Tringular Talks Presented By Raj Neervannan

Great, nice to meet you all. Thanks for having me, Martin. It's a pleasure and an honor. Many congratulations and best wishes to the team members who have built this group. I hear great things about the Agreement Circle and the wonderful work you guys are doing around Triangular Talks. It's a pleasure and honor to be speaking here, and I hope I can add some value over the next hour or so. What we're gonna do is just talk about a few seconds about myself and then get into what I did in the latest venture, which is AlphaSense. Then I'll take some Q&A after that. We can go to anything else you want. Does it work?

Cool, sounds good. So just a brief before I jump in and share a quick intro about myself: I am Raj Neervannan, I'm the CTO and co-founder of AlphaSense. My other co-founder is Jack Coco. He and I met when we were doing our MBAs at Wharton between 2006 and 2008. The idea for this product came about out of our joint experiences. When he was a Morgan Stanley analyst - if you guys know, Morgan Stanley is an investment bank company - he was an analyst working on researching stuff, and we were doing something similar as part of an MBA, you have to write a lot of essays and stuff. So that's when this idea came, and that's what I've been doing since 2008. I'll go into that a bit more. I'm going to go back to before 2008. What happened? I was born and raised in India in Chennai. I don't know how many of you are familiar with it, it's the southern part of India. It's pretty hot, weather doesn't get any cooler. I then did my bachelor's at BITS Pilani, I don't know if you know it, it's in a desert called Rajasthan up north. It was always my first foray into leaving the house and going to a strange land. This was in the late 80s, so the writer - that was the first sort of exposure to trying new things. I want to start off there because that's usually where you want to start thinking of the biggest reason why people don't go into something new: because of intimidation of trying new things and leaving the comfort zone. So leaving the comfort zone of a home and going to college when I was barely 17 into a desert was just another first kind of thing. I spent four or five years there, then the next major change was coming to the United States. This wasn't too popular then, to come to the US at the time; it was early 90s, which is the beginning of a big wave of immigrants coming here. I came from a hot weather to a desert to a place called Ohio, very cold. I was not familiar with the Bowling Green, Ohio specifically, but it was a wonderful campus and I really enjoyed studying operations research and computer science there. So as students, as you guys change from school to school, you should venture into new areas, and this is part of growing up. These things actually played a role in me trying things and being open-minded about going to new places, meeting new people. From there on, I spent several years and went down to Miami, I worked for a few years and got myself a green card. Once I got it, I was able to try what I wanted to try, which was to build products and build companies. That was the beginning of my adventure in 2000. I started my very first startup with some funding and moved to Austin, Texas - another move from Miami to Austin, Texas. I raised a little bit of money based on just an idea. Not many of you may have been even born by then - those are called the dot-com era. The first round of internet companies came out, so I was smitten by that idea, just like many of you should be smitten by the possibilities of AI right now. I was smitten by the concept of computers in the early 90s, then smitten by the concept of interconnecting computers reaching out to a larger number of people in the 2000s, and that gave me a lot of ideas to try things. We did a couple of startups - that's why you call a serial entrepreneur, someone who tries things. Success is not a failure, they try to build. I've been with a few times, and then I got my MBA to sort of formalize my education. That's when I met Michael Florida. So that's sort of a brief intro.

Below, let me get into the product itself. I will do that right now. Let's see. All right, do you all see the screen? I'm just going to go into - yes, see the screen. Okay, so I'm going to go to a slideshow mode so it's easier for me also to talk. Oops, that's not the right one. That's not the one I wanted to share. No, that's not it. It's still too early. So sure, not that one. Sorry about that. Let's see, I'm going to try one more time. I have it right here. Let's see. Okay, this one. Okay, got it. Seeing this one now.

All right, cool. So my topic here is how to build a better search engine. As I mentioned, in 2006 through 2008 when I met my co-founder, the idea was how do we find information that is easier to do our work as an analyst. As I mentioned, he didn't have a great tool. During our work during an MBA, you must have been asked to do some research work, go find information for a particular essay. What you do is you go to Google and look for some information, and that's what business people do at work. We felt that it was not working so well. I'll go into that. So that was the topic. I'm going to explain why we did this and how we did that. The paradigm here is, as Martin mentioned, you want to discuss the problems, then what kind of innovation we did, and then what kind of solutions we propose and how we experimented. There is a slight change to this: the real company is that you don't just have an innovation and then a solution, you go through a lot of iterations in the process. Way too many iterations you go through, but you always get market feedback from real users. So with that as a backdrop, the main problem we were addressing was when we're trying to look for information, Google, when you put in a keyword, gives you top 10 results. It tells you 'I found these keywords found in this document' but they didn't really understand how often you mentioned something about a word. It doesn't always come up with other variants of the word. You may mean one thing, and the word you type may have multiple meanings - these are called synonyms. It doesn't really understand that. So that's one problem: it has a strict keyword limitation. Things have gotten better since the AI days, but this is 2008, 2009, so that wasn't so. Then what you get is top 10 links, you wonder what else is out there, and you don't have the patience to go through each one. That's the second problem. Third, you have to open each document and see what's inside, then come back again, so that slows you down. Fourth, what else has happened since I tried last time? That doesn't really flow into keeping you on a time scale. Something came up in the last two hours, something came up in the last day, it doesn't tell you any of that. Finally, as I mentioned, even if you open up a document, you have to go and find one item at a time. These problems don't seem obvious on the surface initially when you are using Google for fun, for simple projects. But when you are doing like simple things, Google kicks you out into a new zone. Most of the products that people used had used first search only did this. Nobody ever bothered to ask the question 'Could this be made better?' And we decided to ask because we felt there was a pain in the process. The first step is not just identify the problems you have, but you want to amplify the magnitude of the problem. How big is the problem really? How much time does it take? You want to magnify and quantify that problem in terms of the irritation you feel or the time it takes to accomplish something. That's step one. Then you want to come up with ways to solve and try it out. So going back to why - you have to go back before you propose something to innovate or solve, you've got to understand why certain things work. You can't just assume that people haven't thought of this; obviously people have ended up stuck in a certain place because of some reason. So you have to give credit to people who have been working on it. To that end, you understand how search engines work. Search engines work by crawling all the web pages, storing all of them, organizing them by some categories, and then ranking them by certain order. When you put in a query, it tries to match your query to the pages that seem to come closest to what you have, and then based on some popularity contest - it used to be called the PageRank algorithm by Sergey Brin who founded Google in the late 90s - based on how many people link back to it, they ranked based on that and some other factors, and then give you the top 10. It worked really well for internet consumer businesses, but if you go down into it, it doesn't work for businesses because the search index process is the culprit. The main problem is when it crawls and finds the links, it just visits each page and grabs the content and extracts the key important pieces of data here and there - sort of like you go to a mall, you don't remember everything about it, you just remember one or two things. It loses all the other context and all the other information you could possibly be interested in. So it loses vital information about all the semantics, all the information inside the page - just key things that it thinks matter to the world. Then it indexes on them. Index is nothing but a glossary of a book. If you go to the back of the book, you'll find that for the 100 pages you have in the first 100 pages, the last five pages are a glossary where it says for these keywords you can find these keywords in this page. You go back and look. That's exactly how the search index works. You just have to go back and find. Then they store that information, and when you ask a question, they just look it up and go back. As you can clearly tell, this is the one that scales because the internet scaled in trillions and trillions of documents. There's no way for them to index every single line and word in each page, so they have to use this glossary of a page to get to the keyword and the page you're interested in. That's it. That's the way a search engine worked. And that is great for casual, quick one-stop shop, looking for any information at a surface level. Then you are thrown into a link for you to dig deeper after that. This is fine for normal purposes, but if you're trying to look for business information - I'm going to go into the business side. Obviously, you are in school, you probably haven't heard of these terms, but business people do. When I say business people, they work in companies that are looking to understand about other companies, and they go look for - or even if you're in the stock market, if you are trading anything, you want to know what is the size of Apple the company. Well, how big is the market? You find the stock market size: 'Oh, it's a trillion dollar company.' That's not the same as the market. There could be multiple products inside Apple, each one has its own market. The iPhone market is different from the Apple Watch market. So each one has a market or a sub-market. This is called total addressable market or TAM. If you just put in 'TAM' in Google and it says 'TAM Apple', it just says things about Apple - 'you want some apple? TAM is 20th anniversary Macintosh' - it doesn't. That's not what TAM means. That's not what business people mean. So clearly, many of the search engines don't work for business people. So this has been the state for Google for a long time. It's been getting better, but this is a recent search, even now it doesn't work that well. So you think, how many of you have used ChatGPT? I can see you all, but I'm assuming most of you have heard of it. ChatGPT is this new wonderful innovation from OpenAI company which allows you to ask questions like this and it comes back and gives you a long essay answer. It will help you, it gives you an answer not just in terms of links but summarizes the answer for you in a way that is like I'm talking to you. There's a very human sense, a human-centric way. But you think that it would know more? Unfortunately, it doesn't. When I say 'TAM Apple', I'm not sure what you mean by TAM, and then I explained it: 'Okay, it's total addressable market for Apple.' Then it goes on to explain what total addressable market means. So I said 'Okay, what is the total addressable market?' Then it kind of repeats. As I mentioned earlier, even on the most sensible latest interface, it doesn't seem to quite address that. So I'm now going to jump straight into our solution. Then I'll go back to why because I want to give you the solutions quickly.

I'm now going to jump straight into our solution. Then I'll go back to why because I want to give you the solutions quickly. Then go if you put the same thing in AlphaSense, my product, 'Apple and TAM', it will find all variations of TAM, it understands what TAM means and actually give you the addressable market and actual size and give you the specific information coming from experts. It curates all of that to you and says 'I have found tons of documents, but you only need to read one particular line in the specific book.' It's kind of like asking a question to a librarian: 'Where can I find this information?' And then she tells you 'Go to this building, go to the shelf, you may find out.' That's not enough. You want to know the specific book, the specific page, and specific line. Then you may want to email the other information that may be in another shelf, another book, another line. So the real search answer should be like that, or a summarized answer would be just 'The answer is this.' Like an expert would. That's what search engines should be doing, but that's not what it's doing. So that's where the source of our idea was: we want to solve the problem in such a way that it is very specific and very clear. Why is this this way? What's the problem with search engines? Why didn't they do it? Normal search doesn't understand keyword values like 'R&D' means research and development. Some of you may know, and you would expect the machines to understand that 'R&D' actually means new capability, demonstrated potential, commercialization - a lot of different ways to describe the same thing. So all those variants are sort of distributed. So you kind of have an explosion of number of terms you should search for, not just one. You should actually search for more variants of the term. Search engines have a hard time ranking even for one term because they have to come up with the top 10 out of thousands of links. How can they find top 10 out of thousands of words at the same time? They clearly wouldn't. So it's not meant for them to solve. So if you go back into the counter argument for the problem, not only do they need to know the ranking, more importantly they need to know what are the words that look like the similar word but not quite. There are other words like 'internal research', 'research operations' that are not quite 'R&D' or variants of the term but not quite similar to the main term. So you ought to know the exact meaning. So you would think at this point, 'Hey, that's not very difficult. I'm just going to go to a dictionary and find out.' You can go to Wikipedia or something, you'll find out all the various meanings, and maybe you can get to the point. But it really requires an expert to understand what it actually means. The dictionary wouldn't understand the real use of the word. You'll use other terminologies. It's quite a complex problem. If you look at the word 'outlook', outlook may be described differently in a dictionary, whereas in the real world, 'outlook' means 'market share' in the business term. So the way we use terminologies is different from how a typical dictionary applies. That's the main problem. It is really a hard problem to solve. What I'm going to do now is get into how we now that we sort of understood the pain point of the problem and the core problems, and then how we try to solve them. What we decided to do was: what was clear was that the business terms were not understood by search engines. There was a clear takeaway. So we decided, 'Okay, we're going to but search engines are really good at finding a keyword once you give it to them.' So we said, 'Okay, we're going to then tag it. If there is a word like 'outlook', I'm just going to tag the word 'outlook' with a lot of different variants of the term in the book itself. It's almost like going back to a book and you find a word, you go to the page and you scribble on the side, 'Oh, this word means the following things.' Then you go re-index it and go back to the glossary and say, 'Now finally, where that is?' So now not only is the original book being tagged, but also the explanations that you have given tagged on top of it is also being re-indexed, so you can now find the actual meaning. In other words, if you don't know, I'm going to tell you what it is, then you're going to re-index and share me the results. So that's kind of one solution to this problem. The second thing was what you want to find is not exactly the keyword on the thing that you want to find; it could be around it somewhere nearby. That's called proximity. Proximity means close by. So not only do you want to find the word, you want to find related words. Not only do you want to find the related words, but also related words that are close by - not right next to it, but next to it - because this allows you to find the semantics a little better. That's the second part of the innovation. Third, you want to show more of this. People had a problem seeing all information in one shot, so we want to see more information but not overwhelm them. So we want to show them not just 10, we want to show more. Now I'll get to how to not overwhelm them, but first we wanted to make sure that at least we got to the key pieces of information and show as much as possible. These are the three main innovations we wanted to do. There are more, but I'll come to that. These three main ones: the way we tag the related words, the way we make the search engine sort of go back and re-index itself and find the new related terms, and also find the proximity close to it. These two dramatically increased scope. Our goal was to first increase the scope for all the missing ones. It was missing keywords before, it was missing the words around it before, it was missing other documents before. So we want to fix all of that. Now in terms of solving how to then present them better, we then decided to build a much better user interface, and I'll come to that in a second. So as I continue further, the first example of how this problem multiplies: when you look at 'revenue outlook for Apple' type example, if you go to Google and say 'Revenue outlook for Apple', Google comes back and gives you some examples. The real clear problem with Google is that it doesn't use the keyword hits, it doesn't use the actual location in the document. We have to go through each - I went through this before, I'm reiterating here - and you have to download each one, and you cannot generate a summary. There are a lot of problems with getting to an answer as opposed to finding some link. Putting in a keyword and finding some link is not an answer. That's kind of like, to my prior example, going to a restaurant, if you ask someone 'What's there to eat?' and the person just says 'Here is a long 40-page menu.' That's like 'Well, that's too many for me to look at.' Or then they ask you 'Well, I don't know, whatever you want to eat.' That's not an answer either. You want to be able to give them specifically 'What's a special today? What would be great over here?' There's a lot more information you can share and make this process a lot quicker. That's the problem with all search engines: they don't get you to an answer, they just give you even more information. So we want to avoid that. Good that we expanded the keywords, good that we expanded more documents, now we made the problem even worse in some ways, right? So we have to fix that. How did we fix that? Now comes the interface. This is the problem: now we find more important things we were missing before. Now how do we solve going back to what actually mattered? We said, 'You know what, we're going to find relevant pages and first you're going to make it easy for people to scan inside the document without necessarily opening each document.' So we're going to provide them a list of documents on the left side as opposed to what you found in Google which is over like this. As opposed to that, we're going to provide this on the left side. And Google didn't really do anything on the content inside the document; they just gave one line here, that's it. Instead, we said we're going to provide more insights inside each document, exactly what you should be reading in response to your query. If you type 'Apple' and you know what Apple means because it's a ticker symbol that you can find in the business, in any stock market, and you can find 'Revenue Outlook'. We have, if you look at it as a yellow line underneath it, it means that we understand what you mean. That's the yellow line should say 'AlphaSense understands what you mean. I got you covered.' And so it expands that to this type of a 'Revenue Outlook'. It expands this understanding: that's what it means. 'I get what you're saying.' That's what the yellow line means: 'I get what you're saying. Now I'm going to give you exactly what you want inside this document. I'm going to give you the document as well. I'm going to show you exactly where to look and save you time by clicking through. When you click on it, it takes you right inside the page.' By doing these things, we were able to not only show more information, pertinent information, rank them even better, and we are able to cut the time so people that used to take hours and hours now take minutes or seconds.

Now, this is great, but how do we possibly understand the semantics for all possible words? After all, there are thousands, millions of words out there, each one has multiple ways of saying that. How do you possibly teach a machine so many different variants? Well, it has been a difficult problem, clearly that's why it wasn't solved. But over time, in the last seven, eight years, it's been solved better with AI, artificial intelligence. We are able to make the machines not just respond to what you're trying to say with actual code, we are also able to make the machine understand the relationship between words and able to learn by itself. When you learn more math, you will appreciate the algorithms behind how the machine learns. So I would really urge you to pay more attention to calculus next time when you're paying attention to your equations and algebra next time. It really is there a reason why you learn algebra 1, algebra 2, calculus 1, calculus 2. It's going to get really difficult to understand how you can make a machine learn because the machine is just going to say 'You wanted an answer, why I got an answer? It's different from why I want to really get to that particular answer. How do I learn?' You're going to say 'Well, it's kind of like a blind man leading a blind man. You're going to say make a left, make a right, make a left, make a right.' That algorithm is all based on math. It's going to compute what is called the difference between the two words and then it's going to iterate through the whole process. I'm not going to go into the math level here, that would be another topic. But I want to just focus on the innovation portions at the high level, at the interface level, at the product level. Maybe another topic, another day we can look into how the math works. But the idea is that you want to use machines to understand words and their meanings and their related meanings, and that helps us solve the problem of coaching the machine manually one word at a time. That's really our problem. Now we solved it, but we can scale it. Scaling means making it work from - okay, we can teach the machine a thousand words, two thousand words. How can you teach the machine a million words? That's where the AI comes in. Second thing is ranking these words. Ranking is supposed to assign scores where we rank based on how it understood what they find inside, what kind of information is there, how relevant is it based on what's how much information was used to describe my problem, how much information was used, how much importance was given. If you read a newspaper, there's a headline, the thick bold mark, and then there's paragraph headings and stuff. They are using words in the paragraph for a reason because they know that it's important. So if you want to know what's important in the document, you will simply one approach would be to simply start reading headlines and paragraph headlines. It's the simplest way. Same thing here: if a word is important for a particular document, you look at the headlines and look at the paragraphs, first line of each paragraph. Things like that. You start translating what humans understand to be important back to the machine. That's how we started ranking each page on the importance of what it means. Then we looked at the specific hits like a search engine would do, then we rank them to its proximity to the important words lines. So there are a lot of thought processes that go on into figuring out what's important. What's important to us humans should be translated to what's important to computers if it really has to understand what you're trying to say. If you want to use what Google does, which is rank based on popularity - it used to be at least that way, after it obviously understands your content way much better - but then it was ranking based on just popularity, and it was not quite getting there. So those are the kind of - I'll leave you with the top approaches on how to rank, how to index, and how to collect this information, how to present this. So those sort of turned out to be the solution for our innovation big ten for the problems we're trying to address. And how does it actually map into a solution for end users? It makes them save a lot of time. As I said before, they can work faster and smarter. They don't have to control-F. If you use Windows, you use control-F to find something inside a document - many of you are probably using Mac even - for document it's control-F, a keyword combination you use. It's really annoying to have to go back one document at a time. But in our business where time is money, people are annoyed by this and they don't want to lose time. So once we solved the problem, we were able to surface meaningful results quicker, and they can spend more time for higher value tasks. Same thing with - I'm sorry - so it's not just time savings. What happens is that when you have more time, you're able to pay attention to more important things. That's the kind of important side benefit. Second is the missing information is a huge problem. What you don't know, you don't care for, and that can be a huge blind spot. It's like driving and not looking at your rearview mirror at all. You probably don't want to do that. Whenever you learn driving, you would look at all sides and behind constantly. So it's important to not miss critical information that's around you. Same thing with searches as well. You can't just do one search and accept it; you have to be critical about what you're missing. That's the second eye-opener. Third is when you're able to do things better, guess what, you are able to focus on higher order activities which is idea generation. You're not toiling over basic stuff, so you can actually think of what else you can do. That's idea generation. Third is when you're able to do this, you are able to pay attention to each other and make the next person smarter, and by sharing these ideas and collaborating, your productivity goes up as a team, not just as an individual. So these things have orders of magnitude productivity benefits. When you have productivity benefits starting with time savings, then you have typically a winning product because it's going to be sticky. Sticky means it's going to make you want to come back, like Facebook or Snapchat or whatever tools you guys use. It's sticky because you want to come back to listen to what your friends are saying. So it's kind of like that: it makes it better. So when you want to come back and look for it, so that was a story. And then even though in 2008 there was AI, wasn't quite...

As we evolved and built the product and proved the value, AI kicked in around 2014 time frame and was beginning to show its value. We pivoted to doing way more AI-centric search as opposed to statistical machine learning type approaches that were prevalent before then. And that is the foundation of most successful search companies out there, including Amazon, Google, Netflix, etc. So what does AI mean to us? As I mentioned before, finding words and related words would be very important. But AI gets better through interactions, especially through real-world interactions. So we want to apply to those problems that cannot be defined by if-then rules. Like trying to find out what's a cat and what's a dog: you can't just tell a cat is the one that has got furry hair and pointy ears; there are way too many descriptions you have to give, so it becomes very difficult. Image recognition is one of the problems that AI can solve, like thousands of problems. Pretty much everything that human does is a problem that AI can potentially solve. So that is the main sort of breakthrough.

The semantic problem, which means meaning, and making machines understand the meaning of a certain word is similar to making machines understand the meaning of a cat. So it's very similar to that. For us, AI is the way that we make machines understand the intent of a search query, meaning what it is you really mean when you put a word out there. When you put Revenue Outlook, as a business person, what do you mean by Revenue? What do you mean by Outlook? The machine has to understand, and it does understand in a way that doesn't require me to teach it one word at a time. That's a problem that AI can solve. Then you want to have other ways of using traditional filtering and whatnot, but the machines really help you understand the semantics. That's really the biggest problem that we have. So we use machine learning as I mentioned before. There is a specific field of machine learning and machine understanding called Natural Language Processing, because we are dealing with natural language. Of course we also understand more about images, charts, graphs, even videos. Business information is everywhere, so we want to make sure we get that right. The biggest reason we want to use this is to apply the same semantic search, making machines understand the meaning of words, to images or videos over time. Searching by intent is not just a keyword meaning searching; it also applies to the literal picture you see. You want to understand the context of a particular image, be able to say this person is dancing on the beach. So you want to understand the intent. This idea of semantic intent translates beyond words, but it all starts with natural language for us.

So that's one thing you have to understand, that these things are constantly evolving. Second is the typical entity recognition. When someone talks about Apple, we don't know whether they are asking about an apple to eat or Apple the company. Or when someone tells you don't keep slacking, in a business context, are you slacking? It could be an insult, it could be a bad thing. Slacking could now mean you're being productive communicating with your colleague, whereas 10 years back if you said you're slacking, that meant you're goofing off. These days if you're slacking it's actually a good thing. Being able to know the difference between slack the verb and Slack the company is important. Words that have become verbs have changed meaning from negative to positive. It's really important to keep the machines up to speed on the terms we use in the real world. This is one example of why it helps to understand what companies are coming up, how to understand products and services as we understand the real world. If the machine doesn't understand those words, it wouldn't be able to help you. That goes to relevance. I'm not going to go into all the details here, but essentially all this information feeds into the model. I don't know if I'm going too fast; I want to be conscious of time. It's 9:40, so I can go a little quicker to get through the Q&A in the next six to seven minutes.

So I'm going to skip through some of those details. One idea I do want to talk about is words meaning semantics. In reality, if any of you have looked at the stock market, that's where you trade company stock like you would trade baseball cards. You can buy and own a percent of a company by trading on its ticker symbol. The business world people assemble and say 'I want to buy this much of this particular stock', and the price settles based on how many people buy and sell. A lot of it is based on their perception of the world as opposed to reality. Perception is reality. Sentiment is the underlying term. Any negative or positive news can widely swing the markets, making them volatile. When the government says they're going to raise interest rates, that increases the cost of buying a home or a car, affecting demand for car companies. The side effects of that are obvious when interest rates change. But if a company depends on a chip from Taiwan and that chip is delayed by two months, you don't see that connection easily. That information is not understood quickly. Machines are quicker at this, but just like we said how difficult it is for machines to understand even one word, understanding the context of what's happening in the real world and how words affect supply chains and stock prices is not a perfect science. People base pricing on sentiment, and that causes volatility. One of the biggest problems we wanted to solve is taking off some of those false feelings, getting the machine to tell you what's actually happening based on text, showing you positive and negative and exactly why, giving a score. This has been extremely useful. When you see this, you'll be able to understand where it's coming from. If a person says 'we are reporting guidance involving 300 basis points drop', you want to know that something negative is being said without listening to a 30-minute call. Time is money, and sentiment drives the market. What people say is translated and disseminated. You will be decimated if you don't disseminate fast enough. Disseminating information quickly with sentiment allows you to track what's going on and react quickly. That's the value of semantic search. It's not just search or productivity; it has massive implications for the stock market and how people understand things. If you understand better, you make better decisions, making the overall world less volatile. We're not just solving a problem for a few analysts; we're solving a problem for the world. A stock market based on real information is common for everybody. You want to get to the root cause, which starts at the search engine index being understandable. I want to stop right there because this goes into depth on sentiment analysis, and I can go into all that detail, but let's pause right here.

Have any questions?

Good. I would answer that by showing where we are today. Today our product is being used by thousands of financial clients, 3,500 companies worldwide, 70% of Dow Jones, 78% of S&P... these are the top companies. 73% of global consulting. Many of these people use it. Having raised funding in the past, you have an idea. In the early days, you want to show that your idea is worthy of solving and the problem pain point is worth solving. So you want to talk to customers, users, potential users like you and know that it's serious. If you have your own pain point and know for sure it's huge, that's one thing. But I would start there. Once you know that, you want to articulate how big the problem is. Second, you want to have a realistic solution way of solving. Before you know how to get to the solution, you want to know what would be an ideal solution and get the imagination going. When people say 'I want to solve this', it's like saying 'I got cancer as a big problem and I would like to solve it'. That sounds fine, but then how do you actually solve it? You want to provide the key innovations that will solve it and have some proof points. That's where you bootstrap. Bootstrapping, you bring in a very hacked solution. Our early versions of product didn't look pretty; it was very ugly. It was just a talking button and would show answers in the middle. All we did was pull SEC filings and look at free documents, put some important sections from it, and said you can even get value from the most boring, freely available document. So we showed that as a harbinger, an indication of things to come. So you know you have something interesting. That's when you want to go talk to like-minded investors who know the problem. If you talk to an investor who doesn't know your market, they don't understand how big the problem is. That's where I would start. There's a lot more, but when you're starting off early, you want to identify how big the problem is, have some idea of how to solve it, and then go to an investor who understands the problem to take a bet on you and that you are the person to solve it.

Over time, of course, you take it to customers and actually get customers, and then you go back to them and show 'look, I told you there is proof, we believed in it, there's even more proof. Now I want to hire more sales and marketing people. Let's all believe in this and continue.' Last year we raised half a billion dollars to grow. That took about 10-12 years. When we started off, it was just my own capital. We put a small amount of money out. Oftentimes, time is the best investment initially.

What does AlphaSense do? It helps people search for information about businesses. Let me give you an example. Let me share a different screen. Okay, so let's say you want to find out about Nintendo. Do you like Nintendo? I take that for a yes. Nintendo is a company that makes video games. If you type in Google 'what's the latest product release', not only are kids interested, but business people are interested to know if they can trust Nintendo to be creative and attract more kids in the future so they can buy their stock and make money. They go to Google and find out Nintendo product releases. They get some information, but they have to dig in to find what's exactly happening. What we do is when you go to AlphaSense and search 'Nintendo product release', it's able to tell you the top 10 Switch games, exactly what to look for, a lot more useful than that. This information is useful for people to get to it without having to open each one. Make sense? Yes, thanks.

Sure. Unlike Google, this is not free. Google makes money by asking you to put in a query for free, but nothing is really free. What you're really giving Google is that they do it for free, but then they have to match advertisers to know who's interested in their products. If you're searching for Nintendo, Nintendo may be interested in knowing people are searching for their latest product, and they want to show products on the right side. So Google makes money off of advertisers. For companies that depend on that revenue stream, it's an issue. AlphaSense gives it to business people and charges a monthly fee. In return, we don't sell their information. We don't use the information to show advertisements. We don't operate like that. Your information as a business user is not even collected. Even if we have bare minimum information like email ID, it's not shared with anybody. What you search and use for your business is completely held private.

Humans are involved in giving the guidelines and shaping the machines. Are humans involved? Absolutely. The machine doesn't have a mind of its own, not yet at least. It's going to be decades before AI machines become as good as humans. Humans are involved in teaching the machines: if you look for the following things, these things may be important, so rank them higher. In terms of forming rules, telling the machine by giving it data, it's like teaching as opposed to telling exactly what to do on a per data basis. It's similar to how a music teacher would play something for you and say listen and play like this. Once you learn the idea of playing music, you know how to play more music on your own. We're showing you what's important, and once you get the hang of it, you do more. That's where humans are involved, but not more than that.

I'll answer the first question quickly. The second one is more important for students. No, for consumers, they don't see as much value to offer this. The content we get is expensive, and because we don't show advertisements, we need to charge. Companies have to make money to pay employees, vendors, partners, so we pay through monthly fees. If someone wants to buy it, they can, but they would not find as much value in paying this much money for a search engine service when they can do it from Google. So we don't target them because there's not enough value. But if you really wanted to, you can buy it. It's built mainly for businesses. The second question is about failures and how to overcome them. Before you talk about failure, you want to know why you are doing something. That's where the problem and experiencing the magnitude of the problem comes in. If you really feel it's worth it, either financially or intellectually challenging or personal, like someone who lost a loved one to cancer and is determined to solve that problem, they have personal experience to use as motivation. For others solving a problem like this as business, you may want to know if it's worth it. I like to solve problems that are worth solving in a business-centric way, making it worthwhile for everyone: users, investors. The third is someone who is a researcher who just likes to solve because it's intellectually challenging. Regardless of what motivation you use, you're going to have failure. Failure doesn't mean anything except that you are not succeeding as of today. That's what failure means. Failure doesn't mean permanent. It means you try to solve an algebra problem, you get a wrong answer. You can't call it a failure if you would not go to school at all. On a daily basis, you take a question paper, solve it, get a few wrong and a few right. You don't call any wrong answer a failure. Similarly, in the real world, when you're trying to solve a search problem, you're not going to solve every single issue. You'll have wrong answers, so-called failures. But you know it's worth pursuing. You don't stop learning algebra just because you got a few questions wrong. You pursue to solve more, learn more, and as you get better, a couple of years from now you'll say 'that was easy'. Failure is given too much importance because people forget that as a child they didn't pay attention. When you grow up, everything is looked at as a big deal. As long as you maintain the mindset that every day if I get it wrong, so what? Just go back and try again. That's what one should take away from childhood experience. Remember that mindset. The same thing you have to repeat: if you don't get an answer first time, try again or ask for help and try again. Hope that helps.

We provide internship for college graduates only because of privacy and confidentiality issues with customer data. For businesses, we have requirement that people are over 18 or 21. We hope to have an affiliate program for people below that age to work on problems through a non-profit entity. That's why we don't hire interns for high school at this time. For undergrad, we do have interns.

Yes, absolutely. Data ranking does change depending on current searches. If more people are looking at something, we know there's a trend. User searches give us a signal, one of many signals, to figure out what's important. Unlike Google, we don't have billions of users, only tens of thousands, so we can't truly depend on that. But at a micro level, for a given user, their user history helps. Knowing what you've been working on means I can rank even better for you. Your searches play a role in improving that cycle. And yes, it can find patterns in popular searches. That's also called trending. It's not just your own search but search across other people similar to you and across other industries. All that information is baked in. The answer is yes, we do use it, but only to the extent it helps you.

Yes, I think it's great that you're putting this together. I know that sometimes people will be working... I'm much older than you, from the gray hair. But hopefully this is useful. I think one thing that as adults we also like to do is learn from kids. As I mentioned, we adults fail to understand that it's a new day tomorrow and you have to start fresh. In some ways, you have to forget failings, learn from them quickly, and not dwell too much on the negative or positive, and move on. Kids have that tendency. That's one thing I try to learn from kids. As you learn from these talks, it's important not to lose the enthusiasm that kids bring. Keep that enthusiasm alive as long as possible. All right, good luck.