The information economy as a rod. The world is teeming with innovation as new business models reinvent every Intensis is your source of information and insights about how to make the most of this exciting new era. Learn more and Inside Analysis dot Comside Analysis dot com. And now here's your host through Eric Kavanaugh. All Right, ladies and gentlemen, Hello and welcome back once again to the only coast to coast radio show that's all about the information economy. It's called Inside
Analysis. Your host here, Eric Kavanaugh. On a very special episode, folks were going to dive into the roaring hottest topic in the world of analytics and AI and business intelligence as well. These days the large language models, or foundational models as they're called AI. Chat bots, for example, have been around for quite a few years now. Chatbots they were a bit rusty, I think when they first came on the market. Maybe sometimes they weren't
quite ready for prime time. But my how things have changed, and certainly open Ai threw down the gauntlet with chat GPT, but then Google has barred. Of course, GitHub has co pilots. You have data breaks with Dolly two point zero already, So these things are suddenly everywhere and everyone's talking about in generative AI. That's the sort of broader subject area involved here, and we're very pleased to have with us Mike Finley. He's the co founder and
chief technology officer for a company called Answer Rockets. They've been in the BI space for quite some time now. We've been tracking them doing some pretty interesting things, and we're going to talk about Max. Introducing Max, and I have to say, remember the old movie It Wasn't Gunball Rally was one of those kind of movies. Though in the end the guy goes push the button
Max, he goes boom. All these crazy things happen. But Max sounds pretty interesting to me and folks, the extent to which these large language models and foundation role models can change your workflow is really quite remarkable. One of the smartest people I've met in a while. We interviewed him last week down in Austin, Texas, and he was telling me he'll get technical briefs that are one hundred pages long that he has to consume over the weekend. He'll
just dump it right in the chat cheepts. Give me a three page summary, Boom done. I mean, that's like a dynamic cliffs notes engine basically is what you can use it for. So these things are very very powerful, and they're very complex under the hood. They could do some amazing things. Now, there are these hallucinations we've heard about. Maybe we'll talk about that, but let's get down to brass tacks and bring in Mike Finley from answer Rocket. Thanks for your time today. Tell us a bit about MAX
and how MAX is helping with your bi initiatives. Sure, Eric, thank you, and thanks for that introduction and the good launch for the conversation. Yeah. So answer rocket helps businesses have I like to say they have conversations
with their data. Right. If you think of all the data that a business has, and it's expensive to collect it, to collate it, to save it away somewhere, being able to have a conversation with that data, right, to treat it like it's the brain of the business, like it's a simulation of the operations, so you can go in and say what happened and when and why? Right, That's that's what answer rocket does. Right, That's that's our expertise. Um, not so much. I want to
set right that's been done many many times. But answer rocket really is about
allowing that conversational interaction with with your database. And that's what it's been, uh, you know, for ten years now, and of course with the advent of language models, specifically most recently with gpt UM, that's suddenly taken a huge leap forward in terms of that conversational capability, not just to be able to identify what the user wants if they are clever enough to sort of walk their way through the right keywords and the right commands, but just to
be able to vary candidly in any language of their choice, have that conversation with the data, uh, and have have that that agent which is max feel like a coworker who really understands, who really grocks what's going on in
the business. That's what we built well. And what's cool is these are text generators, but they generate syntactically correct tax they generate meaning on having absorbed goodness knows how many volumes of text and the training of these models, and so they are very powerful and they can very quickly ascertain certain aspects of your business, assuming you've connected the data to them or fed the data into them. Basically to feed this data into a chat gpt right, and every one's
out there too, Dolly, whatever, whichever one you're talking about. Of course, cold Pilot is great for coding, right and geth hub it's it's doing the coding side. And Gmail was on this I think first, at least in terms of general availability. You would note in Gmail if you're in the condician your sentence, it's I was like, you know, a year and a half, two years ago. Yeah, I was like, oh wow, this is getting pretty interesting, and it was pretty good zoo.
Right, So these foundational models, and it's basically a language AI bot that can interpret your data and then give you some insights about it, right, Yeah, so absolutely. And there's the thing that the language model is really good at is that interaction with the human. It's good at understanding what the human means, like the subtleties and the intense, and it's good at explaining the answer back to the human. Right, that's the part that it really
was trained to do. That's what it's really good good at is faithfully looking up and using all the correct facts that are in your database. Right. And so that's where answer rocket comes in. That's why this is such a
strong partnership between answer rocket and a language model. It says, the language model can hold that that interaction with the human with all the subtleties, with all the conversational aspects of evolution, of solving a problem, but allowing answer Rocket on the back end to both retrieve the correct data, make the correct decisions and calculations from that data, and then fact check the results that before
the language model ever presents it back. Right, So you really have to You have to treat the language model like like you would treat a human co worker. What would you do? You would train them on your business. You would make sure that they are an expert in the area that you're talking about, and you would fact check their results. That's what you would do if it was were a human that you're talking to. So you have to do the same thing with the language model. An answer Rocket does that by
essentially wrapping that interaction. So I had a great fund one the other day. You know, a customer said, hey, what were the top five products in you know, this category last quarter? And Max sent that to the language model and said, hey, language model, what do they want?
A Language model came back and said, oh, they want you to examine the product table and rank the top five members for the whatever, right and so answer Rocket went away and got that data, produced the answer back and said, by the way of these five, this one's an outlier. Right. That's something that GPT wouldn't have known by itself, but we were able to inject that in. Now GPT gives a great answer back to the
user saying, oh, there these are the top five. Good news, there was an outlier, and it looks like that was due to growth in California. Right. And then the user followed up with um, um okay, which ones were the dogs? And so to those kinds of questions.
This is really where the language model and the bi tool working together is beautiful, because if I tried to write code that said, if then else to figure out what were the dogs, right or which ones were the dogs, I'd be writing code until I'm retired, and um, you know, I'd never have a chance to finish. The language model knew exactly what they meant. It said, switch to bottom that's it, from top from top five,
switch to the bottom five and answer the same question it was. It's it's those moments, are I mean, it's surely goosebump before a guy like me who's been intech long time never really faced a problem that I didn't know how to code. Uh you know, it's it's um. And now these language models are just doing things that are fundamentally different. I mean, imagine if you brought a contractor into your house and you said, hey, I want to add a wall here, and they just scratch their head and said,
I don't know how to add a wall there. That can be done, right. It's almost like like somebody said, make me a floating bathtub. I don't know how to make a floating bathtub. That can't be done. These language models are doing things that are so fundamentally new. And again you hear a lot about this in the press. Not even their creators know exactly how they work. And so I think we're just at the we're at
the beginning of all the possibilities where we've seen that. It's kind of like that old Star Trek transparent aluminum moment, you know, when Scotty needed to save a whale from the twentieth century or something and he made transparent aluminum. Right, we have no idea. We're at that moment. We know we have transparent alluminum, what do we do with it? Right? This is
where all the engineering wrapping the language model is coming about now. And this is the whole ecosystem, the fountain of companies that are starting up around this tech, you know, is so this is very interesting. So I think about data warehousing, business intelligence, even business analytics to kind of stretch it a bit further, all of these disciplines revolve around transactional data and it's really
numbers. I mean, there are tags that assigned to objects like products, for example, services, but really you're talking about the numbers game that's moving all around. Well, what's missing in that is the language to interpret and explain what the ugers mean. And that's the magic that these large language models are bringing into the equation is they can find ways to represent the facts in a fashion that is meaningful and understandable and inspiring to the user. Right,
that's right, and that's really working at the user's level. Right. So, if you go, if you go really back to the beginning of computing, right in the eighteen hundreds, you know, you think of a problem, you express it to the machine as a program. So right away you are having to write your question in the computer's language. The answer comes back to you as a spreadsheet. Right, so you are getting the answer back in the computer's language. That paradigm is broken. One hundred and fifty years
in we finally broke out right. Now now we can ask the question on our language and we get the answer back in our uage as well. Uh. The analogy that I like to make is, um, you know data, That data is digitizing your business. Right, it's a it's you know, it's uh uh, it's the matrix. Right. The data should show you everything about your business. If you just know how to look at the
data, you can see the business right. And this is this is exemplified by seasoned executives who can look at a wall of numbers and spreadsheets and say, oh my god, what are we going to do? Right? Um, and nobody else in the room knows what it is, but they know the numbers to look for. Well, um, if you try to if you try to um uh digitize the business like that, you lose a lot
of the subtleties unless you have all of that depth. Right. So it's kind of like you know, if I if I digitize Romeo and Juliet, well, it's boy meets girl, boy falls in love with girl. Boy loses girl. Right, it becomes a very boring story. Right. I can't I can't digit I can't make transactional records out of what happened to my business. The story is more subtle than that. The story is unstructured, right, And and that's where that's where these language models are helping us.
Take that data, which by definition has to be structured because I have to be able to record a series of things that happen in the business and turn them into unstructured information, which is really where all the subtleties are, right, It's where all the competitive edges, all the opportunities to improve, you know, the whole idea of strength and weaknesses. Right, these are soft things. These aren't digitized. And this is really where the language models is
helping is helping business users shine. Yeah. Well, the data is like the wire frame essentially, and then the lms are filling in all the colors and the shapes and the dimensions and the fluidity for example, to view this
rich experience to help understand what's going on. And you know, in terms of structured data versus unstructured data, well, you can kind of get into a semantics debate about the fact that languages do have structure, it's center acts, and these llms understand that syntax is probably better than people do and can explain things to you. And that's really the other interesting side of the equation,
right is you are having that conversation with your data. We've talked about that for thirty years and it's business, but it's never really been that. It's always been playing with the numbers, looking at visualizations, kind of identifying patterns. But to have an engine then explain that to you in English or French or German or code, that gets very, very interesting because now you can really flesh out your understanding and it is a process you go back and
forth and say, hey, can you explain this? Can you explain that I saw data bricks was talking about with Dolly, how you can use it to explain your code to you? So, you know, let's think about digital transformation, which of course is hugely important these days. And how many of these old programs are written in coball. Well, there aren't that many developers that write coball anymore. A lot of them are retired, and so
this has been a big sticking point. But it doesn't have to be because you can just take corpus of code, PLoP it into your single instance chat, GBT or whatever and say, explain to me what this is doing, and it'll do that, right, sure, Yeah, extracted or translated or do any number of things. Yeah. And you know, keep in mind it's you're still in that, You're still in that. Treat it like you would treat a human in other words, Uh, make sure it's trained on
the thing you're asking it to do, and then verify the results. Right. So there's no free lunch even with the language models, but there is a lot of acceleration that comes from from the capabilities that they're bringing that and I think that's the term. I think we should pick on that term and really drill into it. It does accelerate the process of learning, of designing, of understanding, of sharing, of articulating. All these things are accelerated
by this kind of an engine. And you know, as a writer myself, I'm a little bit leary of just using it to do writing, but I will say it's very good at giving you the contours of a conversation, are very good at giving you the pill is of a certain topic. If we want to talk about data science, what are the top five issues in data science. It will give you five really good issues. That's right for
data science. Absolutely absolutely. It's like a teacher that's helping you all along the way, right, yes, so um, kind of like improv yes and right um. And the end part is that by the same by the same token, that it knows everything there is to know about things that are general to the marketplace. It knows very little about your business. It knows very little about the seasonal behavior of the sales of one of your products.
It knows very little about the supply chain issues that you have in monsoon season. Right. It knows very little about your your competitor and your price wars. Right. So that's where that's where again this one two combination of that deically and how max works is really to say, hey, the language model is talking to the user about the data, but answer Rocket is feeding in those observations, those facts, those insights, not just in the form of
saying January was fifteen million February it was twenty million. No. Instead, where January was an outlier even though seasonally we expected it to be high, it was actually way outside of its even seasonal high parameters. And the reason is these three drivers, and those drivers, one of them, by the ways, that a record high for all time. We provide that feedback back
to the language model. It gets a chance to say, oh, I've seen a bunch of recent highs because everybody's moving to South Carolina, or maybe it says, oh, that is an unusual thing. I get a chance to digest that and provide it back to the user. So it's it's the partnering on on one side, the answer rocket being able to really rock your business and dive in and understand your data, but the language model being able to make sense and have a conversation about those facts back to the user.
Yeah, it provides the contact, It provides the language to explain what's happening, and you do have to kind of iterate, and I think in our next segment we'll get into a bit more detail on how you actually deploy these things, because you want to deploy a chat, GPT or one of these other various AI engines to be able to have some controls around that, right, because this is one issue that you have to watch out for, and it's just a question of knowing what the tool can do and what the tool
can't do what it's designstering to do. And just very quickly here, I remember when the New York Times editor was asking it all these esoteric questions and existential questions and all these things, and it got all this weird stuff back. That's not how the tool is meant to be used, right, So
this is it's sort of an interesting journey to see what happens. But to preclude or I'm sorry to conclude that it must have been sentient because they're saying these things as a pretty significant stretch, right, absolutely, And you could even take it a step further and say, essentially that journalist was basically writing a story and the language model was simply helping him finish that story in an
interesting way. Right. It didn't want to marry. I mean, it had no concept of being married, right, right, But there are a lot of novels out there where where you know, there's conversations that happen like that. So absolutely, it's it's um, it's incorrect to personify the language model. It's more like realizing that actually it's talking about a cultural issue that
the journalists tapped into. Yeah, no, that's exactly right. So you do need to understand what these tools are good for, what they do, how they work roughly to to really get the most value from them, and again trying to trick it into revealing that it's really sent in is probably not a very good use case, at least not in the business world. If maybe in the end, back you were listening to inside analysis, what if
you could own a piece of the future. What if you could build your next castle not on sand, but on the bedrock of a modern blockchain ecosystem. The first Internet gold rush made millionaires, The second wave is mine billionaires, but the third wave is just gathering now and anyone can get in on the action. Hop online to crowdpointtech dot com to learn how you can secure a foothold in the blockchain revolution. Whatever your passion, wherever you want to
go in life, there's an opportunity awaiting you right now. Go to crowdpointtech dot com to learn how the blockchain will fuel the next generation of innovation in this globally connected world. That's crowd pointech dot com, your trusted agent in an untrusted world. What's the longest running radio show in the world focused on data DM Radio? Want to be a guest sometime Send an email to info at dmradio dot biz. That's Info at dmradio dot biz. What if you
could own a piece of the future. What if you could build your next castle not on sand, but on the bedrock and modern blockchain ecosystem. The first Internet gold rush made millionaires. The second wave is minting billionaires, but the third wave is just gathering now and anyone can get in on the action. Hop online to crowdpointtech dot com to learn how you can secure a foothold in the Whatever your passion, wherever you want to go in life, there's
an opportunity awaiting you right now. Go to crowdpointtech dot com to learn how the blockchain will fuel the next generation of innovation in this globally connected world. That's crowdpointtech dot com. Your trusted agent in an untrusted world. What's the longest running radio show in the world focused on data? DM Radio. Want to be a guest sometime? Send an email to Info at dmradio dot biz. That's Info at DM radio dot biz. Do you need to get your
hands on some extra money right now? Maybe twenty five thousand or more. If you're a homeowner, now is a perfect time to get cash out. While homes in many neighborhoods like yours have gone up in money for anything, it's yours. You can buy an investment property, payoff higher interest debt, or make home improvements if you need twenty five thousand, fifty thousand or more. Now is the time home values are up and so is your equity. We offer you a way to use it. No need to use your savings
called New American Funding. Now and see how much cash out you can get. Call eight hundred seven one h three seven three nine, eight hundred seven one h three seven three nine, eight hundred seven one h three seven three nine. That's eight hundred seven one h thirty seven thirty nine. NMLS sixty six oh six Www dot MLS, Consumer XS dot org. This is not an offer or commitment to end subject to borrow or improperty qualifications. Not all
borrowers will qualify. Terms and conditions apply equal housing opportunity. When a player's sudden cardiac event brought a national football game to a halt, it's shown a spotlight on the importance of CPR readiness. Now, with youth sports in full swing, the American Heart Association is rallying parents and coaches to be ready in emergence. To be ready learn hands only CPR. It's a skill anyone can learn in minutes. Just visit haart dot org slash hands only CPR and nationally
supported by Elevant's Health Foundation. Now you can fly anywhere in the world and pay discount prizes on your airline tickets. Book a flight today to London, Paris, Madrid or anywhere else you want to go and pay a lot less guarantee. Call the International Travel Department right now at Local eight hundred two nine eight five seven eight three eight hundred two nine eight five seven eight three. That's eight hundred two nine eight fifty seven eighty three. Welcome back to Inside
Analysis. Here's your host, Eric Tavanaugh. All right, folks, back here on Inside Analysis. At fascinating conversation here with Mike Finley, co founder and CTO of Answer Rocket. We're getting some great answers about these large language models and how to fuse them with your traditional business intelligence environment. You know, always we talk about these things. Don't throw the baby out the bathwater
or no, this is not going to solve all your problems. It gives you a very powerful context onto and infuse into your traditional business intelligence or analytics environment. So, Mike, I'd like to ask you just to kind of walk us through, especially this instance conversation, how you go into maybe as your some cloud platforms, saying this is what I want, kind of walk
us through how that all happens and what it means. Absolutely. Yeah, So look exactly what does a GPT or a palm or a barred look like if you if you walked up to it with the physical thing, what would it be. It's a bunch of computers. It's a bunch of servers, right, that's it. Uh, And they have storage and they have whatever.
But it's basically about a bunch of computers. So if you want to use the same instance of GPT that all the teenagers are using to write their papers, right, then you go to the open AYE website and go to chat GBT and you're hitting the same computer that all those other folks are hitting. Um. And when you do that, it comes with some warnings. Right, anything you send to it is going to be used by it to learn how to do things better. That's how it got as good as it
is, right. But if you if you don't want that to happen, if you want the benefit of all that learning that that thing did, but you want your data to be kept secret on the way, to be used only for you, um, for your benefit in the future, and never revealed to anybody else. Well, then all you need is another copy of that big server UM. And that's exactly what what Microsoft offers through Azure.
Same same as if you had a database or a compute server. Now they're offering they're offering instances, private instances of the language models, and uh, it's as easy as adding one of those resources through an open an Azure account, create that model added and begin using it. And at that point it's it's going to it's going to be a transfer of all the knowledge in the
publicly available model that's out there. Latest one, I think is June thirteen is the last kind of finished to published a GPT three five turbo, and then it'll be used for your benefit and again it won't be shared with anybody else, with all of the same credentials and the same assurances that come with saving your data into an Azure database or using an Azure compute resource to process your your web server. So it's it's as secure as all the other cloud
services that you're already using. Yeah, and then so once you get that instantiated, you can start throwing your information at it. Right. I joked about my friend who just dumped dumps these two hundred page technical briefings into instance and he gets back a summary. So from a client perspective, would your clients then start loading their reports into this thing? Would you connect the data
warehouse to it? How does that work? Yeah? Yeah, So once it's up and running, there's a there's not a way to dump your data into it. It's still very much like the thing that you see that everybody uses online, the chat GPT. You you give it a prompt and it gives you a response. Right. So the key then is to have a piece of software um that essentially is acting as the front end. That's that's bridging the semantic gap between the business use and that back end. So let
me give you some examples. Um, if you let's say that your your your your users are gonna be talking about some obscure part numbers. You know they're they're gonna they're gonna look like numbers. So if if you just go to the public GPT instance in reference number six two five nine, five, eight or whatever. Um, that public instance is going to think that's a
number. It's not going to know anything about it. If if there's a semantic layer between your business user and the language model, then that layer can augment that number with something indicating to the language model. By the way, this is a part number, so that when the language model sees it, it knows, oh, that's a part number. Now now I know how to talk about that thing as a part number, not as not as a number. Right. Um, this happens with zip codes, right, happens
with um you know anything like excuse me, like quantities? Is it a percent or is it a dollar? Value? In business? That's really important in the in the largely numbers with suggested formats right in the in the business world, we need for the language model to treat it very precisely like what it's intended to be and not sort of run away and imagine a different future
for that number. Makes sense? Yeah, And so what you're doing then when you talk about this bridge to the semantic layer basically is answer rocketing. Your instance is going to be the front end, and it's reaching into chat GPT to generate context around specific discussions and topics and things, supply chain issues, or whatever the case may be. So basically it's it's now augmenting. It's part of your information management program, and it is a text generator,
context generator really designed to facilver already doing. Is that right? That's right, that's right. You can you can very much think of it as it's like the concierge. Right. So instead of instead of going in and saying, hey, business user, here is Excel. Please get your answers right, imagine you had you had a concierge who would say to the business user, what do you need to know about this spreadsheet? Um? And then the user would say, I need to know where the you know where the
problem areas are. I'm maybe I'm managing a bunch of franchises and I need to know today where should I go? Right? So you could you could do that in a spreadsheet by saying, give me the most recent day, rank those by the labor ratio or whatever, and that would tell you where you need to go today. Or you could say to the concierge, please look this up for me. GPT plays that that that role in answer rocket,
Right, So you've got all your data in the system. The user is very much in a conversation with GPT, but GPT is getting all the facts fed into it from from answer Rocket. So if the user says, hey, who's in trouble, GPT has no idea who's in trouble, but it knows that that means find me the bottom stores by labor ratio and shrinkage,
right, or whatever those those key parameters are. Because it's AI is goal driven, so you give it those goals and then answer rocket will find the way or the products that are ranked a certain way, or whatever the thing is the users looking for, and then the GPT will communicate that back to the user. Hey, um, you said five, but really two will really focus on these two the other three or these three, but they're much better off, so don't worry about them today. Right, And this
is how the music is made, right. It's it's really a magical experience because in the past, the user would have just gotten the dump of all this information. Um, you know, they'd be looking at a bar chart
that they're going to have to interpret. Instead, they can just get some very straightforward kinds of instructions and conversation and they and they can they can continue that line of thinking and questioning and drilling down and getting more well, I guess one question I have is in the typical setup, what is the actual front end, because like what chet GPT, you're in this environment, you've got to prompt you throw things at it, etc. Is that still how
you would do it ideally with primarily in the answer rocket environment. And then the the chat GPT is kind of the background. It is. The chat GPT is definitely in the background, and the reason for that is because it needs to be under security um and needs to it needs to have you know, this user is authenticated for what data they have credentials to access? Right, So so it's really you know, it's a it is a business intelligence
enterprise tool, right. It comes with all the sophistication of role management security, you know, row level isolation of that data, um. All those things are very important even through this chat conversation. So the experience though, the experience, the actual visual experience the user has is very much like a chat GPT, augmented with some charts and some other drop downs and click clicky type things that you would see you know in a in a dashboard or something
like that. But but it feels very much like a chat conversation UM and that that helps obviously if you're if you're on a you know, if you're on a desktop. It helps the user to feel like that's a live, interactive chat. But it also supports all sorts of other paradigms that you might use, like a like a cell phone based or even embedding this capability in a third party API so that it can be surfaced inside you know, Microsoft Teams or some other some other chat a slack, you know, some other
chat type interface. That's interesting, So would it makes sense to feed in reports like your annual reports, let's say, or your corarely reports. Would it makes sense to feed that into your instance of chat NPT and then over time? Because what one thing I've noticed that's really quite interesting is these large language models, these foundational models will sort of seemingly intuitively understand certain constructs,
even market dynamics and things of this nature. So you know, you could look at it and say, hey, I have to lay off ten percent of the people in this company because things do not look a good. Where should I look to first do that? And it can come back and give you some rationale for cutting back here cutting back the area can be very specific in those recommendations. Right, yeah, absolutely so. UM, so the UM you are correct. But let's talk about the way that you said it.
So you said, can I feed in my annual report to chat GPT. The short version of that is, UM, you can't feed it in in the sense of saying, can I make the GPT instance change because I've given it this report, so it permanently knows that information that that's not a thing yet. All right, So that's that's something that may be coming in the future. UM. We hear a lot of things. But what you can do, you can do. What you can do is say, okay,
now the user has just asked me a question about layoffs. What documents do I have that um, that talk about layoffs? Um, from the past. Oh, it looks like I've got this annual report from from five years ago when we went through a down sort downturn that talks about layoffs. Now, let me augment what GPT has right now for this question. Let me augment what it knows by giving it that document, and then it will answer that question in the context of that annual report that it has from the
past. Right, So, it's this is the discipline of prompt engineering. You may have heard that term. That's very much about It's very much about saying, what all do I need to put into that prompt so that the language model has what it needs to answer, but always keeping in mind I have to be able to fact check it. I can't I can't leave that. I can't have the question be completely open ended because it will hallucinate things above all else, it wants to answer, so it will make up an
answer. If you say, you know how many McDonald's are on the dark side of the moon, it's going to give you an answer. My favorite, my favorite test for language models when we're playing around with them is I'll say Hakuna and the answer always comes back matadha, and it wants to have a celebration about that. It really wants to fill in a blank. It will not leave Yeah, it will not leave a blank. So so you have to be careful when you ask it a question. You're going to get
an answer. If you want that to be based on fact, you have to number one and ensure that you fed it those facts, and number two verify that whatever it extracted is still true or holds true. Or is accurate. You're reminding me of a lesson I learned when we went to Europe as teenagers, and in certain countries they would tell you don't ask for directions from the locals because they'll be embarrassed that they can't answer, and so they'll just
give you some answer. And that's kind of like what check If you do right, they're gonna be hereing what ask they ask it for? Right, bild to verified and that that gets back to traditional business intelligence data warehousing and similar practices, because that's where you verify, Okay, is it really true that we had this many sales and you go check the warehouse and check and balance. You have to watch out before, right, Yeah, validation and
audits absolutely yeah. And so we we build that into the kind of the real time aspect of what the language model is processing and is returning back. And we also put in the guardrails that say, look, don't answer if you don't have this information and so, and that's in the answer rocket layer. Is that right? So basically it's it's a layer of instruction. You've built in some guardrails that say, Okay, no matter what the question is,
if you don't know the answer, don't come up with something. Yeah, I mean, the language model thinks it's taken the sat you know, it literally thinks I must answer every question and because it's it's better to take a guess than not to right. So so you have to be you have to be careful. It doesn't know. You have to tell it. Hey, the right answer if you don't know, the right answer is I don't
know. And so these are these are the very subtle. You know, one of the one of the best engineers I know the other day said, look, let's stop calling it prompt engineering and let's call it what it is, rhetoric, right, Like a lot of the instructions, Yeah, a lot of the instructions to the prompt are really just getting the rhetoric correct, right, getting the rhetoric that because at the end of the day, it is a language model. That's what it understands language. And so it'll do
what you tell it. Um, it will generally not do what you don't tell it. But you have to verify. You have to verify that before you you a business user to trust it well. And you know, so let's just kind of talk about rubber meets road with analytics and business right in the day. I mean, I get all excited about technologies because I'm in the space and I'm a geek. But at the end of the day, the whole point of all these exercises is to run your business better, and
a big part of that is the storytelling. Is to get the data and then explain to the board of directors, your boss, a partner, whoever, explain to some other party what's happening and why you want to take X,
Y or Z decision. And that's also where it can be very helpful because you can say, I have to present to the board of directors tonight, what are the five things I need to be sure to say, and please be sure to focus on the treasurer who always asks the hardest questions, and that it will give you a whole bunch of stuff to work with. You go through that and choose kind of cherry pick what makes sense for you right absolutely. You know it's funny. I my undergraduate was in physics and
I went into business. And one of the reasons why is because I think business is the hardest problem there is. And the reason the reason is the hardest problem there is. It's very straightforward as soon as you know the right answer, somebody else copies you and runs away with it. Right, That's that's how competition works. So so you have to get to get it right again and again and again. Well that's why these language models, basically intelligence
on tap is the ultimate power tool for business. Right And and they're they're very new now, like literally, you know, six months old. But imagine if after six months of its existence, it was unreliable, it was expensive, it was I still remember you know, I was in the cash register business back then, and a chain of pizzeria's own name, the CTO said I will put the cash register in the cloud when you put the pizza oven in the cloud, right, uh, and and so you know that
that kind of fear exists. But now you know, look at Square and all these other ones that they're very much cloud based and and so that's where these language models are going. There, they are they're going towards that that point. Yeah, so yeah, this is this is fast that extel folks don't touch up. That will be right back. You're listening to Inside Analysis
Ray stand up to the next financial crisis. That our top economists are saying is that our doorsteps By allocating a percentage of your IRA into physical gold and silver with a tax free rollover, you can diverse the holdings from turbulent markets and economic downturns by putting your IRA. Find out how to safeguard your assets with a tax free rollover with a Genesis Gold Ira, the only IRA that can hold physical precious metals. Call now for your free gold and silver report.
Protect your IRA today with one simple phone call and learn how to quantify for up to ten thousand dollars in free silver called Genesis Gold Group. Empowering faith driven Stewardship. Eight hundred six four four eight six one one eight hundred six four four eight six one one eight hundred six four four eight six one one. That's eight hundred six four four eighty six eleven. Can your IRA stand up to the next financial crisis that our top economists are saying is at
our doorsteps. By allocating a percentage of your IRA into physical gold and silver with a tax free rollover, you can diversify in safeguard your holdings from turbulent markets and economic downturns by putting your IRA back on the gold standard. Find out how to safeguard your sts with a tax free rollover with a Genesis Gold IRA, the only IRA that can hold physical precious metals. Call now for
your free gold and silver report. Protect your IRA today with one simple phone call and learn how to qualify for up to ten thousand dollars in free silver called Genesis Gold Group empower Ring Faith Driven Stewardship. Eight hundred and six four four eight six one one eight hundred six four four eight six one one eight hundred and six four four eight six one one. That's eight hundred six four four eighty. Do you own a timeshare? We'll face the facts. You
made a mistake, You made a bad purchase. A timeshare is not an investment. It's a money pit that continues forever. If you use your time share, that's great. But if you don't and you want illegally get out of your contract, call my friends right now at the Timeshare Exit Hotline. They're an experienced team of lawyers who help good people like you get out of a timeshare contract that they just don't want. Don't throw away your money on
maintenance fees, use it for things you really want. We can help you end your time share contract and stop the money drain. Immediately. If you are ready to move on with your time share, call our team right now. Yes your timeshare now with a free call eight hundred two eight nine O four one three O four one three, eight hundred two or one three.
That's eight hundred two eight nine zero four thirteen. When a player's sudden cardiac event brought a national football game to a halt, it's shown a spotlight on the importance of CPR readiness. Now with youth American Heart Association is rallying parents and coaches to be ready in an emergency. To be ready, learn hands only CPR. It's a skill anyone can learn in minutes dot org, slash hands only CPR. Hands only CPR is nationally supported by an Elevant's Health Foundation.
Now you can fly anywhere in the world and pay discount prices on your airline tickets. Book a flight today to London, Paris or else you want to go and pay a lot less guarantee. Call the International Travel Department right now at low cost airlines eight hundred two nine eight five seven eight eight hundred drew nine eight five seven eight three. That's eight hundred two nine eight fifty seven eighty three Express Welcome back to Inside Analysis. Here's your host, Eric
Kavanaugh. Hard folks, back here on Inside Analysis. They fascinating conference from Answer Rocket. We're talking all about foundational models, large language models. What do they mean? How do you use them? Well, there are lots of different ways, folks, I mean you can you can't even count the number of ways that you can use these things. But they're very good for, of course, generating That's what they're supposed to do, is generate text.
There are other models that are used for generating graphics. Mid journey. I think it's one that's through the discord channel is just shockingly amazing. But we'll focus on the text. And we just took a quick break there and the break might get a really good point talking about dark data. So what is dark data? Well, large enterprises have had massive amounts of data.
Most of it is unstructured. What do they say, ten percent, maybe fifteen percent is structured, by which we mean sitting in a relational databases,
structured in a way to facilitate reporting and access and so forth. But all the emails, all the PDFs, all the PowerPoint presentations, there is a tremendous amount of knowledge based into these things, and people who are old like me remember something called knowledge management and that was going to solve all our problems like twenty five years ago, and it really didn't get too far in part because we didn't have the compute, we didn't have the semantic models. Yet
now we have all that stuff and more. We have these engines and that's what they are. They are engines that can process information. So what do you think about the future of dark data in the enterprise and why does that matter money? Oh? Absolutely, yeah, So look, dark data is the future, is what I would say. Dark data is becoming the becoming the light data. The fact is we've thought of is the terminal state for information, Right Like once a human reads a spreadsheet, draws some conclusions,
writes it out, it goes in a folder somewhere areage. That's all we ever get to do with it. But now these language models are able to They're able to read that and make use of it and answer questions from it if we can find the right information to answer the question. And this is where the technology area. It's not it's not well understood, but it is
really important. Embeddings and vector databases, right, So, so these these embeddings and embeddings is just a a sophisticated way of saying what the computer or what the model thinks about. So it might be it's two o'clock in Tokyo. That's a statement. The model has an embedding for that. It might be you know, last quarter sales were down three percent and that's because,
uh, you know, we had a supply chain shortage in Asia. Whatever, whatever that statement, anything that you can say that you can write a two hundred page document has an embedding. And that embedding is just a bunch of numbers, a bunch of numbers that means something to the language model. So the way that the way that dark data becomes light is you index it with embeddings, You feed it to the language models UM, and you get from the language model, you get back kind of an index. Right.
And now anytime a question comes in, a new question comes in, Hey, um, when if we had supply chain problems in um, you know India, Right, Well, it's going to go match up against all the embeddings that it has, and it's going to India in Asia ut supply chain shortages, transportation problems, it's going to find, you know, anything related
to causing shortages of inventory. Why Because the language model has such a deep understanding that that it will associate any problem with with UM, apply with supply chain management, any problem with transportation, with supply chain management, any kind of outage, it'll it'll group all of that and and it really starts to show truly the power of what these language models are doing when you look at how how similar the embeddings for two things are, even though they might be
in different languages, they might be very different ways of saying the same thing, they might be only tangentially related, but the embeddings will light up. The embeddings will be similar enough that the language model will will tell you, hey, this is this is meaningful content to answer that question. That's really where where the superpowers are. It's almost like the like the Dewey decimal system where we get you numbers that the library uses. Imagine there's fifteen hundred numbers,
right, that's the index for that content. And if you can take all of your enterprise documents and index it like that, then you can access anything you need to access really quickly, get into this embeddings things. That's very interesting. So what are basically saying is and I have read up on this that one of the things they do in these large language models is that they provide a numeric value to every statement that we sent it to paragraph all
sorts of different things. Is that kind of where you're going with this? And that works? That's right? That that is that is the the the idea of embeddings. Look until until language models, So like this time this time last year, right, language models were kind of a cool thing.
Peripherally, we were all focused on machine learning, right and classifiers, right, and those classifiers they were really good at saying which customers are like other customers, you know which, But which products should I recommend to Eric based on the things that Eric already knows? Right? Uh, which are my stars and which are my you know, lagging solutions? Right? Um, that's as good as the language models could do. They really not sorry,
not language models. The machine learning models could do. They understood the idea of what's like and what's not like. The revolution, the giant leap forward are now able to generalize on that and say, which of these paragraphs are like each other? Which of these statements are like each other? Um, you know there's that like I like to when I'm training my engineers about this. I'll say, look, there's a million different ways that I can tell
you what time it is. I can say it's a quarter to three. I can say it is two four five right right, right, um, and a million different ways of saying that, And yet it's really just the same numbers at the end. That's how the numeration system in the language models works. It produces these these numeric vectors that match the meaning of the of the words not not that's itself, that's interesting. So and this is how they align things, right, finn or a dog? Right? I mean
that classification is good at YEP, but that's a fairly limited thing. And it's first segmentation, you know, and I think about you know, primarily done a couple different things. One classify, segments organized basically recommend you with optimize and optimize on a decision point, which is based on lots of different
factors. But this is a much different environment. I mean, it's it's predicting text, but it's also giving you all this context right to help you understand and so you know, I mean, the bottom line is you can just get about stuff. It just prompt that thing all day long to get information about how to ramp up. If I just started as a product manager for the software company, what should I get focused on? You know,
that's right, that's right, yep, yeah. And if you rewind back to two thousand and nine, is the very first paper from from Jeff Hinton that described essentially the precursor to all the language models we're using now. In that paper, you know it was it was a little eight line completed. It was absolutely fascinating. Um and and all that All that a little eight line program did was teach itself. That's it. It just at leased on
everything you gave it. It taught itself that specific one taught itself to recognize handwritten digits like you know, I draw a seven with a dash, and you don't where I do my ones, you know, kind of crooked, and you do your ones like a slash. Right, That's all it did. But but the thing ones and which ones were eights and which ones were threes. It separated them by itself. And that was the chilling moment that if it could do that, then all we have to do is get it
to the point where it could separate annual reports from baseball scores. Um and and you know from you know, the Declaration of Independence from uh, you
know, Google reviews. If it can separate those because it reads them and knows the difference and starts spacing them out, then all of a sudden, in order to do that, in order to solve that problem, it turns out, it has to understand in order to know whether you know, whether these this list of abs and season d's is a report card or a you know um uh and and and nfl um you know, seasonal training report uh. For it to know that difference, it has to understand the principles behind
it. And it was able to teach itself all that, so we didn't have to train it on those individual parts. We win. That's crazy. I mean, that really is amazing. So now you're teaching it to you you've instructed it to learn to teach itself things, and then you're building upon these. That's why they call it a foundational model, Right, there's foundation
upon which it's learning. What this guy say to me the other day about how doctors can benefit from these things, because a large language model trained on medical data can be a very very powerful thing. And today is the dumbest it will ever be. Gathering podcast segment is coming up next year. Listenings Inside Analysis, case AA where every Day is a Great Day, casey AA, Lomalinda. This segment sponsored by the generous support of the Dream team.
Looking for the keys to something bigger and better, downsizing or relocating to the perfect spot. Oscar Ramirez from Century twenty one Lowest Lower real Estate and Matt Flores from Secure Choice Lending are here to help you sell or buy with their trusted and experienced knowledge and advice. People are called Oscar and Matt at nine five one seventy five one three two four nine. That's nine five one seven five one thirty two forty nine real estate and loan advisors. Oscar and Matt
can give you a no cost consultation. You don't have to buy anything. Matt and Oscar can help you figure your way through the complicated real estate market. Email Oscar at loislower dot com or on Instagram if Oscar romeiraz Garcia and Matt Flores at Secure Choice Lending dot Com. Don't let today's real estate pitfalls
stop you from dreaming. Make your new home dreams come true. Dri number zero two zero seven zero three four four zahabla espanol Khebot Club's original pure pouty Arco SUPERTA helps build redcorepuscles in the blood, which carry oxygen to our organs and cells. Our organs and cells need oxygen to regenerate themselves. The immune
system needs oxygen to develop, and cancer dies in oxygen. So the T is great for healthy people because it helps build the immune system, and it can truly be miraculous for someone fighting a potentially life threatening disease due to an infection, diabetes, or cancer. The T is also organic and naturally caffeine free. A one pound package of T is forty nine ninety five, which includes shipping. To order, please visit to eebotclub dot com. Tohebo is spelled T like tom, A H, E, B like boy Oh.
Then continue with the word T and then the word club. The complete website is to hubot club dot com or call us at eight one eight six one zero eight zero eight eight Monday through Saturday nine am to five pm California time. That's eight one eight six one zero eight zero eight eight toebot club dot
com with sixty years of fascinating facts. This is the man from yesterday and back in time, we go to this time in nineteen sixty nine, Senator Edward Kennedy pleads guilty here believing the scene of a fatal auto accident which killed Secretary Mary Joke Packney. Ted Kennedy is given a two month suspended sentence and is placed on probation. This morning, I entered a plea of guilty to
the charge of leaving the scene of an accident. Prior to my appearance in Chord, it would have been proper for me to comment on these matters and from this time. In nineteen seventy six, the FCC expands citizens band radio to forty channels from its present twenty three. Currently here in nineteen seventy six, there are twenty million CB radios in use. Radio Shack has the hottest
thing on wheels today, realistic two ACB radio. We introduced our realistic CB line way back in nineteen sixty and today we have sixteen low priced mobile and waukee talkie models to choose from. And from this time. In nineteen ninety seven, Bulworth's announces it's going to close its four hundred store chain, bringing an end to an era. Over nine thousand will lose their jobs. The
store has been around since eighteen seventy nine. It's world worth Stereo Spectacular seventy one, a fantastic once a year's Salem albums, LPs tapes, everything in stereo, unbelievably low priced with more Atman from Yesterday dot com. Do you like to safely leverage bank money to earn double digit returns income tax free, with guarantees and no downside market risk? How can you do this? This
is Fern's host of the Your Personal Bank Show. One. You fund a high cash value policy one time to earn dividends and interest to establish a bank line of credit using the cash in your policy as collateral. When you earn more in dividends from your policy than the interests the bank charges, you keep the difference, and the difference is average two to five percent annually in your favor for the past forty plus years. Three the bank funds contributions years two
to twenty plus. Each year the bank adds funds your rate of return, increases your average rate of return and grow to strong double digits annually within a few years. Contact me at your Personal bank dot com, Your Personal bank dot com or eight six six two six eight four four two two eight six six two six eight four four two two for more info or tune in to
the Your Personal Bank Show. Your Personal Bank Show airs Tuesday's at four pm right here on casey AA ten fifty am and one to six point five FFM. The station that leaves no listeners behind. Ivy League Auto Transport and Flavors Mobile Detailing would like to remind listeners. Drunk drivers remains one of the nation's most serious issues. Forty five percent of traffic fatalities recording the US caused by an intoxicated driver. When you get on the road, do yourself, family,
and your community a favor. Remember the three CS of safety, Caution, courtesy and common sense. If your hand has been on the bottle, keep your foot off the throttle. This is a reminder courtesy of Ivy League Auto Transports and Flavors Mobile Detailing, serving our area with pride and integrity. Edge Manufacturing would like to remind listeners. Drunk driving remains one of the nation's most serious issues. Forty five percent of traffic fatalities recorded in the US are
caused by an intoxicated driver. When you get on the road, do yourself, family, and your community of favor. Remember the three cs of safety, caution, courtesy, and common sense. If your hand has been on a bottle, keep your foot off the throttle. This reminder is courtesy of Edge Manufacturing, serving our area with pride and integrity, specializing in truck rolloff, welding, steering and suspensions. Called today for all your trucking problems at
nine O nine six five three five seven one six. That's nine O nine six five three fifty seven sixteen. That's Edge Manufacturing on the air because they acare. This segment is sponsored by press Print, Southern California's best full service union printer and mailhouse. They offer their lowest prices around with unmatched service and
reliability and free delivery throughout So cal. Press Print can print anything from letterhead, business cards, and campaign literature to mailers of any size, lon signs, banners, doorhangers, are just about anything you might want. Press Print promises to save money for you, your business, or your campaign. If you'd like to learn more, contact Mike Croombrint at press Print seven one four
three nine nine eight seven zero eight. Get the Union Bug. Do you want to learn and get answers to questions not address in the mass media. Then you want to hear my show business Game Chambers with me, Sarah Westall, I have conversations with thought leaders who have the courage to address off limit topics, so you and your family have the tools to make the right decisions. Join me Wednesdays at four pm on ten fifty AM and one oh six
point five FM. Right here at casey AA, that station that leaves no listener behind. Case A Radio has openings for one hour talk shows. If you want to host a radio show, now is the time. I'm make CACAA your flagship station. Our rates are affordable and our services are second to nine. We broadcast to a population of five million people plus. We stream
and podcast on all major online audio and video systems. If you've been thinking about broadcasting a weekly radio program on real radio plus the Internet, contact our CEO at two eight one five nine nine ninety eight hundred two eight one five nine nine ninety eight hundred. You can skype your show from your home to our Redlands, California studio, where our live producers and engineers are ready to work with you personally. A radio program on CACAA is the perfect work from
home avocation In these stressful times. Just type CACAA radio dot com into your browser to learn more about hosting a show on the best station in the nation, or call our CEO for details. Two eight one, five nine nine ninety eight hundred. You're on board caseyaa's Inland balk Express CACAA ten fifty am, the station that needs no less. You're behind and be