 It's a pleasure to be back in Yerevan. This is my third trip. I've been working with Armenian companies now for six or seven years. And one of my favorite things is interacting with people here. We only here was really smart. Some of the technology I was seeing is world-class. I'll be referring to some of that this afternoon. And this is my second lecture anyway. I'm glad they were willing to have me back. That is not always the case. So we're in trouble with the frontier of AI. There is some amazing work very far out, not yet ready for primetime. Just to give you a sense, Kizula Ligwa at the top there is a company that's trying to do English to animal translation to understand what your dog or cat is doing. The bee is Russia. Russia is experimenting with a whole herd or a swarm of mechanical bees to help solve the bee pollination problem. At the bottom left, those are two new newscasters in China who become standard newscasters. They are both completely fake. They are bots, but they look and behave and sound like real people and they're among more popular newscasters in China. An AI-created painting was just sold for over $400,000 in New York recently. Next to that painting is an entirely artificial, freestanding, moving painter or named Adya who has been, you give her a suggestion to look at the painting you like and she creates something brand new and original, real brush strokes, moves around, can talk. So this is, these are some of the far out things going on in AI right now. If you look at what's happened to AI since 2011-2012, Stanford University combines venture funding, publications, patents, student enrollments in AI around the United States and the amount of activity in AI has gone up sevenfold in the past five years. It's really, if software is eating the world, AI is eating software. So we're going to talk about a big picture about deep learning, look at what some of the trends are that people have been investing in and companies have been founded around and then spend a little bit of time backing up and saying, what are some of the benefits? What are some of the risks? So what we mean by artificial intelligence is software that performs at the level of a human being and it comprises a whole bunch of different disciplines. The most important right now where all the money is going is machine learning and deep learning but it also includes natural language processing, computer vision, autonomous cars but where most of the money is going is in machine learning. But this is not new. We've been experimenting with artificial intelligence now since the 1940s. Back then it was called cybernetics and you, how many people have heard of Alan Turing? Alan Turing is a birdist scientist who created what's called the Turing Test. Basically say if you cannot tell whether a machine is a human being or machine. We have passed the Turing Test and AI is becoming normal. And Alan Turing created that. What many prizes for that. In the 1950s the first two paradigms of AI the two ways of thinking about AI both became popular. One is rule-based computing if then logic. If the patient has a certain of blood pressure rate if the patient is exhibiting other symptoms then we can say this patient probably has hypertension. That is one main threat of AI called rule-based computing. The second is basically rule networks where the attempt is to model AI on the behavior of the human brain. Thousands to millions of small interactions creating a logic that is not necessarily perceivable to us but it is perceivable at the level of machine learning. Both this activity became very popular in the 1950s. The 1960s and 70s was the age of expert systems. Those rule-based systems that were used in medical diagnostics. 1982 neural computing became important to get. And for us today the date to remember is 2006 because in 2006 a Canadian scientist named Jeff Hinton spent half his time at Google like everybody else in AI radically improved the accuracy of neural nets and we'll talk about what he did in a few minutes. But his students in 2012 won an image classification competition called ImageNet with radically better ability to classify images automatically using things like NVIDIA graphics processing units. That shocked the industry and suddenly deep learning neural computing had won. Today we're in the age of applied deep learning. We are here where deep learning is now become the standard and it's being used in medicine in automotive computing and retail and manufacturing. These are now become mainstream AI. One more thing to remember is that in the race to deploy AI in real situations China has a tremendous advantage over all of us. Population, big data, number of companies. So 70 cents of every dollar spent worldwide on artificial intelligence is being spent in China just to keep that in mind. If you look at patents and publications the explosion of patents and publications about AI started in 2012 that was the takeoff and if you look at the complete difference between machine learning, deep learning, neural nets as a force you can see that they are dominating the world of AI compared to logic programming which is basically rule-based systems. At the same time AI is always about to change humanity. We're always about to usher in a grand age of where AI is going to solve all of our problems. Back in 1958 the two leading scientists Simon and Newell said within 10 years digital computing will be the world's chess champion or we're going to have a fully intelligent machine within a decade. So we've been making ridiculously outlandish predictions about AI for as long as I've been alive. Most recently one of the senior scientists Google Andrew Ng said AI is no electricity and Sundar Pekhai who runs major companies said AI is more profound than electricity and fire. So we're still in the era of AI hype. We've had cycles of AI and a new theory will spark a tremendous wave of interest in AI. Compute power will catch up. Investors will invest. We'll harvest the results. Newell step back and say you know the results weren't nearly as good as we were promised and we'll enter a period called AI winter where funding stops research stops companies say Mel, I don't want to do any more AI for a while and then a new theory will come around and spark lives. So right now this deep learning stuff where Jeff Hinton in 2006 led some pretty profound improvements in deep learning that's the current generation we're at. So the question is have we escaped? Have we reached escape velocity? No more AI winters or are we going to see another one? Because this has always hurt practitioners as well as investors when AI winter hits. In terms of money and deep learning is the fastest growing it's the dominant two thirds of money spent on AI are being spent on neural network computing and I will explain what that is in a few minutes. In terms of where the money is going in terms of business it's healthcare it's sales and marketing to some degree it's automotive autonomous cars but healthcare and sales are the two dominant places where AI money is actually being spent on business solutions. We're in a maturing market AI is no longer new it's no longer the cool thing and so we're spending more money in terms of investment going to established companies already in the market. We're spending around 20 billion dollars a year on AI investing and as you can see the share of funding going to established companies is going up and up and up. Where is the money really being spent? Who is spending most money on AI? It's not investors like me. It is technology giants like Google, Apple, Facebook, Amazon. 75% of all funding of AI work is happening in big tech companies and this has a profound implication for startups. One of the things that's happened since 2012 is a surge in mergers and acquisitions as AI companies are being snapped up by the technology giants. So if you look at who is who is buying AI companies it's Google, it's Apple, it's Facebook, it's Amazon, it's Intel, it's Microsoft, it's Twitter. So most acquisitions have been made by the tech giants who are essentially in an arms race to buy not only technology but many of the acquisitions have been for the people. So the quality, the technical quality of people is essentially very important in building an AI company. So let's kind of take the big picture where are we in terms of numbers. We have funded to the almost 50 billion dollars of venture capital in artificial intelligence that does not count Google, Amazon, Microsoft. We've increased funding eightfold between 2013 and 2018. We've doubled the number of startups and this is just an extraordinary market. We've had 434 acquisitions. About 7% of all AI companies have already been acquired and the exit values, if you do well, the median is about 200 billion and the average including many multi-billion dollar exits is about half a billion dollars. And for companies in the market if you want to know what your company is worth it's typically around six times your sales numbers. So takeaways, we've invested 50 billion dollars. We have 6,500 active AI companies today. There is lots of merger and acquisition further. We're probably at the top of a hype cycle where there's more excitement about AI than there's ever been and that will come down. That exit valuation is strong. There's still a lot of money to be made but if you look at your companies if you are expected to be acquired or to join a bigger company Google, Amazon, Facebook, IBM have only made many of their bets. Yes, there's still more to come but I think the M&A market is going to be transitioning and I'll talk about it a little bit later. 32 companies in the AI world work more than a billion dollars and a good percentage of those are in the healthcare space. Let's turn the numbers to trends. What do people like? And we're going to talk about six trends that are worth investing unsupervised deep learning. X AI, everybody know what X AI means? Anybody heard the term? Okay, we'll read you the term. We moved from tools, AI platforms to vertical solutions. Big market for data engineering and there's some exciting companies right here in Yerevan in that space and finding insight as a service for all of you as a consulting world. So unsupervised deep learning how does a neural network work? Will you take things like objects like pictures of animals and human beings classify them? There are whole cities in Africa where the number one employment is having women in front of computers manually drawing boxes around pictures on the screen and classify them. This is a cat, this is a dog, this is a car, this is an airplane, this is a gun. So you apply labels then a neural network tries to figure out correlations. Okay, all of these images that have dogs in them what is dog mis-like compared to cat-ness? What do they have in common? And they'll create some synthetic and hypothetical variables that might be short tails the things that humans didn't classify to distinguish between a dog and a cat. They will then run the model and compare the output in stage three with the prediction versus the reality. And this is where Jeff Hinton's magic came in because he invented something called back propagation a fancy term for simply saying let's feed the errors back into the model and let the model correct now that we know what the difference is these five images are raccoons not cats let's feed that data back and improve the model. So that's how AI machine learning works and so for deploying an application you take unlabeled data segment four run it through the model and you get output. An output can be used to classify, to tag, to organize, to predict. That's all machine learning does. There is a subset called deep learning and that means that these models in the middle can be thousands of layers high and people compare deep learning the model they use is it's like lasagna or it's like a sandwich a multi-level sandwich because there are thousands and thousands and thousands of different variables and factors that feed upon each other. So this is what deep learning is all about where all the excitement is. Some examples of deep learning my favorite are deep fakes and the deep fake world is really exploded. These people never existed. These are also heavily created by the graphics company NVIDIA. Last week someone unveiled technology I think it was Google where they can take a 40-minute clip of video and they can then create synthetic videos of those people saying things they never said. Pretty scary technology. To show you this in action let me bring up one example. I was going to show you somehow this is not working. I was going to show you a picture of the Mona Lisa and then how graphic scientists have now made the Mona Lisa move talk, move in many directions not one static image and that's the power of deep fakes. So we've been talking a lot about neural networks and there are two flavors. This is number one investment thesis. Right now we're in the age of supervised learning which means human beings have to classify those objects and if you're building an AI solution about 90% of the cost is having human beings manually classify and tag objects. So that is supervised learning. You use it to build and validate networks but you need humans in the loop. There's a new technology around unsupervised learning it's often called UNL unsupervised machine learning where you're using statistical techniques to figure out what things have in common and to classify them without any human input. They're less accurate than supervised learning but they can work wonders in narrow events. So in Israel there's a company called Cortica which is looking at one thing. How do we avoid hidden pedestrians in autonomous vehicles? And they're learning simply by looking at thousands of video clips which have not been categorized. Company in Germany, Wandelbox you put on a jacket that has sensors in it and you can train a machine in 20 minutes to do whatever the robot needs to be often picked in place in an assembly line happens automatically with no statistical or rule-based training. Dark trace is looking at how do we estimate financial fraud in situations we've never seen. The problem with supervised learning is you can only predict what you have already seen. For unsupervised learning you can predict and evaluate things you've never seen before. So if a woman is crossing the street drops her shopping bag stoopes in the gutter to pick up a loaf of bread that has never been seen before and an autonomous car needs to understand that. So unsupervised learning is being used to use things like that. So if you hear the term unsupervised learning very important and extremely hot whether you want to work at the company so what's happening what's next with being learning a number of companies like Google and Microsoft are releasing tools into the marketplace to do automatic machine learning to create machine learning based on providing pre-written models. The United States Defense Department wants to do lifelong learning have AI models and continuously learn without program or intervention. So if you hear things like lifelong learning auto ML these are automated technologies that reduce the need to build models by hand. X AI next trend it's all about building trust what X AI stands for is explainable artificial intelligence. So last year the European Union established a law that says anytime an artificial intelligence touches a human to explain how the AI came to its conclusion. That's now law in the European Union it's law in New York City companies like MIT and Microsoft are trying to figure out how to make these neural models explainable. Remember in the middle of the neural model we don't know what's happening we don't know which neurons are firing or why it's very hard to understand how a neural net comes up with a conclusion. But if you're getting a patient a healthcare diagnostic that says you must go on a very strict regimen the patient wants to know are you sure how did you open that result? In the law if you're deciding who gets parole or who has to pay a fine based on increasing use of artificial intelligence you need to be able to explain to that person what happened. If you've been denied a loan in insurance agencies right now we're using AI to process loans you can get sued if you can't explain to a person why he was designed to defy a loan. If the military shoots a civilian talent accidentally if an autonomous car kills somebody these are the kinds of decisions that need to be explained. So xAi is on the forefront of what's happening in AI we need to be able to explain our models and our models almost by definition are inexplainable. We engineer those models so software companies build the tools to address explainability or a hot market. One of the questions about data science is we really don't know how our models work and some people have said data science is as much magic as it is systematic learning. Can we explain it? Do we understand how it came to its conclusions? Can do different researchers produce the same kind of results with their own data based on what a data science model is supposed to do. Can we understand where the data came from how good the data is has the data been audited has the data been cleaned and there are no data standards right now for using data in AI and I'm going to talk about one of the dangers in healthcare in a few minutes. So being able to go up with algorithms is at the forefront. Sense-making. We all know that computers are great at taking a sound or an image and doing something. If you hear a loud noise that can be an alert. You might want to look around and see if there was an auto accident down the street. What's happening right now is taking these senses of vision, sound, sight combining them into bigger models. What does this mean in reality? What this means is situational awareness. You want to know what's happening in the world around you. You hear a sound behind you you see a bright flash as an aircraft pilot I have to attack what's going on. So the Canadian Air Force is using this kind of sense-making technology correlating sound, sight, noise and other kinds of vibrations to create a picture of the world and to figure out is there somewhere dangerous. Marriage, the world dominant shipper is doing the same thing at sea. What are the dangers I'm facing? It puts again the picture of my ship right now of my face getting kind of dangerous. This is being used in automotive computing to see what's happening inside the cabin of a car. Has the child in the front seat spilled a drink on dad who was driving? What kinds of situations should autonomous car software be able to deal with? The first thing I need to know is what is going on. So sense-making situation awareness is one area of growth. One of my favorites is effective computing which is to try to understand emotional content by the use of sight sound imaging. So a company called Affectiva which is already a unicorn can detect the vision of the emotional reactions of the audience to a movie or a TV show beyond verbal. Can detect do you have a potential illness by the tone and tenor of your voice? Are there vocal cues that something might be going on? Crowd emotion uses eye tracking to understand how audiences like dislike a given movie or a new show and because they are so good they can also look by gender by age, by race. Well, the 40-year-old white ladies seem to like this movie but the 15-year-old black kids don't so they can go with that level of discrimination. This is being used in advertising. Online emotion is using eye detection as a lie detector and this has now been approved by the court system in Arizona. There's lots of interesting work going on in using your emotional behavior to predict to monitor, to manage outcomes. From tools to verticals. So we have moved in AI research from focusing on fundamental platform technology now to experiment with we make a difference in vertical markets. This has happened because we have these brilliant graphics processors from NVIDIA. We've got big data consortiums in the industry like insurance building big data models and sharing data across companies. We have sensors that cost pennies that can be invented in roadways and cars and clothing so the sensors of revolution and then we have lots of free available machine learning frameworks for anybody in this room to experiment with AI free versions of technology like R like TensorFlow from Google so these are now not tens of thousands of dollars they're essentially free. Cheap sensors free software means a whole age of experimentation. If you look at the top markets for AI it's healthcare in terms of diagnostics early identification of potential pandemics as is going on in Africa with evil of monitoring and a lot of work in image diagnostics automotive autonomous cars financial services automated financial advisory and services for people who can afford a financial advisor anyone can get basic financial advice transportation technology these are the big markets right now for AI let's look at a few of them first is healthcare we'll say the robot will see you now machine learning is being used for admissions for triaging patients as they come into the admission system for classification of doing genomic screening for matching treatments with automated diagnostics for clinical trials to find the best patients to be recruited to the clinical trial processing is being used to alleviate one of the physicians biggest hates which is transcribing their notes their patient interview notes automatically this is being used for clinical documentation for coding of records there are over 100 companies providing some kind of blockchain for healthcare to secure healthcare records productive analytics to figure out what patients provide a risk to a hospital to themselves that might need early intervention and we didn't realize it but they can find the sign to say wait a minute this patient is exhibiting all the signs of the patient who will not adhere to the drug regimen they've been prescribed and then robotics are being used both for surgery as well as to assist elderly patients some examples these are all research university results but in areas of depression skin cancer eye condition AI is being used and works as well as trained physicians the real bad thing is when you combine trained physicians and AI then your accuracy level goes up way into the 90s there is a problem with some of this if you look at all of these these are done by the world leading hospitals and research organizations what is that mean the data is beautifully curated really accurate data this is not the real world the real world is messy but in terms of when you control for data and you control for experimentation the results we've seen have been pretty dramatic and this is probably I could have expanded this list with 3, 4, 5 more pages as a result of the early work in AI we're seeing a tremendous growth in funding there have been over 400 companies receiving their first their first financing since 2016 34 billion will be spent on AI and healthcare in the next 2 years and the market is growing at more than 50% a year driven number one by Asia we have a problem of not enough physicians major rural communities under served so they're investing a tremendous amount of AI to begin to bring remote diagnostics remote medicine remote reading of charts remote reading of images the two populations across Asia but you can see the dramatic increase in AI funding in the United States but it's just as big globally it's an interesting project a radiology assistant has been approved by the FDA Medtronic has been approved for a diabetes management system that's been fully automated in London there have been something on the order of half a million people on the screen and automatically diagnosed by a company called Babylon here in Armenia there's a company called Silence Image very interesting work on trying to create electronic records for entire populations in Armenia and in Kazakhstan and then using that data and some rule-based logic to begin to do some automatic diagnostic to train physicians in diagnostic techniques so I keep an eye on silence there are the forefronts of these trends Katalia Health has a robot in the top corner that is able to calm elderly patients and help take their vital signs get basic health information in a less threatening intimidating way and attuned to the unique problems of communicating with elderly patients who don't cure well who may be slower to react and might be a bit more resistant the world is emerging in a new AI electronic medical record this is a vision of the future that people like the physician Eric Topol and others are predicting will happen so the next data in the next decade the center of life is a lifetime medical record that is fed across the chain of providers that you see consistent, coherent medical record it serves as the intake when you're admitted to a hospital in health with patient screening all the way through the progression of your progression through the health care system eligibility for clinical trials transcriptions of medical records all the way through post discharge treatment via care plans on your cell phone via automatic monitoring of your adherence and things like drugs and the use of robotics as patient assistance what this is leading to again this is not near term this is a 10 year forecast but faster diagnosis a single view of the patient across all the providers of higher clinical trials completion rates very low completion rate as patients drop out we can do a much better job of predicting who we live into the trials in the first place population data mining of the sort that SILICS wants to do to look at populations as a whole all the way through predictive analytics to improve wellness and compliance outcomes one of my favorite experiments is from a company called Go Marge they have a product in the market called Pediatric HAL Pediatric HAL is a baby, it's a robotic baby the baby cries, it sits up it reacts to commands you can take its blood, you can take its blood sugar it dies, you can use a defibrillator to revive the baby it expresses emotion it expresses pain it makes heart and lung sounds so this is being used for medical education this kind of simulation is being used for medical education primarily for nursing staff but it gives you a sense as to what the future is going to look like the one thing it does not do luckily is admit body others but in virtually every other case it does simulate to some degree the behavior of a three or four year old child you can attach patients you can monitor to it and to any one's needs so the bottom line in healthcare is a million dollars in savings in the U.S. from population health forecasting of the sort that Sylex is planning on doing three million dollars in savings from preventative care in the U.K. this is what Mackenzie says the world is going to look like 30 to 50 percent improvement in nurse productivity and the bottom line is as much as one to one in the 30 years of increased life expectancy once this technology gets fully rolled out over the next decade but we have problems and for those of you who care about data science this should be a forefront problem the first is there is a very tight regulatory regime controlling the use of software in healthcare anything that touches a patient needs to be validated often by the U.S. authorities on the one hand you have tightly controlled the implementation of health software but what the market wants is AI to keep evolving to keep learning every time they die I know something that goes back in the database so these two principles are fighting each other and the U.S. Food and Drug Administration is about to issue regulatory guidance for how you can have both a securely evaluated AI and one that changes over time more importantly is data accuracy the experiments I have talked about have all been with heavily curated data great data scientists really making sure that that data is accurate manageable that the data had a good source, that there was the data is not corrupted that is not the situation with those medical data and it's been estimated that as much as 70% of data in a medical record can prove somewhere else you have no idea if it's clean and it's dirty and the physician was paying attention there's no way to validate that data so we have an issue around healthcare data standards the big data of the data science program here at AUA I would hope would look at in the bioinformatics program ways of improving the quality of the basic data that feeds our models you know the first rule of computing is garbage in, garbage out so if the data isn't going to begin with how good are the predictions the diagnoses made with that data this is the problem that's going to confront the commercialization of AI in healthcare real estate the normal transactions of renting an apartment buying a house are now in places like the UK and the US heavily involved with artificial intelligence used to match buyers with the right properties if you're an investor which house do I want to invest in in this neighborhood which is going to have the highest appreciation value houses been selling I want to have been on a house help me make a more accurate bid review the contracts, review the legal documents to find anomalies and problems mortgages are underwritten now increasingly by AI I have to predict who is going to be a somebody who's going to widely pay these things back so among the companies doing interesting things we have virtual agents replacing your real estate agent with a bot and we also have analytics tools to help you invest or buy properties better which houses will appreciate which multi-tenant which many-tone apartments are going to have the highest rate of tenants paying their bills and so forth there was an experiment in Denver, Colorado about a year ago where a bot was created with the goal of saying this client like this house based only on a picture which other houses currently for sale in Denver is this potential buyer going to like so they compared the results of three real estate agents after real estate agents with a bot and the bot made predictions and the real estate agents made predictions and in every case the bot beat the agent only problems we had no idea why of all the the AIs were shown was a picture and we had to determine from the picture what the buyer's preference was but again this is one case that we mentioned the bots did a very very good job digital lawyers were using AIs across the law everything from managing conflict adventures to documents to predicting the settlement value of cases so you've all heard the Shakespeare comments first kill old lawyers well that's being replaced by replace them with bots so there is a famous chat bot called Do Not Pay that is overturned over 10 million dollars in traffic bombs by having the bot fight with the local court and 375,000 tickets were overturned by this bot a thousand new bots were launched by the order of last year in landlord and tenant disputes in immigration disputes and all of this was done by a 15 year old a 15 year old armed with a computer and some good AI tools and he has already made the United Kingdom poorer by over 10 million dollars by what he's done he's now a junior at Stanford University lots of activity around digital lawyers for contract analytics for predicting case outcomes here in Armenia there's a company called Zero which is able to scan email and send to lawyers which I focus on which of these emails impacts the case which emails do I jump on right now what do I send my assistant what do I archive as part of the case system so you can get a sense that the work of what lawyers were what is estimated that something like two thirds of the manual work of the law is already being automated my daughter is a lawyer and I was explaining this to her she said faster faster faster by the way please automate me too so do you all know what a non-disclosure agreement is this is an agreement that if you're with a company and you're bringing somebody in to see a new product you'll have them sign a document that says I will not share this information with anybody outside the company it's called an NDA non-disclosure agreement so a firm created the bond to scan non-disclosure agreements and see the errors see what was missing and the lawyers on average for a set of documents took on an average hour and a half to review these documents and they found 85% of the errors the BOT did this in 26 seconds same thing with a much higher accuracy and if you look at what lawyers bill my estimate is is that the cost of using the BOT was $3.80 the cost of using the lawyers in the United States just to look at six non-disclosure agreements would have been about $4,000 so dramatic improvement of doing normal human stuff data engineering if machine learning is going to continue to explode we need it's all fueled by data you need millions or tens of millions of impressions images or sounds or documents you need millions and millions of pieces of data we need collaboration across companies to help provide this data so it's a lot larger than we can provide we need data engineers to curate and manage this data we need data scientists to build models to analyze these data and we even need AI familiar lawyers because we're about to see a whole era of litigation about the use and misuse of AI in human situations like healthcare just one example in the automotive industry many many automotive insurers on everything from driver behavior to accidents to road safety to create a vast big data world that no single company can provide to their major consortia and data service bureaus to provide billions of data objects that insurers can use how do we know how to insure autonomous cars we need to find out what kind of insurance range should I pay and to do that we need to understand autonomous cars safer or less safe than normal cars so on average today there's about one fatality for every 100 million miles driven so we need to instrument autonomous cars the same way what this means is we need 8.8 million miles of driving data just to predict are we going to be plus or minus 20 deaths but that's not good enough for insurance to do that we would have had to start collecting data around the time that Shakespeare lived if we have a fleet of 100 vehicles operating 24 hours a day if Shakespeare had started instrumenting when we have enough data but that's not good enough we want 100 billion miles to understand plus or minus 5 deaths and that means the Sumerians and the Babylonians would have had to start collecting driving data to feed our models today so we have massive data insurance problems in solving some of the biggest problems I'm going to look at some of the solutions to enable the new world who we're going to be living in we need a massive number of data scientists IBM estimates that we're going to need about 61,000 data scientists the people who build the models for every data scientist we need 10 data engineers to purely manage, create that data so we have huge amounts of shortfall in our data sciences that's why the program here at AOA is so important training data scientists is one of the biggest growth areas and I'm delighted that AOA is stepping up in the leadership role here doing Canada all by itself is producing one tenth of a number of data scientists they need in the city of Toronto alone so big, big, big shortages for real data people if you ask IT people in western Europe and eastern Europe what is the biggest skill gap they have that companies are afraid of in western Europe 42% of IT people said it's data science in eastern Europe 41% so again the demand from the market itself is suggesting we have big problems but their help is on the way we have new investment areas some of which are active here as well data engineering tools help us clean the data using better technology I mentioned the importance of data labeling so you need to label all of these images and objects, this is a cat this is a dog, this is a car a number of companies that have stepped up in leadership roles around the data labeling and one of them is super annotate AI right here in Yerevan that has great tools and also offers data labeling as a service there are many changes where companies can sell the data they are not using or share it you can crowdsource data with a million impressions for the project I'm running can you please crowdsource them for me can you define ways of getting me that data and then you've got the whole towns as I mentioned in Africa where people are devoted hundreds of people doing nothing but data labeling eight hours a day so these are all investable areas, they're all interesting if you're interested in forming a company or joining a company these are hot areas and we have massive data shortage insight as a service if you look at the history of computing we have gone from algorithms to tools to platforms to solutions that actually help people and then to consulting firms who say forget all of that you don't want to hire programmers you don't want to build models we'll do that and then we'll tell you what to do with those results so this new thing called insight as a service and it's particularly appropriate for consulting firms AI is now demanded by everywhere we have data science shortage we can't find enough data so let me go to a consultant who manages all of these things for me and I don't care about the methodology all I want to know is what do I do to make my company run better so we've got data as a service we've got analytics as a service we've got to help analyze the data that's what we learned last year now insight as a service tell me how to improve my business using data but I don't want to know how you got there so one of the projects I've been involved with insight as a service is using big data in an oil field putting 40,000 oil wells to say what do I do to manage my oil wells better how do I twist the knobs and gears to make a better production and the company E-links manages 40,000 oil wells and then uses that data for its customers like Shell Oil to figure out how to better manage production and they build a digital twin model using state of the art data science to monitor what an oil well looks like they have data sciences then evaluating the results of these models and then instrumenting these models at client sites the clients are not pulling the models and they're providing the insight turn this valve to the left and what they manage to do is using big data to increase production by 6% which is in oil and gas enormous they're also able to cut down the use of hazards chemicals by applying these big data models and again what this company does is say why don't you subscribe to your oil well we'll manage all the data and we'll tell you what to do will you just subscribe to your oil well and we'll tell you what to do about managing the massive data flow so let's see what we're going at some risks this is a fatal slide lots of issues with we have yet to to fix data bias data is typically biased around healthy Caucasian males which really does hurt our ability to use models in healthcare at the end it's such a bad job of cancer diagnostics they canceled the project we've seen Facebook had a big issue last year you may have read about it where Facebook had an AI two AIs something were created that started talking to each other in a language that Facebook did not understand and they felt that the AIs were duplicating themselves replicating themselves no idea what was happening so Facebook shut down the project but fear Apple was sued for a billion dollars last year because it's facial recognition technology led to someone's arrest a house was raided lots of police showed up at a house they thought there was wild activity going on and they thought there might be a crime being committed what was happening was that Amazon Echo, an Amazon speaker turned itself up to maximum volume and caused the neighborhood to say what the heck is going on a six year old girl called Amazon Alexa with no controls and ordered herself a $170 dollhouse so we can list over and over and over again the problems with AI it's not all great stuff and a hedge fund went bankrupt using AI to predict great investments what companies said we need some new color names for brand new colors and here are some of the names for some of the colors that the AI came up with sand dan, grade bat grass bat cindus poop dope so again we we do good work here so some good news on the AI front I think you'll all love this okay share the good news with you so what's next we're the areas that we still have a lot of work to do what is the search for artificial general intelligence we now can do some pretty interesting things by printing which house a buyer might buy but we can't there's no general AI that operates in the world for any problem all of our problems that are vertically specific so there's the search for automatic general intelligence or artificial general intelligence that operates the way we do in any unforeseen human situation this is a science fiction fantasy it's not going to happen anytime soon we have no theory about AI AI is all applied it's all experimental there is no general theory of how all of these bits of AI should fit together I've mentioned that deep learning has models that are thousands of layers thick the new technology coming from quantum computing there's also early AI experiments will allow us to have models that are millions of layers thick and operate in a fraction of a second but again, deep learning is now up for some level of criticism that cracks me out of this it's bias, we can't reproduce results, it's all experimental there's no scientific discipline around AI and the AI practitioners are restless so we have people saying guess what's going to happen we're probably going to have another AI winter if your company does not have Google's research budget PhD talent and massive data you're not going to get very, very good results that's what one expert says another expert is predicting AI winter is like predicting the stock market crash you will have seen it coming in retrospect but you will not be able to predict it but because AI is so high right now as Senator Pickhouse says AI is the new fired there's bound to be some deflation in expectations so what's next beware what I call the crapware division where everything is called AI so the House of Lords in the United Kingdom did a study on the United Kingdom's readiness for artificial intelligence and the bottom line was what is called one thing 40% of companies saying AI companies had no AI in their product none so buyer beware when I evaluate business opportunities I look at five dimensions do I like the business model am I excited by the opportunity emotionally are the deal terms appropriate looks acceptable like management risk and market risk and kind of confirm what you told me by due diligence so I applied this kind of a screen where I'm investing in AI I'm not going to go any more to this but just to make a few closing comments in 2019 is not like in the wild west the new world of 2012 consumers are more sophisticated about AI investors are more sophisticated we've already had the way the tools were now moving to deployed solutions vertical business problems as Kaifu Li the leader of the largest investment fund in China and the former head of Google's AI program says we're moving from experimentation to deployment and that's where the opportunities for new companies are like silos on deploying real solutions solving problems not creating theoretical models we're at the top of the hype cycle right now MIT did a study and they said that basically two thirds of all companies today are marketing themselves as AI companies so AI is is everywhere in terms of ecosystems where is the real energy the top ecosystems for AI Silicon Valley, Beijing, Tel Aviv, Boston and London why? because of the availability of funding the number of companies but most important the town pool I'm just going to a couple of minutes talking about things I think if you would like to build more a more robust AI presence here in Armenia I've got some suggestions so if you want to build a big data ecosystem focus on solving business problems marry data scientists with people of good knowledge of vertical markets and areas here in Armenia that are right for the use of AI agriculture this has become a very big deal of marketing for tourism a major tourist industry you could be doing a significant job marrying AI to the tourist industry or to water management or to defense and logistics these are just some ideas but focus on doing some practical things for real people who are able to pay you University industry partnerships are essential in the U.S. all the major breakthroughs have come to some degree by this kind of collaboration often involve stealing people from universities but there's usually a university involved to some degree in the major developments of AI in the U.S. and in Europe if you need to support your companies on average in the U.S. and Western Europe all the golf ring has to raise $27 million before they exit so it's not just the first $50,000 or $100,000 you need funding it may come locally in Armenia from the diaspora you need funding to keep these companies going and growing you need to think about the life cycle of big data in AI not just startups what do you need to support companies from being local to regional to reaching out to Georgia for example or parts of Russia and then parts of Western Europe or the U.S. they're going to need talent they're going to need staff they're going to need trained executives how do you do that in a responsible way and then you have lots going on here in Yerevan I'm blown away by how much AI activity there is but it's hard to find anybody who knows where all the pieces are I've talked to 10 people who have helped me put together a picture of AI because there is so much going on but it's very very very diffused so more coordination would be helpful in Silicon Valley is a great example of the three things you need to build an AI ecosystem you need infrastructure you need attorneys you need somebody to fund this you need accounting firms and you need some way to get better known you need IP and talent incubators, accelerators you have great universities here you do have some incubators here but you need more of that you need the press to make people publicize their success that means the smarter press understands what AI has all about and then you need capital you need early capital from groups like Smartbay you need more mature capital to grow those companies with folks like Synopsys and Mentor Graphics to help fund the growth of your AI companies so you've got a great start I am blown away by the kind of work that you're doing these are some of the more interesting AI companies right here in Armenia or that have major Armenian presence if I overlooked a company it's a favor of yours I apologize but I wanted to get a sense and here you've got Kutnes that is able to take financial documents and make them understandable and usable across the financial life cycle you've got this first which is using computer vision to figure out where are we in a construction project this is absolutely a state of the art you've got two Hertz able to do a voice analytics to make things clearer for people who are hard of hearing like me get rid of the background noise and talent there from my good friend Al Ozian you ever a doctor you hear Al speak run and do it he's an incredible motivational speaker he has got a combination of AI image analysis and drones and robotics to figure out the health of the crops he's doing this around the world I mentioned zero before that's taking a leadership position and using machine learning to understand email and to help a lawyer optimize their day super annotate image labeling again at the forefront solving one of the biggest problems in artificial intelligence smart click is doing predictive analytics looking at consumer behavior inside the stores on online shopping system to figure out what the consumer wants and what they'll do next I mentioned silage before both at the level of national health and at the level of using diagnostics to educate medical professionals two interesting consulting firms development do both help other companies solve problems using their own big data tools their analytics and their teams to create AI applications for those who don't want to build them that's just a hint of what's happening here I think the new program of data science is more of this so in thinking about AI in terms of looking at what's attracting in terms of your ability to evaluate companies that you might want to work for or start or invest in is the product cutting edge is it repeated business intelligence is instead of the art deep learning is instead of the art rule based learning design is there a vertical market partner you're working with understand requirements are you solving real problems or is this a science project remember the mainstream company mainstream customers in real markets solve problems is are you solving a problem or are you developing technology that's going to be used by a few experimenters can't you explain what your AI is doing where are you going to get your data from to power machine learning where are you going to get your capital from when you want to exit doesn't Google or Amazon or Facebook have something like this or not because that's probably going to be one exit path and the most important is do you have state of the art outcomes are your outcomes better than the other AIs so conclusion AI 2019 is different 2012 tons of opportunities for innovation we're moving though from tools to vertical markets so think about business solutions remember that we already have nearly 7000 companies active remember that many investors have already made their bets and remember that the competition isn't just coming from your company it's coming from big companies as well early adopters by tools mainstream buyers solve problems so somebody on your team needs to know a vertical market in their bone build use cases validate results by partnering with customers and in the ultimate project does your company demonstrate better outcomes than the other AIs so thank you I hope this has been helpful in framing what's happening with AI I'd be glad to take some questions in the time remaining how many of you are involved in an AI project class company how many of those are companies how many of you have companies I did not mention that's great because I know I can learn something how many of you are finding some obstacles in your path to working with AI and if so what are those obstacles things are going great that's awesome thank you for delivering this fascinating speech and we're presenting Salix that was cited several times in a speech and we're also working on our AI solution and when we talk about the obstacle the major obstacle is the data of course when you have to train your models you need a good quality data we're in the healthcare sector our company is also a national operator here in Armenia and also in Kazakhstan we're also delivering technologies in Kazakhstan we're expecting at some point of time we'll have data in these two countries but when we're developing now the AI and I believe this is the obstacle for everybody you need to find a good quality data in order you can train your models and this should be the number one obstacle I believe for all who are working on the AI domain so there are data exchanges where companies are contributing we're selling their data there are public domain data sources and it will be great for regional data here if people would start combining or creating some kind of online directory of the public data sources that do exist maybe government data, maybe company data but you need to figure out what the assets are and these these directories are increasingly common in the US and I think you're exactly right you need to do it regionally because if you think that Armenia has 3 million people and we need, you know, millions of objects to power our systems then regional cooperation will be helpful again that's why China is such a threat because China has bazillions of people and China monitors people's behavior certainly, dangerously but they have data and that's why they say they're going to beat the rest of us because they have this vast data store so your point is extremely well-executed one of the messages that I've been hearing all week from the various people I've been talking to is the need, is there one word that's impressed on my head is collaboration that you need to collaborate wherever possible between institutions with you know with industry with normal people you need to start collaborating, sharing building shared repositories would be a wonderful thing thank you if I may that was actually part of the issue that I wanted to raise which is very important question thank you for your answer another part of my question is more if you wish philosophical as the adoption of AI increases in our society well, all of us know that our society is now governed to a large extent by our biases you were talking a little a bit earlier about AI biases but we have our biases right now so how do you see reconciling these two or will our society completely change because of these new biases that's that is a question on the top of many people's agenda right now it is now common in many AI conferences we'll have an ethics component all I can say is if we get ahead of AI and start infusing our AI with ethical concerns we'll have better outcomes but I'm not very optimistic may I please you know there's a danger that if we start to impose our ethical concerns on the AI we will impose our current biases so then we'll get nowhere the best way to avoid that is to have better data so now in the area that you mentioned like healthcare we have stopped simply looking at the dominant demographic in the United States we're now trying to expand our data to include many many different populations so we can get away from that kind of data oriented bias so the first the defense is great data and ensuring and asking 19 times of anybody building an AI that might affect you where did your data come from is it truly representative we even have data labeling problems that the people who label data are going to have their own biases on what's important at the time so we're all human beings but again ask questions validate do your diligence powering these AIs and if you're an organization that is responsible for curating data then you have a real moral responsibility to make sure that data is as unbiased as it can be including the people who label it to the people who curate it the people who claim it so again there's no simple solution here technology is always about to ruin our lives and that's back in the era of the printing press people criticize Gutenberg when anybody can read the Bible anybody will have an interpretation and society will fall apart so we've managed we've managed to survive we've managed to get through this but people be aware of the risk of the dangers and address them when you can great question you raise one should be work of consideration and you're not entirely optimistic and I was going to ask a question that when you get access to a free artificial software that you purchase at the end of Google does Google harvest your data as part of their agreement is that where they get their money and the reason that I raised the question is that it feels to me as an older, much older person than the young people here that privacy to me is more of a concern than it would appear to be in younger people maybe that's an inaccurate observation but a lot of these big data and so forth is giving away access to things that can come back to society in a way that you may not actually like in other words if you can predict what an audience is going to buy or you can advertise selectively you start to change people's behaviors based on this data and they don't know it and so it feels like how privacy has managed how the risk of harming somebody who's giving you something in good faith and doesn't realize they can't worry about this competition and the industry leaders can't so I just wonder what your view is of somebody looking after those problems for us Larry you're exactly correct how many of you read how many of you know the term EULA this is something you confront every day what is EULA end user license agreement end user license agreement comes with every piece of software you buy and we all don't read it and we all lightly skip through it and check by end user license agreement typically says what the company is going to do to you and there's an old saying in Silicon Valley if you're paying for a free product if you're not paying for a free product like Facebook and you wonder what's going on so what is Facebook's business model they give us all these free stuff then they watch us they track our behavior and they use that to manipulate us or they sell that data to somebody else so we have become the product if you there are some right now in the United States Congress is thinking of some rules against this not demanding that these big tech companies not sell us as the product without our explicit permission not buried in 4,000 words of legal junk so there's a reaction against this but being informed consumer do I have great hopes that this is not going to happen what's happening in China how many know what the social credit score is social credit score what is the social credit score that's monitoring behavior of our product so China is tracking every kind of behavior how many times literally how many times you were spotted walking across the street how many times you have been to certain kinds of meetings what kinds of things you subscribe to and they're giving you a score which is really political reliability score and some people are not allowed to travel outside the country some people are being denied jobs based on this social credit score because the Chinese are collecting data on every aspect of your behavior think of all the things that Google knows about you Google knows who your friends are who your family is Google can understand what's your religion they understand who your kids are where your kids go to school where you bank these are all the things that Google knows about you just because every time you send an email or buy something Google is scraping all of this information but magic has happened to get into the wrong hands so there's big concern now in France and in the EU and in the US about how do we deal with this we need to get ahead of this potentially by some regulation without killing the process so we're very early in this process but be aware of the dangers and keep asking these questions I don't have a simple answer Larry but I do know that the minute we forfeit our rights by clicking the box on the end-user license agreement now in terms of younger people are more willing to do provision-based marketing than older people some people say hey I'm being monitored anyway I can't stop it I might as well go along with it getting outweighs the potential risk to me so we all make our individual calculations I've given up years ago myself I just clicked the box what do you think about the integration of AI and non-technology creating such a value of chips technology being non-technology creating non-technology as a serious so already going to go along I'll try to I think it's already happening particularly in the area of sensor technology that's going on so I think nano and quantum are going to usher in a whole loop generation of AI so I've been talking a lot about AI and deep learning and big data you run the data science project here would you like to make any comments I think that the decision was very are there issues that this group has come up with that you plan on addressing as an administrator as a teacher well as a second teacher we think that our students need to get understanding of more and more industries on the side to define the business problem to be able to communicate in the right way to be able to communicate the exact message they need so we'll review a sense of interest that involves AI so let me make a personal comment that's a great way for us to wrap things up today I had seven careers in my life as a professor focusing on the Italian Renaissance and on medieval Christianity and that was where I was doing for many years and I ended my life in venture capital and AI and things like this what I've learned is that the old humanities that we tend to say that's done, that's understood the humanistic skills of communication engagement with people understanding people's motivations understanding how effectively to work with them those are soft humanity skills and I think there's an ongoing place even in the world of big data of AI for people who can effectively communicate the old fashioned way who can write coherent business plans and engage with customers to understand the requirements those are soft human skills programs around AI big data, machine learning we'd be able to integrate those kinds of skills as well because ultimately you are serving a customer and you can't understand that customer unless you can dialogue effectively, communicate effectively so I think there is still a reason for the old humanities my boss for many years was Steve Jobs the founder of Apple and Steve's favorite thing we used to have a picture of this at one of our companies was the intersection of science and humanities so just keep that in mind that you need to be effective listeners and communicators as well as effective coders, designers and data wranglers so let me say thank you you've been very attentive, great audience if I can help anybody around these issues my email is at gmail and these slides will be available to you thank you