EP 007 - Securing You Company's Future: AI Compliance with Michael Silva

00:00:02: Hello and welcome to the episode number seven.

00:00:08: Lucky number seven.

00:00:16: which is in, Michael.

00:00:20: Today it's artificial intelligence.

00:00:22: There's other stuff I've moved on past.

00:00:23: IT I still am involved in but AI is current focus.

00:00:29: Michael, introduce yourself.

00:00:33: So my name is Michael Silva.

00:00:34: I am founder and CEO of Emily AI.

00:00:37: We have an AI system that sits in Microsoft Azure to deliver secure

00:00:41: information real time in a very highly regulated way.

00:00:45: I also run an IT organization.

00:00:47: I've had that for 14 years.

00:00:48: That's Archon One.

00:00:49: We focus on cybersecurity compliance for law firms in New York City.

00:00:53: And I've got four boys under eight years old, live out in New York.

00:00:57: and got other interests that don't have anything to do with AI, so we can save

00:01:00: those for the end of the podcast.

00:01:02: But we're here for AI, so we'll stick to that for now.

00:01:05: awesome.

00:01:06: Busy, busy home.

00:01:12: Hello.

00:01:13: Yeah, we were just we actually had a full start.

00:01:16: Oh, I want to say two weeks ago.

00:01:18: So we're going to met each other and we've done these intros.

00:01:20: But fun fact, Mike has three, four kids.

00:01:23: I have three and then Edgar has two.

00:01:26: So we have a good diversity here of of kid representations and busyness.

00:01:35: One say diversity, one say laziness on my side, right?

00:01:41: But yeah.

00:01:43: So, Mike, can you tell us a little bit more?

00:01:47: So one of the things that I think as it relates to AI, just a very highly talked

00:01:54: about topic is building trust with AI.

00:01:58: And I think AI compliance, you know, building trust and they kind of go hand in

00:02:03: hand.

00:02:04: One one of the things that I want you to kind of ask, since you work with a lot of

00:02:08: clients in the highly regulated industry of

00:02:11: kind of, you know, legal, not even regulated, but I would say, where trust is

00:02:17: key in delivering or kind of outputs behind these LLMs.

00:02:23: So how do you go about kind of addressing and helping roll out or helping implement

00:02:30: some of these LLM systems while building trust with your legal customers?

00:02:36: And I'd love for you to dive into any use cases you could speak about.

00:02:42: sure.

00:02:43: So trust is earned in general, right?

00:02:45: That's between people and machines.

00:02:48: And I think that one of the easiest ways to ensure that you're getting a trusted

00:02:52: answer is to start with accurate data, right?

00:02:54: This is commonly known.

00:02:56: industry as rag, but that's a technical term.

00:02:59: If you're not in the industry, you wouldn't know what that is and that's

00:03:01: fine.

00:03:01: Basically what it means is that you're asking AI to give you an answer from data

00:03:05: that you know is true, right?

00:03:07: Also known as garbage in, garbage out.

00:03:09: Okay, so there's a couple of ways you can go about that, but at the end of the day,

00:03:13: it really depends on what you're storing and also how you're asking the question.

00:03:17: So the nature of AI systems today, and I'm not talking about like the 1980s, 1990s,

00:03:23: because yes, AI did exist back then.

00:03:25: I can give you guys a couple of you.

00:03:26: if you like.

00:03:27: The way that we know AI today is generative AI.

00:03:30: So a lot of people are focused on GPT 3 .5 and GPT 4 and nobody tends to think about

00:03:36: GPT 2 or GPT 1 or GPT beta or anything else.

00:03:39: It's just the starting point is 3 .5, right?

00:03:41: That's how everybody mentally thinks about it.

00:03:43: And the model that everyone was exposed to in the last year and a half is a

00:03:47: generative model, which means that it's taking a statistically probabilistic

00:03:51: outcome from a data set and repeating back a version of that to you based on

00:03:57: So where this can go right is if you have data that you know you trust, you can get

00:04:02: back a relatively accurate answer.

00:04:04: When I say relatively, because it's generating it, right?

00:04:07: Where it can go wrong is if you are asking it to answer a question and you don't

00:04:11: understand where the data is.

00:04:15: a system that does not have the capability to say no if it doesn't understand the

00:04:19: answer.

00:04:19: So an acceptable answer from a human in building trust is, I don't know, but I'll

00:04:25: find out.

00:04:26: That's something that you wanna hear, particularly if you have a job interview

00:04:30: with somebody.

00:04:31: What you don't wanna hear is the wrong answer, or a fake answer, or a BS.

00:04:35: That's the biggest, we're not gonna continue this interview type answer.

00:04:40: So unfortunately with machines, there's the double standard of we're very

00:04:43: forgiving of human beings when they make mistakes.

00:04:45: We are extremely unforgiving of machines.

00:04:48: So the odds are already stacked against the machine to begin with, right?

00:04:52: So the question of building trust is kind of a two -part thing, right?

00:04:54: Just to close up on the thought, I mean, the first part is data, of course, right?

00:04:58: That's a very important part of the strategy.

00:05:00: But the second part of it is the human interaction with it and the expectation

00:05:03: setting and really how you're using the tool in as much as the tool and the data.

00:05:09: think expect to...

00:05:12: top of it?

00:05:12: Because I know you talk about data and the importance of having a data strategy a lot

00:05:17: on your LinkedIn posts.

00:05:19: And as you mentioned, garbage in, garbage out.

00:05:21: So how do you ensure you're putting in the right data into the system?

00:05:26: Because if you're putting in the wrong documents and you're prompting the system,

00:05:30: so does AI know any difference?

00:05:33: So if you've loaded that data, does it just query everything that's possible?

00:05:37: Is it?

00:05:38: on the humans to kind of clean up the data and to embed trustworthy data into the AI

00:05:46: system?

00:05:46: Or can you just throw all of the data that you have basically and then let AI figure

00:05:52: it out?

00:05:53: What are kind of some of the best practices?

00:05:56: Okay, so this is a really loaded question.

00:05:59: So AI is not going to know the correct answer, right?

00:06:02: It can't.

00:06:03: What AI can do at best is give you an accurate estimation of a correct answer

00:06:08: based on data that you have.

00:06:10: Now, if you're looking for a degree of confidence, again, to go into the

00:06:14: technical end of things, if you're building it, there's a system called

00:06:17: RAGAS, R -A -G -A -S, that you can look up that will give you a statistical

00:06:20: probability for how close the answer should be to being accurate.

00:06:24: However,

00:06:25: like talking to a person, unless you just trust that person to know what they're

00:06:29: talking about, you cannot assume something is factually true.

00:06:33: And assuming what something is factually true is being actually true is like a

00:06:37: philosophical question, right?

00:06:39: Because it really is by consensus.

00:06:41: That's what we consider to be true, right?

00:06:43: If you look at like the philosophical definition of truth, it's consensus.

00:06:47: The way we use words, which is how LLMs work, to describe a concept.

00:06:51: is the same way that we use words to describe things that we see in real life

00:06:55: that we consider to be true.

00:06:56: Right?

00:06:56: So my shirt's white today, right?

00:06:58: Yesterday it was salmon, which is one of my other favorite colors, right?

00:07:01: But today it's white.

00:07:03: If everybody started calling this a different color with a different set of

00:07:06: words, we would say it's a different color, right?

00:07:09: So the question is really what is true and it's what's commonly accepted as true.

00:07:13: So LLMs aren't any different in that regard, right?

00:07:16: What they're doing is they're taking data that it sees as being true.

00:07:20: and it's returning back an answer that it thinks is true because of what it's

00:07:23: seeing.

00:07:24: So the question is how do you define truth?

00:07:26: That's really the question that you're asking.

00:07:28: And the answer to this question is how are you managing the data behind the scenes?

00:07:32: There isn't an easy way today for a machine to understand what is true and

00:07:38: what is not simply by looking at the data that's behind it.

00:07:41: It just doesn't exist, right?

00:07:43: Because of the reasons I mentioned before.

00:07:45: it's also if you have like this concept of level one and system one and system two

00:07:50: thinking and someone means basically intuition and intuitive thinking and

00:07:55: system two is basically thinking about stuff a bit longer and taking your time

00:07:58: and retrieve the more just use more energy to get the answer basically and like

00:08:08: modern systems they aren't technically capable of system two.

00:08:13: still out there.

00:08:17: That's why your intuition is not always right.

00:08:29: Mm -hmm.

00:08:31: They have a similar problem that they only can rely on intuition if you don't give

00:08:37: them any knowledge.

00:08:40: Like you said, it's a lot about statistics.

00:08:44: From what I learned, it's also you don't have to have 100%.

00:08:50: We already had a complete episode about this.

00:08:54: for it to be useful, but of course you have fields like you're somewhere in the

00:09:00: legal space where it's more than crucial to be precise and to not...

00:09:08: I think there were even lawyers using JGPT for their defense and quoting some random

00:09:19: case that never went by.

00:09:22: So that...

00:09:23: That kind of, that's kind of serious.

00:09:27: Yeah, well, the funny thing that you mentioned there is that specifically the

00:09:30: law and that use case, the attorneys have a very clear directive.

00:09:36: They're responsible for it, right?

00:09:38: In the fine print, and actually the New Jersey Supreme Court Commission on

00:09:43: Artificial Intelligence covers this quite nicely in one of the publications they

00:09:46: produced.

00:09:46: On the first page, it basically says two things.

00:09:49: The first thing it says is, if you're not using AI, you're going to get out competed

00:09:53: by another law firm that is using AI.

00:09:54: That actually says it.

00:09:55: I have it highlighted on LinkedIn post.

00:09:57: The second thing it says is just like a junior paralegal that doesn't know what

00:10:01: they're doing, you're responsible for it.

00:10:04: Right?

00:10:04: So like in the eyes of the law, they see very little difference, right?

00:10:08: It doesn't matter if it's poor note taking, if it's an intern, if it's chat

00:10:13: GPT making the mistake, it doesn't matter, right?

00:10:15: You're still responsible for it.

00:10:16: So this is where I think there's a little, you know, overly embellished concern

00:10:22: around AI taking people's jobs.

00:10:25: because AI, it's not going to.

00:10:28: I mean, some things it might come close to, right?

00:10:31: But these are like really basic routine tasks that are just forcing evolution,

00:10:35: right?

00:10:35: It's not like some threat that's gonna come in and destroy humanity, you know,

00:10:40: unless it's being misused.

00:10:43: Yeah, of course.

00:10:44: Have you heard about the case?

00:10:46: So there is a a new story that came out June of last year where a lawyer was

00:10:54: basically representing a man for an airline, you know, in some some personal

00:10:59: injury suit.

00:11:00: And he used Chad GPT output.

00:11:03: I think verbatim, probably from from its output.

00:11:09: And it cited fake.

00:11:12: use cases as defense too and things like that.

00:11:15: So how do you, so do you feel that, is this kind of what you were talking about?

00:11:21: It's still the problem that the lawyer is really behind whatever the output that

00:11:27: they're putting in front of the judges.

00:11:30: So they're responsible, not chat GPT.

00:11:32: Is that kind of a good?

00:11:35: defined by the law, right?

00:11:36: I mean, the law is not gonna hold OpenAI or ChatGPT responsible, but let's look at

00:11:41: that exact question that you asked.

00:11:43: So you said that ChatGPT was used to cite a case, okay?

00:11:48: Now going back to what I said earlier about where the data comes from, where

00:11:52: would ChatGPT be getting that data from?

00:11:57: I mean, it could hallucinate as well, right?

00:11:59: So if it's citing fake use cases, you could either retrieve information that has

00:12:04: been trained on or hallucinate if it doesn't know the answer.

00:12:07: even do one more.

00:12:10: It most likely hallucinates if it doesn't do any research beforehand, like a Google

00:12:17: search or anything that extends the context of the question.

00:12:20: Right.

00:12:23: Svetlana, you're correct.

00:12:24: But Svetlana, you hit on, right in the middle of your description, you hit on the

00:12:27: real reason.

00:12:28: You said what it's trained on.

00:12:30: So again, the AI systems are drawing from data that they know about they're trained

00:12:35: on.

00:12:35: Do you know what information Chatchie BT35, as an example, was trained on?

00:12:41: Of course not, because it doesn't have sighted sources.

00:12:43: It's a closed model.

00:12:45: There's just no transparency behind.

00:12:49: we know for sure that Chatch EPT originally was trained on Reddit data.

00:12:55: Okay, so think about the average Reddit user as described on Reddit.

00:13:01: Just go onto Reddit and look at how people describe the average Reddit user or the

00:13:05: average Reddit mod, whatever.

00:13:07: They make fun of each other all the time.

00:13:09: Think about the average demographic.

00:13:11: Okay, now you're asking ChatGPT a question from a data set that it was trained on in

00:13:17: large part from that data set.

00:13:20: So does it really surprise you that the information's not accurate?

00:13:23: Like is it really a hallucination?

00:13:25: Like I would argue that you're using the wrong tool, right?

00:13:28: And that the data is coming from the wrong place because that's originally where it

00:13:31: came from.

00:13:32: So about a month ago, $60 million was invested by some nameless company into

00:13:39: Reddit, right?

00:13:40: Recently, who didn't turn a profit for forever by an unnamed company to get

00:13:47: access to the data.

00:13:48: It turned out to be Google.

00:13:51: So why are people suddenly interested in all this public data?

00:13:54: I mean, it's not rocket science, right?

00:13:56: I believe most likely Google is using it to see how people interact versus the

00:14:00: quality of the data.

00:14:01: But even when Sam Altman came out and talked about the release of GPT -3 .5, he

00:14:06: said it's not a finished product, it's a proof of concept, it's an example, it's

00:14:09: not really meant to be public.

00:14:10: And everyone was blown away when they looked in the mirror of their own data and

00:14:13: said, oh my God, it's alive, and it isn't.

00:14:17: And this is the biggest problem, is that people do not understand fundamentally.

00:14:21: how the answers are being generated.

00:14:23: They don't understand where the data is coming from.

00:14:25: They don't understand that things are missing.

00:14:28: It's like the fourth unknown unknown quadrant.

00:14:30: If you don't know to look for a cited source, how can you justify whether or not

00:14:33: the information is accurate based on a cited source?

00:14:35: If you don't know the source is supposed to be there to begin with and nobody's

00:14:38: even thinking about it.

00:14:40: So that's another huge problem, right?

00:14:42: So there's just so many issues with using this as the main interface to answer your

00:14:45: questions that approach is destined to fail, right?

00:14:49: That's why...

00:14:50: this is this fallacy.

00:14:51: That's why these attorneys are getting in trouble because they're starting without

00:14:55: any degree of education.

00:14:56: They're starting depending on accuracy and they really don't understand how the

00:15:00: questions are being answered.

00:15:01: It's like, would you hire a paralegal intern off the street and do this work for

00:15:04: you?

00:15:05: No, but you're doing it with Chatchie P .T.

00:15:07: and somehow you're expecting magic.

00:15:09: I mean, that's where the problems are.

00:15:11: I think in general, I have talks to people expecting the AI solution to work out of

00:15:20: the box with their specific use case.

00:15:23: Why?

00:15:23: Why should it?

00:15:25: It's not trained on anything you've ever done, so why should it just work?

00:15:32: And yeah, that's a general understanding.

00:15:35: But...

00:15:36: I would love to emphasize one point because today is about data privacy and

00:15:41: security.

00:15:43: Public data is one thing, but I've seen also in a lot of your posts you actually

00:15:50: warning people that there is a potential misuse of their data if they are not

00:15:57: careful where they use ChetGPT or AI solutions in general.

00:16:01: Would you love to elaborate on that?

00:16:04: Yeah, I mean, this is unfortunately buried in the privacy terms and conditions of

00:16:07: every AI system that you interact with.

00:16:10: So if you've ever upgraded iOS, this is one of things I always say is the biggest,

00:16:15: you think about the biggest lie that you've ever told, right?

00:16:18: No one ever wants to admit it.

00:16:19: The biggest lie that everyone who owns an iPhone has ever told is when you get the

00:16:23: iOS update and it says this is for security reasons, did you read the terms

00:16:27: and conditions?

00:16:27: I agree.

00:16:30: Okay, I mean, show me honestly, anybody that's actually read it.

00:16:34: So we fool ourselves in our ambition to deploy the latest and greatest and

00:16:39: innovate.

00:16:40: And truly, experimentation is a big part of that.

00:16:43: But unfortunately, when you're looking at the terms and conditions of a lot of the

00:16:47: major AI systems, all of them have frameworks in place, whether it's for

00:16:54: insight or data collection or just outright privacy violations.

00:16:59: We see it.

00:16:59: So you've got to be really careful.

00:17:01: I'm not an attorney.

00:17:02: Okay, so this is where I'll draw my line.

00:17:04: I'm not gonna tell you what AI algorithms you should use or shouldn't use.

00:17:08: I will tell you that if you're getting involved with any system, you need to

00:17:12: understand the algorithms that are used behind it.

00:17:15: And then as a CEO, you should be looking at, okay, what does this mean for us in

00:17:20: terms of privacy?

00:17:21: Now there's some level of implicit trust that you have to have, right?

00:17:25: The minute that data is exfiltrated from your environment, it's somewhere else,

00:17:30: okay?

00:17:31: So you give it to Microsoft, Microsoft's got it.

00:17:33: They're gonna go through and handle it with their terms of service.

00:17:35: You give it to ChatGPT and OpenAI, they're gonna handle it through their terms of

00:17:39: service.

00:17:40: So you have to be comfortable with that.

00:17:42: It doesn't necessarily mean that there's risk, but when you have systems like

00:17:45: ChatGPT where on the front homepage it says, we are training other people with

00:17:49: your data, you gotta know that, right?

00:17:52: that would that be the same?

00:17:56: Would that be the same consideration for people who subscribe to the Enterprise

00:18:02: plan?

00:18:03: according to them, no, but you gotta remember the enterprise plan is different.

00:18:09: It's a different use case, right?

00:18:11: So if you wanted to use it for that, then yes.

00:18:13: But this depends on what you're using it for.

00:18:15: If you're just going in and you just wanna use it for that, then I'm sure it's

00:18:19: probably not bad, right?

00:18:20: You'd have to read the terms and conditions.

00:18:22: I'm not 100 % sure.

00:18:24: My understanding of the enterprise is that they go to great lengths to keep that data

00:18:27: private.

00:18:28: It's still another system though.

00:18:29: You're still transmitting your data off premises, off site.

00:18:33: All right, so if you're looking at like GDPR and all the rest or like any of the

00:18:36: other major compliance requirements that require you to have data processed in your

00:18:40: country on systems that you maintain and control, any third party API call that's

00:18:45: not gonna work.

00:18:46: So how do you tackle this at Emily?

00:18:49: So if let's say, you know, for organizations that really, really take it

00:18:54: to heart, and they don't wanna use, you know, the regular $20 subscriptions for

00:19:01: Chad GPT, and they're not comfortable with the enterprise plans that Chad GPT offers.

00:19:05: So I'm curious, how do you address that with your work at Emily?

00:19:10: It's so simple you wouldn't believe.

00:19:12: You ready?

00:19:14: It never leaves Microsoft.

00:19:17: That's it.

00:19:18: That's the answer.

00:19:20: It...

00:19:21: say if you trust Microsoft and Microsoft's cloud, then you can also trust their AI

00:19:26: implemented on it.

00:19:28: And a lot of people trust all their business data to Microsoft 365.

00:19:34: one of, if not the most valuable company in the world.

00:19:38: Okay?

00:19:39: This is like almost a, question's almost a fallacy, right?

00:19:43: Because you say, well, you know, I don't trust them because they're so big and they

00:19:46: can do whatever they want.

00:19:46: Okay, so you go with a smaller guy.

00:19:48: But the smaller guy's not gonna have the resources that Microsoft has to provide

00:19:52: the level of protection that Microsoft can't.

00:19:54: So do you want a company that's got a greater likelihood of surface area of

00:19:57: attack?

00:19:57: Do you want to take an open LLM, like a truly open LLM, and run it locally?

00:20:02: but then you don't really understand what's going on in that code, right?

00:20:05: Because you don't monitor and deploy and install code for a living.

00:20:08: So you don't really know what that code is doing.

00:20:10: But hey, it's running locally.

00:20:12: My data is not leaving my premises.

00:20:14: Where is it?

00:20:15: You don't know, right?

00:20:17: So this is the problem, right?

00:20:18: Unless you are the guy who's doing it, it's the same problem you have with

00:20:21: encryption, right?

00:20:22: On the security side of things, you know, is it better to have an algorithm that

00:20:25: nobody knows how it works, but you know, once somebody figures that out once, it's

00:20:28: completely correct.

00:20:29: Or is it better if everybody knows how it works and you depend on the security of

00:20:32: the key?

00:20:33: Most people would argue the latter.

00:20:36: It's the security counterpart to the AI argument about data privacy, if you're

00:20:40: really wanting to do it.

00:20:41: So it really is a chain of custody of trust, right?

00:20:44: Where do you want your data to be?

00:20:45: Who do you trust it with?

00:20:46: And then can you put reasonable security controls in place to prevent exfiltration

00:20:51: and manipulation?

00:20:52: So you hit the confidentiality, integrity, and availability that's required.

00:20:56: That's the CIA triad, right, for security of your data.

00:20:59: And of course that's going to involve a third party unless you're the guy writing

00:21:01: your own LLM.

00:21:03: But good luck with that.

00:21:06: Even Sequoia Capital won't get behind you.

00:21:08: One of the largest Silicon Valley investors in the world will not get behind

00:21:11: brand new AI engine algorithm.

00:21:14: They're getting behind companies that are deploying AI.

00:21:17: By the way, we have no investors for the record.

00:21:19: We're not looking for any.

00:21:20: But they're not doing it because the battlefield is just too competitive and

00:21:25: people are getting slaughtered left and right.

00:21:27: And at the end of the day, it's going to be about people that are able to use AI

00:21:32: are going to be the guys that are going to win.

00:21:34: Yeah, and it is a resource issue, you're right.

00:21:38: That's why I personally look like more positive on Facebook being in the open LLM

00:21:45: space because at least they have the resources to provide proper compute.

00:21:51: But yeah, it's definitely interesting.

00:21:54: From a privacy perspective, if I want to deploy my solution, like I have an

00:22:00: internal use case I want to solve.

00:22:05: Having it on Microsoft is one...

00:22:07: but more in general, what are the things I...

00:22:13: to be sure about my data.

00:22:18: Are there some key points you would suggest?

00:22:22: Since you brought up Facebook, let's just take a moment of silence to remember that

00:22:27: Mark Zuckerberg started his company by saying, they trust me dumb schmucks.

00:22:32: Okay.

00:22:32: But he didn't say schmucks.

00:22:34: That's the beginning of his company.

00:22:35: Okay.

00:22:35: If you don't know that, you know that.

00:22:37: So go back and look.

00:22:39: Now they've got a ton of money by collecting people's data and they are now

00:22:43: selling an AI system.

00:22:45: Okay.

00:22:45: Might be open, but hey, we trust them.

00:22:50: Okay, so I'm not huge into companies that start like that.

00:22:55: It's no surprise, I've talked about this at length.

00:22:57: Now, the question that you're asking Ed in terms of the use cases, the biggest

00:23:03: problem that I've seen with deploying AI systems and the easiest thing to get

00:23:09: around in terms of a successful deployment in this regard is to succeed, you have to

00:23:15: start with the use case and the data, not the tool.

00:23:18: And I've said this a few times, but...

00:23:20: it's worth clarifying.

00:23:22: Starting with the product first or tool first approach will fail.

00:23:27: You can't come in and say, I've got AI, I'm going to deploy it into the

00:23:31: environment.

00:23:32: There's this picture I'll have to get to you guys.

00:23:33: It's a cartoon where it's got a guy in the back of a boat with a little martini glass

00:23:38: and he's talking to a girl and there's a flag flying off the back and there's like

00:23:41: 10 people on the front of the boat and the girl says, man, how'd you get money for

00:23:44: this project?

00:23:45: And the guy goes,

00:23:46: I don't know, it was the same thing I was going to do except I just called it AI and

00:23:49: the back of the boat says AI.

00:23:51: You know?

00:23:53: So, I mean, everybody wants to figure out how to do it because they're enamored with

00:23:57: the potential but nobody really understands what they're doing unless

00:24:00: they've thought it through.

00:24:02: So the number one thing is the use case.

00:24:03: What problem are you trying to solve?

00:24:05: Because AI is not the solution to every problem.

00:24:08: You know, it's the old hammer and nail question.

00:24:09: You're running around with the nail and everything looks like a hammer to use, you

00:24:12: know?

00:24:12: Like it's not, it's just not...

00:24:14: it's not always the best solution.

00:24:16: So I think the use cases are really key and you mentioned those before.

00:24:19: So to get into some of those that you asked about at the very beginning of the

00:24:23: chat, what are some of these use cases?

00:24:25: Anything that's generative that you need creative feedback on is a great use for

00:24:30: AI.

00:24:31: Anything where you need to summarize content is a great use for AI.

00:24:36: Anytime that you need to extract out bullet points, rephrase, correct your

00:24:41: grammar, these are good uses for AI.

00:24:44: today without a properly trained set asking it for advice.

00:24:49: Bad.

00:24:50: Asking it to do math for you.

00:24:52: Bad.

00:24:53: Asking it for advice that involves math.

00:24:55: Probably the worst.

00:24:56: Last week I posted up a question, if you guys saw about the towel drying.

00:25:00: It was a very simple question.

00:25:02: I said, this is a very simple question.

00:25:04: If it takes 15 minutes to dry, or sorry, if it takes one hour, pardon me, to dry 15

00:25:11: towels in the sun.

00:25:13: is a one hour to dry 15 towels.

00:25:15: How long does it take to dry 20 towels?

00:25:21: What do you guys think?

00:25:22: Like hour 15.

00:25:23: No, actually, no, it's an hour.

00:25:25: It's an hour.

00:25:25: I'm sorry.

00:25:25: Why am I being stupid?

00:25:28: Yeah.

00:25:28: you think Ed?

00:25:29: no, I think like that's debatable because that's that's one that goes around a lot

00:25:35: of times often with shirts a drive five shirts How long does it take for for 20 it

00:25:40: depends how do we have enough space to do it in parallel?

00:25:43: Then it's still one hour Do you have it sequential then it takes longer?

00:25:47: So yeah, that's that I think an odd example, but yeah in general enumerations

00:25:53: stuff like that's not

00:25:55: AI is not that big in that.

00:25:56: It gets better, but sometimes you give it a more complex math problem and it gives

00:26:02: you the right steps and the right thought process, but gives you the wrong result.

00:26:06: So yeah, that's definitely not a strong suit.

00:26:09: I always see like a rule of thumb, like does it involve language understanding?

00:26:17: Is your task with automated language understanding improved, like automatable

00:26:22: basically?

00:26:24: and could it bring you a benefit in understanding unstructured data?

00:26:30: And extract something structured from which you can work with.

00:26:33: So that's basically how I always viewed it.

00:26:37: yeah.

00:26:38: So I think you're right, and I think that the fact is that there is a clear answer

00:26:43: if you make some assumptions.

00:26:44: The sun is the same, you're not putting the extra five towels in the shade, blah,

00:26:48: blah, blah, right?

00:26:48: All that stuff, the conditions don't change.

00:26:50: So that's where you will get back what I call the confidently incorrect answer,

00:26:55: right?

00:26:56: because you get back the math formula, so Elena, you're correct.

00:26:58: It's exactly how the AI started.

00:26:59: It was your first thought.

00:27:00: It started writing out the math formula, except it didn't know any better, right?

00:27:05: So it just kept giving out this incredibly authoritative looking answer.

00:27:08: And this is where it can kind of fall apart.

00:27:09: So you want to talk about the not so great use of AI if you don't know what's going

00:27:14: on.

00:27:14: And this is where a lot of people get confused.

00:27:17: You ask AI a question.

00:27:18: When I say AI, I mean most major AI LLM systems like Copilot and Perplexity and

00:27:23: others.

00:27:25: the answer they give you back looks so authoritative, particularly if it's not a

00:27:28: topic that you understand, you assume it must be true.

00:27:31: And that's kind of like what Mark Twain said, right?

00:27:33: It's easier to fool somebody than to convince them that they've been fooled.

00:27:36: That's actually not exactly what he said.

00:27:38: It was part of his book, but I have it written out in my Elephant in the Room

00:27:42: post from earlier.

00:27:43: But it's the general sentiment, right?

00:27:45: So the idea is that if you're not comfortable with the subject matter that

00:27:49: you're asking it about, and you get back the answer that looks correct,

00:27:54: you're misleading yourself and you don't even know it.

00:27:57: And I think that's the biggest challenge because once you lock that information in

00:28:01: you go, this is the way it is, you start making other decisions based on that.

00:28:06: And it's hard, it's really, really, really hard for people in general to go back and

00:28:10: to remember where they learned things from.

00:28:13: It's a big challenge.

00:28:16: kind of talked about one of the things is, you know, with AI is that it doesn't know

00:28:23: any better because you've prompted the system to provide an answer and whether or

00:28:29: not it has information to back that information up, it's going to respond

00:28:33: because you asked it and it's going to come up with an answer.

00:28:38: And one example that we've provided before is unlike children,

00:28:42: So who you ask, you'd be like, hey, Tommy, what do you think the subject is?

00:28:46: And they'll look at you weird and say, I don't know, and run away.

00:28:50: With AI systems, they don't do that.

00:28:54: They don't respond by default to say, oh, this is not data that I currently have.

00:28:59: I'm just going to take something that's maybe potentially related and become very

00:29:04: creative and sound authoritative, as you've mentioned.

00:29:08: So it's hard to decipher when.

00:29:10: you the information the output was in fact based on information that was trained on

00:29:15: versus completely like made up.

00:29:19: But one of the things I think we've kind of talked about before is rag and citing

00:29:24: references and citing documents.

00:29:26: Is that a potential way to again, build, like build trust, but also validate or

00:29:35: kind of understand, okay, well, should I be trusting this output?

00:29:38: Is this something that

00:29:40: does look reasonable as an output.

00:29:43: So do you think that that's a good alternative or do you think that we still

00:29:46: have significant work to do beyond the sources citations and the rag paradigms?

00:29:54: Fantastic question.

00:29:56: You're not gonna like my answer, it's both.

00:29:57: I'll give you an easy example, a very easy example.

00:30:01: Retrieval augmented generation reg, where you're pulling data from a document and

00:30:05: you're getting a sighted source, helps a lot.

00:30:08: But here's where falls apart.

00:30:10: You have two versions of the same document with similar content.

00:30:15: One's a month old, one's a month new.

00:30:17: Okay, so now your mind starts racing.

00:30:18: Well, I know how to solve that problem, I'll take the newer document.

00:30:20: What if the newer document has...

00:30:22: have been validated, what if somebody uploaded it and they weren't supposed to?

00:30:26: And it's not, the changes there aren't supposed to be reflected.

00:30:28: You need to use the old information.

00:30:31: Okay, well let me go through change control process.

00:30:33: Okay, well what's the change control process?

00:30:35: Again, now you're getting to data.

00:30:37: Okay, so now the question is it really, is it ragged?

00:30:39: And this is why a tool -first approach fails.

00:30:41: Because you can't go in and say, I'm putting a tool in and it's gonna solve my

00:30:44: problem, because it won't.

00:30:46: Right, the problem you're dealing with is a data problem.

00:30:48: So until you can solve for that human element of data qualification, the

00:30:53: business process optimization, right, behind the scenes, you're not going to fix

00:30:59: anything with AI.

00:31:00: In fact, you may complicate it.

00:31:02: And the worst part is that humans in general, like trusting things, you know,

00:31:06: it's like glass.

00:31:08: Once it's broken, it's very hard to put back together again.

00:31:11: So you start looking at an AI system and it starts giving you, you know,

00:31:14: incredulous answers.

00:31:16: And suddenly you say,

00:31:17: I can no longer trust this thing for anything, right?

00:31:20: Even if it's not the AI's fault, even if it's bad data that you're putting into it,

00:31:24: you know?

00:31:24: So I think that's really the challenge.

00:31:26: So, you know, unfortunately, I would love to tell you it's one or the other.

00:31:29: It's both, for those reasons.

00:31:31: So what would be the best practices?

00:31:33: So I know that, as you mentioned, so you do need to have, again, a strategy for

00:31:38: managing data, managing versions, accuracy, probably taking care of bias and

00:31:44: things like that.

00:31:45: So what are some best practices for?

00:31:47: doing it right.

00:31:48: So should you have a data team that really strictly just focused before even

00:31:55: implementing AI or even thinking about the solution, really focused on the data to

00:32:00: figure out what's relevant for my use case?

00:32:04: Is this data relevant to the problem that we're trying to solve?

00:32:08: And how do you ensure quality of the data?

00:32:10: And then is this a team that you maintain over time that kind of maintains and

00:32:15: oversees the quality of the data?

00:32:17: Is there a way to automate kind of that process with some, you know, monitoring

00:32:22: tools and things like that?

00:32:23: So just curious, what are some best practices that you recommend to clients

00:32:27: that you work with?

00:32:29: that's a really good question.

00:32:30: I know we only have limited time, so I'll try to condense this down.

00:32:33: The number one thing is you have to talk to somebody who knows what they're talking

00:32:37: about.

00:32:38: I've mentioned this before, but there are a lot of people out there that claim that

00:32:42: they're AI experts who don't know what machine learning is.

00:32:45: who don't know what a knowledge graph is, who don't understand that the fundamentals

00:32:47: of AI existed prior to the 1990s.

00:32:49: They see GPT and they go, oh my God, I can connect to the Zapier and it works and

00:32:53: take a look at this in magic like a chat with your PDF and this is magic.

00:32:58: Those are not the guys you wanna talk to.

00:33:01: You also don't need to talk to somebody that's super far in the weeds that's like

00:33:05: a complete LLM optimization gear head.

00:33:07: I love those guys to death.

00:33:08: I have several of them on my team, but they're not the people that you wanna be

00:33:11: leading the charge when it comes to business requirements.

00:33:14: So you need to find the right person or the right team to talk to who understands

00:33:18: what AI can do and what AI cannot do.

00:33:23: What AI is good at and what AI is probably not going to be that good at.

00:33:28: And then that same person needs to understand what problem are you trying to

00:33:32: solve and is AI a good fit for it at all?

00:33:36: So several organizations are hiring chief AI officers.

00:33:39: They're giving part -time AI officer roles.

00:33:42: Some folks are hiring.

00:33:43: you know, part -time AI consultants, that's all wonderful.

00:33:46: I would just stress that leading with the business process, leading with the

00:33:50: strategy that you're trying to solve for over a period of time, which is not gonna

00:33:54: be a week or two, okay?

00:33:56: I've been in the room with so many executives over the course of my career.

00:34:00: You come in and say, oh, we have this really important problem we need to solve,

00:34:03: and they start talking in weeks or months, and you're like, there's no way.

00:34:07: There's no way that a problem that's this big with this many horrible tentacles all

00:34:11: through your organization.

00:34:12: to be solved in three weeks.

00:34:14: So right off the bat, as an executive consultant, you need to understand that

00:34:18: that's a flag, that's a red flag.

00:34:21: So these are the types of pre -qualification questions and the right

00:34:24: person will ask them everything I mentioned and there's a lot more to it

00:34:28: behind the scenes.

00:34:29: But it starts with the right relationships.

00:34:31: So I would say that's probably the number one thing.

00:34:34: Unfortunately we already chewed through half an hour so it was kind of fast.

00:34:43: Michael, really thank you very very much.

00:34:48: you

00:34:53: actually works with that stuff.

00:34:56: Thank you very much.

00:34:57: we, we, you know, we don't see maybe on my side, kind of the ins and outs, the

00:35:03: details that you kind of provided, especially on the data.

00:35:06: I think I've learned a lot.

00:35:09: So I appreciate your time and working with us because again, we had the full start

00:35:13: before and we we figured it out, but really appreciate your insights.

00:35:19: I think that our audience will definitely benefit from your advice and kind of the

00:35:24: stories that you've you've told us.

00:35:26: to our audience and the advice that you've shared.

00:35:29: So truly appreciate you and we really enjoyed having you on the show.

00:35:34: it was a pleasure.

00:35:34: Thanks for having me guys.

00:35:36: Thank you.

00:35:37: Bye bye.