment then they can deploy solutions so I think that's some of the things we have in comfort and I think so I think just to wrap it up building an IQAI company it is about of course having a really interesting problem in their solving it's really about trying to build a really good product but at the end of the day there's a lot of other things as well when it comes to really a company that we need to really solve and it's really about solving this problem there every single day and hopefully getting a kind of priority can be in the presence of hype and of course for us you know if you guys are really interested in question and answer problems very interested in our key company you can send an email as well so we are trying to really focus on the chat animation problem and really excited about what's going to happen thank you yeah so with respect to chat there are a lot of thoughts and there are a lot of frameworks that are being supported by most companies to solve specific questions so with you as a company if you are planning to do like this would you post each one of them in a different way or would you aggregate all the data so that the rest of your chat also gets smarter would that be a good thing to do or would that be better okay this is actually a really good question so you said about every company has kind of a different need I think maybe to answer that question the first part of question is what are the needs so for example let's say maybe let's say a company in the shipping space they have this like SMS system that goes out to people but they want to automate that so they have this specific need and they need to be on-premise and private products and then you have another company who will be an e-commerce company who needs it to be a post-purchase shipping experience right so definitely we are seeing some things that are customized so we work like integration partners kind of stuff but some things that are very common which is actually the FAQ part so the frequently asked questions or the questions that are essentially how do you train the models to be able to understand what the user is saying and then actually answer the question that actually is something that everyone so please but at the same time the solution is do not there yet and also on top of that the software component like the management controls and also the department processes so from a software part of perspective we are able to provide that so I think that's the first part of the question about hey when you have these customizations how do we then ensure that it's still a product and it's not just like one single department every single time and I think the second part is then sort of choosing the right people to work with so it's really a business problem it's like hey there's always going to be a ton of people putting in different directions to work on so I think for us we pick custom areas where we are interested in unfortunately not areas that make a lot of money so we are interested in for example support we are interested in BUB to C kind of use cases we're not exactly super interested in the BUB entirely BUB use cases we like that the end result of whatever we do on the chat automation side can help people to sort of maybe have a better experience with the companies right so that's kind of either proactive self-experience or ends up being a reactive support experience thanks so let's see maybe you know all the frameworks around chat but it's an easy one to get started with easy one to get started I mean I mean it's not being biased but we are Microsoft farmers and we are we work with Microsoft but I think Microsoft BUB is very good to get started with or you can to get for example maybe you are I think Microsoft BUB is trying to to write the chat bot in like whatever language that you are comfortable with being a jazz whatever let's do something that is used I think that's really the best way to get started for us we were like hey let's just like you know writing you know we were familiar with that and just like whatever and we didn't really like use any of the string that we guys really wanted right instead of thinking about being an SK I just wanted to quite a bit and let's just get to a faster way to do that I think you can go with companies for those kind of right if you are trying to explore yourself you want to understand the fundamentals then it's better for you to just have a little thinker whatever you are most comfortable with I think that would be the best and then I think personally I'm really excited about a bot called digest.ai so essentially it's a select summary bot that sits in Slack and summarizes the conversations and sends back the keys along with digest or an email and when we launched it like last year we decided this is a fun project but then since then I've worked with about 300 companies it's a free product so about 300 companies including NASA they start using it so like oh shit this is actually something that the media talked about alright show me your email guys thank you