The Real AI: Unmasking the Hype | Kevin Surace

24:20All of a sudden, I can code 10 times faster than I was ever able to code. Now, I've still got to debug it, and it's still got problems, and so it still needs me, but I'm the equivalent of many, many, many brainpower now instead of one brainpower. This is incredible, right? And everybody's gonna have more than one brainpower. Now, that has an interesting outcome. It says that IQ overall will get somewhat depreciated because everyone can have 3, 5, 10, 20, 100 brainpower. And EQ and collaboration will become more important. And because if you've got all this brainpower and you're going off in a direction but you didn't collaborate with anyone else, you could be down in the wrong direction, right?

25:02So collaboration between team members is going to be really, really appreciated. and IQ, anyone can use these tools, right? Once you learn to use the tools, you've got the IQ of the tool, which is really high. So I wonder, that's an interesting question. I wonder if per industry, it'll be possible to measure the difference between someone using AI as an assistant versus someone not using AI as an assistant. So here's an example, like let's say, I'm gonna take a random one like tennis. So let's say an AI assistant can watch you play tennis, see the arc of your serve, you know, when you swing and study all your health and your workout regime and your training and your opponent and your games.

25:45And it gives you suggestions on how to train, how to swing, what equipment to use, what to eat, how to play against an opponent and so on. I wonder how much better the AI helped person will be than the non-AI helped person. And if that could be done per industry, like writing, music, chess, business, and so on. Well, that's a great question. So I think it depends on what you've got the AI doing. So in this case, we're asking the AI to be my tennis trainer as opposed to a professional trainer, right? And then you get to who was the professional trainer and how much experience do they do and how good are they interacting with me and all of those other things.

26:18But augment it with a human trainer as well. Oh, it's got to be better. I mean, as long as it can be multimodal, that is, as long as I can give feedback and it says, oh, I saw your stroke and it's this. it's a little bit like, can I evaluate ECGs better with AI than I can move cardiologists? Well, of course I can. That was solved five years ago. We know we can because a cardiologist will miss things that are ever so nuanced that AI will see. And this is true with x-ray readings, right? I mean, the AI will see things that the human eye just won't pick up. Yeah. In fact, they passed a law then that a human has to be the one to tell you the news.

26:56Yes. Human has to be one to tell you the news, even though you're using these systems that know better. So then the human evaluates that human doctor says, okay, it's your radiologist says, well, okay, well, I'll agree with this or I disagree with it or whatever. The AI is right. Now, the problem with AI is a bunch of hidden layers, right? So when you've got these neural nets, you've got a bunch of hidden layers and people get angry. Well, that's not explainable. Here's my explanation of why it's not explainable. Like when I look at you, I can say, oh, I know you curly hair glasses, probably has the headphones on, has a microphone.

27:26I get it, right? Like, I know that, right? I know how to recognize you. But AI will recognize you in a thousand different ways. And there isn't an English word for each of those methods, all those hidden layers. They don't mean to hide it. It's just there's no way to explain what that was. I saw a little thing here, a little thing. There's no explanation. So that's sort of the interesting thing, is that you want AI to tell you exactly what it found. And it says, I can't, but after looking at a million x-rays, I know this person has cancer. I have no way to tell you why I know that, or even maybe where I see it.

28:02It's just, this is a pattern I've recognized before. And the AI is unlikely to be wrong in those cases, if it was well-trained. Right. Now let me play the devil's advocate a little bit, because I've spoken to a lot of people and a lot of smart people, and I've almost been disturbed by how negative they are. And just as an example, I was in New York City the other day and I had dinner with a bunch of hardcore scientists. Like these were professors, well-known physicists, well-known chemists, biologists, a Pulitzer Prize winning science writer and other science writers. And they were universally pessimistic about almost every technology, but particularly AI.

28:40Like, oh, you know, AI is going to basically turn the guns on the humans and decide to take over the world and all this other stuff. AI is going to make viruses that kill everybody. Like, okay, I get it. Some people tend to be more negative than positive. But what are the negative aspects of it? Like, okay, let's take another one. I had a conversation with Vernon Reed, who was the head of the band Living Color. Had the song Cult of Personality in the 80s, one of my favorite songs. And he was like, what are you going to do when an AI version of your father's voice calls you and says, you know, son, I'm stuck in the middle of nowhere.

29:18I need you to wire this money right away. You know, please help me. Like, is there going to be more, what kind of problems are going to result? Sure, sure. Great, great, great question. A great conversation piece. So first of all, everything AI is doing today in terms of deep fakes from voice to visual has been available in Hollywood for 25 years, right? This is not, like, we know that Hollywood has been able to de-age people or put people in movies that have been dead and all kinds of things, and they're incredibly convincing. and put people into, you know, change their face and do all that, right?

29:49In fact, create people out of nothing, as Pixar has been doing for 25 years. And this is a really fascinating thing. All this technology is doing is democratizing access to those tools. So those tools, you know, used to render like one frame a minute and take six months to render, right? They were hugely, you know, painful to do, but the models for them were understood. and they weren't the same as AI models are today. They were rules-based. But nevertheless, Hollywood's had these tools. But we knew that Hollywood wouldn't use them to call you and try to get your money, right? We knew that. So we trusted that.

30:26So none of this is new. Deepfakes are not new. We could make Obama say anything we wanted to 20 years ago. However, it required Hollywood tools and millions of dollars. So nobody did it. Now, we democratized access even to this. And by democratizing access, it means we can use it for free or a dollar or two. And so can bad guys, right? So, um, but this has been true with every tool in the world that has come forth forever, right? Bad guys use them. Bad guys use everything. There's bad guys who use Excel. I mean, it just is what it is, right? They use banks, they use the banking system. So here's another set of things.

31:03And, and, and it is true that already people are getting calls from their mom or their grandmother, and it sounds exactly like them and said, please wire the money and they think it's them, etc., etc. So we're going to have to be more diligent on what we believe because you will no longer be able to believe that watching that president in that park exchanging money with that person actually happen. You have no way, in fact, you may have no way to tell whether it actually happened or not. And for all of history, we've always believed it happened unless Hollywood told us otherwise because they've told us otherwise for decades.

31:38Now, that's not new, right? But now, this isn't Hollywood. It's just Joe Schmo down the street, some kid who generates it, right? So we're just going to be more thoughtful about what we believe and what we don't believe. And you will not be able to believe every video that you see. Let me say one other thing. People have been Photoshopping, you know, things together, at least in still frame, I don't know, for three decades, right? So fakes, deep fakes on Photoshop have been around, but you needed to be a good Photoshop jockey to do it well enough to hide the fact that you put the face on someone else's body and maybe you hid that with a necklace and all these tricks, right?

32:14So that's been done in Photoshop for 30 years, but you needed to know Photoshop, which wasn't that except for$100 you could learn Photoshop. Now people can do it by just typing text. Put this hat on this body and there it is. and go, okay, well, now I put them in a place they weren't, change the background, put them in front of an oil plant, put them with money in their hand. I can do that with anything, right? That's sort of democratized access to it. So I think this is the real problem. Like this is the real negative thing. If you want to find something negative about AI is that we're going to be dealt with a lot more, let's call them, you know, scam.

32:50Yeah, fake news, phishing attacks, scams. And it's going to be harder and harder to verify whether something's true. So verification is the key. If someone claiming to be your mother calls you and needs money desperately, you just got to verify somehow that this is a true thing. Yeah, call your mother back, your actual mother, on her actual phone. This is not hard. It takes an extra minute. Right. And that's fine. You know, that is not a bad thing. The other thing is throughout history, when technology comes forward, new technology, whatever it is, you notice, and I love scientists. We need scientists.

33:27You know, it's usually scientists. Oh, it's the end of humanity. Oh, it's going to be all over. Okay, look, this is a language model that is doing for text and language what the calculator and Excel did for math. That's all. And we solved it in math decades ago, and now we're doing it for language, which is a much bigger problem. Do not hook the language model to the nuclear arsenal. Okay, this is a bad idea. Right. Right. And by the way, we've had the ability, because I was once offered a job to work on this at MIT Lincoln Labs, we've had the ability to recognize using AI which objects in space were junk and which objects in space were nuclear missiles heading towards us.

34:09We've had that tech for 30 years. So never then did we hook up the AI to make automatic decisions. Hey, in fact, we signed a treaty with Russia back in the 80s to not use any form of AI to automatically launch weapons, to give it to people who would then make the decision. There's actually a treaty against this, against using AI and hooking it up directly, because we knew in the 80s, even with the AI we had then, which was very simple, that this was a bad thing, right? Right. So it's not like suddenly we're going to change our minds and say, you know what? I want to kick back and have donuts. I'm going to let AI make these decisions for me.

34:46It's not like magically we're going to give AI the power to do something they could have done 30 years ago. Right. That's right. No, nobody's going to hook it to the nuclear arsenal. Nobody's going to hook it to crazy things. Everybody relax. It is not sentient. It is not dangerous. It's just a large language model, just like we have large math models. And by the way, a large language model is actually a large math model. It's only math. I mean, in the end, all we're doing is we're signing probabilities to neurons, right, to neural nets to say, well, the probability to put this word after that one based on what they, you know, asked me is here.

35:21It's easy to, well, use, you know. And that's all it's doing is forming sentences. Everybody relax, right? These scientists, God bless, are wrong. Now you got a lon out there going, oh, but AI is going to kill everyone. Then he starts his own company. So I think that's a little disingenuous. I think it was all positioning. No, I agree. Whenever one of these guys says something, even like Warren Buffett for his entire career, if he says, you know, there might be a chance terrorists could attack the Super Bowl with nuclear bombs, what does he do next? He's an insurance company. He sells insurance to every Super Bowl.

35:56Yes. Yes, he does. So you always have to look at what the agenda of these guys are. And, you know, Godspeed of the most innovative people around in some cases. But, you know, Elon Musk certainly created a lot of innovation. But why do you think that Google engineer thought the AI he was working on was sentient? Presumably he's tech savvy. He knows everything any other tech person knows. Why would he think that it was suddenly sentient? I could only guess, and someone's going to hang me for it, I can only guess that it was kind of his subconsciously his time to get in the limelight. And today you're not in the limelight.

36:30If you're going, listen, it's just a freaking math model. Everyone calm down. It's, oh, it's going to kill us. Put you on every podcast. Right. So I'm here clearly stating, because I understand the underlying math. And so do you. It's freaking math. Everyone calm down. Don't hook it to the nuclear arsenal. We're fine. but if you're him and you're one of lots and lots of important people at deep mind and google ai and everything else you know you're not getting any press by saying this is really cool in fact you're not getting anything so it's better for you to leave and say i have interacted with this thing and i think it's actually sentient okay well now you get all the press in the world and now you're really really important right but it's just bs it's just math if you understand the math it can't to be sentient.

37:15It does not know what it's saying. It learned from trillions of phrases, and it put them back together for you in some random way, but, you know, thoughtful way, from novels and from truth and from fiction and from everything else, with a huge rules engine around it, again, that took a year and a half to put together. Everyone calm down. Right.

37:49It's sort of like a simple way to think about it is the Google search box. If you type in the letters S-A-N and you live in San Francisco, Google is smart enough to know that Francisco is probably the next word you were going to type. It's just using the same statistics. But like you said, it went from words to phrases. You've got a much more sophisticated and larger model. Now, what about cybersecurity? Obviously, AI is now going to be used by the bad guys. to figure out where there might be holes in operating systems and so on that they can exploit. So here's the interesting thing. Since the advent of ransomware, everything in cybersecurity has changed.

38:26So it used to be bad guys got into your network and messed it up because they hated you or because they could. Then ransomware came around and it was like, I can charge$100 ,000 to not reveal their data on the web or to give them back their data or give them access to their data, whatever. Okay, then we started encrypting our data. And that meant that people getting into our network, maybe they could do some network damage, but they couldn't do anything with the data because all the databases have become encrypted. Except our applications, all of them know how to decrypt the data. So now you want access to the applications because then that's how you get the data.

39:03So then they go, well, I'd have to come in the front door. Well, it turns out that's freaking easy because passwords and IDs are dumb. People use their kids' names and 1, 2, 3, 4 and use the word password as the most common password, right? So what happened is there's an entire network now of identity access brokers. And what they do is if you're a ransomware person, you go to an identity access broker and say, I want access to someone at Bank of America. I'm making up some teller, someone who has access to accounts or whatever you want to do, right? That comes in through the front door. And they go and get you that access, including MFA.

39:44So all traditional MFA has been hacked. What's MFA? Oh, multi-factor authentication. So the codes that come on your phone, ridiculously easy to hack. They're easy. And I can show you the social hacks right now. I won't tell everyone how we do it technically, but technically it's harder, but actually pretty easy, like in minutes, because I have your cell phone. I can certainly get all your codes this afternoon. That wouldn't be hard. Okay. But the easiest way to get MFA is actually simply a social hack. and social engineering is very easy. So let's say you're at a company, but you're at a big company, not a small company.

40:20Okay. In the big company, you're one of 50 ,000 employees. IT calls you, says, hi, this is IT. This is Jim from IT or Bill from IT. James, we think we've got to break into your account. What I'm going to do is I'm going to initiate a login and password that's going to send you a code. I don't have access to that code. So I need you to share your code once. By the way, just to validate it, your employee number is this, your social security number is this. You go, oh, okay. You're on the phone with them, right? You get your code. You say it's 1234, whatever it is, right? They then put that in. They now have access to your account.

40:57They say, you know what? We're in. We're in together. Looks like everything's fine. Thank you very much. Goodbye. You hang up. They now have access to that data. They're sucking out data. Takes them 20 minutes. They've got everything they want. You handed them the multi-factor authentication from your phone or from RSA card or whatever it was, and now they're going to hold the company ransom. I know in one case,$22 million recently. Oh, my God. Did that company pay? Oh, yeah. Oh, yes. What was the data, roughly? You don't have to say exactly. It had to do with running machines. And what they did is they essentially went in and changed all the data.

41:35And they didn't have a backup for several weeks, and they wouldn't be able to operate the machines. it basically shut them down. And so it was cheaper to pay the 22 million and bring the place back up. I have example after example of, by the way, everybody pays and almost no one knows. They don't publicize it, right? They try to just bury it. This is true with universities. It's true with hospitals. We've, you know, every hospital pays. They cannot let that patient information out. You know, that's it. So then you say, I got to close the front door. So one of my companies that you're not an investor in, but it's a great company.

42:07It's called Token Ring. Token ring is literally, I don't have one on, but it looks like this. It's a ring that you put on your finger. And when you put it on, it's tied to your fingerprint. So you've registered your applications with this ring. Okay. That ring now only works on your finger. No one else's finger. They could steal it. You can lose it. It doesn't matter. Right. And your bio information is stored inside the ring in a secure element. So if you try to get in there, it destroys the information. You know, in the movies, they're going to just cut off your finger and use it. Well, yes, in the movies.

42:36but it turns out we're not trying to solve for people cutting off your arm. That's a different problem, right? That's a different level of crime. As you know, most of this ransomware happens from people far, far away in Russia and China and in North Korea. So this is fascinating because now your access is tied to that ring and it goes behind your application server to something called a FIDO2 server to validate that you are you. 128-bit encryption happens in one second, so we can't decode that today in that speed. And it's different for every application, that validation. So it makes your access 100 % foolproof.

43:14And no one can call you and say, give me your number. You go, I don't have a number. What are you talking about? I have a ring. There's no information on the ring. There's no serial number. I can't give you anything. And they hang up, right? They go, well, you're not helpful. I can't do anything, right? So the solution is to close the front door. the solution is with, I think, wearable biometrics. And anyway, so that's one of the other companies I have because I think cybersecurity has to be solved. And in this case, AI can't beat it either. There's nothing AI can do. Because they can phish you all you want, but they can't get your ring.

43:48It doesn't matter. They can't get the ring and get access. Phishing's all about, hey, this is easy. I'm going to steal your password and your login ID. But if they can't steal your password, if they can steal even your password and login ID, but they don't have the ring, they still can't get it. Will quantum decryption eventually be there for decrypting, or are we decades away from that? So as you know, quantum computers are coming along, and we're looking at things that are getting into the hundreds and then thousands of qubits now. So this is really fascinating. However, we're talking about two or three quantum computers in the entire world that live at Google and IBM, right?

44:23So no one has access to these things right now. So I think we're at least a decade away from bad guys having maybe two decades away from bad guys having access to a billion-dollar quantum computer, right? This access is highly limited, and it's for researchers right now. And these things are only doing sort of one thing. A quantum computer could certainly decrypt 128-bit encryption if you could do a man-in-the-middle attack. It could decrypt it in a second or less. However, you've already completed your transaction, and it's gone. And the next time that, you know, those encryption keys will be different.

44:59So it still may not solve it because that rotates around. It's not as simple as I've decrypted it and now I've got it, right? By the time you've decrypted it, they're already logged on and they're doing their thing and there's nothing you can do, right? It's a little bit late. So I think it'll still be hard. I think it'll be two decades before bad people have quantum computers. And by then we may do 512-bit, 1024-bit encryption. Yeah, there are algorithms also for when there is quantum encryption or decryption, but I just don't know how far advanced those algorithms are, and I don't know how fast we can employ them.

45:36problem is the speed of the processors, right? So think about the processor I have inside a ring that I have to do decryption on. So I can do 128-bit encrypt-decrypt, but if I had to do some quantum encrypt-decrypt, that's just a massive amount of processing power that I might not have in a bio ring, right, which has a tiny little processor in it. So that's the problem, but that's a problem 20 years from now. Again, no one has access to these things but researchers at IBM and Google. So you can't log on and say, I'm using a quantum computer as a bad guy in Russia. You don't have one. And Russia's not going to develop one.

46:12They lost whatever brainpower they had. And I guess the other thing that's very interesting in the AI space is how many AI companies have started overnight. There's like hundreds of thousands of AI companies have started in the past few months. And my own gut feeling is when I look at these companies is that there's nothing special. Like it's so easy now to just hook up to the API of ChatGPT and make a pseudo company that anyone else could also create overnight. So on the one hand, the price of making an AI company has gone to zero. But on the other hand, none of these things are worth anything because they're not really doing anything of value.

46:52That's right. Where's the moat, right? Do you remember when mobile, once the iPhone came out and then there had to be a hundred thousand mobile first companies. I'm doing mobile. I'm doing, and then we're, I don't know, easily a million app companies that were writing apps for iPhone, right? Are there like 20 that ever made money? Some number like that. It's amazing, right? It's a very, very small number of apps that ever made any money. Yet all of these companies got VC funding. Tens of thousands of companies got VC funding to do some app for the iPhone of which nobody made any money. In fact, all the downloads are Facebook, Twitter, you know, you can just go down, you know what they are, There's 10 downloads that everybody has, and then usage on their phone goes pretty low.

47:31So most of those app companies didn't get anywhere. And that's going to be true with AI. Everybody's going to do AI, but if you can write an AI interface in a weekend, your moat is too small. Now, there's other companies. AppVance is an example. It's taken years of AI effort to get to where we are. Yes, we're also leveraging large language models in unique ways, But even that is another year of effort to make that work across the plethora of mobile and web applications. It's not something you do in a weekend and hook up in a weekend and go, I'm in the test business. Now, there's a few startups that have said, oh, that's what we're doing.

48:09Good luck. When do you have to test every application with every library, with every different set of code that someone has created? It turns out it's very hard. It's really, really hard. And you've got to use visual information. It's got to be multimodal. And then people give up. And I've seen this happen. People start and then a month later, they're gone. And two months later, and they go, this is really freaking hard. I thought it'd be easy. So I think the easy ones, you're right. And they probably shouldn't get funded. And there's a rush to fund them. And the hard things will be hard. And some of those are going to succeed.

48:40Yeah. And again, I think just to wrap this up, just as a mindset, for every possible problem where you get scared, oh, AI is going to replace me. This is going to happen. This bad thing. Think of AI as an assistant instead. Like AI will be a very good assistant and will make your life better. And it's not that hard to figure out and start using. You could do it today. You could just go to, you know, whatever, let's just say you go to openai.com, chat.openai.com, sign up for free and start using it. And you'll see, you can use it as an assistant for most of your jobs. Copilot, that's now been out two years from Microsoft, by the way, it's in GitHub, right?

49:18Two years we've had Copilot that helps you code and it's getting better and better. It's based on GPT-4 now. That is making coders' latest stats 55 % more productive. 55 % more productive. You can't use it without a coder. Like, it would be ridiculous because it's just giving you snippets of code and hints of code and it doesn't always work and then you have to modify it to make it work for exactly what you're doing, etc. But that means I don't have to write every line. It'll spit out 20 or 30 or 40 lines of code just from my typing. just like Word does. You know, you start to type and it starts to finish the sentence.

49:53That is exactly what Copilot for coding is doing. It literally finishes your code. It's not exactly right all the time. Sometimes it doesn't compile, but it's better than what I used to do, which is go to GitHub or go to a variety of open source places and start to look for open source examples because you'd want to start with an open source example often today for this function. Now it's just giving you that example. This is, it's a 55%. That means, in theory, we might be able to double the productivity of coders using this. And so it should be a given in every coding shop that, of course, you're using Copilot.

50:27Do you still code? Not as well as I, you know, here's the problem, right? As you move up in management and you move into architecture and things like that, you lose the ability to do it syntactically correctly, right? So I could still sit there and write pseudocode and people go, okay, I know what you mean. Let me go fix it. but to sit there and do it, you have to know the syntax and you have to remember the syntax exactly. Exactly where does that bracket go and exactly where the comma goes. And so if you don't do that all the time, you lose that. Do you think age has something to do with it? Like the last time I tried to code was about two years ago and I just was really unbelievably strikingly bad at it.

51:08And I used to, you know, I've put in my 10 ,000 hours on coding and I really just can't do it anymore. I don't know if it's a great question whether age has to do it or we simply just graduate to higher level thinking. Right. And you get to, again, architecture and systems level and thinking about where the world, you know, and then you just it's just not a skill. OK, could you do long division? You might not be really good at it, but if you've struggled with it for a while, you could do long division. But you haven't done it since, I don't know, fifth grade, right? Nor have I. But we were taught it, but we've kind of forgotten how to do some of that because the calculator came about.

51:44so I don't know I think if you said I am going to learn Python and I'm going to take a month to learn it or two months at the end of two months you'd be every bit as good as you ever were and you might be more thoughtful about it you'd probably be really fast but you might not enjoy it you might say this is this is dumb because I could be at a whole other level so I look I think Copilot is a miracle for coders it's fascinating but no I wish I I wish I coded as fast as I did when I was at IBM and I could just rip code out thousands of lines, you know? Yeah. But those days are behind me. And, you know, you got to give up something, right?

52:18You just can't do everything all the time. I guess so, yeah. Sometimes I try to do everything all the time, but it doesn't work. It doesn't work. Kevin, Sarais, you know, thank you so much. There's been so many questions and angst and anxiety and curiosities about AI just from everywhere. And I've had my own frustrations just seeing how people have been abusing the words AI and what they do. But look, they're trying to make a living too, so I don't blame them. But thanks so much for helping clearing the air on all this stuff. And good luck with all your entrepreneurial efforts, particularly the ones I'm invested in.

52:50So yes, exactly. Exactly. Well, we'll keep plowing along and we'll get back together maybe in six months and see where AI is then. But I think all this hype is going to settle out and it's going to settle into, like I said, what Excel is today. It's a great assistant. It's a great tool for us. It doesn't run itself, nor will AI. You have to prompt it. You have to get stuff back. You have to edit it. You have to be thoughtful about it. But what a powerful tool and what a great time to live. Like, I can't imagine living in another time. I'm sure our parents thought, oh, there's color TV. Look at that.

53:23Okay, this is way better than color TV. Right. No, and, you know, because of the increase in the speed of computers, it's created exponentially growing industries in many other industries that required the computational power. So initially, computer technology was the only industry exponentially growing. But now that it's exponentially grown, it's unlocked the power for other industries to exponentially grow. So that's what makes this super exciting time now. So I agree. It's amazing. It's the best time to live. Yeah. So thank you so much, Kevin. And I'll talk to you soon. Yeah, a pleasure.

54:05Thank you.