Developing Generative AI Applications with Dmitry Shapiro, CEO of MindStudio

e incomplete and reaching out to people asking for more information and trying to figure out how to rank them and route them to the right sources, right?

It's like full time job for a bunch of people. It doesn't need to be. You can, in whatever, half an hour, without any engineers involved, build an application that watches the inbox and does all of those things. And so those poor people can now go on and do things that are more interesting and more productive than doing that.

Um, or, you know, lead ingestion, you know, every, You know, website in a sense as a contact form, some are quite busy of like, this is the way people try to contact the business, get more information, try to buy a product, et cetera. Those are typically monitored by individuals to sort of look at the form and again, potentially might need to reach out and get more information and disambiguate and then take that data and put it in some other system like CRM or something else, pass it on to a salesperson, an account rep or whatever.

Really? All of that can be automated. You don't need to do any of that as a human. So it's like automation, complete automation. Uh, then sort of next batches of things are like partial automations, things that still need humans involved, but And AI can be built to do a lot of the heavy lifting. So the human now is able to just live dramatically more than they ever could and sort of operate much faster with higher quality audio.

So there's tons of applications like that. Um, uh, and then beyond that, uh, for things that still require a lot of human interaction, Uh, uh, organizations are saying, well, could I now build custom AI powered applications for our employees that are specifically built for doing what they do, not using like off the shelf sass that we've all used for years now, like it was a staff that came out recently that, uh, I think medium organization.

Has 130 some odd SaaS products that they, that they use. This is insane. Doesn't make any sense. Why? Because those are all sort of generic products that we cobbled together to create these like weird products. information systems that we have today. Well, AI changes all of that. So now an organization can very rapidly, again, even without developers being involved, just like regular business users can learn to use MindStudio and very rapidly build for themselves custom business applications that are AI powered, made to fit like a glove.

And like dramatically make them more productive, you know, and get, do the things they don't want to do, like make them happier, happy employees are more productive employee. And then, you know, many other things like, uh, you know, there are still times that it's valuable to be able to just like chat with a, with a large language model, right?

You're just doing some enumeration, like some ad hoc research or thinking, well today, you know, a lot of employees are using chat GPT or cloth, like these basically sort of commercially available. Consumery type things, right? And, and so what MindStudio lets enterprises do is, is to, uh, sort of replace that kind of usage and give employees enterprise grade, you know, made by the enterprise for each job function of a, of a team member.

Uh, personalized assistant that uses MindStudio is actually, I probably should have mentioned this, model agnostic. We support all the models. Open AI, Anthropic, Google, Meta, Mistral, Open Source. Organizations are connecting them to private models in their own private clouds or on premise. So any model. Uh, and so these specialized assistants can use any model but are managed by the enterprise, can enforce rules of the enterprise, understand the job function, require much less prompting, can reference data from the enterprise.

So they're not just like making stuff up, but using real data. So it's like a really comprehensive tool set. That again, has over 50, 000 of these things that have built, but again, mostly fall into these like categories. And then by job function, depending on what you are, if you're sales, you're doing lead injection.

If you're HR, you're, uh, ingestion. If you're HR, you might be doing resume ingestion. But the pattern is kind of similar. It's like stuff that an AI can do. You can just build and, and make your organization radically. 

Richie Cotton: Okay. Uh, I like some of the stuff you're mentioning, like happy employees and making your organization better.

Those both seem like, uh, excellent goals. So is First two categories are particularly interesting because I think there's been a lot of question about are AI applications going to replace us as workers or are they going to enhance us? And it sounds like the answer is both. So your first category was hyper, well, automation.

So it's replacing humans. Your second category is sort of doing things interactively with humans and adding to them. Actually on that second note, you mentioned that, um, these more interactive examples are kind of replacements for SAS web apps. So can you talk me through how is AI changing SAS applications then?

Dmitry Shapiro: Yeah, well, first of all, all SAS applications are now implementing AI features. And so whatever apps the enterprises are using now have AI built in. Slack has AI, Google Docs has AI, you name it, it's got AI. Um, so AI is going to be everywhere. I mean, generative AI is sort of going to be like Clippy, for those old enough to remember, that's always sort of around.

and ready to help you, much more intelligent than Clippy. Um, but this is not what we're talking about, meaning that will exist no matter what. Um, what we're seeing in market now is that while organizations tend to start because, you know, sort of ChatGPT was the first big gen AI thing that gave us all sort of the mental model of what is AI.

organizations or individuals, you know, that they're saying these organizations start off thinking about a eyes like, oh, a eyes like a chat bot and great. Maybe my organization can build a chat bot that like knows about the organization. I can chat with it and I can ask a question. What's our policy on this?

Or please help me edit this thing I've written or whatever. Right. And so that's sort of like the mental model that people have. Um, and, uh, and then whatever they evolve into, uh, yeah, that's cool. But really what we're trying to do is we're trying to be more productive as an organization, right? Like all enterprises are always trying to be more competitive.

more efficient, more productive, right, more, more flexible, move quicker, be able to take opportunities. These are all like competitive advantages of any type of a business, right? And so very quickly after this, like thinking about AI was sort of like a chatbot. Um, they start to realize, oh no, we can actually sort of, uh, look at our processes, at all the things our organization does, sales, marketing, HR, whatever, uh, through this new lens of like, of the modern world, that there is now software that's intelligent, that can do things that we used to do manually.

Like it used to be that an email came in and a human took it and entered it into a CRM system. This doesn't make sense to do anymore. This is bad. Don't do that. Don't do that, right? Uh, or you had project managers that watched a spreadsheet and then talked to users, you know, to, to, to teammates and collected information from them and then sort of processed it and created reports back out to the organization of, You know, timeframes and misalignment and well, you don't need to do that anymore.

AI does a marvelous job of doing that kind of stuff. Okay, great. So like this last generation of tooling that enterprises have, you know, used, let's say these 130 SaaS products. Uh, now if you look at it as like a blank sheet of paper and say, well, Knowing what I know now, is this the way I would architect my enterprise tech stack?

For many organizations, the answer is of course not. It's like a bunch of things we don't need anymore because AI is just going to do them. And then a bunch of things, because AI is assisting, we really wanted to work very differently than this. Like, out of that CRM system, all we needed were just a couple of functions.

We had to buy the whole CRM system. But what we really need is a couple of functions. And this afternoon, You know, somebody that's learned to use Mind Studio can build us a custom application that has those functions and, and, and then all these other goodies that are made. So it's like a, uh, yes, I believe what, again, we're seeing kind of the early stages of this, but, but I'm certain that, uh, we, we have an inflection point now.

of enterprise tooling, um, that, uh, is going to refactor all of those old, uh, uh, patterns. Complete refresh for enterprises of, of their, of their tech stack is absolutely warranted. Those that do it slowly are going to be at a disadvantage to those that do it quickly. 

Richie Cotton: Okay. So, um, it's interesting. It sounds like what you're advocating for is having more smaller applications that are tailored to specific use cases.

Okay. Uh, so, uh, in that case, um, suppose you're organizing, your CEO goes, okay, this is a cool idea. Let's change our tech stack. Where do you 

Dmitry Shapiro: begin with that? It depends. Uh, again, we're seeing organizations take both of these paths. Uh, oftentimes it's, uh, bottom up where somebody, an HR manager, At a government agency, uh, learned about MindStudio and had a bunch of these resumes that are coming in and said we could just automate that.

That just sort of happened bottom line. Or a project manager at a, you know, giant tech company, uh, Learned to use MindStudio, brought it in with a bunch of project management stuff, impressed two other project managers. Now they're a team walking around using an AI magic wand, and just automating things and fixing things and optimizing things throughout this giant organization, right?

Or in other organizations, you know, it happens from the head of IT. You know, giant retail brand, global head of IT shows up, says, look, we're seeing this strategically. We're going to refactor the whole org, doing it, and we're just doing it top down. Or had a product that a 150 year old, you know, company shows up and says, listen, we're not early adopters.

of any tech, but we're going to be early adopters of AI, because it's a no brainer, we're going to do that more. And so the product is here because it's foundational to everything. Um, so it's kind of all over the place, and it kind of doesn't matter. That's the other sort of thing I think is important to understand, you know, for people watching, listening.

Um, we kind of skipped this thing that maybe they know what MindStudio is, but like, MindStudio is this, is this platform that allows anyone to do anything. Uh, like literally anyone in the organization, you don't need technical skills. Anyone can learn to use it really quickly. Allows anyone to build AI powered applications to use any and all of the models and create any kind of applications for, you know, employees to use or automations that run on their own or like many other things.

Um, and again, because the dynamics are so different, because it takes 15 minutes to maybe a couple of hours for really, really sophisticated incredible applications for one non technical person to do. The best people to do the work are the people that are actually doing the work now. Like, they're building kind of the digital themselves.

Like, today, I process all these resumes and then do all of this stuff. Or, I do all of these steps in project management. I should be the one that articulates for the AI, here are all the steps I want you to take to do that, and here's how I want you to do it. And then push a button and it's done. So they should replicate themselves.

Senior person trying to do it comes at it from a very different perspective. Now, in some cases, that's the valuable way because these people are too close to the ground. We don't need to redo them. We need to think strategically and replace them, move them somewhere else. And they might have a hard time doing it.

So again, just depends on sort of dynamics in the culture. of the organization of how it happened. 

Richie Cotton: Okay. Um, so, um, it sounds like it doesn't really matter whether you start at the bottom or start at the top, just, uh, pick something useful. So, um, I'd love to get into, um, uh, like who, who's involved in this and how roles are changing in a moment, but before we get to that, um, I imagine every business has lots of processes that could be automated.

How do you decide like which one to automate first? 

Dmitry Shapiro: So, uh, I'll answer you from standpoint of, uh, best practices, meaning my opinion, uh, uh, I think is irrelevant. And also, I think it's different organization by organization. But the general patterns that we see sort of bubble up, uh, goes this, uh, you start by automating the lowest hanging fruit.

What things are obvious you can just right now automate, just automate them. But again, you can do it in the next 15 minutes to a couple of hours and just automate the thing. And, and then your business has changed. And, right, and then you can look at now how is the business operating. And then you might realize, well, actually, that's not the thing to automate.

I should actually automate that thing, and that'll make this thing sort of irrelevant. So you can do that, and because again, it's so quick to do it. You're not blocked by engineering. You're not blocked by spinning up big projects or costs or weird stuff. Like, you can just play around with things. Um, there's sort of like one approach, which is low hanging fruit and rapidly play around and like see what works combination, right?

Um, another approach, uh, is, uh, again, kind of a more strategic approach. And some people are like that, where they're like, look, this is cool. That's not the right way to do it. The right way to do it is to architect everything to work sort of together in an intelligent way. And yes, we can build some prototypes and maybe you guys do that, but really anything production, we want to step back.

And now it's a great opportunity for us to rethink how our organization works. Like, do we have a marketing department? in a sales department? Like traditionally, the answer is of course. With AI, the answer is unclear. In fact, kind of points to the there's not much difference in marketing and sales. Talk communication.

That's fascinating, 

Richie Cotton: and I think that's quite a radical statement. A lot of people are going to be shocked at the idea that maybe you don't need separate sales and marketing teams. So, um, okay, so once you've gone through this path of automation, what what do these commercial teams look like? 

Dmitry Shapiro: Yeah, so again, all of this is, uh, depends on the organization and how their teams operate.

So I'm going to talk in sort of general patterns that we see. Um, so first of all, sales, uh, and marketing tend to, uh, lead often in organizations like who's looking for AI. It's oftentimes like the operations people that think there's like data analysis stuff that they might be able to get, like business intelligence or some automation stuff.

And it's sort of revenue generating functions, right, because everybody wants more revenue. Uh, and so sales and marketing are well represented, although all other functions are as well, HR, finance, whatever. Um, Sales, uh, again, let's separate them, but then we'll bring them back together, sales and marketing, right?

Marketing is a function of product packaging and messaging and, and sort of articulating the brand and, and, you know, creating, uh, segmentation of potential customers and communication to them, like all that stuff, a lot of things, marketing, right? And then all the growth stuff and all of that, uh, in sales is this thing that people do.

Typically better than a website. Why? Why can't the website sell just as well as a salesperson can? It's because humans can read other humans and can see, you're not understanding what I'm trying to tell you. And so I need to paraphrase it, or you're asking questions and I'm answering them, but you're not asking the right questions.

You're not looking at it from the right angle. I need to change your angle. A website has a really hard time doing that. It's like solutions for enterprises, for solopreneurs, whatever. It's not, it's static. Human is better. Um, but AI can replicate that. And so you can create now, um, uh, again, and that sort of gets manifested in a number of ways.

Thanks. Uh, a common way is like everybody's recording their communications and transcribing them using things like Otter and Firefly and all that stuff, right? Mostly they go into some buckets or just sit there and they're there for reference purposes. But now you can easily build an AI that watches that bucket and and then takes that.

And then sort of asks all the questions that the salesperson doesn't have time to ask. Is this customer really understanding this, or are they not? What are their objections that they've stated explicitly, but what might be their other objections that have not been stated? Sort of read between the lines.

Are they not asking the questions that they should have asked, etc. And then, by the way, can, can do the job of marketing. This is where they get connected. Where it says, I'm just going to create the materials for this, customer that you can just send them with a push of a button. Or I can send them for you automatically.

And so every salesperson simply wants to talk to other people. That's what salespeople like to do. They want to be in front of customers closing. They don't want to do all the follow up work and all the deep thinking about each step. AI can do all of that. And that's extraordinarily valuable. It can take, like, you know, there's always a scale, you know, caliber of salespeople, experience and talent, and like great salespeople massively outperform average salespeople.

Well, AI can take average salespeople and make them great. And that's, that's crazy. That's incredibly valuable. The ROI on that is spectacular. You know, be in front of more people because you get all this automation, you get all of this intelligence, you get custom content for every prospect, because every prospect is different.

Like I've built now three venture backed companies and, you know, I worked at Google for, for four years, I was CTO of, of MySpace Music, so I've been around the block. And so I've seen and, and cringed the countless. Uh, exercises that marketing teams do in defining these personas. There is no such person as any of those personas.

Those are all crude caricatures of bundles of humans. That's not how humans are. AI doesn't work in personas. AI can meet each human where they are and lead them to the realization that they need your product. And here's how they can use it. Because all sales is, great sales, is education. The ability to educate the student to see that your solution is obviously awesome.

Take my money. AI can do amazing jobs. Yeah, and by the way, sort of many other things in sales that are there. But that one's super exciting, crazy. 

Richie Cotton: I love that, um, these ideas that it's really, really common problems in sales and marketing to defining like who are your customers or what objections do they have to buying your product or solution.

These can be tailored and just because you probably have that information somewhere in the organization, somewhere in a transcript with a conversation with someone and yet it's not. The rest of the sales team doesn't know about it, so you can sort of bring 

Dmitry Shapiro: that forward. We compile these things called frequently asked questions.

How do we get these frequently asked questions? We rely on things we hear as a marketing group and sales people out in the field to report these frequently asked questions. Because everything's getting transcribed, we know exactly what the questions are that are being frequently asked. But not just frequently asked questions, in, in what combination are they frequently asked together that show that these are types of wrong mental models that people have, that our prospects have, that we need to get them to see it in this way?

That's not like one frequently asked questions. A bunch of questions is a bundle together. And if you see that, chances are the prospect is seeing it from this angle and they need to see it from that angle. this angle. And so you can now do that. And again, human marketing teams have a really hard time thinking in this like multi dimensional way.

They just think, oh, somebody asked this question. Well, that's not a context. And so it's a frequently asked question, but answering it actually doesn't do much good because that wasn't really the question. They had a different question, they just asked it on this one. 

Richie Cotton: In order to take advantage of, um, like creating assistants, it seems like, um, there at least needs to be a mindset shift and probably some education for a lot of, um, these sales and marketing people who maybe don't have that much of a background in AI.

What do you need to know in order to get started creating assistants? 

Dmitry Shapiro: Uh, you need to watch, uh, one or more YouTube tutorials. We now have tens of thousands of people that have learned to do this by watching an 18 minute YouTube tutorial, and then we've got like an hour and then two hour YouTube tutorial.

Some people come to our live webinars that we host. Some people get really serious about it and come to an eight hour certification class where they really dive into it. So you need to get educated. Um, you don't need any technical skills whatsoever, meaning, uh, you need to know how to turn on a computer and use a mouse and things like that.

Like you need to be generally, uh, uh, able to do computing, but that's it. You don't need to know how to write code. Never, you don't need to know anything about code. The skill set that's needed is for you to know the job that you want the AI to do. And if you know that job, and once you've learned to use MindStudio, you can very quickly create these multi step workflows, we call them, where you like put these components on a screen and connect them together and configure them.

And that's how you tell the AI how you want it to behave. Input, output, what to do in the middle. Um, the real skill is just, uh, uh, clear thinking of like, could you describe that to another intelligent human, how to do the job? If you can, solid, that's all you really need. And just basic, you know, language, you, by the way, you can misspell everything in like, in regular code, you put a comma out of place, the whole thing breaks.

In AI, you could misspell every word and the whole thing works. There might be some ambiguity in something, so you got to be careful there, you know, but like, but, uh, misspellings and typos don't get in the way. 

Richie Cotton: I have to say I'm a lousy typist, so, uh, , so it's been a, a challenge writing code, but um, yeah, uh, that, that sound incredibly useful.

Alright, so, um, a lot of the stuff you're describing about, um, just building workflows to sort of, uh, perform task, it reminds me a lot of, uh, of Zapier, which is designed to sort of connect different applications together. Can you talk me through, um, is mine studio a similar idea or are there some differences there?

Dmitry Shapiro: Great question. Uh, the most common way that people tend to describe MindStudio in sort of it's, it's like this for that, is it's like Zapier for AI. So this comparison is made frequently and so it makes sense that you saw that. Uh, in reality, uh, it tends to get used along with Zapier or Make or any of these other things.

Why? Because, uh, those types of systems, uh, one have been around for a long time and have thousands of integrations with various applications. They have hooks into any of these applications that the enterprise might be using. And, and you can very easily take Zappy or Make or any of these other things and basically create these triggers, as they're called, and some sort of event listeners, as they'd be called.

in software development, that if something happens, if an email comes in, trigger in AI, or if something changes on the web, trigger AI, or if something gets saved to this directory, or a form comes in from our e commerce form, trigger in AI. And so they make it really easy to create those triggers. And then we make it really easy to build these, like, really sophisticated things that those things can trigger.

And, and do things, and we are much better at that than Zapier is, and they are much better at triggers than we are. We don't even do any triggering this time. So you have to use something like that to trigger us, or write custom functions or whatever. But, but the concept is similar. And by the way, much easier to use.

Zapier is extraordinarily powerful. It's awesome. Uh, as is Make, which I think is a little bit easier to learn than Zapier, Extraordinaire. But MindStudio is radically simpler, like, to learn to use than either of those platforms. Dramatically less time, less mental effort to be able to become a master of doing that.

Richie Cotton: All right. So, um, I'm gonna, um, I think there are some, um, hesitations many organizations have from the using AI. And so one thing you mentioned was that, cause basically anyone can get started creating these AI assistants, AI applications. That means you could potentially end up with hundreds or thousands of these within an organization.

And it's going to be very difficult to track what's going on. So I think that might scare a few, um, people involved in governance. So, uh, how, are there any guidelines you can make to make sure that a particular assistant conforms to certain, has, to certain properties or rules? Or how do you make sure that, um, you don't have some rogue AIs within your organization?

Dmitry Shapiro: Uh, so, um, the way Mind Studio, uh, works is anyone in the organization can create an account and create a workspace. Uh, and that workspace could just be for themselves or they can invite other people to the workspace and make it a team workspace, right? Uh, and then an enterprise can sort of go in and take a bunch of workspaces, put them together and say, okay, you're all part of an organization together, right?

So, uh, you absolutely, theoretically could in your enterprise have some person that uses Mind Studio, doesn't tell anybody that they're using Mind Studio and is doing stuff. Probably, you know, breaking the rules, but again, generally really easy for an enterprise to know all the AIs that exist in the enterprise.

who they are created by, how they are being used, have analytics for them, um, and, um, coming soon, AI powered business intelligence that helps you understand not just analytics, but like gives you insights of like, Oh, you could be doing this a lot better. Or like this thing you're doing, you thought you sort of created the right business application for this, uh, uh, uh, job function here, but they're still doing a bunch of like ad hoc stuff.

You should add this capability to this application and then they won't have to do this ad hoc stuff. Like, you can upgrade your applications like that stuff. That's coming soon. Um, but yeah, so an organization can, can manage users, can log and archive and enforce policies, um, uh, you know, route to compliance systems, grant and revoke rights, obviously know about all the AIs, et cetera.

Richie Cotton: Okay. Um, and I guess, uh, the next biggest concern is going to be around, uh, privacy of data and the security of, um, of these AIs. So, uh, What sort of worries do you think managers are going to have around this, and how can they be dealt with? 

Dmitry Shapiro: Yeah, so obviously we're already at scale, so seeing a lot of these.

Um, there, um, there's a whole spectrum of, of enterprises. There are some enterprises that are perfectly fine, um, just allowing their employees to use ChachiBT, you know, getting their accounts and using it, not monitoring anything, not logging anything, not sort of imposing any rules on it. Um, uh, and on other extreme, there are enterprises, you know, oftentimes in like regulated industries have compliance mandates or highly sensitive information where they are, they don't trust the cloud at all.

And even if the cloud is Google and Microsoft and like reputable companies, they just can't do it. They can't run inference on any of these models. And so they want to have a private model, you know, in house on premise or in some private isolated cloud. Um, so it really depends on the organization, uh, uh, and, and those are the issues typically is like.

Are you going to get, like, data leakage from, uh, the enterprise and to, like, the model and, like, either publicly leak out and your data sort of train the model and the model could respond to other users with your data, uh, uh, you know, which should not happen if you believe the way these models are architected and the terms of service of calling these models via API, which is the way we do it, that they explicitly say we don't use any of this data for training, like, it never happens.

It's training? Um, uh, but again, so I, so that's like, that's like one category of, of, uh, you know, privacy concern is like, is data going to leak out, uh, the, uh, um, uh, from the enterprise into the model and, and then others are, you know, uh. different, obviously, like sort of granular rights to enterprise users of what data sources do they have access to and like what kinds of things they can build and like those kinds of questions.

Um, and look again, our, our answers to both of these is you get to choose, right? You get to choose whether your organization or this team or this person get to use publicly available models or not. We support them. You can use them. Or if they must use your own model. Um, again, we can enforce that. Um, and on the other side, on data sources and all of that, again, today, that, that is on a per AI basis, so we don't have sort of organizational data sources yet.

So that's not a problem. We will soon have those and you'll have granular rights of being able to. Uh, to specify, you know, who has access to, to do what with them. 

Richie Cotton: Okay. So I guess in the meantime, um, the control of like who has access to what data that's going to live on the, I guess the data warehouse side, right?

Exactly. 

Dmitry Shapiro: Yeah. Does, does the user creating the AI have the ability to call the relational database or this data source? And the answer is no. Problem solved, the AI can't do it. 

Richie Cotton: I guess the other question is around, um, maintenance of these things. So if these things are very cheap to build, you can get lots of them.

A lot of these AI's are probably going to be quite disposable. How do you like manage all these like, um, AI's that may or may not be relevant in six months time? 

Dmitry Shapiro: Too early. At this moment, like that's clearly going to be a thing. Uh, I'm sure there's gonna be multiple patterns that people take. We will certainly, uh, again, via this business intelligence layer that I was just telling you about, be able to detect, uh, things that don't make sense.

Uh, you know, things that aren't being used or being used inefficiently or whatever, and sort of bubble them up and allow the architects of these AIs to refactor them and change the way they work. Um, uh, and then I'm certain that, uh, earlier in the conversation, as we were talking about like, what's the right approach, should it be like bottom up, everybody creates the thing that solves their own problem, or does sort of, do you do it strategically, where you really architect the thing?

Again, today, it's mostly the, the, the first where it's bottoms up and, and, and the sort of top down comes in later. I'm certain that long term that's going to shift where, uh, enterprises will be very thoughtful about not creating tons of these AIs and we'll create, uh, you know, some limited number. That are meant to work together and the architect is aware of, you know, architects are aware of this.

Richie Cotton: Okay. Yeah. I guess, um, my mental model for this is probably like dashboards, like companies often have thousands and thousands of dashboards. But there's maybe only a few that are important and regularly used and a lot of them are a bit more. Disposable. Okay. Maybe, maybe we'll get you back in a year or two to see how the problems.

All right. Um, so I guess, uh, the other big concern organizations have, um, with particularly with generative AI is the cost of it. Like, um, I think a lot of companies built prototypes last year and they're sort of suddenly realizing you put these things in production, they're quite expensive. So, um, can you talk me through, um, how you manage costs of generative AI?

Dmitry Shapiro: So, first of all, the way our platform works and our business model works is that the customers are responsible for paying for the usage, it's metered usage. And so depending on what kinds of AIs you build, how many people use them, and sort of how much work the AI has to do, how much you have to spend on inference or, or calling vector databases or like these various other things, um, You, you, you one, you can understand those costs.

We make it sort of easy for you to itemize them and see, uh, them accruing. And so you can see how much your AI is costing. Um, uh, and then we also allow you to make choices in architecture of the ai, of, uh, because look, different models, uh, cost different amounts of money for usage, right? It's a very broad spectrum.

Uh, you know, a lot of these large language models from OpenAI and DropIt, you know, Google, et cetera, um, uh, they themselves have a great, uh, you know, spectrum of cost from very inexpensive to very expensive. Um, people that are just starting out to build AIs tend to think, well, I should just use the most expensive model at all times and pay the price, uh, and, and that's going to work best, but in, in practice, that's actually not even remotely true.

And so, again, once you get a little bit more sophisticated in creating these AIs and you're optimizing for cost and latency and all of that, you start to realize that the right way to architect AIs is to, uh, very rarely call large models and allow smaller models to do the job as part of this workflow.

And then typically, at the very end, to take all this data that's been sort of compiled and, you know, And, and, and created by smaller models past that to a larger model that can take and sort of package it all in a nice little package that you like to read as a human, right? And, and so, uh, cost control in AI is, is, is, is beyond that.

That's a function of, of latency. Again, there's a very broad spectrum of how fast or slow these things are for various types of tasks. That's a variable as well. So, uh, so you want it to be highly performant for the task. You want it to have a quality level that's at least good enough for the task. And then you might spend a tremendous amount more and slow it down to get another 1 percent that you might not be able to notice in most cases.

You might choose not to do that. Yeah. Uh, and then, uh, so, you know, it's cost, quality and, and, and performance. So those are the three factors that typically are used. And again, we make it easy to be able to use a bunch of smaller models. By the way, again, we are seeing it already in market and we feel certain that as AI evolves, large models, uh, are not going to be the dominant thing that, that enterprises use at all.

They're going to be smaller models, more and more are going to be private because this is important for them. They can train them with their own data. They can retrain them with their own data. We're about to make that dramatically easier. Uh, to do, uh, where you can like, we, we will help you automatically build your own private model and train it and retrain it and connect it and all that, not here yet, but, but, you know, coming shortly, um, because there is already a need and we're certain there's going to be more of a need for that.

And, and so again, the costs when you architect these things in the right way can be mitigated. And certainly the ROI, like if you're not just building toys, you're building things that are automating and things like that. Well, the ROI is amazing because no matter what, AIs are cheaper than humans. 

Richie Cotton: Okay. Um, that's interesting because a lot of people tend to ask me like, well, what's the best large language model?

And I'm like, well, go and look at the leaderboards. But actually the answer isn't, um, a single model. If you're going to build an application, you probably want lots of different models and you want the simplest model that gives you a good result for any given task. 

Dmitry Shapiro: Yeah, absolutely. Again, I think there's a lot of misconception because of how quickly AI, generative AI showed up and chat GPT showed up where, where people are like, Oh, well, GPT 5 is going to have a larger context window.

And so I'll just be able to give it like all my documents and say, figure it out. And, and it's just going to do that. Then it's going to call like a bunch of functions and research the web and all that. And again, while conceptually it's a cool idea and, and. Uh, these context windows are getting bigger and bigger and also all of that is true.

But that's not the right way to get the job done, especially in an enterprise. Because you have no control there of that, of like, go do my job. That doesn't make sense. It doesn't make sense for humans to prompt large language models either. That doesn't make sense. You know, computing didn't take off until we got Windows.

DOS didn't work for people. Most people have never learned to use Google search operators. And you can put commands into the Google query string that, that allow you to have mastery of Google. Nobody uses them. Yeah. And, and so, uh, to think that the employees are going to like type prompts and copy and paste prompts, this is just silliness.

It's just silliness. Uh, that the right way to do it is for people who want to take a process and make it more efficient to learn how this thing works and sit down, put the right framework in place to make that process efficient. And when you look at it in that way, there's very little use for large models.

Again, they're just generally needed at the end, where a lot of this sort of stuff has been done, you just want to package it in some comprehensive thing, and there's like a bunch of pieces that they can put together, and they're great at putting a bunch of pieces together. And you shouldn't have them making decisions, certainly everybody understands like this concept of hallucinations, and this is what they're specifically good at, that's why we call them generative, they're creative.

They make shit up, and that's wonderful, and in many cases that's good, but in most cases in enterprise, that's not good. And so you want to give them as little sort of leeway to be able to get creative. You really want them to be not creative, but like really good at putting things together. 

Richie Cotton: Okay, uh, this is, uh, quite a refreshing contrast having as all people like, Oh, well, you know, once we have GPDA, you know, it's going to be better than humans, everything we've got AGI and it's going to take over everything.

But actually you're saying the future is actually going to be lots of smaller models working in tandem. Okay. Uh, so in that case, are there any limitations to what assistants can do at the moment? 

Dmitry Shapiro: Hard to answer the question. The answer is obviously yes. Uh, what those limitations are, I think depends on the dimensions, right?

If your AI does not have access to some kind of data, it could need, and you don't have access to that data, uh, you know, the AI, you could ask it to make up synthetic data to do stuff, but it's like, it's like that. There are those kinds of limitations. If the AI needs to connect to third party services, uh, and, uh, whatever, for whatever reason, That connection is not available.

So again, sort of like in the enterprise, there are limitations that can be created. Uh, obviously the capabilities of these generative AIs continue to evolve. And so we now have like multi modality that AIs can sort of understand images and output video and, you know, do text to video and video to text and like vice versa.

Um, And so we believe that layer will continue to sort of accelerate and in its capabilities. And so hard to say what like A. I. s can do. They can do amazing things now. But again, I think the From an enterprise standpoint, the right way to think about this is not be fascinated by like all of these new amazing things any one model can do, because that's not the right way to use AI anyway, but to, um, again, really realize that like anything you want to do today, the current AIs can do.

Like pretty much anything that anybody is doing in any enterprise that requires humans to do it, AIs can help do. Now if you're whatever, a researcher and you're doing a bunch of, you know, chemistry, a large language model might be of some help, but that's not the thing you need. Now there are models that are very good at chemistry and so you might need those models to be able to do that.

And then you could create automations where you have those models driving. You know, manufacturing processes and other model supervising and sort of analyzing, like you can do crazy, um, uh, again, because it's multi step, you can connect these systems together. Uh, you can do incredible automation and, and sort of IOT integration and, and obviously connection to, um, uh, to, to, to the public internet and, and partner services and countless things.

So. 

Richie Cotton: Okay. Um, so it sounds like, um, there's possibilities for using this almost everywhere. I know it's still fairly early days, but you said you've got 50, 000 of these, um, uh, systems created already. So, um, do you have any examples of success stories where. Companies or people have built these things and they've had some big win.

Dmitry Shapiro: Well, uh, lots of stories of, again, sort of real digital transformation with incredible ROI, ability to take people out of the loop, to do functions that they never liked doing anyway, and now that function just runs automatically. The organization never has to worry about it. The quality goes up, cost goes down, people can do other things, like many examples.

Um, uh, and then whatever, many other types of examples as well, uh, one is a website academicinsightlab. org are, um, Kimberly and Jessica, they're both, uh, PhDs that, you know, last year, their business was a consulting business where they advise universities on building. you know, Ph. D. programs and operating.

So working with, you know, districts and faculty, staff, etc. Um, but since then, since they found MindStudio, uh, they've taken like all of this work, their consulting work, and transformed it into over 80 AI applications. that are made for academics to use to do this thing that Kimberly and Jessica used to do for them.

Well, now it's self service where the AI can do it for them. And so the cost is lowered dramatically, the productivity is raised greatly, and they, Jessica and Kimberly, no longer sell their time. They sell access to these 80 AIs as a bundle to this universe and now have a multi million dollar a year business.

To completely non technical people, neither have ever written any code, are now have this enterprise. 

Richie Cotton: Yeah, that's absolutely fascinating. Um, it's just swapping your own, uh, or your own consultancy, um, capabilities for an AI, outsourcing yourself. And charging 

Dmitry Shapiro: access for 

Richie Cotton: it. 

Dmitry Shapiro: That's the right way to do it.

That's the right way to scale yourself. People are afraid that AI is going to take their job. I propose to them they're thinking about it completely the wrong way. They should say, ai, please take my job. But they are the ones that created the ai, so it's their AI that can take their job. And that's fantastic because why do you wanna work?

You can sleep, let the AI do the work. You, you are blocked. Serially AI can scale, you know, horizontally and, and in parallel service customers like, uh, Jessica and Kimberly could never service the number. of, you know, people that their AIs can serve. Um, yeah, so I think it's just a, it's an incredible opportunity for anyone who is, uh, you know, an expert in something that, that sells their time, sells their expertise.

Um, they should absolutely consider simply creating AIs that do that. And then charging people to get access to these AIs. It's a marvelous approach. 

Richie Cotton: That sounds like it's going to radically upend the consultancy industry. I'm sort of McKinsey's and Young's and all that sort of thing. Yeah, certainly. Um, all right.

So, uh, exciting times, uh, just to wrap up, do you have any final advice for organizations wanting to make use of generative AI? 

Dmitry Shapiro: Yeah, don't get, uh, don't get fascinated by the shiny objects and, and, you know, waste your time, you know, uh, trying to keep up with innovation and AI and like, is it, is Claude Opus better than GPT 4 Turbo?

Last week it was, and now Mistral is better. Like any of this stuff, I mean, it's cool if you're fascinated by it. You know, awesome, but that's not the productive thing. Uh, don't spend your time on that. Learn to create automations using AI. Learn to use a product like MindStudio and, and look at your organization and transform it now.

Like, don't wait, because there's no point waiting. There are no things that you're waiting for. You can do all of that right now. And all of that, something can get better and better and better as those models get better. Your applications can get better because you can easily switch to the new model.

Like we make upgrading trivial easy, we'll pick a thing and you've upgraded, congratulations, woohoo, soon it'll be automatic if you want, you know, but for now you got to do it manually. Um, and so yeah, now it's a, it's a, it's a sprint, uh, and I think it's, it's ultimately clear that, uh, success in the age of AI, uh, meaning competitive success, competitive dynamics in the age of AI.

are all about, it used to be about like, what's your team like, who are your people, and like that was like the most important thing because humans have to think and do whatever. That's not the case anymore, I mean, becoming less and less the case. Competitive advantage in the age of AI is, are you, are you faster than other companies in being able to, to take advantage of opportunities?

Are you more intelligent than they are? Can you see things that they can't see that you can take advantage of? Um, Are you efficient in how you leverage your resources, financial and human money is expensive now, and so everybody is like slimming down and trying to get more efficient. Uh, or can you squeeze out productivity and sort of more value per employee that you have?

Can you keep those employees happier? You know, these are all competitive dynamics, and all of those are addressable right now. with the current generation of generative AI plus MindStudio. And organizations that take advantage of those dynamics sooner are going to have an unfair advantage over those that are going to be late adopters.

Certainly, I can't imagine, uh, sort of non adopters, like, those will be seen, you know, have the fate of dinosaurs, like, unless you've sort of optimized for these, you know, dynamics that I'm enumerating as an organization, if you're still doing things the old fashioned way. Well, I don't know. As they say in the South, bless your heart.

Uh, it's uh, it's not a very good approach. 

Richie Cotton: Alright, that's a pretty, uh, harsh call to action there. It's like you've got to get automating things, start adopting this stuff now, or, you know, go in the way of the dinosaurs. Um, alright, super. Uh, thank you so much for your time, Dimitri. Yeah, I had a great time.