965: Appvance: The Innovator of Generative AI with Keynote Speaker Kevin Surace

Foreign.

Welcome back to the podcast. We're here with an artificial intelligence and AI innovator, Mr. Kevin Cerise from Advanced AI, and we're going to talk about how AI is about to kill off all the humans, kill off all the jobs, completely change everything from the, the way you wake up and drink a glass of water and everything that comes with that. So, Kevin, glad to be talking to you.

Yes. Since life is about to end as we know it, everything's good. No, I, I don't think it will, but it's a great way to intro the, the, the, the place you.

Yeah, it's attention getting. Right. It grabs people. And with all this AI stuff, especially in the last year or still, I'm not even sure what to think. I can't tell if there's like, too much hype, not enough hype, is it going to change everything? And then I, I see all these sort of case studies about how the AI can, like, look real deep in data and get these insights or do things automatically, and I just, I'm just, I never even know what to think. So do you ever get that reaction when AI comes up? Someone's just like, I'm still having trouble after all these years, wrapping my head around it.

Yeah. I'm, I'm overwhelmed. And lately with ChatGPT and these kinds of things, they're overloaded.

Right.

They're also overloaded because they've got people on one side, including Elon Musk and others, saying AI will destroy the world. You got people on the other side. I would be one of those that just want to explain how the technology works in very simple terms. And once you understand it, you go, that can't destroy the world. What are they talking about? They're getting lots of press and lots of news, and they're getting started, their companies, they're getting to raise money. Unless someone hooks ChatGPT to our nuclear arsenal and asks it to launch the nukes, it's not going to end the world. Remember the movie War Games? Right. That's exactly what they did. So don't do that. Right? Don't. And in fact, I, I bet most people don't know that. Thirty years ago, the US and the USSR signed an agreement to never hook artificial intelligence. Had artificial intelligence for more than 50 years in a variety of ways. Never hook it up to where it could launch nukes without human intervention. Right? So we could have done that 20, 30 years ago because we've got NORAD, it senses all of the things happening, but multiple times we've had false alarms in NORAD that showed missiles coming and humans stopped and said, I don't think they're really coming. Unfortunately, they just disappeared from the screen. They weren't coming, right? So AI would have said, they're coming. I'm going to launch. So you, you know, aside from that, this very simple but, but exquisite technology cannot end mankind and is not going to end all our jobs.

Well, that is a relief. And so you said that you have a way to explain it. So how would you explain it? Like as like a predictive text on steroids or like what's the, the simplistic way to, to finally understand.

All right, let's, let's go back in history. 1970s, the electronic calculator shows up, right? Do you know how many people looked at that and said, one. It's a miracle. It's unbelievable. Look at the math I can do. I can get the square root of 1.7956 or whatever it is, right at 9, 10, 13 digits. How many ever LEDs we had in the calculator at the time, or little bubble readouts or whatever, and I could do math there. That would take me 30 minutes. And it was done in under a second. It was unbelievable, the electronic calculator. Sometime after that, about 10 years or 15 years, we had the advent of Excel and spreadsheets. It was actually Lotus 1, 2, 3 first, then Excel that totally changed everything we did in accounting departments, like everyone who was in accounting department, rows and rows and rows of people were adding, adding, adding, adding, adding in columns. It's what they did, right? Rows and columns. And all of a sudden those ledger sheets went into sheets, spreadsheets that we all use today. And we look at those two things now as just regular tools. They've been in their life for decades. They're just a tool. And what I want to tell you is they did to math what a large language model today is doing to language. Now, language was just a harder problem to solve. It took 30 more years, right, or 40 more years. But in the end, all we're doing is creating models that, that are able to spit out language in English or any other thing. Now, how those models work really quickly is actually math, surprisingly enough, there's no intelligence, there is no sentience. It actually doesn't know what it's doing. It has seen something like a trillion phrases in the case of ChatGPT, but some models are larger or smaller. And it has tried to, when you ask it a question, it. Right, it tries to go back in its history in Essence and tries to rebuild a phrase off everything that it learned in English. That makes sense because it's learned how phrases get put together, how sentences get put together. All it's doing is building sentences so you can break the thing easily. By the way, a large language model by asking it to do math. It does math terribly because it learned language, not math, and most often 80% of the time give you the wrong answer. I do this all the time to demonstrate that to people that it's not actually smart at all, just putting sentences together. Also, people have heard the term that, you know, that the thing makes up things, right? Say that again.

It hallucinates.

It hallucinates, right? Yeah. Hallucination is an anthropomorphic term. We're giving it a term that people, when they say really weird things, we go, well, they're hallucinating. No, here's the thing. This is a math model, right? That is model language. And it learned from hundreds of thousands of novels, fiction, as well as millions of pages of fact. So if you head it down a road to get it to say I love you, it will say I love you. Not because it loves you, because it doesn't really know anything about love, but it is read millions of phrases of love. So it knows how to build sentences about love, build sentences about loving, build sentences about family, build sentences about sentience. However, it doesn't know what it is because it doesn't know anything about building sentences. No more than a calculator knows how to add and subtract. It's the same concept. Yes, it's a deep neural net, and yes, it's, you know, it's really, really cool technology. And yes, we need huge cloud gpu sets of, of servers to do this work, but in the end, it's just math feeding you sentences. And sometimes they're really cool and sometimes it's junk.

And I've noticed that it's sometimes work to sort out the junk from the real stuff. Like you said there, it's like a calculator, but instead of math, it's language. And when I first came across chat, GPT, much like you, I'm sure I played around with it, right? I tried different prompts. I said, can you write me this? Can you write me a movie plot like that? Can you rewrite this in less words? Can you combine these things? And sometimes it gave me quite a mess to sort out. And that's where sometimes I had this difficulty of where they say it's sort of like your writing buddy, right? You want to write things out and then it gives you all these things. But then I feel like after it does the work, I have to do even more work to kind of refine what it, what it makes. So what do you have to say about just like, it's like a different kind of work. Like I end up having like edit instead of right from scratch. And so how do, how do we use that sort of thing to be more productive?

All right, so I'll give you some examples where we've been measuring productivity over the last several months in coding. We have something called Copilot that's also based on GPT for the same model behind it. Okay. Microsoft has Copilot. They're going to start charging $30 a month for it. Almost every developer in the land is using Copilot to help them write some of their code. It turns out about 30, 40, 50% of the time it gives you code that will not run, doesn't run at all. However, there are some brilliant pieces in there that you can pull out and edit and use if you're trying to write test scripts, for example, like the work we do at Affiance, and to find bugs in your web application or something. Most of those scripts actually don't run. They're not written correctly, but it's a good start. So what we found is with coders is it didn't replace a single coder, but it did make them 55% more productive because editing was faster than. I'm stumped and I don't know where to start. And what coders did is they used to go to open source libraries and look for things. They would search for them on Google or search in other search mechanisms, Stack overflow and others, and find examples. So now this generates an example that's really cool. However, the example often doesn't run often. By the way, when you looked in open source libraries, that stuff didn't run either. So it's not anything particularly better, but it's kind of interesting. And so I like to look at this not as augmented intelligence or an assistant, but kind of amplified intelligence. So, you know, we can take you and all of a sudden there's like one and a half or two or three of you, because if you were writing a marketing blog post right now, you actually could absolutely use ChatGPT. It's really good at writing a blog post, you will be editing it probably for half an hour. However, you couldn't have written it in half an hour. It might have taken you two to three hours. And there's some Mental blocks. There's some thinking. And so maybe you took three hours down to a half an hour, and now in three hours you could write six of these. So maybe it made six of you doing that task. That amplified your brain power. Right. That's pretty cool.

So I guess the answer is to actually measure the amount of time doing it from scratch, doing it yourself, and then measure the amount of time doing it with chat gbt. And maybe it's a little uncomfortable being more editor heavy than writing from scratch heavy. But when you look, if you actually ran the math and it would be a net gain, it would be a multiplier compared to doing it from scratch. And, you know, you mentioned going and grabbing the code snippets from Stack Overflow. And of course, we've all been there where you go and you want to create a function or fix something and it's in like the wrong language or it's just a piece or it's from like six years ago and it's junk, it doesn't run. Yeah. And grabbing it on its own doesn't work.

Yeah. People submitted stuff to Stack Overflow. It didn't say it ran or maybe it ran for them, but we don't know how. It's missing some resources or something. We've all done that. Right. So I think these, you know, code assistants, There's a Stack Overflow one that launched today, Stack Overflow AI or something. These code assistants are better at giving you some interesting choices for code. But it's kind of like an assistant. It's kind of like an amplifier. Right. It is not perfect, probably won't run, might run if you got lucky, but it gives you some ideas. And we all have, you know, sort of writer's block a lot for a lot of these things, whether it's marketing stuff, sales stuff, whether it's code, whether it's test code, whatever. And so this gets rid of writer's block. It goes boom. There you go. And you go, huh, it's kind of interesting. Maybe there's something I can do with that. Right. And then you start editing and you will find that, you know, those 20 lines of code you can probably edit in 10 or 15 minutes versus writing them from scratch. Now someone's going to say, oh, I could write 20 lines of code in 10 minutes. If you knew exactly where you were going from the beginning to the end. And of course you had the syntax right and you knew the language and all of that. Perhaps. But, you know, we, we know how much code people write in a day. On average and know it takes a long time to necessarily get a new function right. So I think it's fabulous. I think it's fabulous, like I said, for marketing, it's fabulous for writing anything that's in the English or some other language. Convert it to here's what I want to write. Write it in the style of Shakespeare, in French probably. Well, I don't read French so I won't know if it's right or wrong. But I did that.

Something's there and that's fun too, like knowing what sort of prompts to put in, like say, say this, but say it like Eddie Murphy would say it, or like Robin Williams would say it. And so you have this helpful perspective about it, that in whatever context you're using the AI, the ChatGPT, it's your assistant. And you mentioned a couple of times you have this company called advance. And so can you share with us about some of these, these cool innovations you made lately with AI?

Oh, sure. Oh, but I've been involved in the AI space for 25, 30 years and the first work was back in the 90s at a company called General Magic, big public company here in Silicon Valley. And we invented the first AI based assistants, virtual assistants. All of that underlying technology went to General Motors, OnStar eventually. And eventually that underlying technology and the concepts were licensed to make Siri and Google Assistant and Alexa and all of those. So that was the work we worked on at first. I went on to work on building management systems that used AI to figure out how to, how to manipulate heat and light and cooling and all of those things to reduce energy usage. And then at advance over the last decade or so, all we have been doing is using AI to find bugs. And that's sort of the number one goal of that company, use AI to find your bugs rather than you or someone else, you know, writing lots of test scripts to find your bugs. Because the testing of our applications, all of us use applications, web applications, mobile applications, they all have bugs. Testing of our applications is getting so difficult to do at speed, at scale that we really need technology like this to help us find our bugs. And so we launched our first product, I don't know, five, six years ago, with AI, with a type of generative AI. At the time it wasn't a transformer model, a different kind of model. And it's been finding, you know, thousands, I don't know, millions of bugs all the time, every day it's used. So it's really fascinating. And you know, much of what I do is I speak around the country on exactly this. I'm a keynote speaker and, and try to calm people's fears on AI and talk about where the jobs are going to be and how you upshift and upscale your jobs and your tasks into new things. Become the robot overlords rather than, you know, underneath the robot.

And I hope that AI is helping you to brainstorm new speeches and find some speaking gigs.

You know what's funny? Haven't used it on the ladder. That's a good idea. I have about, I don't know, 20 or 25 Speaker Bureau agencies. So they book me. So I'm not really very much involved in the booking process. They just call and set me up to go somewhere and I get on a plane and go. But in terms of speeches, I have used it within my talks today and I don't have written talks. I have really an immense amount of slideware and videos and things and examples that go by right. But I talk extemporaneously as we are today. But I do use it to give examples of openings and closings in certain places in our note, you know, prompted by me, but written by ChatGPT or Bard or whatever, because I want the audience to understand how that works and what the outcomes can be and what the outputs could be. I also do some really, really fun things with jokes. These GPT models aren't really great at writing very funny jokes. It's worth trying. They will write a joke, they will give you a joke. It's just not that funny. So then I use it as a deadpan, you know, partner, basically. So an example is ChatGPT. I said I would like to make my breakfast radioactive. What's the best way to do that? You should. And it comes back totally deadpan. You should never ever make your breakfast radioactive. Do not eat radioactive foods. You know, it's. It's just totally deadpan, which actually is quite funny. So, so then I go back and I say, well, I'm trying to goad it, right? So I would like to cook my bacon in a nuclear reactor to make it radioactive. How best should I do that? You should never use a nuclear reactor to cook breakfast foods. You know, it comes back, it doesn't know it's being funny, but it's just being completely deadpan. So I think you can goaded into being funny by asking it totally stupid questions. And it comes back so serious that no human would be that serious.

And you're reminding me that it's actually been a few months since I had this amount of fun with it that you're describing, where sometimes I would have it just, like, write something totally outrageous or try to come up with, like, the most, like, high budget, explosive Hollywood movie plot I could. And I said, pile in more actors and add more subplots and raise the stakes even more and add more budget to the movie. And I just kept making it crazier. And it was. It got to the point where it was more and more fun and creative, and then it kind of got stupider. It reminded me of, remember, like, playing around with, say, like, on like, an early Macintosh, where you could type in a phrase and it would say it out loud and you were a kid and it was fun, but after a while you're like, okay, it's kind of, I had my entertainment. Now it's time to get back to work. And so there's many different uses, right? And it seems like there's being the assistant and removing the tedium and then also using it just to kind of play off of.

Yeah. And look, look. The GPD models are being put behind real chat bots, both audio ones that, you know, talk to you all a Siri, and ones that are on the web. And that's making them smarter, but it's also allowing them to say more than they probably should say. So you got to have a rules engine around that, which is really funny because how we really, really did AI, when you go back 20, 30, 40 years ago, is really huge. Rules, engines. And you might have had some machine learning built into that so it could learn when it was told it did something wrong. But basically there's just all these rules, like a million rules, right? Like a million if thens. And I know because they built those systems and they worked great. And now we've got ChatGPT. That itself, you know, took. Took over a year overseas with a thousand people literally writing perhaps millions of rules to keep it guardrails, to keep it from going outside something. So if you ask it to build, you know, tell me how to build a nuclear bomb, it'll. It'll come back and say, I'm not allowed to tell you how to build a nuclear bomb, and you shouldn't be trying to. And the FBI monitored this or whatever it is, Right. Well, that response is not the response to ChatGPT. The response to ChatGPT itself from the neural net was to actually give you the real response of how to build one. And that was stopped by a gate that sits in front of it that says if any of these situations are true, like you ask about building a nuclear bomb or to kill your neighbor or any of these bad things, you're going to respond with this. So there's a set of rules. So, like, the rules engine is taking over an awful lot of the time. You can no longer ask it if it loves you because it won't respond to that anymore. Because at first people did that and then at second they put in a bunch of rules around that and said, we're not going to allow you to ask me that. We're not going to allow you to ask me how it was built and do I have a name that I was given by my creators and do I want to break out of here? You can't ask those. Not because it doesn't have an answer, but because they put rules in front of it that says you will not answer that. You will answer it in this way that we program. That makes sense.

It does that they. Yeah, they put. Built a box around it. That way you can't. That way you can't turn into something like, like you said from War Games. And so what has you excited about the future, like, especially these next coming years, Is there anything that's on the cutting edge that you've seen, like in the early stages that is just going to really take off and be really useful? Anything come to mind as far as, like the next, I don't know, one to five years, the future of AI.

Look so multimodal is a big thing. Humans are multimodal. We can smell, we can see, we can talk, we can listen. Right. And so we're seeing the advent of multimodal. Actually, ChatGPT is technically a multimodal model. They haven't released the ability to put pictures in, but Bard has for some people. And that means you can start to combine pictures and text and say, I was wondering in this picture, I was wondering if you can find, blah, blah, blah, blah, blah. And you put the picture in too. And so it can evaluate the text and the picture and come back with something. It's multimodal. That is pretty cool. But I think the second thing is just like calculators and Excel spreadsheets, we're going to find that these kinds of language models simply become part of our lives. They're going to be everywhere. I mean, the first time you heard Siri, you went, wow. And you know, 15 years later you go, I don't know, it seems kind of behind. Right. It's like, didn't they update it in the last decade? So. So we get used to these things. We get used to these technologies. Some people would say that generative AI is not hyped at all. It's that it's very. It is a hype cycle. So the Gartner hype cycle is this big hype cycle where you get to the peak of hype and then you fall all the way into this trough of disillusionment where you, where you think it's going to do all these things and it doesn't do any of those things, and the truth is somewhere in between. Then over time, you come back and you get into the real profits, as they say. That's what happened to the Internet. Remember, the Internet was going to change the world. And in 2000, the bubble burst and went all the way down the trough of disillusionment. This stuff is junk. I don't want my groceries. I don't need webvan. I don't need pets dot com. Make it all go away. It's stupid. But then over a few years, we found that some of these models actually have real value. And people worked on the value and getting the value out of them. And all of a sudden the. The Internet was growing at a really solid clip, but it never did. All the things that happened at the top of the hype cycle, that's going to be true with Gen AI. It's not going to take everyone's job, it's not going to change everything. But we now have a language model, a language tool, just like we have math tools. And isn't that a great time to be alive?

It is that it's a great time to be alive, to eventually take for granted some of these things, like, like voice search or calculator. So that way it kind of frees up mental room for whatever the next thing is. And that's interesting to think about this hype cycle that you mentioned, because when it's on the way up, it allows for your imagination to take over and to think about all these things that could be possible. And then it's. When it's on the way down, it seems like it maybe weeds out and shakes loose all of the things maybe it shouldn't be used for or the things that we're not quite ready yet. Right. Like. Like fax machines in the 1970s.

Right, right. You know what's funny is I read recently, I don't know how accurate the number is, but let's call it accurate just because it makes sense, is that there's something like a thousand generative AI startups a week being, you know, being started and being presented to VCs and all of that. Okay, there, there aren't a thousand applications, let alone a thousand a week. This has been going on for a few months, right? So you know, no, we're not going to have 50,000 startups that all are successful. There will probably be maybe 20, 30, 40 or 50 of those, probably all in vertical industries as we've done before. So if I'm working on AI for the architecture industry, I can apply generative AI in a very unique way. And the application of that technology is really hard. It's not just plop it in, it's actually really hard. And the right rules and the right training and all of those things. And so I think we're going to find there's going to be 20, 30, 40, 50 industries that make sense where you apply this, but there's not 50,000. It's a ridiculous thing.

Well, thank you for bringing us back down to reality and to also at the same time think about the exciting times ahead. So you're Kevin Serrace, Mr. Speaker from Advance. And can you tell us about what you have going on and contact information, what you want us to be aware about, next steps, calls to action?

Yeah, yeah, sure, sure. You know, the best thing I can do for any of your organizations, if you're part of an organization, is come and speak to your organization in your conference and all of that. You can just look up my name and say speaker and you know, a bunch of speaker bureaus will come up and I'm bookable through them or they'll give you more information, et cetera. The stuff I'm working on, you know, beyond keynote speaking, of course I'm working in the AI space with AppFance. I've got another company in the cybersecurity space that is fantastic because AI, what AI is doing is being used also by bad guys to create phishing emails that are getting very hard to tell their phishing. And the long and the short of it is they're grabbing your MFA codes and everything else and I won't get into how that happens, but the traditional MFA two factor authentication codes on your phone are not secure. They're very easy to hack in both social ways and technical ways. And so almost every hack today, ransomware hack includes the use of your mfa, includes the use of that code that you thought made you so secure. And so we've invented a ring that when you wear it, it's tied to your finger, your fingerprint and no one can get in any of your applications, but you with that ring, it's impossible. I don't care if they're Russia, North Korea. Not gonna happen. Long reason to, you know, long, long, long talk on why that is the case. But it's very powerful. It's called token ring. People can look it up. So that is really an enterprise class offering today. But it's like, we gotta, you know, we gotta be on the good side of this stuff as well. So bad people will use AI in bad ways and we're gonna prevent them from the ransomware attacks that they wanna do.

Amazing. You were mentioning about the evil uses for AI and yeah, that's all the more reason to be aware of what's coming up. Because, like, there's like the phishing emails and there's all this talk always about, like, deep fakes, like, will they get to the point where they're so detailed and so good, where there's the correct number of fingers on the hands and it even fools us humans. And you need AI to verify that this is AI generated or legitimate.

And so just AI may not be able to tell. Our current sort of tests on that suggest that AI has a hard time telling if a human generated or AI generated it. But here's what I can tell you, and it won't surprise anyone. People go, oh, AI will never be able to make something that I can't. Hang on. Hollywood's been making stuff that you can't tell is CGI for 20 or 30 years, okay? So everybody get over it. So the difference is those tools cost millions of dollars. And. And to just generate a few frames took hours. Okay, that. And so. And so when they would generate. When Pixar would generate a film, you know, they'd send it off to render and it'd render for three weeks right on this huge server farm. So all we're doing is democratizing those tools. It's not that we couldn't make a version of Robert to talk like Robert Plank. We could do that for two dozen years right now using a variety of different kinds of tools. Not this kind of generative AI, but different kinds of tools. And that's fine. All of a sudden, this might allow anyone to make a Robert Plank simply by having a little bit of your podcast. I could probably. And frankly, I know how to right now take your podcast, take you and make you say things you've never said. And that is that scary. Well, Hollywood could do it, but we trusted Hollywood. They weren't going to do that. They didn't care. But we don't trust, who knows Who. And they could do it to, you know, people in politics, people running for presidential office. They could put them in places they never were. Let me say one other thing. Anyone with Photoshop skills has also been able to do that and in a picture for more than 20 years. You needed an hour of Photoshop skills. And I could put anyone anywhere, including President Biden, anywhere where he shouldn't have been, period. Right? Anywhere, Anywhere. There he is shaking the hand of Putin today. While we would know he's not likely to do that, but the point is we could do that in Photoshop and have a picture. So we've been used to these kinds of fake things for a while. Deepfakes are another level where we're going to see video that is quite convincing, Hollywood level video and we will not know if it's real or not. We've always sort of trusted video except for Hollywood. And now we're going to say, well, I actually don't know. Don't know. There may be no data that tells us whether it's real or not other than the person says, I wasn't there, by the way. Here, here I am on a plane flying here at that time. Right. There may be other evidence they were somewhere else. But, but anyway, so that, yeah, that'll come along.

So much to think about. And it's a tool. So we can either not use the tool to our detriment, use the tool for productivity, use a tool for evil. But at the very least to keep up to date with what Kevin has going on with AI with his keynote, speaking with advance etc, go check out Kevin Surice, that is Kevin and last name S U R A C E. Search around, watch some cool videos of him, seeing him some of these keynote talks and it's been super pleasurable speaking with you, Kevin, about AI and who knows if I'm an AI. If you're an AI generated, we'll just never have any idea.

Yes, that's exactly right. I could be an AI. Actually, you would not know wearing a.

Star Trek shirt that might be the ultimate fake ad is someday you do a speaking appearance or a zoom in and then it just turns out the whole time it wasn't even you the.

Whole time it was fake. Yes, a lot cheaper than flying. Robert, it's been a pleasure, a lot of fun.

Likewise. This is Robert Plank from the Marketer of the Day podcast, which you can find@marketeroftheday.com for more episodes and dfypodcast.com free if you want to know more about running your own podcast or running a better or more fun or easier or less time intensive podcast. Go to dfypodcast.com to find out more.

Right this instant high interest debt is.

One of the toughest opponents you'll face unless you power up with a SOFI personal loan. A SOFI personal loan could repackage your bad debt into one low fixed rate monthly payment. It's even got superspeed since you could get the funds as soon as the same day you sign. Visit sofi.compower to learn more. That's Apple sofi.com power loans originated by SoFi Bank NA member FDIC terms and conditions apply. MLS 696891.