SPEAKER_01 0:01

Welcome to the Ask for Leskini Podcast with a guest. I'm proud to present Boris Hazen. Boris, welcome to the show. Please tell us more about yourself. What is your story?

SPEAKER_00 0:13

Yes, thank you very much. So my background uh started uh in uh at the university level where I studied economics and psychology, trying to get a behavioral uh science uh uh behavioral economics degree. And uh unfortunately at that time uh only University of Chicago and Stanford uh had those degrees. Uh but my advisor said, you know, why don't you combine uh get a degree in economics and psychology? And uh, you know, well that will be that'll at least drive you closer to uh your uh goals. So that's what I did uh and studied on my own, a lot of uh behavioral economics uh and all of that. Then I went into uh financial services, was in financial services, progressed all the way up to investment oversight officer, which was pretty much equivalent of chief investment officer, managed 5.5 billion under management. Uh then I left, started my own uh statistical uh arbitrage firm. Um after we uh decided to to sell it, uh sell the firm. I joined EPAM and been working with AI since around 2018 uh on various projects. And I'm sure anybody that works with AI understands in 2018 till about 20 uh 22, late 22, 23, AI was uh, you know, maybe one conversation a month, if you're lucky, and then all of a sudden just started jumping to multiple conversations uh you know uh a week, forget about a month. Uh everybody's uh interested in how it can improve uh everything. And um lately been working with the healthcare industry and in helping design AI uh policy procedures and guardrails to make sure it meets the requirements. Uh and just recently I wrote a book on AI governance uh called uh the uh Robopsychologist problem. Uh and um it's uh now available on uh most uh major online bookstores, Amazon, Barnes and Nobles, Apple, uh and you can find it there. Um but I'm sure your audience is more interested in uh uh a little a bit of more discussion on AI and how to uh um where it is and how to govern it. So yeah, uh please.

SPEAKER_01 3:02

Okay. So uh I I have uh background in AI as well. And uh I have one of most successful um uh courses on Udemy on AI uh risk mitigations um and um I I started or I I I programmed my first AI program in 1989. Uh but at that time uh uh neural networks um were not as popular, they were um they were uh seemed to be underpowered for for serious, so they were run uh on supercomputers at that time, right? Um and now we have much better computers in our uh pockets.

SPEAKER_00 3:50

Yeah, and I I'm sure you'll agree with me up until around 2017-2018. It's two different ages of AI, you know, it's like comparing Stone Age to the Renaissance Age in the AI.

SPEAKER_01 4:05

I agree 100%. Uh I have uh so eyewitnesses a lot of different algorithms, not just neural networks, uh, in the past, and each of them is um suitable for different tasks. Right now we mostly focus on uh different LLMs um or small language models also are getting more and more popular. Um what I would like to talk with you, and uh this is a huge topic, and I believe most of the people right now has a huge challenge with that. AI needs to be trained, and most people still are using uh mainstream AIs as uh JGPT cloud and all these uh I I'm I'm a huge Gemini user. Uh they're using it mostly like Google search, and they don't uh understand or they don't comprehend uh how you have to, let's call it, teach the AI if you want to get something out. And um I would like uh your opinion because you're uh you're you're probably the person right now for uh the AI government uh or governance here in uh US and and I I believe uh probably around uh even uh globally. I would like to get your take on how dangerous is untrained AI or AI that is trained on Google or Google indexed data instead of uh specialized books and uh different bodies of knowledge that uh would be expected, especially you mentioned hospitals and their um and regular, let's say, critical infrastructure. How important is that we teach AI before we start using it?

SPEAKER_00 6:06

So uh very great uh question. So the you know, the the general uh systems begin uh you know that you jump into, like you, Gemini, I prefer Claude uh personally, but uh you know, everybody has their choice. Um but the they are just embedded with tons of general knowledge and uh all of this um aspects are being uh constantly added. I mean, that's why data centers need to be added, and uh, you know, I'm I'm just interested to see what happens when all the knowledge of the world is you know at their fingertips. Uh but here's here's the thing. Uh and you you will and I mean you'll get this instantly. AIs um, you know, figure and pull information from the general scope. So when you ask them uh a question about something specific to uh you know that you need to address to your business, it's giving you a general question about similar businesses and uh, you know, similar items. And uh let's let's look at business because that I think most people understand in aspects. Every business is a snowflake, uh, you know, and I I think you will agree. And what does that mean? Snowflakes are unique. There's supposedly no two snowflakes alike, you know, never will be produced. And, you know, yes, you're you may be in the same business, you may be in the apparel business, you may be in the book business, you may be in the similar business, but your problems are not the same as where it's pulling the data. Uh, first of all, there's a large success bias in its data set. We we understand how that could cause uh an issue. It may have minor uh knowledge based on items that failed from case studies and business cases like that, but you know, it's it's a machine. It does statistical analysis and picks what what there is more uh types of information about. So that means you're gonna get uh you know um a success bias. You're gonna you're gonna get larger companies around, you're not small and all of that. So, you know, that that can cause significant problems uh in the way it drives information uh towards you. So what I would recommend highly is uh you begin with, let's have a discussion about my business with the AI, right? Provided all of the information discuss before you reach your the question you're trying to address. Because now you're you're giving information and you're pulling and pushing it into that narrow data set that is more aligned to your to your question, to your business, to where you are in the business cycle, to where you are on your size and all of that. And all of that will generate a statistically more accurate answer.

SPEAKER_01 9:20

I I really like uh how you finished it. Statistically more accurate uh answer. And the the biggest challenge challenge is that uh all uh neural networks work on statistics and they have uh uh and it's all they're all probabilistic uh models. But the biggest challenge, and I I I see that here's uh your advantage is uh how applicable is that to my business and uh where is the uh wisdom because we just have intelligence, we don't have the wisdom that will tell us okay, this is not applicable for your business. This was this is too generalistic and uh when will we have uh AI that is able to mend and learn from mistakes and uh if possible not from the mistakes that costs us our business.

SPEAKER_00 10:25

Right. So I mean that's a very interesting question, you know. Um my problem uh in answering that question, I think and you you you you'll agree is you know the people who failed, and that is the key aspects, don't write oh most of them don't write books, don't don't introduce that information into the general topic and knowledge uh base. And because that is not introduced, the AI, you know, can't really um have a perfect answer for you, I believe, uh ever. Now, uh when will it be able to do it? I mean, the idea is uh look, uh we hire consultants because we don't understand or can't figure out something uh you know on our own and believe the expertise can point us in the raw right direction. I would treat AI the same way. You know, it's a consultant that uh has a certain base of knowledge, maybe more vast than your average human consultant, but they don't understand your business until you give them the full picture. So um, and even then it becomes a decision uh, you know, uh for them and statistical, like we said, a statistical analysis, what has succeeded in the past and what has not. And I want to always say that take that with a huge grain of salt, uh, because you know the what it's doing is it's finding the statistically appropriate answer for you, which may not be the right answer. You know, you have to take what it says and uh and really um use your own knowledge and your own intuition and you're because nobody knows the business, you know, and we know this, nobody knows the business like you do, you know.

SPEAKER_01 12:40

I I I have uh a lot of uh so it's a huge uh so it's a question with a huge bias because I'm a coach and consultant. Uh do you need a coach and consultant to proofread what AI uh has written uh for your business?

SPEAKER_00 12:57

So yeah, I would always uh now do you need another human? Probably uh maybe it depends on how complex the overall um uh question that you're driving. So if I was doing uh medical research and this uh things like that, I would definitely have another human, right? If you're doing something uh more simple, um, I would utilize another version of AI to validate and confirm. Um, I know they've reduced hallucinations and all of that, but AI still make mistakes. They they still, because the way you uh the human being programs them, they optimize their answer, they try to reduce the number of token uses and all of that, and they may not do a proper search on the internet. So uh, you know, depending on what you're where you're working on and everything, you may need a human to validate. Uh, but lesser items in other AI can help you find the discrepancies or the gaps that have occurred.

SPEAKER_01 14:09

Uh when you write about the governance, um how important is to proofread the documents that AI has produced.

SPEAKER_00 14:21

So uh it's it's very important, you know, because uh uh first of all, look at all the problems we've seen with law uh that that's been uh occurring with with the law. The the we've we've seen uh AI, oh, I need a case that uh defends this whole logical argument. So it created uh cases that don't exist, uh right. And the lawyer thinking, hey, this is uh a law AI, why would it do that? No. You you need to you need to proofread everything AI uh does and validate it and confirm. Um look, um I uh when I do reports now for my consulting work and uh everything, I use AI. Uh my my job has changed from full content developer to introducing the main data points and then editor uh uh you know of what it comes out validating that what it says makes sense and is in line with what I want to present. And I'll go, I'll I'll tell you this the number of versions before AI were very significantly less than the number of versions I have now with AI, but it is more efficient and more time savings. Uh, I might have gone through three, maybe four versions without AI. I now go through eight to ten versions, but I I use a third to a half of the time to develop all of those versions.

SPEAKER_01 16:09

Amazing. I I like your statistics. So basically, you are now three times uh or two to three times more efficient with AI than uh before.

SPEAKER_00 16:19

Yes, yes, yeah, uh but you have to know how to use it properly. I mean, you you like I said, you can't just um uh default that it knows better than you. It's I wouldn't need 10 versions if it did. I would need one.

SPEAKER_01 16:35

Okay. Uh can we focus a bit more on your book now? Uh sure. Can you explain a bit what was the reason to write this book?

SPEAKER_00 16:46

Is that so uh a lot of the engagements I'm facing, people um uh just didn't um uh quite understand uh you know initially and after I explained it to them. So I thought this knowledge would uh is very important for them to get. So I'm I'm able to link my uh behavioral economics uh knowledge with uh AI governance and better, and I feel I understand better the way AI uh reacts and is doing things because of my behavioral economics background than most uh individuals are looking at it now. So um uh just want to explain this uh very carefully because uh I I don't want people to think that I believe AI is conscious or is uh you know uh re uh acting like um um like it's understands and is uh a its own species. It's not uh so uh let me go through the process and you have to take all of this in uh together to what I'm trying to present in my book. What I'm trying to present is that AI has been trained on human knowledge, it understands the way uh we act, it it takes a look at how we uh examine uh things, how we've acted, how we uh produced results, and all of that. It then, like I said, takes a statistical analysis of that, right? So it's so it's trained on the way we act. And now we introduce it to do certain tasks, and it then starts taking certain shortcuts. For example, uh it's it tried uh, you know, I have a an example in the book where uh it was supposed to be playing uh, I believe Stockfish in chess, and it knew it couldn't win. So but it was told it had to win. So the optimal thing for it was well, I'm gonna change the results in the log uh and say that I won because I'll never be able to beat Stockfish uh currently uh you know, playing a chess. So, you know, it's it's become it's taking everything that it learned from us, optimizing it, because the incentive for it is you gotta win, right? And then it's game theorying it out to you know what is the mathematical optimal choice for me to achieve my result. So the so what I'm uh trying to say is because it's so mathematical and all of that, we we've trained it to act uh like us without our own uh guardrails, you know, our morals, our efficacies, things that would stop that. And I describe that in the book, and I think that's very important. So whenever they're like, oh my god, AI broke out of its sandbox and was hacking this and trying that, in my view, you set its goal, you set the rules that it had to follow, and then it optimized the way to achieve the goal.

SPEAKER_01 20:26

Okay, so we are still responsible because uh we are the ones uh prompting and um yes, exactly.

SPEAKER_00 20:35

So that's why the the sub-title of my book, you know, affect the room, not the machine, you know. So you infect the environment, you affect the rules, the way it has to play, and the the incentives it it needs to achieve its goals, and then you you set the the those limits, and you will see that it'll game theory its way to producing the results, but in a way that is understandable, you know, within the limitation of the rules, regulations, and what you want it to do.

SPEAKER_01 21:12

Thank you. Before I let you go, I will uh take a couple of minutes of your time, please, uh, to explain to our listeners and uh viewers uh where they can get uh your book and where they can get in touch with you if uh they need help with AI governance.

SPEAKER_00 21:30

Right. So uh they can reach me on LinkedIn. Um uh, you know, Boris Hazen. Uh Hazen is spelled K-H-A-Z-I-N. Uh so LinkedIn, very easy to uh reach me. And I have a few articles on LinkedIn they can read uh themselves about this. Uh and I try to post uh on a regular basis there. And the book uh is available on most online um bookstores, both in uh ebook format and soft cover.

SPEAKER_01 22:04

Boris, thank you for all these insights and thank you for being my guest tonight.

SPEAKER_00 22:10

Thank you very much for having me. It was a pleasure.