This is an interview with Dmitry Shapiro, previously working at Google and CTO of MySpace Music. Now, Dmitry is building something super ambitious with the goal of democratizing artificial intelligence. We talk about his platform, YouAI and Mindstudio, but also give super applicable tips to build better AI apps, such as the model selection, prompting, using RAG, and more.
We also talk a lot about the user perspective, democratizing AI, and the future of AI. Dmitry also has another goal of indexing the minds of everyone. And I for sure talked about that in this episode.
This episode is for everyone. Builders or users. I tried to orient the discussion towards applicable tips but also to make us think about the future of AI and its potential. I hope you enjoy it!
Dmitry Shapiro: [00:00:00] And today, ChatGPT is great for nerds, but for the average consumer, you know, they don't have time for that. They never learned Google search operators, which is a really valuable skill to have. And we nerds learned it and we wield it, but they're not going to learn it. And so they need simple applications, just like you have on mobile.
Dmitry Shapiro: The people that do index their mind will be able to leverage AI radically better than people that do not, that still have to type into ChatGPT.
Louis Bouchard: This is an interview with Dmitry Shapiro, previously working at Google and CTO of MySpace Music. Now, Dmitry is building something super ambitious with the goal of democratizing artificial intelligence. So that's pretty much the main topic of this episode. We talk a lot about his platform, MindStudio, but also give super applicable tips to build better AI apps, such as the model selection, prompting, and more.
Louis Bouchard: [00:01:00] We also talk a lot about the user perspective and the future of AI. Dmitry also has another goal of indexing the mind of everyone. And I for sure talked about that in this episode. I hope you enjoy it. And if you do, please don't forget to leave 5 star review. a like and comment what you think it helps my work a lot and tells me if you enjoyed this episode now let's dive in.
Louis Bouchard: Well i first wanted to to ask you about YouAI basically what inspired you to create The MindStudio and everything around YouAI.
Dmitry Shapiro: I think I, like, you know, so many others now am sort of profoundly inspired by these new capabilities that we have. In the form of, of, you know, these consumer accessible, generative AI models and the ability to leverage them to kind of transform all facets of our, of our lives, [00:02:00] whether they're professional, personal, creative, whatever.
Dmitry Shapiro: And on one hand, it's amazing that I can simply, you know, go to ChatGPT on my phone. And start typing, you know, you are a blog post copywriter with 20 years of experience. Please write me a blog post about how AI will disrupt medicine. And it responds, and it writes an amazing blog post on how AI will disrupt medicine.
Dmitry Shapiro: But that just seems like the wrong way to use the power of these types of new technologies. And so our insight was that all of these foundation models, whether they are language models or image diffusion or, you know, video models that are coming online now, code, whatever, that all of these foundation models should be treated as backend services.
Dmitry Shapiro: And not accessed by users directly, meaning not have front end interfaces. They do. [00:03:00] And it's cool that they do. And for us nerds, it's, it's really cool that we can sort of go and use these command line interfaces to interact directly with this, I call it the intelligence layer. That I can interact with it directly, that the average sort of end user can't be expected to think like a prompt engineer, understand the nuances of these models, choose between these models, and then do a lot of typing and mostly on their phone, because that's how we compute these days, right?
Dmitry Shapiro: We're mobile and so like, it just seemed like all of that was a problem. And so anyway, so that was the idea for MindStudio is that there's a need for an abstraction layer to the. Intelligence layer, we call that abstraction layer, the application layer, and MindStudio is an integrated development environment and sort of a platform that allows anyone to show up and build AI powered applications that consumers can then engage with directly [00:04:00] and not have to worry about prompt engineering and things like that.
Dmitry Shapiro: So prompt engineers do the prompt engineering and package it up as a front end interface for consumers or business people to use, right? And so today we've got thousands now of these prompt engineers that have signed up and are creating these AIs we have over 5000 AIs that have already been created.
Dmitry Shapiro: You can find them, you know, on our website because these are just web apps. If you just go to YouAI. Y o u A i . ai you can say explore apps and you can see what other people are building. It's extraordinarily diverse already. There are many, many categories for large enterprises, for small businesses, for consumers, personal use, parenting, whatever.
Louis Bouchard: And it's all built by those people that we call prompt engineers, which is basically a skill that I, from what I understand, you would compare [00:05:00] as a backend engineer or like. back end developer. It's, it's a skill that would require someone to, to learn and practice, which you believe that is not important for the vast majority of people to, to understand.
Dmitry Shapiro: Right. Yeah. We believe that the average You know, end user, whether they are a consumer or a business user, should not have to do a lot of typing and should not have to understand how to think like a prompt engineer. Yeah. That, that doesn't make sense. And so, yeah, this, this term we use prompt engineer. The skillset that's required is not any knowledge of coding.
Dmitry Shapiro: In fact, sort of knowledge of any coding doesn't necessarily help you at all here. The skill you need is the ability to natural language is the ability to articulate clearly what you want the system to do, how you want it to behave, how you want it to [00:06:00] format responses, what its constraints are. And so it's, it's the ability to write a spec, if you will.
Dmitry Shapiro: And for this thing, and if you do that, then the intelligence layer does the rest.
Louis Bouchard: Basically, I've seen that, and you just mentioned it, that you have over a thousand applications built for, from, from such prompt engineers to facilitate the discussion with models like GPT 4. For example, and do you think that having so many applications to do specific things go around, go around the final goal of, for example, having one general foundational model where, where you can just ask it any question it will answer based on.
Louis Bouchard: Your own preferences, because I can compare it, for example, as the early days in image classification, where we, we would just train a model for, for finding cats and images, another one for finding [00:07:00] boats, another one for all different objects. And now we can. We can even use ChatGPT to do that now, but we, we can just have one foundational model that, that can classify all images.
Louis Bouchard: So don't you feel like trying to create super specific AIs and thousands of them go against like a, an optimal goal?
Dmitry Shapiro: No. So it is clear that these foundation models one are becoming multimodal. And two are extremely broad in their capabilities, right? Can do many, many so, so they are actually not the problem.
Dmitry Shapiro: The problem is the interface between the model and the human. For the human to be able to articulate to the model what needs to get done. It requires the human to say many words. And to say them in the right way. To explain it. Because again, otherwise everything is ambiguous. And so [00:08:00] dealing with that ambiguity and the need to craft the right prompts to be able to then get out of the model, what you want, we believe can't be put on the shoulders of regular end users.
Dmitry Shapiro: Like there's a lot of reporting that the traffic, you know, the, the usage. Of ChatGPT for the last three months has actually been declining, and that might surprise a bunch of people because it looks like we're all using it all day long. And that's amazing. And so I think the problem there is sort of a crossing the chasm issue for, for those that are familiar with the book, you know, lots of things get early adoption, but for them to get massive adoption.
Dmitry Shapiro: They need to be for more than just nerds. And today, ChatGPT is great for nerds, but for the average consumer, you know, they don't have time for that. They never learned Google Search Operators, which is a really valuable skill to have, and we nerds learned it, and we wield it, but they're not going to learn it.
Dmitry Shapiro: And so they need simple applications, just like [00:09:00] you have on mobile. On mobile, there's a concept of there's an app for that. And that's a really powerful concept. And instead of having a Swiss army knife, you've got a real screwdriver, a real knife, a real saw, right? And the Swiss army knife is a cool tool, but it's not a very good tool.
Dmitry Shapiro: And when you've got the best tools for everything, then we think that the end users get what they need.
Louis Bouchard: And do you think that if we compare it with the internet, for example, the internet is. We use it for everything. So it's like a Swiss army knife, but which contains all the specific tools. I don't know if artificial intelligence in general is different than the internet as in, it's a new, very powerful technology.
Louis Bouchard: And We are just at the beginning and I wonder if for instance, will, will it ever be able to be a Swiss army knife? Like, do you, you think that it will never be able to do everything [00:10:00] specifically and in a, in an accessible way to everyone?
Dmitry Shapiro: There's a lot there in that question, so I think that if we unpack it, then I think I can answer it.
Dmitry Shapiro: The capabilities again, let's call it the intelligence layer. Okay. These are these can be models. These could be other web services, for example, APIs that the models can use to be able to make their decisions, you know, real time data that they need in order to be able to questions. There's a lot of things that exist and sort of this layer that can gather knowledge about the world and then be able to then respond.
Dmitry Shapiro: To any sort of conversation, let's call it, or transaction or interaction that humans have with it. So I believe that the capabilities of the intelligence layer already are way, way, way beyond the capabilities for regular humans to be able to leverage them. There is no [00:11:00] prompt that you can write to be able to get this thing to really wield its power and, and because like the, you know, that's just not the right way to do that.
Dmitry Shapiro: Not to mention that I think it's clear that we will want multiple models to be involved simultaneously in fulfilling. Requests of users, some model might do something faster, cheaper, better than another model, and that other model might be better at something else. And so you will be using them in sort of in parallel or serially in multi step workflows.
Dmitry Shapiro: And so I think just sort of waiting for whatever, AGI, I guess, to just kind of do everything. But I don't think there's any point waiting for that now. We can already take advantage of amazing capabilities that exist. If we, again, create an application layer, make it easy to [00:12:00] unblock that bottleneck. The bottleneck is the human to computer interface.
Dmitry Shapiro: And especially now that we have to do it in language. In, in, by typing with our thumbs. It is really, really hard. You know, most people can't articulate, you know, can't, can't sort of craft sentences that are nuanced. And so to be able to then expect them to do that on, you know, with their thumbs on their phone, I think is is crazy.
Louis Bouchard: And what do you think are the required are the best skills to have to be a good like prompt engineer to actually craft those different applications or more specific applications.
Dmitry Shapiro: I think you need a number of skills. One is you need to generally understand what it is that you're creating the AI for, like, so you need to understand the human on the other side and what do they need.
Dmitry Shapiro: And, and so that's one, then you need to understand what the AI needs to do. In order to get the human what it [00:13:00] wants, what, what they want, and, and so you need the ability to be able to understand the steps required and to be able to disambiguate those to be able to explain that to something else to either another person, or in this case, an AI, you need the ability to explain things in a nuanced disambiguated proper sequenced way, and if you can do that, then you don't need to ever think about code or anything else. And by the way, that's the amazing opportunity now presented again to all of us is, you know, I'm, I'm a developer. I started writing code in 1984 when I was. 14 years old in high school. And so I know the power I feel knowing that I can just build things that if I want to build something, I can go build it.
Dmitry Shapiro: Okay. That's an extraordinarily powerful, you know, sense. The problem is like all of those [00:14:00] things used to require a lot of work on my part in order to be able to do like, even if you are a great coder, the amount of work required to write code is, is a lot. And Now that's changed where you can actually write the concise set of, you know, the spec, these instructions.
Dmitry Shapiro: And the intelligence layer sort of handles everything else. And so it went from months to hours, if even hours, less than an hour. In fact, mostly like less than 15 minutes to build most of these AIs. You're spending more than 15 minutes. There's probably some very fine OCD going on, which by the way is valuable.
Dmitry Shapiro: But you you're now you're tweaking the last sort of. 5%, maybe 2 percent of response. You get 95 percent value in 15 minutes.
Louis Bouchard: And I feel like there's also a difference in the way we use [00:15:00] language models. For example, you are well, from what I understand, you are mainly referring to the general public or like anyone using those models to, to help for general tasks that we do on a daily basis or like specific tasks.
Louis Bouchard: But what, what do you think of the, of using, for example, ChatGPT. To be a better programmer or to be more productive in your work as a prompt engineer, do you still think that learning how to use ChatGPT or another model to improve coding and to improve your work is valuable or is it still more valuable to use ChatGPT?
Louis Bouchard: Something a better prompt engineer built for, for example, for coding or for like a specific task.
Dmitry Shapiro: Look, I think again, it depends on who you are. If, if you are whatever, an accountant or, you know, sort of any other normal worker that uses technology, but it's really not a [00:16:00] technologist, you do something that.
Dmitry Shapiro: sort of just leverages technology. Do you need to learn how AI models work? I don't think so. You just need tools to do your work better. And you want nerds that do want to learn how models work to do their job and make amazing, simple to use tools for you to do that. Is it now? infinitely more accessible for you as some accountant to all of a sudden say, Oh, I'm curious about that.
Dmitry Shapiro: And I'm going to learn how these models work. And I'm going to spend an afternoon and I'm going to learn how to be a prompt engineer that can, one, leverage the power of these models and two, actually package them into apps that my coworkers could use or other people can use. Yes, that's very accessible now too in an afternoon.
Dmitry Shapiro: I think any normal, intelligent person can show up and learn all they need to know about generally, at least to get started and feel capable, right, about [00:17:00] how models work. And then how to take again, how to create the interface between regular consumers and these models and how to write that logic and create multi step workflows and configure various system parameters and things like that.
Dmitry Shapiro: So, and that's amazing that in an afternoon. You could get good at something, like you can master it. Now, Ultimate mastery will take a lifetime. You can be, you can be proficient at it. Yeah. In an afternoon.
Louis Bouchard: I just want to circle back to the, the early adopters. I really like your, your hypothesis, your theory.
Louis Bouchard: Just because I, I've been wondering why most of my friends that are not in the, in the developers field, at least, are like, incredibly surprised at how powerful ChildGPT is. But they are still all not using it. Like they've just logged in once, tried, found it super funny or super powerful. And then [00:18:00] they've actually never tried to use it for their real work or their real job.
Louis Bouchard: For example, I have a friend that is in finance and he, he lots of his time is used to create like Excel functions, whereas. He could basically just ask ChatGPT to create a function that does X and Y, and it will generate it for him. And so, from what I understood, is that you believe that in order to be like, more accepted by the general population, a model like ChatGPT will need to be extremely easy to use, which is...
Louis Bouchard: kind of already easy to use, but like even more easy and accessible as in, it should be for, for my friend, for example, it should be built in Excel or like Google sheets. Or do you, do you think that like, basically my question is what stops the person that could be more productive using AI? What stops this [00:19:00] person from using such a model that is already accessible and can Help this person be like much more productive.
Dmitry Shapiro: Let's, let's, let's take an example, maybe as a as an approach, I have a blog and I heard that ChatGPT can write blog posts and I feel kind of, you know, creatively stuck or lazy or whatever, or busy. And so I show up and I log in. And I get the chat interface. I then need to figure out what to write to get this thing, to write me a blog post.
Dmitry Shapiro: The things that come to mind are what is the blog post about and about how many words do I want it to be? and by the way, if I simply prompt it with that, it will return a blog post. And that's, that's amazing. But again, it's probably not the way that we want people writing blog posts. We probably want the.
Dmitry Shapiro: To get a bunch more information, right? Like, there's other things that matter. Like, [00:20:00] who's the target audience? And what is the tone and and what is the style? Is it writing it in my style, which tends to sometimes be verbose for the sake of being disambiguated precise? Or is it writing more in a style that's whatever visionary?
Dmitry Shapiro: And that's actually extremely imprecise, intentionally. And by the way, a bunch of other things that one could think about that you could. Tell ChatGPT, and it would write in a very different way and you would get a very different piece. And so the problem is that, again, the average person, 1, can't do that just naturally and 2, even if they can, doesn't have time to do that.
Dmitry Shapiro: So a better scenario would be if they showed up and said, I want to write a blog post and says, great, here's the blog posting app. And the blog posting app starts asking the question. What is the blog post about? And they type it [00:21:00] in. How many words should be? They type it in. Who's the target audience? They type it in and they just fill out a form.
Dmitry Shapiro: They don't have to figure out what the form fields are. They just fill them out. Do you see? And today, ChatGPT is asking consumers to figure out what those form fields should be. And ain't nobody got time for that. And so that's the difference is, is that that piece is missing and it needs to be built by people who want to do that kind of work because it requires thoughtfulness and it requires some tuning, right?
Dmitry Shapiro: Because models are finicky and weird sometimes. And so somebody's got to work out the kinks to really make an awesome blog post writer. And so we've got a bunch of them, you know, for example, actually on, on the service now already. So people are experimenting with dozens of sort of experiments with, with various blog post writers.
Dmitry Shapiro: You can go play with them. They're all different. You see, and some are going to bubble up and be your favorite and others might be somebody else's [00:22:00] favorite. But without that, I just don't see how any user, even a power user. It should show up and talk to ChatGPT. By the way, there's another problem. The interface of ChatGPT, the text box, actually doesn't allow you to tune all kinds of other parameters that the model has, like temperature, or response size, or, you know, system message, things like that.
Dmitry Shapiro: And so even if you're like the best prompt engineer in the world, whatever that means, ChatGPT is still not the right interface for you to use to, to wield the power of AI. So you need something else. And that's what MindStudio provides. It gives you access to all of those, you know, deeper parameters that models facilitate and allows you to create front ends that are dead simple for the most casual consumer to just show up, get the most amazing response back or responses back from the intelligence layer.
Louis Bouchard: But don't you feel like this [00:23:00] is both the parameters and the, the prompting is kind of a temporary solutions as a solution. As in, if we take back the example of a professional copywriter, where you ask him a blog for a blog post, the person, the professional will definitely ask you like about what and how long and like.
Louis Bouchard: The person will ask you all those questions that the prompt engineer currently has to figure out. But the more the AI are intelligent, like the more intelligent they become, shouldn't they soon figure out that they indeed need more information and Figure out exactly how to, what to ask and how to ask, and basically allowing us to simply use ChatGPT to do anything.
Louis Bouchard: Like if we ask it to do a, a blog post, if it's intelligent enough, it'll ask us about what and all the important questions that should be relevant for
Louis Bouchard: a task.
Dmitry Shapiro: [00:24:00] I, I think that that is a not a crazy. Sort of thought way of thinking about this. That's the obvious one. But I, I believe we're going to leapfrog that.
Dmitry Shapiro: And so I think that that's not going to be the case for. Very long at all. If it even exists, is that case? Because if you assume that that is true, I think you go to the scenario. Maybe I'll describe it in this way. Let's say you meet a genie and the genie says, thanks for, you know, freeing me from the bottle.
Dmitry Shapiro: You can ask me. I know everything. You can ask me any question you want. 1 question. What's the question to ask the genie? And some people might say, well, what's the meaning of life, you know, or good question. I believe the right question to ask the genie is the following. What question should I ask you?
Dmitry Shapiro: That's the problem is we don't even know as humans what questions to ask. Yeah. [00:25:00] Cause we don't know what we don't know. And anything that can do what you're suggesting it can do this type of an intelligence, in a sense does know. And so the right thing it needs to do is to start asking us a bunch of questions About ourselves.
Dmitry Shapiro: In order to understand what our strengths are, what our weaknesses are, what our gaps in knowledge are, what are our neuroses and various constructs that create our reality. And once it can figure that out, then it can be the driver and you're in the passenger seat. And it can just say, you know what, buddy, this is, this, this is an opportunity you should take advantage of. Because it'll help you do this, or here's some learnings, you think you understand this topic, but I've detected that you are missing a core component of understanding the topic. I've now made it clear to you what that is. And you're like, so I believe we get there, uh, [00:26:00] rather than we again, talk to ChatGPT and ChatGPT answers all our questions.
Louis Bouchard: And this cannot be done only through prompting, it will need to study you as well, or to see you, to hear you. What do you think of the interaction needed between the AI and the human, the user?
Dmitry Shapiro: We actually in May, June, July, before we launched MindStudio in August we did sort of a quick alpha experiment of something we called the YouAI Mind Indexer.
Dmitry Shapiro: And what it was, was think like TikTok is a never ending feed of videos. This was a never ending feed. It had a limit, but felt like a never ending feed of, of interfaces. Think like questions, open answer questions, multiple choice questions, a large grid of images with a label that says, tap [00:27:00] all that look yummy.
Dmitry Shapiro: And then another grid tap more that look yummy and another grid tap more that you look yummy, right? And so many different interfaces that are presented to human in order to tease out of the human, their preference, bias, gaps in knowledge, mental state, whatever. We had over 5, 000 people show up in alpha and engage with the median was 177 of these prompts.
Dmitry Shapiro: So a lot, and so we collected over a million data points, um, and certainly that's already a data set. We kind of stopped it because we're a tiny team. We didn't have time to continue focusing on that. We wanted to first build MindStudio. But, but yeah, we believe that through prompting human in the right way, with the right types of interfaces, not text.
Dmitry Shapiro: Text is like one type of interface. Most interfaces, I think, need to be visual. Kind of this or that. [00:28:00] Pick all that look yummy. Like, that's much easier than explain to me in text what you think is yummy. Right? And so so yeah, I think you can, I think you can index the human mind. And you've got to sort of continue to do that because we're constantly changing.
Dmitry Shapiro: And so it needs to be something that becomes a sort of a habit, you know, a part of the thing that we do it periodically. Pings us and says, Hey, here's three more experiences for you to give me a signal of how you're feeling right now, or how you're changing, et cetera. And so I think that's going to be pivotal to that next phase to your point of, of doing it.
Dmitry Shapiro: I think it's going to be much more like that. Then again, more chatting with ChatGPT ChatGPT is, is, you know, terminal. It's, it's DOS, you know, and, and while that's powerful and I like using the terminal on my Mac to be able to use my keyboard and drive things, but like, I'm a nerd, [00:29:00] the average person does not use terminal should not use terminal.
Louis Bouchard: Isn't tikTok already indexing our mind in some way for, for the people that like what we call the algorithm that just basically starts understanding everything you like and want to consume? Isn't, aren't they currently indexing our minds?
Dmitry Shapiro: Yes. They are, they are, but they're doing it in a, in a limited dimension, set of dimensions.
Dmitry Shapiro: What we were doing was, in fact, I described it like that when I described it to people is just like TikTok, except on all dimensions. Yeah. Yeah. TikTok, you know, TikTok's algorithm could sort of be described, sorry, let's take another step back. I'll add another one. Large language models can be described as you know, predictive machines that predict given a string of characters. What is the, you know, tokens of words? What is the next word? The probability of the [00:30:00] next word, right? And so that's what these things are doing. You could describe in a similar way the TikTok algorithm as a prediction machine that's trying to predict what video should I show Louis that he will find interesting enough to spend time on, push the like button, comment on, et cetera, right?
Dmitry Shapiro: That's what it's optimizing for. is getting engagement from you in any of those dimensions. If you were to describe our algorithm in the same way, although it wasn't an algorithm at the time, it was just a dumb feed that everybody got. But if we were to build an algorithm to do that, you would describe it as the following.
Dmitry Shapiro: What, given this prompt, given any situation, how would the user respond to it? How does the user feel about any situation? Give it to them. Which things are yummy? generally understanding what the user will choose. You know given this [00:31:00] decision to make high probability, they'll choose this rather than that.
Dmitry Shapiro: And so it just does the same thing as TikTok does, but all other dimensions. And if we wanted to then of course also recommend videos based upon those dimensions. Recommend videos to you based upon all of these other things rather than videos that you've chosen to watch and like before. So this is much more powerful than that.
Louis Bouchard: This is just a quick interruption to remind you that if you are enjoying the podcast, please don't forget to leave a five star review or a like if you are watching on YouTube. It helps a lot to support the channel. Let's get back to the discussion.
Louis Bouchard: And we've seen a lot of problems with TikTok's algorithms, just like privacy concern and just how intrusive it can be or who controls it.
Louis Bouchard: Do you see any such concern with indexing your mind where it's even, as you just said, way more powerful and also [00:32:00] much more intrusive as in it will just basically understand YOU and not just what you like to see, but it will understand YOU so is there, like, what can we do for privacy concern? Do you suspect lots of people won't want to use it?
Louis Bouchard: Or like people will just want to be indexed? What are your thoughts on this?
Dmitry Shapiro: That's a good question. My general. Approach to building technologies that might be disruptive is to go in to it with an optimistic point of view, being conscious of, you know, issues that could be, you know, dramatic issues and realizing that those things might exist and, you know, mitigating potentially for situations, uh, but generally following the, you know, the, the vector of, What, what [00:33:00] positive things can this do?
Dmitry Shapiro: And I think this is the same way that consumers will think about it. Some people are early adopters and they're not concerned about this vulnerability that you're pointing out. And perfectly fine for something to know them intimately. Well, you know, better than anything else has ever known them or anyone has ever known them, right?
Dmitry Shapiro: Like that's where you will get to very, very quickly with something like this mind indexer that we were in alpha with. By the way, I'm one of those people I do not fear telling whatever my, my weird secrets to, to people or machines and, and machines seems even less scary. Other people will be terrified and will never do it.
Dmitry Shapiro: But those people will have a problem because the people that do index their mind will be able to leverage AI radically better. Then people that do not, [00:34:00] that still have to type into ChatGPT, you know, the people have like plugged into the AI, because again, if you really want the, again, to leverage all of those sort of parameters of AIs, right, or not all the parameters, but many more parameters of the AIs, then you're able to articulate them language, then you need to, again, sort of create a digital model of your mind.
Dmitry Shapiro: That's what this mind index does. And when you then give that. Data set to AIs in order to understand you and personalize themselves for you and take the lead again, watching the world figure out opportunities for you and and recommend things for you. Well, then you will be in a radically better place, or at least more powerful place than, you know, a lot.
Dmitry Shapiro: I'd let's call them who is someone Afraid to do that, but I'm not pushing anyone to people should live their lives however they want to live their lives [00:35:00] and those that don't want to index their mindset. I've got no issue with that at all. I think that's great. But I hope they're not competing with the people that do because that's just not going to work.
Louis Bouchard: Yeah, it's just, well, I, I'm wondering what are your thoughts on Neuralink? Because I feel like it's extremely similar yet a bit less intrusive, but do you think Neuralink is kind of the future?
Dmitry Shapiro: I Think that these are different things, but yeah, let me see if I can frame it. So there's the ability to understand the brain and the ability to understand the mind.
Dmitry Shapiro: And so the brain is this physical electromagnetic chemical thing in our heads. And the mind are a bunch of thoughts and beliefs and constructs and preferences, you know, sort of conscious and subconscious. These are all sort of like weird words to throw around, but the things that make you think the way you do.
Dmitry Shapiro: Yeah. And then there's the thing that's like firing off a bunch of of [00:36:00] a bunch of neurotransmitters, right, that are, that are going off what Neuralink and another company called Kernel, K E R N E L so this guy, Brian Johnson is, is the head of Kernel. And then, of course, Elon Musk founded Neuralink.
Dmitry Shapiro: What those 2 companies are doing is focusing on the brain. And what we are doing is focusing on the mind. So, in a sense, we're similar in that way, but we're radically different. In the use cases. For these things, right? And by the way, kernel and Neuralink are different that Neuralink is, is sort of an implant and kernel is this like helmet that you put on and, and basically it gets sensors from.
Dmitry Shapiro: Electromagnetic impulses in your, in your skin, in your, in your skull.
Louis Bouchard: And would, but wouldn't they be ultimately able to also process and ideally understand our mind, thanks to those [00:37:00] sensors on the brain?
Dmitry Shapiro: We will see, you know, what can be seen from those types of sensors. There are some interesting things that like they've been able to do, like, you know, show pictures of, of cats to people and then sort of be able to recreate those by having them think about, you know, a cat and things like that.
Dmitry Shapiro: But again, I think the first I'm not qualified to, to talk about it sort of at the scientific level of, of what they're doing. So take that with like a trillion pounds of salt, but it feels to me a much easier way. is To have a relationship with a feed just sort of gives you a bunch of opportunities to give it feedback rather than plug an implant into your head and walk around with a, you can't even walk around with a helmet, just sort of be sitting with a helmet periodically. So like, I just, I think they're for different purposes.
Louis Bouchard: I want to get back to [00:38:00] like building those tools that index, not index our mind that, that we interact with. So basically they're the creators of all the different applications on MindStudio or the different AIs on MindStudio.
Louis Bouchard: For example, I know that well, I assume it's still a feature, but that you can select the model that you want to use for, for the application you are creating. So GPT or Cloude or, or Llama and I wonder if you have any. Do you have any insights or tips on which model to use when, like when to, to select an open AI one versus cloud or when to, to try to use an open source one?
Dmitry Shapiro: I Do, but I caution people from taking my, my insights here as, as being sort of canonical you should experiment. I think the fundamental thing is each different use case. [00:39:00] A different model might do it better, and especially, and then tune the parameters of that model. And for other use cases, you should have multiple models that you're using in the right sequence with the right things.
Dmitry Shapiro: But generically, sort of what we've seen is this. GPT 3.5 is faster than GPT 4. And so if you care about that, GPT 3.5 Turbo can do most of the things GPT 4 can do. And there's no need to use GPT 4. In fact, we have GPT 4 sort of isolated as, as behind the premium tier, one, because it's just more expensive, and two, I think because people sort of misunderstand that they don't need it.
Dmitry Shapiro: They should be using these other models. If you're dealing with large documents. Then Claude's, you know, context window, token window is you know, much larger right now than GPT. And so using a Claude model in that case, [00:40:00] you know, is preferred. We support retrieval augmented generation, so you can take documents, upload them, and we turn them into vector embeddings.
Dmitry Shapiro: And in this case you, a lot of the work is done, sort of the preprocessing work is done. In that retrieval phase where you're not passing sort of giant documents to the LLM and having it figure it out. And so that generally kills that thing of you should use Claude versus GPT in that scenario.
Dmitry Shapiro: But I'm certain that all of that is a moving target, very fast moving. And so by the time somebody might be watching this those things might have changed already too. And, and so I, I wouldn't. bank on them. By the way, this is another reason that like, we can't expect the consumer to choose. Do I go to GPT for this query and Claude for that query?
Dmitry Shapiro: Like that just can't be it.
Louis Bouchard: Yeah, exactly. There's another [00:41:00] concern that we have when we create those kind of model is to use a fine tuned model where we take for example, an open source one, and we just train it on, on more specific data of the task that we want to achieve versus using the most powerful one, like GPT 4, for example.
Louis Bouchard: And it's, you kind of need to know if the fine tuning fine tune model will work better because you need to build a data set and you need to train it. And it's, it definitely has more costs than just calling an API. And so would you have any insights on when should someone take a fine tuned, cheaper model versus GPT 3.5?
Louis Bouchard: And if they should fine tune, do you have any recommendations on the amount of data that they need?
Dmitry Shapiro: It is different case by case. Yeah. All of these sort of neural network models [00:42:00] are non deterministic, right? They're not consistent. Right. You, you can so, so it's hard to sort of make absolute statements, uh, about them and not to mention, again, that things are changing so rapidly.
Dmitry Shapiro: So I think at this moment and for the near foreseeable future, certainly for this next year, I think sorry, I'll only pause because things are moving so fast that even a year seems like a really, really long time, but I'll go on record saying that at least for this year, you still need to do a bunch of experimentation as the prompt engineer to find the right.
Dmitry Shapiro: To find the right thing, right? So like really align kind of like tuning an instrument, you know, you tune too much, you go sharp and you're like, oh, that's wrong. And I got to tune back and you go flat. And then sooner or later, you finally got it. And like, it's on the right frequency. And this is the same thing with building AIs today.
Dmitry Shapiro: You can get very generic again, you can sort of not worry about those [00:43:00] fine tuning and they work pretty well, still amazing, but if you really want to tune it off, like it's a bunch of experimentation.
Louis Bouchard: What should you do to mitigate any risks of productizing those models? Because of course they are powerful and they are super useful to many specific use case, like being a copywriter or whatever you want to use them for, but they still have limitations, even if you prompt them like incredibly well, for example, hallucinations or just biases in general. And yeah, how, how would you try to, is there a way to not fix, but to, to mitigate hallucinations and what should the prompt engineer do or the user do to, to ensure that either the language model is not.
Louis Bouchard: Providing completely false facts or to spot that the model is providing false facts. [00:44:00]
Dmitry Shapiro: Yeah, great question again. I, I will give you answers, but I encourage people to again, take everything I say with a lot of salt. This is. Sort of best practices at this moment, rather than sort of science, right? It's, it's an art.
Dmitry Shapiro: Exactly. It's an art. And again, these aren't even my sort of own insights, but these are things that I too have found to generally work better than just writing sort of. Regular prompt one is to break down whatever the thing is that the machine needs to do into discrete steps and get the machine to do those steps.
Dmitry Shapiro: In the right sequence and it sort of at every step evaluate bias or hallucination and you can then sort of detect [00:45:00] how a hallucination is getting introduced potentially into the process, right? It's like, it's like signal process and you've got to make sure that your input is clean. Because if it's got a bunch of noise, it's going to get amplified and you're going to get distortion on the other side.
Dmitry Shapiro: And so one is this, like, step by step approach that people talk about. Again, nothing that consumers should have to think about doing. That would be crazy, but we prompt engineers should be thinking in that way. By the way, that's the right way to think about building any system is you got, you know, before you got to ultimately understand it and extremely nuanced to be able to write code.
Dmitry Shapiro: Now you can cheat. You can just generally understand it, but you have to understand it in some way. Now that's fine. Another one is you can use again, for example, retrieval augmented generation to keep the responses within. A corpus of allowed responses versus any response from learned. You [00:46:00] know, behavior before and so that can dramatically limit the sort of the, the the spectrum of hallucinations that could happen rather than from the entire world.
Dmitry Shapiro: It could be just in your set and then, you know, playing around kind of with this temperature parameter, which is kind of a weird proxy parameter that models have. But lowering the temperature a bit so you get sort of consistency, those are the ways that you can sort of generally deal with hallucinations.
Dmitry Shapiro: Again, a lot of this feels like if anybody has ever done, you know audio production work, you know, working with a big mixer that sort of enumerates all of the various you know, spectrum of sound, and then having to find the right, way to balance everything to create room for various frequencies.
Dmitry Shapiro: Like it's very similar type of work that you do now with these [00:47:00] models, except the output is not audio, it's language or images, you know, or soon audio and video too.
Louis Bouchard: And do you think that I'm just, I just had a thought about the recent news that. OpenAI released with ChatGPT where you can now scan PDFs and it like removes tons of companies and, and plugins that allowed you to do that.
Louis Bouchard: And are you afraid of, of basically the companies that, that, that own those models, like OpenAI releasing the tools themselves to allow them to, to allow prompt engineers to create better apps for users rather than another, a third party company like MindStudio, for example, where you try to do that. Like, are you either concerned or scared from what the owners of those models can do?
Dmitry Shapiro: I'm not. [00:48:00] And I'll tell you why but that's also sort of generally my approach to things is I, I never think about potential competition and, and kind of focus on, on the problem and solving the problem, but I'll tell you, I believe that that's not the case. 1st of all, I think they should continue if they want to be.
Dmitry Shapiro: Consumer facing, you know, and user facing that they should continue to work on interfaces to make things much easier than this DOS command line ChatGPT interface, right? That's a prototype that we're all hacking to do amazing things, but open AI can do much, much, much, much better than that. And and so, and so introducing the ability to talk to a PDF, for example, is one tiny step in them, you know, doing that. So I fully expect they're going to do that and everybody else is going to do that. The problem is like, I believe that that's not what end users need, though. End users don't need, I believe. [00:49:00] A, a better single interface to lock them into one model in this case, you know, GPT OpenAI or model ecosystem, let's say, because not one model, it's multiple models, but in one model ecosystem they need Switzerland, they, they need something that works across all services, all models is not stuck on any one model and, and facilitates the ability to use, The best model for the right part of the job and to abstract all of that again, this is, it's a whole layer, this application layer, there's a, a lot of decisions that need to be made and a lot of systems that need to be built to make that easy to use for prompt engineers, for developers, you know, even though, like, for example, you can, you can use your own PDF, which at GPT, but you can't package that experience as an app so that your mom can use it.
Dmitry Shapiro: You have to teach your mom to use [00:50:00] ChatGPT and now upload this thing and good luck with that, you know, and so like, there's a, and you can't charge for that if you did the work and so great, you'll teach your mom, but like, so we facilitate this ability for people to show up no matter what these models do, it doesn't matter.
Dmitry Shapiro: To us in a sense, the more capabilities they have, the better for us, because we just allow prompt engineers to leverage them so that regular end users can benefit from them. So Tide, you know, rises all ships in this case.
Louis Bouchard: So right now, all the, the prompt engineering skills and the, the things we learn about, for example, when to use cloud or when to use OpenAI or when to use Llama or Mistral, it's all completely, well, not useless, but it will soon be all irrelevant.
Dmitry Shapiro: No, no, it's relevant for those people that are building AI applications for everybody else to use for them. It's relevant, meaning [00:51:00] this is the new developer. The new developer is not somebody who writes code. The new developer is somebody who writes instructions for orchestrating the intelligence layer and packages it and, and delivers it to end users.
Dmitry Shapiro: And so those people, all of these skills you're learning now are valuable as long as, again, the, the, the technology itself, the models are changing so rapidly that like a lot of these things like this temperature parameter, or these multi step prompts and things like that, that might become much less relevant or disappear altogether.
Dmitry Shapiro: And, and so all of these like, technology moves so rapidly that the interfaces to be able to. Wield the technology for developers are also changing very fast. So developers have to constantly adapt and learn new ways of doing things, new patterns, new technologies, new approaches. Prompt engineers will have to do the same thing is to stay current and learn all of that so that the regular end user doesn't have to do any of that because they can't.
Louis Bouchard: Do we still need developers or do we only need prompt engineers?
Dmitry Shapiro: It's a good question for now, we still need developers for some near foreseeable future. We will still need developers. We need developers for fewer things. And so a lot of things that would have had to have been done by developers certainly don't need to be done by developers anymore.
Dmitry Shapiro: Yeah. But other things, you know, [00:53:00] infrastructure things, for example connectors and other things, those still require developers to write code. But even that is like radically changing. Like for those people who've played with like GitHub Copilot, you know, or other copilots, like, it's, it's amazing. I started writing code in 1984 when I was in, in, I was 14 years old in high school.
Dmitry Shapiro: And, and so I've seen all of this come up. Copilots are profoundly impactful. To software development, uh, mostly taking sort of very, very junior developers and giving them the ability to. Do real work and then for, you know, advanced developers sort of making them radically more productive.
Louis Bouchard: Yeah. So this leads me to another question because I for myself, I learned much later than you to, to, to code. I think it was like when I entered university, just because. Before [00:54:00] that, I was in science and like, I knew nobody involved with programming or internet or whatever. So it's, it's like, I was really into math and physics and gaming, but not programming.
Louis Bouchard: And so anyways, I started learning about programming to actually develop an app, a game that I had in mind and I wanted to, to productize, to create a product. So I learned to create something just as I believe you did at 14 years old. You learned to code, to do, to create something. And so I wonder if right now there's a 14 year old listening or just anyone with no programming skills.
Louis Bouchard: Would you suggest them to start to learn to code to create something or to start to be a better prompt engineer? If like, for example, their end goal is to be able to, to start a new company or, or just create a product for themselves.
Dmitry Shapiro: Yeah. So I think [00:55:00] again, very much depends on the person and what it is that you're curious about and you want to create.
Dmitry Shapiro: If, if you want to, if you have an idea for something that you want for yourself and it can be created by leveraging AI, and again, it just depends on what the idea is then the right approach is to use AI to create it because that's going to be much faster, easier, you can do it now, and there's a lot of things that you can do with that.
Dmitry Shapiro: Many, many, many, many things. If it requires you to write code, Then you might need to write some code, but first start with the AI piece of it. do I believe that learning to code is a valuable skill to have? I do. I think by the exercise of learning to code, you learn how to think and that also helps you be a better prompt engineer from the standpoint of, like, realize that Where the ambiguities need to be and how to abstract [00:56:00] things into systems and things like that.
Dmitry Shapiro: So I think that's, that's valuable, but in a similar sort of dimension, I see people now get into AI. And so, like, I have friends who have, like, reached out to me. Over the last few months now, I'm so excited about AI and like, I'm taking this like course that's like MIT on like how to build neural networks.
Dmitry Shapiro: And like, I'm sort of studying my math again, like linear algebra. Like I'm studying gradient descent and I'm like, that's cool, but like, why? And like, they think that's going to let them leverage ChatGPT better. if they understand gradient descent. Yeah. And that's just silliness. Like that, that is, that does not make any sense.
Dmitry Shapiro: So if you want to use ChatGPT better or leverage AI better, learn to communicate in words in a nuanced, structured way, disambiguated way. If you want to learn how [00:57:00] to do the math behind neural networks and help as a scientist in that way, then by all means go and brush up on your, you know, on your math.
Louis Bouchard: And communicating better will also help you just in your personal life and just with your partner in general.
Dmitry Shapiro: Indeed, indeed. Yeah, it's the most look these. So my degree is in electrical engineering. I, I graduated with a bachelor's. Yeah. I've never done a day of it. I have five kids. And so people periodically will ask me, well, what do you want your kids to study in college and. My answer is like, I'm not going to tell them what to study.
Dmitry Shapiro: I want them to choose what they study. But if they sort of came to me and said, well, should I, should I study like a science engineering or should I study sort of like liberal arts and communications? I would point them to communications, versus the science, um, which is crazy because of, you know, the 18 year old me certainly would have thought that was nuts.
Dmitry Shapiro: But now that I'm 54 I know that the most valuable [00:58:00] skill to have is the ability to communicate with other humans. And now also machines and that that's the most important set of sort of valuable skills. And you'd learn more by studying literature than you would learn by studying whatever computer science university.
Dmitry Shapiro: You can learn computer science whenever you want. go, go read some books.
Louis Bouchard: Yeah, I definitely agree. Just communication. And even, even just simple, like knowing I, it will sound dumb, but knowing how to tweet and to use social media is also extremely relevant just to first build an audience. But also if you have an audience.
Louis Bouchard: If you build a product or anything you want to sell. Well, you can sell it right away. It's like invaluable how, how much communication brings you just in all facets of life, both professionally, but also personally is [00:59:00] definitely the, the right choice. And I will, I think I would also right now say that the same thing as you, even if I also did an engineering degree, I, I don't know, it's, I guess it's hard to be sure when looking back, but I currently.
Louis Bouchard: Kind of assumed that I could have learned everything I, I learned on the internet or just by myself, which is also kind of what I did. I, I did follow online courses while I was doing university classes just because I felt like they didn't teach me enough. So I guess going to school to learn.
Louis Bouchard: Things that actually need practice and expert supervision, such as a better communication and like just I don't know how to say that in English, but like just appearing better and when, when you talk and allocation if that's a word in English, but yeah, I, I feel like everything around communication is so important and it also [01:00:00] needs supervision.
Louis Bouchard: It's, I guess it's somewhat possible to learn it online and just by practicing such as going on to podcasts and, and things like that. But it's, It's kind of a skill that is hard, harder to, to learn by yourself compared to programming, for example, where I feel like it's even easier to learn by yourself at night when you are super into it and just like coding.
Dmitry Shapiro: The as I look back on my career. I graduated in 1992 with my degree in electrical engineering. I didn't want a job as a, as an electrical engineer. And so I got a job as a salesperson selling phone systems to small businesses. I could sell up to 20 phones. It was 21 phones.
Dmitry Shapiro: I had to get a, a major rep to do it instead of me. And I, I was one of. Eight people in this like cohort of new salespeople that they hired, a company called Executone. And every day at 7:30 AM, 7:30 to 8:30, five [01:01:00] days a week, we had a role playing sales training where we would pair up and somebody would be the customer and somebody would be the salesperson.
Dmitry Shapiro: And, you know, we would train, and then we would be on phones prospecting, trying to get appointments with small businesses to go talk to them about their phone systems. And that job sort of the 1st, 3 jobs I had were sort of salesy jobs. Again, looking back, we're just some of the most important things that I needed.
Dmitry Shapiro: You just need time communicating with other people and understanding I think people misunderstand sales, meaning they think like used car salesman thing that like you're pressuring people to do it. That it's not sales. Sales is education. The sales is the ability to take a person who's obviously interested in something, because they're listening to you.
Dmitry Shapiro: And to be able to get them to understand that whatever it is that you think they should buy is good for them. [01:02:00] And so you need to be able to read people. You need to be able to understand how to meet them where they are, so that you could then lead them to where you want them to be. And that's a teacher, so that's the skill you need and most people, you know, never had a chance to really practice that.
Dmitry Shapiro: And so if you can find an opportunity to practice that, then you can do it. By the way, speaking of that, so I recently interviewed, uh, on a podcast like this a guy Michael, who is a sales enablement person inside of a large enterprise. And his job is to make sure that the salespeople, sales engineers, and customer support people in this cybersecurity platform as a service company are well trained and equipped, right?
Dmitry Shapiro: Like, that's his job is make sure the sales force is in top shape. And so he's used MindStudio [01:03:00] to build trainers that the salespeople interact with. And so the AI is training the salespeople, a different AI is training the sales engineers, a different AI is training the customer support people. And then he's also built like a bunch of automation things for the thing.
Dmitry Shapiro: So like this training thing, it used to require a human. Now it can potentially be done with an AI. There you're missing the eye contact. And body language and tone, which are really big signals to understand if somebody's understanding you or not and where they are, but still like that, that's the skill that needs to be practiced.
Louis Bouchard: Yeah. It's, it's really cool how I say that in, in, in many episodes, but I find it crazy that artificial intelligence was kind of seen as evil back in the day, like 2012, I don't know where like [01:04:00] people basically assume that larger companies will. will use AI to even better control the population and to do like, it will increase the gap between reach rich and poor and et cetera.
Louis Bouchard: Whereas I feel like it's completely the opposite where AI democratizes a lot of stuff and allows you to do allows anyone normal or just like Michael to do way more just by leveraging this technology. It's basically super accessible to anyone. And it allows people with less resources to do more. So it's like the complete opposite of what we first initially expected.
Louis Bouchard: I, I just find it crazy and really cool.
Dmitry Shapiro: I, I think that both of those things are true though. Meaning organizations that have sort of more sophisticated AI capabilities [01:05:00] than the average consumer. Are going to be able to do orders of magnitude more, you know, have a competitive advantage, whether they use it for good or evil.
Dmitry Shapiro: So those tools not only empower the population, they also empower the, the overlords. And so we're not out of the woods yet, but certainly the democratization of the, of being able to leverage AI, which is again, kind of another way of putting what mind studio is trying to do. It's to democratize the creation and accessibility of AI.
Dmitry Shapiro: That, that's extraordinarily important because that other side is also true.
Louis Bouchard: Yeah, it's also my exact goal of democratizing AI, but in more of an educational way. But I, I find it very cool that you basically remove the need to learn how to, to prompt engineer and to, to tune these models to allow for, well, [01:06:00] first to allow for anyone to create and monetize an application, which is kind of really cool without having to learn how the software development and monetizing process and et cetera. So that's first initially super cool, but also to, it makes us able to create very specific and super well built, well prompted apps that someone can use right away to do a task and ask for a blog post, but it already knows.
Louis Bouchard: It's everything you want without you needing to think of like, what does it need to know? And will it hallucinate and et cetera. So I, I really like what you are doing and I feel like it's, it's really aligned with what I'm, I'm trying to do. And I invite everybody listening right now to just try and play with with the platform.
Louis Bouchard: It's YouAI.ai so the platform is MindStudio and you can. You can play with, with the models right away. Well, the, you can play with the AIs right away or build yourself your, your own [01:07:00] AI and even monetize it. And also if you have anything else you'd like to share with the audience, feel free to, if you have anything to promote or, or anything else to share now, now's your time.
Dmitry Shapiro: We, what depends on when they listen to it tomorrow we're hosting a hackathon. That's a weekend long hackathon. We'll be hosting other hackathons if you're listening to this later than November 4th, 2023. But tomorrow's our first hackathon and so you can join it and, and create things. And, and so that, that's sort of one announcement.
Dmitry Shapiro: There are a lot of new capabilities coming. So one of the things that you should expect from us is that we move very, very rapidly. In expanding the capabilities of the platform, so a lot of people in our discord are sort of commenting on that. So if you like sort of bleeding edge stuff, you can come and watch us do it [01:08:00] and play with it and give us feedback.
Dmitry Shapiro: If you are. Excited about this technology and start doing things, please start recording videos and tutorials and writing about it. People need help understanding it. And so all of that, we're all pioneers here again, if you choose to sort of engage in this and we welcome you to be. You know, part of this sort of movement that, that, that we're doing here.
Dmitry Shapiro: And, and please join our discord. We'd like to meet you and, and uh, yeah, appreciate the time.
Louis Bouchard: Awesome. Well, thank you very much for what you are building and also for your time for this amazing discussion.
Dmitry Shapiro: Yeah. I had a lot of fun.
Louis Bouchard: Me too. Thank you.[01:09:00]