[Music] Thank you everyone!, Thank you Rod for the invitation My name is Johnny Saffar, I am the CEO and founder of OrNsoft So OrNsoft is an AI Software House We have been around since 2009 working in many different countries and helping many different types of companies The objective of this presentation is to make you understand what is this technology We've been hearing about AI since the beginning of the year, a lot and uh some stuff were true, some were not really but there's a big buzz and uh and the the girl my goal today is to make you understand what is AI and how it can help you "really". So first let's talk about demystifying AI, so the the the first thing I would like to uh to say is about a the survey that was done by the leading scientists, leading AI scientists, and according to those AI scientists, 50 of them are saying that there is one chance out of 10 that we are actually all going to die with this AI. That's uh that's quite a statement and just so you understand now, it's like saying there is 50 percent of the engineers who did build the plane saying that if you ball the plane you have one chance out of 10 to die so I I don't know if you if you I will not take the plane to be honest so we we did start really to speak about AI since 2020 people did.. the hype let's say started but yeah I started a long time ago, it started in 1950s when the basic concept of where I was created uh the the first person who fantasized about AI was called Alan Turing and he said that there is a basic set of questions and if the machine can answer those questions the same way human being will answer those questions it means that the reasoning and the.. the reasoning of the machine is matching the one of the human being, so that's great, a very very nice Concept in 1956 the real birth of AI happened with the first machine learning, and I will I will get soon into details of what machine learning means, and then the first neural networks in 1957.
So Since the 50s all the way up to about 2017 there was progress but not that much, right each field of AI was, was really individually developed and the progress was really segmented per type of AI, it was AI for audio, there was AI for images, there was AI for voice... many different types of AIs and when you were studying and learning about one type of AI it was a not the same books, not the same way to learn as the other types of AIs. Those types are used in many different fields right sometimes they are combined and what happened in 2017 is called, there is a paper that was published by a researchers at Google regarding a technology called Transformers okay, Transformers are very important and that's what did made the big change because that allowed the convergence of all the AIs into one main AI, so I'm gonna gonna formulate it a different way, imagine you have images, and you have text and you have video, the text for us it's a simple to understand there is a text that can be understood by AI, and then and then processed by AI, now if we talk about video, video is a completely different type I mean different type of processing Etc... so what does the the Transformer technology do is convert this file into its basic element, it transform this file into its basic element basic element is binary file I'm not sure if you're familiar with binary but it's like a basic computer language and the Transformer technology will do this for all the formats so when it does this to all the formats then the principle of of AI to learn and understand and uh it will not be different from one type of AI to another, they are all the same now so the the I will get more into exactly how it's working soon but now you have to understand all those incremental uh advancements for each type of AI that was very slow since the 1950 are now all converging to one and when one scientist does an advancement in one one field, it's benefiting all the fields, so that made the first real stimulation of AI, so let's go a little more this is actually the graph you can see from 1950s to 17 was not that big right 2017 the Transformer model, Transformer model, you can see this as an engine a new engine to process the information was born, the scientific Community took this technology and used it across all the types of AIs, 2017 to 2018 you can see the curve goes up and then something happened in 2018 a new type of engine again called generative pre-train Transformer, and that was released by the openai community, because back then it was an open source and an actual community and this uh pre-train Transformer was a very very uh.. I'm going to explain because this this is very important to understand, this is not a New Concept "pre-trained Transformer" You've been using most of you I'm sure, cell phones seems at least the 90s the text messaging, you know where starting to type a word and it was finishing the world, or suggesting the next one, this is the technology that's used for the generative pre-trained transformer Just at a much higher scale right, because there's billions of informations.
So what does it do, it's trying to guess the best way possible, what will be the next word, and this is the way for it to generate the the end of the sentences for example, now if you take this into its basic element, like we said earlier, you have the beginning of a text it can finish it for you.. you have a beginning of an image, it can finish it for you... same thing for a video, same thing for an audio, we know today that with a few seconds of audio of your voice, AI can generate much more, and do whatever it wants, same thing for images right, so that's really the technology where we are at today 2018 that was released that stimulated even more the community the adoption was very high and you can see 2018 to 2022 it went even faster, now 2022 November 2022 that was not so long ago, ChatGPT was released, chargpt was released and something even greater happened, and we are not talking about adoption by the scientific Community or, or really only technical people, we're talking about human adoption of the technology November 2022, ChatGPT was released. January, two months later, they already reached 100 million user, in two months so to give you an idea, the fastest one before that we reached 100 million was Instagram and it took them two years, so in human history it never happened something like this This was a big signal to all those smart people doing business, saying oh! maybe it's time right, and I'm sure you you've had the same reasoning now let's break it down a little bit, talk about the basic concept of AI Obviously there is much more, but I'm gonna try to make you understand the logic behind it.
we have four different elements that will compose the AI, the first thing is called the data set, so the data set, it's very simple and those are the information that we will use as a training, so we want to learn we need information right. The second one is called a neural network, so the neural network I like to see it as the neurons and, actually that was the idea behind it, neural networks is the neurons that are in the brain where you store the information and connect those information together The third one is called machine learning, this is very common World also, you've been hearing this one most likely a lot. Machine learning is compared to the synapse, the synapse meaning the the things that will make the information work together, that's also what mimic the reasoning of the human being, I know this, I know that, maybe and then I can extrapolate, now by combining those things, that will give birth to what we call the model, and the model is the actual thing that will be used to do the job, so the intellect, now you know things you can do things.
I would like to do an analogy, with the first human being, and to make sure that's uh very clear for for everyone, I'm giving this analogy a lot to some clients, a lot of clients actually, and they seems to like it so, this first human being, imagine it's night and he's seeing fire, first thing that come to his mind is "something shiny" or first information that stored in his neurons in his brain is that there is something shiny. This human being is getting closer, it's getting closer and he realize that "fire is warm". Second information, fire is warm, now what happened with the "machine learning" ? A third information is born saying "fire is good". Who gave him this information that fire is good ? Nobody! He did end up realizing that this is good because... and this is the the what
we try to mimic with the machine learning. Now you can go even further, he's getting even closer trying to touch the fire, now what happened? "fire burn", "fire can cook" "fire is dangerous", "fire this", "fire that"... a lot of other information are going to populate the brain. Those information have not been given by anyone, they have been produced by the brain and where it's interesting is, you remember at the beginning, we said fire is good... now fire is not so good anymore, so fire is not so good, so it can also correct the information it had in the brain.
That's the main difference with other technology. Before that we had a great deal of automation, We had great deal of of trying to do things that looks like human being were doing it, but the goal of AI is really to mimic human processing, without the constraint of course, because we know that AI don't sleep, AI don't take vacation, it's not sick, and of course, work way much more faster than human being. We'll get back to this information later. Now! There's different type of AIs, and all those types are basically combining different Technologies always in AI, and by doing this you can achieve different things. The first one is called Curative AI, so the curative is the one that will help you find a solution and do things to fix things, so that's the most common one in business, we have a problem here, we need an AI to do this, and that's what it does. The second one, and this one is very popular now is the generative AI. The generative AI is capable of creating new content, ideas, images that's the one we talked about a little earlier.
The third one is called predictive AI. So this one is capable of seeing through a lot of information, things that happened in the past, not just basic events, but also go into details thousands, hundreds of thousands of parameters, compute all those information and give you an idea of what will happen next. And finally the third one, and that's my opinion, but I think this is the most powerful one, is the prescriptive AI. So, in addition to being able to tell you this what may happen, or this will happen, it will tell you also, how you can avoid it, how you can take advantage of it, what's the best way and suggest what you should do next.
Those things could be used in many different areas of business. As you can see there is no an industry that that will benefit more than another form AI it can be used in any department, it can be marketing, accounting operation, administrative of course, legal, sales support, etc.. everyone can benefit from AI. Now, the idea, the most important thing is to identify really if you need it, and where you need it, because it's not just to say oh I want to do this. No there is an actual logical way to identify and say okay this can be done by AI, these can be fixed by Ai and we have an AI implementation action plan if we can say that you can download, after a presentation I will make those documents available to everyone, you can download it, just go through it and see with your own company, you will see that it can help a lot.
Now we've done a benchmark with our own AI, obviously we have our own softwares also . And this is the latest Benchmark we did, it was at the beginning of the year, and to give you an idea where we stand right now, our AI can execute a week of work, done by 50 highly trained people, in less than two hours, with the five time better quality so, just realize the amount of work we are talking about, the AI can do it in about two hours the benefit that you can get from this is not just, okay obviously the main one is money, right, but also, time, the client satisfaction. You know when the client doesn't need to wait a week, or a few days to get his answer, and he can have it right away within two hour, that can be huge, also Imagine, when you have to approve a loan or process a project or whatever, this can be done in few hours, usually it's taking half a day, two days, so that we are talking a lot of benefits from this kind of technology. Now If you're not convinced yet that you should board the plane I'm gonna give some examples of first AI implementation that went wrong, and then I'm going to give you examples of AI implementation that went very well, with a little more details so first the implementation that went wrong, there is quite a lot, I did select a few some from big companies some from smaller one, and in I try to find things in different Industries. Mainly when it goes wrong, there is a reason. The first reason could be because the AI used
the wrong data to learn. So imagine you have a new employee, and you tell him you know what I'm going to teach you what you have to do, and you teach him the wrong way, this employee is going to do a bad job, are you going to tell him : "hey, you did a bad job!" You don't train me well, how do you want me to do a good job" you know it's a logical thing right, so it's the same thing with AI first thing and most of those problems are coming from bad data set, the second reason is that limitation of the technology, so sometime and that happened a lot, we I've seen it mainly during the hype of AI on 2021, 2022, people saw ChatGPT, or Bard, or some other generative AI and they said oh great that's it, I'm firing everyone from my Support Department, and I put the chatbot oh, okay, you can try but that may be a problem. The AI ended up talking to clients, or you have AI : "ho, you have a problem with your car? Guarantee we replace the car!!" The client comes in, okay where is my new car ? No sir, we don't replace the car, you know! The agent said you replace my car! I don't care it's your problem! But it is an AI ! I don't care... and and you end up with a lawsuit..
So you see, the risk the liability it's not because you use an AI that you're not liable right, you have a business, you have to be very careful what you do with this technology, I'm not saying it it cannot be done, I'm saying it has to be done the right way. You see in these examples, we have the Covid-19 that was a very famous one. The technology was, I really believe the technology was working, and the main problem was the data set, like we just said, we have some chat that were telling patient "kill yourself, it's better, you will feel good", that's uh.. that's a problem also some were biased, because there is also a problem with the data, the the AI sees that men is always at this position, has been trained with hundreds of thousands of information, it's gonna think that, okay maybe men is better for the job you know, while it's not true.. today we can say this with certitude, and the latest one I saw is from a lawyer in New York, who used chatGPT to prepare a filling and the AI made up information, it was using a information that for the appealing everything that went after that that did not exist, and when the judge saw this, he said wait a minute let me check, he checked, and this information didn't exist so the his career finished on the spot right, but you realize that it could be used, but the right way, generate AI is very powerful, we believe it can be used in a certain manner I'm going to give the next example, and we'll see how it can be used the right way so The real world application of AI "the right way" I'm going to give an example, first how we did implement, to be honest it was difficult to get the approval from clients on sharing information that may be confidential, the first one is regarding our company how we did it, because I have no problem sharing this information, and the second one is from a local business here in Miami and we'll go over it, and I will explain what type of AI was implemented and how.
So, the first example is regarding customer support So in our company, we have clients, and those clients, when they have a contract, can raise a ticket when there is an anomaly, or whatever question they may have. Now at the beginning they were sending email to the company, and to handle those emails we had two human agents, those were technical support people, Junior, two years experience, paid between $20 and $25 an hour, and they were to be honest very efficient, but the amounts of tickets grows as you acquire new clients right So, what did we do, we did try to centralize those requests into an Extranet, so the extranet is like a client portal, where clients can go, open a ticket, until now, nothing special right clients where uh okay with that of course, but they didn't want to go there and open the ticket, and they were still sending emails, so we did add an extra layer, and that was not AI, it was automation a lot of people out there are trying to sell you things or tell you about things that are AI, but they are not, this is automation. We did build a bot that was monitoring the mailbox and when there is an email, taking the email, go in the platform, open the ticket, that's not AI right now that helped a lot of course, but something happened because you reach a certain point, where you are at capacity and the next step is to hire another person, you cannot just say okay I'm gonna pay a little more and I can I can handle no no you need an extra salary or at least part-time right, so what did we do we had about 300 tickets a month for two full-time employees, that's a $6 000 to $8 000 dollar a month for those two employees, and average about $80 000 for the year I'm giving numbers to show you really the impact okay what happened uh once we did change this step of human agents, Just so you understand why we have this step, i went a little fast on this one, we have this step because when the ticket comes in we don't want the engineer to work on it right away, maybe there is not enough information, maybe it's not a bug, the engineers are very expensive right, we don't want them to lose time trying to guess what the client is trying to say right, so that's why we have this layer in the Middle with this technical support agents. Before we used to have those technical support, Junior, two years experience agent, between $20 and $25 an hour, two full-time employees so that's about 80 000 a year for those two people.
To take care of those tickets, there can be a delay about 24 hours, based on the SLA, based on the contract, but when the company is closed, nobody is going to answer, you have to wait tomorrow morning those two people were at capacity, and if we want to handle more, we have to hire more people now after we did Implement AI in this department this is what happened, instead two people we still kept one part-time and I will explain soon why this person is still paid between $20 and $25 an hour that's about $19 000 a year, now with this, there is a near real time answer to the tickets unlimited capacity, there is no limit anymore, it's not going to cost more, and to show you, you see on the right the implementation cost of the AI uh for for this department was $20 000 for the license, there is other fees, the processing fee because each time the AI is thinking it's costing money right, of 50 Cent per question so the total cost for the first year was $40 800 if we compare this to the $80 000 when we had two people, this is already 50% saving on the expense, and I'm not talking about the user satisfaction, to augment the capacity people are very happy, and they're engaged, and if we want to take on more clients we'll not be worrying about, holala, we have to find someone and hire people, no that's covered now also you see this, I'm saying here the cost for the first year because the initial $20 000 implementation this is only the first year, so the the next year we're going to make even more savings, so it's very interesting. It's a simple example, but I think it's very effective to make you understand how it can be done, now regarding the technology, when I say the right way what does it mean in this scenario the AI is going to communicate with clients, the ticket comes in, what does the AI do, the AI will read the ticket, understand what the client says or want it's going to go check who's the client, what project the client did with us, the history of all the tickets that were opened by this client, who is taking care of this project, and based on this reply to the client, we need more information, is it the same as this this and that, or thank you for raising this ticket, we'll take care of it as soon as possible, and the AI will assign the ticket automatically to the right person in the company, so you all this job usually require the human intellectual job, you see it's not automation, it's really thinking, seeing, understanding and then taking action, that's what's done by the AI, in here Now, let's move on to the next example this company Dayoris Doors, so those are based locally, in North Miami, they offer Modern Luxury interior doors and they manufacture those doors now when they get an order, this is the process they have today. I'm telling you I will not be able to share financial information, I'm not allowed to do it they did not accept, but I can give some details on what we did, and what was the gain out of it.
So usually the client requests an estimate, when they request this estimate, they send a plan, or they give the requirements, what they want, I want twenty doors, ten doors, usually you don't order one door for your house, you change a set right and the agent, the person on the phone, will take the information and prepare the estimate they send the estimate, the client approved the estimate and when this happened, they generate what they call a master sheet, a master sheet is a document where there is all the information detailed information, the size the way the door should open Etc... the material that should be used, there is a lot of information in this document now they generate these documents and they have one person going to the client, on the field, going over the master sheet with the client and based on what the client says, adjust maybe this door supposed not to open like this, but like that this material he wants a metal instead wood we don't know, there is always change orders now they take all this information in the in the master sheet, and then go back to the office when they get to the office they do what they call a Reconciliation, whatever information that has been added or modified in the master sheet need to be reflected on the estimate, now that's what I call "the pain point" the pain point meaning, this job was heavy, risky, and for this company was a real problem because there was about 5% to 10% issues with those reconciliation, so once this reconciliation is done, then there is a change order that is sending out, update the estimate, client sign and to be honest, when you see it like this, it seems pretty solid, bulletproof almost, and uh and and supposed to be fine they used to have four people to do this reconciliation and per project, it was taking up to three hours, so just to make you realize, it's a lot of work for a lot of people, it's not a simple task, now after we implement the AI, this is what happened before, they had the four people full time they were at capacity, not everyone could do the job, and you cannot use someone that knows the job, that you hire and they can work right away, you have to teach them, make them understand why and how it works, and all, this so it was difficult to increase the capacity. 3 hours per project, there were weeks behind on a reconciliation, that's a problem because reconciliation is not done, the actual final estimate is not signed, a final estimate is not signed you cannot start the project, so it's not the minor problems, it's a huge problem after we implement the AI there was one person full-time, we kept one person, the capacity now is unlimited because it's AI right, the time for this person went from 3 hours / 4 people to 20 minutes 1 person and for the the reconciliation, at least the AI part is near instant and if they want to do more, they don't have any problem anymore, so just to give you an idea this is in term of time and resources 97 gain that's huge they are very happy, and now we are working on other departments but uh I I can tell you the the when AI comes in it's not only a 10% gain that you will have, it's way much more, remember, the AI is working way much more than the human being, if trained properly, can do a better job so it's a big gain the last time we saw such a big uh benefit and advancement was the last Industrial Revolution in the late 1800s beginning in 1900s, when human labor was replaced by machines in the factories Did people suffer from it? yes, okay Did we benefit from it? civilization I mean.. yes, Was it avoidable ? no, and what is sure is that the one that did jump in the train, those are the ones that stayed after that, the other one disappeared so I think it's pretty clear that to have an AI today is not an option, you will have to take it, don't wait too long because your competitor will not wait for it, and just be very careful on on how you do it, always ask yourself what's the liability, what will happen if it goes wrong, I'm not talking about just it doesn't work while it can cost a lot of money so you see that's one risk there is no real liability behind it unless it kills someone right, but when there is an AI doing something for someone you must be careful Now if we want to talk about the future of AI The future of AI, I think that's my opinion is the convergence between machines and AI the body with the brain right, the body with the brain, actually the future is already here and it's moving very fast, there is a company, and I did find this example very interesting who did a point as its CEO an AI, and this company it's a big company, large company, they in the Stock Exchange and they have thousands of employees, is now managed by an AI between the time they did appoint the new CEO, and about six months later, the stock went up by 10 percent, profit skyrocketed, without them having to sell much more or firing anyone so you see that that's a pretty impressive, just by operating better without causing any damage to anyone, if I can say you can actually turn things much better, there is some more examples about that but I did find this very interesting, I'm pretty sure soon, we will see much more of those, helping us in our personal lives and also in the business. For now, replace the employee completely, no, I don't think so assist the employees, yes. Thank you very much [Applause]
[Rod] Can you answer some questions ? sure if any ones have questions yes [Person in the public] How long from like when you started working with the door company until you train the ai ? and then how do you monitor it once you put it in place to make sure very good question okay so for the company to prepare the data set because that's really they took more time than than us actually uh to put in place so they took about amounts to get the documents ready because uh we are dealing with documents we have a software that is specialized in what we call hyper automation, so intelligent document processing we can just once those documents are ready give them to the AI, tell the AI what to verify on those documents, you have the estimate you have the master sheet, this should match with this this should match with that once this is done the training goes very fast the training takes about an hour or two after that you need to use the AI, so that's the time that took about maybe three to six months, and uh within that time, but don't get me wrong when you put it in there it's deployed, it's going to be maybe at 50 capacity, I mean capacity of doing the job this is why we keep a human being in the loop, the human being will see what the job the AI is doing correct if something is not okay, the same way you will do with a new employee, you have a new employee you go over, you say okay here you did a mistake, here it's okay, and after six months you tell him, okay now leave me alone, do your job I have no time, you know, so it's about the same thing yes [Person in the public] What's the risk when we use article intelligence in yes political campaign ? The main risk I can think of is hallucination, so hallucination the the technology the AI technology and again in political, it could be used for prediction or prescriptive, tell us what should be done, or where we should go, or there is different ways to use it, but I believe the main way today people will use it is more to generate content, for example generate content, make sure to use a technology that will be evidence-based, meaning the generative technology should be used to understand the request, to generate an answer but not as the base of the knowledge, because it has so much in the brain, if you ask it about something and it doesn't have the information, because it has so much, it's going to make up information and that's when problem happen, we have another technology called GVK not GPT this technology will take a set of information that we will provide get the question, generate an answer based on this knowledge that we did provide, and give the proof the source of the information along the answer, that's the right way to use it so yes the the I believe that would be the the highest, the highest risk, yes [Person in the public].. kind of slim down our overhead and I'm extremely fascinated with AI and the future if you just have a couple questions that I want you to uh just to you know give me a quick answer to or some information, um I've I've been told the tractors have told me AI is too limited right now AI does not give access to current information to limit dangers, AI databases are one to two years outdated they're pre-loaded with certain information certain access to this information and for me I just made it to you on what AI is learning and also that AI has no access to Legal databases, so if you ask AI about a particular case it won't know so I was curious about that [Johnny] sure sure that's, that's a very good point if if you don't have the knowledge of the different ways technologies are working the the AI is not one one type, there is different types of AI and different ways to process the information, what you just said is very true, for example chat GPT again, they are not gonna like me a lot but when you go in there and, I like this one, I go in there and I ask it not so long ago, who won the the World Cup soccer and it was still saying it's France, so for me that's okay [laughter] , but you can imagine if you ask it about something else that you need in your business and current information, that's not going to work this technology I was just mentioning the GVK is working a different way you take the current information you provide it as a document PDF, the AI will ingest it in a vector format, so I'm not going to get too much into the technical, but basically what it will do is when you ask it something it will understand go in the base of knowledge that you did provide look for the information, we are doing right now an experiment with, and that will tell you it's not so far from the example that you have, but we do an experiment now with the city of Miami website, where you know when you do construction, it can be very complicated very quickly, not just for us people trying to change something in our house, you know, but even them, when you call them sometimes you know, they need time to find the answer, and that can be a problem, so we went on the website did an experiment, took some of the information, just by printing them as PDF we did put them in the folder, and we used the generative AI we tried with ChatGPT, Bard some others that could be installed locally in your company if you have privacy issues and uh we plug those technologies with the GVK technology now what happened, when you ask the question, the the GPT technology will be used to understand what you say look for the information get back the information generate the answer and give you the link to the the actual file where it did find your information for you so if you use it that way you will not have problems because if it's not in the base that you provide it cannot make up an answer [APPLAUSE] thank you thank you so much
2023-09-19