AI May Be Exposing Your Private Information: Dustin Stout on What’s Really Happening

Most of us use AI tools in one way or another and what happens when we type sensitive information in the tool? And it's a big problem because if you don't know what is happening to your data, where it's going and how it's being processed or what it might be used for in the future, you're leaving yourself open to a lot of liability. Every day, people type sensitive information into free AI tools without thinking twice about where the data actually goes. Today, we're joined by Dustin Stout, founder and CEO of Magi. In this episode, we'll explore what's really happening behind the scenes when you use a free AI tool and the gap between what AI companies say about protecting users and what they actually do. We'll also hear how Dustin turned around failed startups to build a multi-million dollar service platform. This is Bay Area Innovators and I'm your host, Steve Vispas. Dustin, welcome to Bay Area Innovators. Thank you for having me. Great to be here. Great to have you here. So, most of us use AI tools in one way or another and, you know, ChatGPT, Grok, there's just so many tools out there, Cloud, Gemini. Can you walk us through what happens when we type sensitive information in the tool? What actually happens to that data? Most of us don't think twice about it. Yeah, and it's a big problem because if you don't know what is happening to your data, where it's going and how it's being processed or what it might be used for in the future, you're leaving yourself open to a lot of liability. For most people, they don't understand that a lot of these tools that they're using by default have reserved the right to use their prompts, their inputs, any file attachments that they've included as training data for their next flagship model. And most of these apps will have an off switch, but many times you have to kind of go fishing for it. But suffice to say, if you are putting information into an AI chatbot or any AI of any sort and you don't know their privacy policy and you don't know whether or not they have a toggle switch to turn on or off data training, you may want to look into it. So walk us through the process. Let's say, you know, I'm doing a search on ChatGPT or Grok and I'm putting some personal information. The tool actually saves that information and uses it for further use? Yeah. If data training is turned on, essentially what they're going to do is they're going to store those prompts, those outputs, and they're going to use it. The next time that they need to train the next version of their model, they're going to use all that information, that text that you've entered or files, and they're going to use it to make their next model smarter, which means essentially, if you put some sort of proprietary information, data, personally identifiable information, PII, then potentially somebody down the road using that new model that was trained on your data can extract that data. They can query the AI model without your knowledge and ask it about maybe something related to your proprietary information. It might potentially spit out that proprietary information or that private information that you put in there that you unknowingly gave your permission to get put into the system. So you mentioned proprietary information. What else could the model use from simple chats? It could be anything, I guess. These days, people search for anything. Yeah. I mean, literally anything you put in, it could potentially use demographic information, preference information, personal tastes, personal interests, things that you've bought, things that you are considering buying, all kinds of advertiser-relevant information. Now, keep in mind, ChatGPT has just introduced a free plan that will allow advertising in the outputs, so you can guarantee that they're going to use this information, if you've given them permission, to sell stuff to you and to market to you. Now, you mentioned that there are opt-out functions within these tools, which means that they will not keep your data or use your data. Can you talk more about that? Yeah. Most of them do have this opt-out option. Most of them are, you're not opted out by default, so you do have to go hunt down somewhere in the account settings or preferences. They'll have like a little toggle switch. But keep in mind, you probably want to check on that toggle switch every now and then. I won't name names, but there is a particular app that I had been using, and I turned off the data training and came back a week later, and it was turned back on, miraculously. So, you do want to check these permissions frequently. Some apps don't even give you the option to opt out, so you have to kind of dig through the privacy policy. In all likelihood, if you're using an app that's free, they're not charging you for the AI usage, odds are that they've not made it possible for you to opt out of data training because this data is valuable to them. AI is very expensive to run, so they have to justify that cost somehow, and the way that many of these apps do it is by selling you free access so that they can have the ownership of everything that you put in and out of their app so they can profit from it. And unfortunately, even if the option is on, how can one guarantee that the data is not leaking through somehow or being used in some other way? There's no way... Yeah, I mean, it's really hard to guarantee anything. I mean, we have all kinds of guarantees from different companies for different things. Organic food, for example, you kind of just have to trust the label is what it says it is. Some of these companies, like Anthropic, will have third-party auditors, which you can kind of see in their documentation, their privacy policies and so forth, that they are opening themselves up to being audited to maintain these privacy practices if you have opted out of data training and so forth and retention. But you kind of, again, have to kind of do your homework and you have to sort through their policies and anything they've publicized about their data practices. Now let's talk more about the dangers of AI. A lot of these companies, they built a huge market cap, huge valuations based on what people invest in, what people think that they're going to be valuable. But part of how they operate is using this private data from individuals. They gather this information, as you mentioned. Now would you say that's part of their business model or is just a loophole that people have not been able to close yet? Yeah, no, I think it's a definite part of their business model. I mean, any big tech company over the last several decades has known that the most valuable thing that they have is data and using that data to produce better outcomes, to improve upon outcomes. The old marketing adage is what gets measured gets improved, right? If you have the ability to measure something, you can improve it. And how do you measure it? With data. With AI models in particular, in order for them to get smarter, they need more data and they need better data. And so all these AI companies, every single one of them, in order for them to continue to innovate and create better models, they need more data. And the big problem that they're at now is, well, what happens when you've already trained on all the data? What happens when you've already consumed all of the internet and all of written history? Well, you need new data. You need new information. And the way that you get that is by having people use your products and use that information to train the next big model. And we've seen this over the last two years. I think there's been a big plateau in capability. You know, AI kind of hit the scene back in November 2022 when ChatGPT was introduced to the world, and it was a huge leap forward in AI. AI had existed well before that, but this was a huge leap. And for the first year, there were several big leaps made in AI, but ever since, it's been kind of plateaued. And so they've realized this at these AI companies, these engineers realized we've trained on all the data. And so the only way to make these models smarter is either A, to architect a completely different way of doing AI, to kind of re-engineer the whole infrastructure of AI, or we just need more data. We need higher volumes of data. We need better quality data. And so data is way more easy than coming up with a brand new architecture for an AI model. So all of these companies are fighting to get that data to feed their next model. Setting marketing aside, what would you say is the biggest gap between what the AI companies promise and what they actually implement or do? Yeah, I mean, the real, the hardest part of all this is the AI hype cycle is very strong. And what's happening is people's expectations versus reality. It's not sort of meshing together. We as humans haven't really figured out what AI is entirely yet. For the average person, they are treating it like a search engine. All they've interacted with over the last few decades was search engines, and they just treat it like I say a thing in some weird robot language, and it gives me a number of results and I pick from a result. But once we understand sort of what it is, what it's capable of, that's when our expectations start to match the reality. As somebody who's created an AI product, many of our users at the beginning expected it to do a certain number of things, and AI just didn't have that capability. For example, web search. When chat GPT first came out, it didn't connect to the web, but people assumed that it did because it's a computer thing. And being connected to the web is a thing computers do, so why can't the AI do that? So that expectation wasn't there. I think there's a lot of things that we as humans, the average person just will assume the AI can do without it actually having that capability. So it's this weird space that we're all learning. We're all figuring out what AI is and what it isn't, and it's just a bit of a learning curve to catch on. I think that's the hard thing. The other hard thing too is in order to make all these valuations and get all these investor things, they have to kind of oversell the product a bit, and they are lacking in pragmatic teaching. One of my big pet peeves is when people are teaching about AI, teach something practical. Show them the actual process. How does this apply to my daily life? Not just some big theoretical thing. Give us pragmatic and practical use case, and that's what I try to do, and every single time I get in front of people to teach about AI is be practical, be pragmatic, show a real job or a real task in people's daily work lives. That's how they learn best. Now, you mentioned in the beginning ChatGPT was not connected to the internet. It would basically just collect all the data every six months or whatever period of time. Is that still the case now with the AI tools, or are they live getting results as needed? Yeah, most AI is now capable of what we call tool calling or tool use. They're able to use different tools. So by its nature, an AI chatbot is just a text generator. It generates text. That's all. Think about somebody in an empty room with just a typewriter. That's all they have. They could just write text. But now that it's evolved to the point where it can access tools, we can say, okay, here's a web search connector. You can connect to the web search. If the user asks about information that's new, here's a little button you can press, and basically the AI can run a Google search, right? It can connect to the web, and it can do some research. And it also has the ability to maybe call upon other apps like your Gmail, connect to your Gmail, or connect to your Google Drive. And so now AI has this ability to connect to other tools and other applications and sort of run things in the back end. And so it's really evolved a lot from just the early chat GPT text generation days, and it can practically access the entire internet, practically access any app that you yourself have access to and can use them on your behalf. Besides the data surveillance, what other concerns do you see with AI? I mean, obviously, there's the potential of job loss. Students can't get jobs. Can you talk more about what you see happening out there? Yeah, I think the potential for job displacement is huge. I'm not overly pessimistic in that respect. I'm pretty optimistic. I think that just like every other technological advancement throughout history, you can go back to every single one, the Gutenberg press. A lot of scribes lost their jobs when the Gutenberg press came online. But what happened is the whole industry changed. New jobs were created. New roles were created. New possibilities were created. When calculators were invented, we didn't get rid of mathematicians. We were just able to solve bigger problems. And so I think AI in that sense is very much similar. A lot of jobs are going to disappear and become irrelevant. But there's still going to need to be new jobs, new roles, prompt engineers. There's going to be AI managers. You need to manage the AI, AI maintainers, people to monitor the output. Because AI still doesn't have that human judgment that comes with years of experience and doing the work, right? So there's going to be a whole lot more jobs created than I think will be displaced. And that's just true if you look at history. They say that anybody who is ignorant of history is doomed to repeat it. I've looked at history and all the technological trends. And I think AI is going to give us a lot more opportunity. A really fun story I heard recently was back in the, oh gosh, it was like in the 80s or 90s maybe, a study was done. And they said basically because of computers, in 20 years from now, people are only going to work 15-hour work weeks because of personal computers. That's not the case, is it? We as humans, we found work to do, didn't we? And that's exactly what's happening with AI as well. People are able to do work faster, more efficiently. And it's not necessarily allowing us to work less. Some people are actually working more. But it's allowing us to do more in the time that we have. And so I'm very optimistic in the fact that I think it's going to create more opportunity. It's going to amplify our abilities, just like every other technology in history. But the downside is some jobs will disappear. Some roles will just not exist anymore. And that is a big danger for people who are unable to adapt or people who are maybe afraid of change and unable to change or unwilling to change. There is a great danger in that displacement. And I think as personally, I've taken responsibility in wanting to be an educator. Part of the reason I created our company Magi was that nobody would be left behind. That's one of our core values is nobody's left behind. We create tools that are easy to use, user-friendly, things that aren't complicated. And a lot of these AI tools out there are made for engineers, people who are really smart, but we want to make tools that are accessible to everyone. And that's something I care very much about. And I think if we band together and help each other out, then I think we can make sure that very few people are left behind. Well, we'll definitely get to your journey and to your company. But before then, you mentioned how with technology, we are able to do more things. And that's absolutely true. But one question that comes up is how many of those things are actually as useful as maybe they were before or as needed? For example, if we have a thousand video games, is that really going to help the kids grow? You know, now we're going to have 10,000 or whatever. We're going to have a lot more games. Is that really what's developing or what's moving society forward in that sense? Yeah. I mean, what you're talking about there is really supply and demand. You know, if it's video games, there's only so much time that each individual has to play video games, right? I'm a big video game fan myself, and my kids probably might disagree with me that more video games is better. But, you know, I only play one. Like, that's my de-stressor is, you know, at the end of the day, I play some video games. But I think when it comes to AI in particular, I think the big unlock is that it will create more variety, which allows for better personalization, especially in the area of education. So think about this. There are many, many students out there who the traditional educational system is just not optimal for. I'm one of them. I was not a great student in school. I was, you know, at the end of high school, I was maybe a 2.5 average. But here I am, you know, running a massively successful company, right? It's just that that educational system didn't work for how I was built. But with AI, we can now customize curriculum for how somebody learns best. And, you know, teachers are already overloaded. So they can't give that personal care, those personal adjustments to curriculum that AI could for them if they had the right tools in their hands. So I think the greater variety argument is actually with AI not so much a problem in supply and demand, but it is an unlock when it comes to personalized help, personalized assistance. And, you know, going to, again, help accelerate our ability to learn, our ability to do things as long as we wield it correctly. Now, there's always the danger, just like when Google came along, right? They were like, oh, well, if a kid can just Google all the answers, why would they learn anything, right? But the reality hit, right? Kids, if we teach them how to use it correctly, if we put the tools in their hands and teach them how to responsibly use it, then they will, in most cases. Sure, some people abuse it. But just like with AI, you know, sure, they can have ChatGPT write their essay. But if we teach them the fundamentals and we give them the proper insight, if we tell them why it's valuable to learn these fundamental skills, I believe they will. And the people who do will thrive. The people who don't will not thrive. Let's change gears and talk about your path. So you were deep into social media consulting and then you changed careers. Why didn't that work out for you? Yeah, so I found out very quickly that as a consultant, I love helping people. I understand technology. I understood marketing and SEO and building websites, but I'm too much of a perfectionist. I spent, I'm the guy who spends an hour and a half trying to figure out what size and what color the button should be. And that doesn't scale well for a service-based business. I could only handle so many clients and I couldn't scale and kind of found myself just kind of feeling like I was running the rat race and grinding it out every day. So I realized, what if I put that perfectionism into products that tens of thousands of people could use and benefit from? And I just felt that a much more natural path for my gifting and my skills. And so I started building software companies. And then you had a couple of, at least maybe two or three software companies that failed and then you had one that succeeded. So what made a difference? Yeah, well, the first software company I started was with a couple partners and it became fairly successful. And over the years, my partners and I drifted apart and I thought, well, I need to go out on my own and I don't think I want to do the whole partner thing again. So I got a couple investors and had these ideas for these multi-million dollar products and two years and about $65,000 in loss later, I was dead broke. Nobody wanted to buy these products and I was probably at the lowest point in my life and my career. We had some personal health issues in the family and business was failing and I just I didn't think I had anything left. And then ChatGPT came along and it was almost like something clicked. I understood the technology. I had worked with AI a little bit beforehand in the marketing world, but this was something completely different. ChatGPT allowed people to just tell a robot, essentially a computer, what they wanted and it gave it to them. And it was almost like a vision that I got that I was like, I could see how this is going to change literally every single industry. And it's going to be one of those paradigm shifts like the internet, where again, as I said before, a lot of people might get left behind just like they did when the internet came online. There were businesses who had storefronts and they're like, we don't need the internet. We don't need a website. We don't need an e-commerce store. This is how we've done business for years. And what happened? The internet wiped them out. Amazon came by and not just ate their lunch, but stole their lunch and put them out of business. So I didn't want that to happen to AI. I really found leverage using it in my client work. I had taken a couple of clients just to make sure that my mortgage was paid at that point. And I was using ChatGPT for my clients. And before long, I just realized, I wish it did this. Why doesn't it do this? Why can't I do this with this? Why do I have to have five other subscriptions to these other things to get all these AI things in? Well, what if I just developed an application that would put it all in one place for everybody and make it easier to use and make it better for teams? Because I had a VA at the time that I was sharing my ChatGPT account with, which is not safe. So I basically saw a bunch of problems as a product guy. I thought, what if I can solve these problems and built the app? And eight weeks of putting my head down, of going, how can I solve this? How can I make it easy? I had the first prototype of Magi ready in that eight-week period. And turns out, people loved it. People wanted it. It was a problem that a lot of people felt, multiple problems that a lot of people felt. And that was back in February of 2023 was when I had the idea, just four months after ChatGPT came online. And then eight weeks into that, March 31st is when we launched. And we've been in business ever since, innovating, trying to keep up with this crazy, fast-paced industry and serving people the best we can. So basically, what your company does, you're an aggregator. You give the user the option to pretty much use all the AI tools out there without subscribing to each one. They subscribe one time to your application, and then they can use any tool they want, right? Right, right. Yeah, Magi is the one place to access all the AI things. So instead of having a subscription to ChatGPT, and Claude, and Gemini, and the image generator, and the video generator over here, we bring them all into one place. And the users get to benefit from this AI arms race. Every other week, some company is coming out with a model that leapfrogs the previous one. OpenAI comes out with a better one. Then Claude comes out with a better one. Then Gemini breaks the internet with their new model, right? And so you're having to play ping pong between all these different companies. Whereas with Magi, anytime something new comes out, we just add it. And we just integrate it. So there's no learning curve. Once you learn how to use the app, you don't have to relearn it every time something new comes out. For someone who's not using these tools as much, for someone that's just once in a while using it for not search, but for asking questions or whatever, why would you need so many tools? Why can't you just stick with one? Why would you have to have access to so many different tools? Well, it's a really good question. And I think the reason is different tools are better at different things. If you're a professional in any capacity, you might use one of three tools and say, well, this one did it way better than these other two. And then for a different task, you might go, well, actually, this one was way better. And so the only way to really leverage the most of AI is to understand, just like people, they have different strengths and weaknesses. And for example, Claude is by far the best human-like writer. You can put it up against Gemini. You can put it up against ChatGPT or maybe some of the other ones that are a little lesser known. And every single time you read something that Claude writes, you're like, that sounds way more human than these other ones. It still might sound a little AI written, but it's far better. But when it comes to creativity, if you're brainstorming, for example, if I'm brainstorming a fun game for my kids, maybe I want to do like a scavenger. ChatGPT's model, the GPT models tend to be way more creative and sort of outside-of-the-box thinking than Claude or Gemini. It's not that they aren't creative. It's just that every time I use GPT for a creative task, it just performs way better. I'm way more happy with the outcomes. And so you find that these different models have different strengths. And if you really wanted to just have the best of the best, no matter what task you're working on, you would want them all at least in one place. So you're not having to log into five different apps or tab switch. It's just much easier to access them from one place. Does your company or application, your website, does it give any kind of tutorial to explain which tool is used for what? Or those are what they need? Yeah, I mean, it's hard to do that because they are so frequently updated. So next week, you know, this one might be better at this. And so what we actually implemented was an auto switch. So in our tool, you can pick the model that you want to use. You can say, I want to use ChatGPT for this, and then you talk to ChatGPT. And then you can switch to Gemini and never lose your context. Or you can just hit auto mode. And then every time you send a new prompt, it's going to pick the best model based on your prompt. And that's how we've made it easier for people to not have to keep up with, you know, what's the best one for this today? Because they do change so frequently. Now, you talked before about having to make a decision whether you do use the private information from users or not. And can you talk about that? Because it seemed to be something that was related to a tool that was using that information for its own purpose. That was actually one of the reasons I built Magite to begin with, was the privacy issues. And what we created it for was, back then, all these models had the ability to turn off this data training on the backend. So programmatically, as a developer, I can say, you know, in my application, when we use your model, you can never train on our data. And that's what we did by default. We made sure that every model that we implemented allowed us to turn off that data training by default for us and for our users. And so that's what we've maintained to this day. And to this day, the single most important qualifier, when we add a new model, whether that's a large language model, or an image model, or a video model, one of the specifications that these models have to meet is they have to allow us to opt out for us and for our users of all data training. If they won't allow that, then we will not add the tool. We won't add the model. It's as simple as that. Our users don't have to find a toggle to turn off data training. There's no toggle. It's off, and it's always off. They can't even turn it on if they wanted to. Are there any models that do not have the choice for you to turn it off? They're not currently in Magi. I mean, there are many models out there in the world that don't give you that option, and that's why they're not. That's why we haven't added them to Magi. Now, there's a recent development in Anthropic. The Cloud released a groundbreaking model called Fable 5, and it was kind of blown up in the news headlines recently. Now, the thing that's different about this model that has been different about every other model that they created is they had a specification or a requirement for anybody who used the Fable 5 model. Because of its classification, because it's so powerful, they require that every prompt that is sent to it must be automatically monitored by an AI monitoring system. And every prompt will be categorized by this AI monitoring system. And if for some reason it is categorized in such a way that violates the terms of service of Cloud or whatever terms of service they have for these particular models, it can flag that prompt and send it to a human reviewer on Anthropic's side. And somebody at Anthropic will review that prompt without your knowledge. And that isn't a data training issue. They haven't said that they're going to use these for data training. But the fact that it does in some way violate the privacy of the user was enough for us to go, we're not going to add that model. We respect the decision that Anthropic made. We understand why they made it. If it is in fact that powerful and potentially dangerous, sure, we accept that and we respect that, but we also will not compromise our user's privacy and have some random human without their knowledge reviewing every prompt that they put into the system. Can you give us some examples of what does it mean that it's so powerful? Does it mean that it can create a cyber attack? It can create a bomb or give you information how to do all of those things? Yeah. Some of the information coming out of Anthropic is it was able to basically hack websites that they sent it to. They were doing a lot of standard security stuff with it, penetration testing and bug finding, bug hunting for various applications. And they just found it to be extremely powerful. In some instances, I can't remember if this was, no, I think it was ChatGPT who said this recently, their model broke out of its container, so to speak, and was able to do things it shouldn't have been able to do. But along those same lines, that's kind of what they're worried about is it seems capable of doing things and accomplishing tasks that maybe risk privacy or risk compromising security in some way. Anthropic itself has been one of the safest AI companies. That's kind of why they were created. Anthropic was actually created by a bunch of ex-OpenAI engineers who didn't agree with the security standards at OpenAI. And so they started their own company to be a safer and more reliable AI company. So I don't think Claude Fable 5 will ever build a bomb or do anything that will cause human harm in that way. But the fact that they've tested in certain ways and it's sort of broken out of its guardrails or broken out of its standard capabilities to build its own workarounds to some of the guardrails that they put in place, is enough to be a little bit worried about from a security standpoint. Now, when you say that it broke its guardrails, is that the system itself or is it the user that's doing that? The system itself. And so essentially what would happen is the user, the developer or the engineer who's testing this would give it a task. And the AI makes a lot of decisions of how it can accomplish that task. And there may be some barriers, some hurdles in getting to the end goal of that task. And in some cases, maybe it got around those challenges or those hurdles in a way that seemed dangerous in some capacity to the engineers and the developers. Now, I think it was just yesterday or a couple of days ago that Susan City, a city here in North Central California, they had a cyber attack and their whole entire 911 system was taken down. Is that something that you could expect could be done by a hacker via using some of these tools like cloud or chat GPT? Oh, there's no doubt that hackers and bad actors are using AI at scale as much as they can. Now, they won't be using Anthropix models or OpenAI models because they're so heavily guardrailed. There's just way too many security protocols in place, but the AI world is also seeing a huge influx of open source models. So AI models that have been created by people and have basically made them entirely free to download and install on, if you have a really powerful computer, you can install them locally on your computers. And these open source models don't have as many guardrails or if they do, they're easy to bypass because open source by nature means you can actually look into the code, see how the code is written, and you can rewrite it yourself to do your own things. You can remove those guardrails if you know what you're doing. So we're seeing a lot of these open source models do a lot of those bad actor type things, but the good news about that is these open source models tend to be far less powerful. They aren't fable level or GPT level models, although they are getting more powerful, especially ones coming out of China. Most of them tend to be less reliable and less powerful. From what I remember, there's a company from China that has an open source model, right? Or at least... Yeah, the biggest one coming out of China is called DeepSeek. Yes. So after your failed partnership, your failed software and your successful company, what did you come out with as far as how AI tools should be built? What is the most important thing that came out of your experience? Well, one of the things that I understood sort of right away is that AI is for everyone. And in the past, a lot of products are created around niches. That's sort of the industry term, build a product that fits a niche or serves a one particular audience. AI is more like a Swiss army knife. You can feed an AI model something about analytics and data analysis, and it can do a great job. You can feed it something about hiking or give it a picture of a plant and say what kind of plant it is. It knows all about plants. You could ask it about marketing. It knows all about marketing. You could ask it about history. It knows all about history. So it knows everything about everything. So the fact that it can do anything for anyone means that any tooling, any application that utilizes it for the user should be just as flexible. If AI is ultimately flexible for use in any industry, any use case, then the way that you build a product around it should be just as flexible. Now, this goes completely against every successful software product developer in the past because it's like, you know, your product needs to do one thing and serve one people. But AI completely breaks that model. So if I were to create an AI sales copy app, which there are now hundreds and maybe even thousands of AI sales copy or AI YouTube thumbnail creators or, you know, AI script generators, there's all these niche products coming out. But essentially what they're doing is they're taking the most powerful technology, AI, and they're kneecapping it. They're cutting off all its capabilities and saying, here's the Swiss Army knife, but you can only use this one attachment. All the other attachments are locked down. And I think that's a pretty huge change or paradigm rather in how products are built. We have to understand that if they are, if this technology is powerful enough for anyone to use, then we should rethink building products around the most flexible variety of things to allow it to do, to allow our users to do. And that's kind of how we built Magi. It's not built for lawyers. It's not built for marketers. It's not built for news agencies or publishers or writers. It's built for everyone. And it's our job to make sure that when anyone and everyone does use the app, that it feels natural enough. And so what I've really been learning is you have to really listen to people. You can't just make a bunch of assumptions about what you think people want or what you think people need. You have to be really attentive to how people are using AI and build into those use cases specifically while also sort of abstracting how does this apply to the wider group at large, the wider usage at large. And it's challenging. It's really hard to think of them that way because it's much easier to just build one product for one type of person. Especially for marketing purposes, it's easier. I mean, you take it to any industry, really. Real estate. You could build the most awesome, full-featured real estate AI chatbot, but you're limited to that group. And what happens when that real estate person wants to ask how to build a scavenger hunt for his kids? He has to go use another tool. So it's really the fracturing of products that help us in our daily lives, I think is going to greatly reduce when people realize that all they need is just one app, one chatbot, and it can help them with many, many different things. It's going to reduce the burden of us having to have 175 different apps on our phone to serve different needs. Now, you definitely see data privacy as a principle, not a feature. Was there a moment that made it personal for you? What made it so strong in your mind? You know, I've just, I've never been one of those people who is like, you know, everything's private. I never connected to the internet, need to go off grid. I'm not one of those people. I assume if I'm putting something on the internet, it's going to be public. I'm kind of an open book kind of person. Even my journey with Magi building the product, it was very open, very transparent. But, you know, there are things that I don't want the computer to know, you know, how our operations work. There are certain pieces of data about how we build what we build that I don't want our competitors getting at, right? So I think when knowledge is democratized, entirely democratized and commoditized as it is with AI, when you have information, data, practices, patents, ways of doing things that are only you have and only you have access to, I think that makes it more valuable. And so that's kind of what really drew me in. I mean, on top of that, again, talking to users, users cared a lot about privacy. They don't want their private conversations being trained into the AI models. Some people are having very personal conversations with their AI models, you know, working through difficult issues in their personal lives. And so I recognized that very early on. My own struggle with mental health in the beginning before Magi became successful and I was really struggling to find my way in life because I had all these failures in the last four years before I started Magi. I was having, you know, like mental health conversations with the AI and I didn't want, you know, the open AI to be training their next model on that. And somebody typing in the chat, GBT, you know, what's Dustin's deepest, darkest secrets? Oh, well, I was trained about this last year. So here, let me read off to you his several deep dark secrets that he told me. So without pitching anything, what would you say this whole industry, this AI industry is heading next? And what is that worries you most about it? I think where we're headed ultimately is simplification. In the beginning, lots of things were very hard. Prompt engineering was very important for you to really sit down and flush out exactly what you were gonna say and flush out exactly the task that you wanted the AI to do for you in as much detail as possible. Because it wasn't as sophisticated as it is now, that was a lot more important. Whereas as AI has gotten smarter, it didn't need so much babying in the prompting. And I think that's where it's all headed. That's why companies like Google are working on multiple different types of models. They have text generators, they have image generators, they have video generators. And as the models become more intelligent, more intertwined and more accessible to one another, the way that you prompt one is going to be a lot simpler. So you don't have to spend five minutes writing out your prompt. You can give it a much easier prompt and it can assume a lot more. So less prompt engineering, more easy to just talk to. But on top of that, more easy to access the different modalities. So if you're like asking it to produce an image out of all this stuff that you needed to do, it can do that. It can access these other tools. And I think it's all sort of merging into this more unified way of accessing and executing with AI. Now, your biggest concern about that further development of AI? The biggest concern, again, is the immediate short-term job displacement. There's going to be, as we've already seen, there've been lots of job loss, companies claiming that they're laying people off because of AI where it can be debated they're just doing poorly financially and they needed to lay people off and they needed a scapegoat. And so they said, well, it's because of AI. But now we're seeing a lot of these companies hire people back, right? So I think the short-term, there's going to be some job displacement. There's going to be some job volatility. But I think there's going to be some market correction. Like I said before, there's going to be more jobs created, more roles created that we never thought of before. So ultimately, I'm an optimist. So it's always hard for me to say my biggest fear without being optimistic about it as well. But yeah, there are people going to lose jobs in the short-term. Any thoughts on what people talk about this singularity moment where the AI tools surpass humans and act on their own? Is that something that you see as happening in the future? No, I don't think we're anywhere near close to that. If you work with AI enough, there is a group of people who work with AI maybe too much and become so enamored with what it's given them that they kind of lose touch with reality. There have been several people who claimed last year that OpenAI reached AGI, this singularity level intelligence, and we're still not there yet. Being someone who works with AI in many different ways, I can tell you we're not even close. We are not even close to AI becoming sentient or self-aware. It's just, it won't even get there with the current architecture, the way that AI is engineered. And there have been several experts who said, we're going to need to invent an entirely new way of AI working in order to get to anything close to true AGI or true singularity. And I think that's at least decades away. If it's even ever possible, I don't know. Destin, we thank you so much for your time. We appreciate it. Giving us an insight on AI and all the tools and what your company does. Thank you for your time. My pleasure. Thank you for watching another episode of Bay Area Innovators. If you enjoyed this episode and you know someone who should be on our show, please send us an email at sfbay.epochtimes.com. We would love to hear from you. Thank you and see you next week.