Share this
dnsUNFILTERED: Michael J. Stattelman, AI For Good
Mikey Pruitt interviews Michael Stattelman about the development and impact of AI models, particularly in the realm of NSFW content detection. They discuss the origins of Stattelman's work at Falcons AI, the importance of contextual awareness in AI, and the ethical implications of deploying AI solutions. Stattelman emphasizes the need for accountability in AI development and the potential for AI to address real-world problems, while also acknowledging the challenges and responsibilities that come with such technology.
Mikey Pruitt (00:00)
Welcome everybody to another episode of dnsUNFILTERED. Today I'm joined by Michael Stattelman. Did I say that right?
Michael Stattelman (00:07)
You did.
Mikey Pruitt (00:15)
Awesome. And Michael and team at falcons.ai has the most downloaded NSFW image model on hugging face right now, which is quite the accomplishment. Let's start there. How did you, you know, the success that you've seen, how do think it happened?
Michael Stattelman (00:26)
So honestly, I feel like with the rise of Gen AI, people weren't ready. And so obviously with any new creative tool, there's good and bad. And so a lot of the bad started proliferating. Prior to that, though, the origin of how we got started with this model was we were at our son's friend's birthday party.
And we're talking to some parents and they were talking about, you know, the proliferation of this type of content on these gaming platforms that kids use like Roblox, Minecraft, you know, things like that and how they were struggling with it. So, you know, being an AI, you know, computer vision engineer, you know, me and Nizam partnered and we're like, hey, you know, we can kind of, you know, put something out to solve this problem. So curated the data and it was horrible. My honestly, thank God for the content moderators. You guys are. Wow. You guys are tough.
Mikey Pruitt (01:22)
I was actually thinking like when I was kind of thinking of questions to ask you like content moderation, a big part of it is identifying like NSFW content, adult content. Unfortunately, humans have had to do that job for quite a while. But now with models like yours, that may not be the case. So like, kind of out of a job in a good way.
Michael Stattelman (01:37)
Right, right, right. So, honestly, we took it from the approach of the ML side, the AI side, right? So it's not just grab up a bunch of illicit images, throw them in and train, right? So there has to be balance and you want the model to be trained appropriately. Yeah, it's a, wow, it has a psychological toll. And for us, we didn't deal with any of the like graphic, violent, you know, that sort of side of the house. ⁓ But so, yeah, we just kind of, you know, curated a broad and deep data set to train the model on to get to catch as many of the edge cases as possible. Cartoons, yes, they're cartoons, all the anime, like, you know, a very, very, very diverse and deep data set, train the model, put it out on Hugging Face. And then I started engaging in a lot of the developer communities, the AI ML developer communities. And honestly, that's where I think the majority of the kind of initial surge came from is because a lot of these smaller developer communities, they talk to each other and they're, know, developers span the globe and they're like, Hey, you know, I'm having this problem, you know, in a stack overflow exchange or whatever.
Hey, does anybody know of a model I can use really quick? We have this customer portal and people are uploading some wild images. And so they're like, hey, this open source model on Hugging Face, it's really good. And so that's where I think the initial traction came from. Spinning around to earlier this year, late last year, the whole Flux model came out, Stability, Excel model, giving people the ability. All the uncensored versions of those models too started coming out, giving people the ability to generate this type of content on the fly. I you know, even with regards to, you know, like, I don't know if you're a parent, I'm a parent, we've got eight, right? Right. So when you start talking about the incident in, I think it was Nassau County, New York, where the high school student got busted for creating deep fakes of, you know, girls in the school.
Michael Stattelman (04:03)
And so I think with the rollout of all of these GNI tools for the good, there's also, know, the, the, the crevice was deep, right? For the, for the bad. And so that's kind of where I feel like a lot of that stuff started picking up a lot of the popularity because, you know, when you start looking at paid API services, they get very expensive, you know, when you start doing a per image, right? So if you're spinning through, you know, directories or something like that, you know, it's cost prohibitive for SMEs, small business, independent developers, you know, things like that to actually justify the cost of paying per image. So if I have something that's 95 plus percent accurate across the board, you know, with high confidence, I'm okay deploying that. And so that's honestly the take on it. And as of today, we are at like 160 million downloads over the last 30 days, as well as our total download amount is 500 million, so half a billion.
Mikey Pruitt (05:01)
And that's across all your models. I have a few questions stemming off of what you just said, but start with what are the other, well, so like you have, you've used your model to identify NSFW content. I'm curious like where the line is between like, let's say like a bikini and like nudity. Is it that sophisticated?
Michael Stattelman (05:03)
So the way we've trained our model, our current version two model that's actually not on Hugging Face, it's the commercial version, is we added a layer into the model for contextual awareness, right? So a bikini it may or may not be NSFW.
Mikey Pruitt (05:44)
Some bathing suit of tires, NSFW.
Michael Stattelman (05:48)
This is true.
So there's kind of two parts to that, right? And we in our human brain, we know this, but you have to kind of spell this out for the model. So a lot of it comes down to not just the attire, but the disposition of the person wearing it, right? So you can have an outerwear, swimwear, lingerie that kind of rides the line, but if it's in a suggestive pose, it pushes it over the line.
So yeah, so that's the benefit of adding that contextual awareness piece into the model is it gives it kind of more of a threshold to say, okay, this is on the line. It's right up to the line, so it's not really NSFW, but if the way this person is posing is suggestive, then push it over the line, right?
Mikey Pruitt (06:41)
So is there like a statistical likelihood, like a percentage that your model spits out? Okay, so you're like a confidence score. So I've got like a site for parents. My threshold level is very, very low, like 10 % confident. It's fine with me. If your model thinks it's, you know, a very low confidence level, I'm still going to block it. Other sites may have a range. That's really cool.
Michael Stattelman (06:44)
Yes. Yeah, it gives you a confidence.
Right, well,
yeah, so you can actually, so the user ends up, you know, gets to set the threshold, right? So that's why the output is JSON and it's just, you know, what the predicted label is with the confidence score. Well, you know, in your solution, you can say, okay, I need that confidence score to be, you know, 90%, either way. Or even if it's 61% that it believes this is block that, right? I don't want any chances.
So where that comes into play is you look at things like school systems, right? So our newer model is able to run on device like Chromebooks, laptops, because it is low power. And so with the school system, you would set that confidence like ridiculously low because you don't want anything, anything even remotely. Now, is it going to flag, you know, probably a Barbie doll in a bathing suit with a low threshold? Of course. But you accept that to prevent any of the other things from slipping through.
Mikey Pruitt (08:03)
So like I work in cybersecurity and there's a lot of vendors out in the space. I actually work in marketing, so I understand this more than most. There's a lot of, let's say, stock that is like AI security, AI base, AI this, AI that. Your model is one of the, I mean, maybe not the first, but one of the early things that I've seen that are like actually AI is being used to combat bad behavior or the bad guys or however you want to phrase it so.
Michael Stattelman (08:34)
So that's our thing, right? If you can my shirt, AI for good.
Yeah, and so for us, for us, it's, you know, these tools can be used, however, you know, we don't use, we don't, so we don't get into any of the physical, the crossover into the physical. And what I mean by that is we don't get into like, you know, robotics or anything that can have physical impact on human life, right? That's kind of like off limits for us.
And so we look at AI for good. How can we deploy these solutions to people that may have this problem, but don't have a way to address this problem. So one of our earlier products is Precise AG. And there's a, you know, there's a problem with bird flu going on right now. People want to know why egg prices are high. Look at the bird flu problem. And that's, that's the kind of leading cause of this. It's not, you know, trade or economics or whatever. And I don't want to get into that, but.
Mikey Pruitt (09:30)
Yeah
Michael Stattelman (09:30)
it's the bird flu, right? Because they're having to cull millions and millions and millions of these birds, these chickens, turkeys, things like that. So our first application was Precise AG, and it was trained on a data set of visual symptoms of bird flu. So it allowed rural farmers to take a picture of a chicken and at a cursory glance, right? It's not, you biological and 100 % accurate. So at a cursory glance, take a picture of a chicken, and see if it's giving off visible signs of bird flu. And you know if it does, then you need to separate that chicken or whatever. And it goes in the realm of contagion management. How many farmers have the ability to curate a data set, understand the mechanics of actual computer vision model training?
Michael Stattelman (10:20)
And so, the ability to, and it was, you know, downloadable through your phone. So it was small footprint, low latency. So if you were out rural, limited network capability or limited network bandwidth, you were able to still use it. Right. So that was just, you know, kind of, that's kind of how we're geared is how can we deploy these spectacular solutions to solve point problems for real people in the real world? you look at, go ahead.
Mikey Pruitt (10:47)
So what are some of the other models that Falcon is working on? Not that you're working on that are done. We'll get to the working on later.
Michael Stattelman (10:55)
Yeah, yeah, absolutely. So another one is from a linguistic perspective, right? Last year was the biggest election year in human history. There were so many countries that were having elections last year. And so, you know, being a parent, you start looking at, you know, what type of language is filling the air of your home, you know, from coming off these devices.
So I started looking into it and you could do like advanced linguistic analysis. It's really eye opening. But a lot of the content coming out early last year, as you would probably understand, was a lot of fear mongering. And so what we did was we curated a data set of just kind of signals and identifiers that would kind of signal fear.
You know, somebody trying to scare someone into something. And so we trained a model, gave it away free to journalists, reporters, academics, researchers, to be able to analyze political speech, any type of speech. No, I don't want to even just, you know, limit it to political speech. Newspaper articles, YouTube video feeds, things like that. Because I look at it like, why are you spreading fear?
Michael Stattelman (12:18)
Right? So it's either one of two things. You're either doing it to protect me or you're doing it to sell me something. Right? And so what is, what is the, what is the need? So I always try to pride myself on, giving people actionable information. Right? And I think, I think being in IT, that is our job, right? Even cybersec you're safe or you're not. This is taken care of or it's not. And here's why, right? You know, you try to give people, you know, legitimate fact-based status of their systems and so for you know from the fear-mongering like I said it was It's pretty rough like I even did a YouTube video where I showed you know somebody how to just use Gemini and spin up a really quick fear-mongering detection that you can just paste in a YouTube URL and will basically catalog everything that said during that YouTube video as fear-mongering not right
Mikey Pruitt (13:09)
This is what I love about AI, the current state anyway. It's like there's a lot of happenings in the open source side of the house. So take your fear-mongering detection model. You can use that to analyze the sentiment of a YouTube video. That's a great option. YouTube could add a little, how inflammatory on a scale of one to five. Or it sounds like you could also use it to say, do ad blocking for ads that are trying to scare you into buying something, which are not always scams, but a lot of times are scams.
Michael Stattelman (13:44)
Right, I got another use case. So I was contacted by a physician and he was saying like a lot of his patients are older. And he was asking, was there a way specifically to your point right now, is there a way to deploy this model like maybe on like a TV or the fear mongering is a little bit too big on like a TV or their YouTube feed?
To kind of show his patients the content that they're watching and its direct impact on their stress levels in real time, right? If they have the Apple watch, right? Because see, we don't think about things like this. What we consume does have a psychological impact on us and psychology translate to biochemistry, right? So if you're constantly hearing, my God, you should be scared of this, you should be scared of that, you should be scared of this, your stress is going through the roof.
And it may be a gorgeous day outside and nothing's really happening in your general vicinity, but you're, you know, you know, locking yourself in the basement, you know, grabbing cans of food because you think it's the zombie apocalypse outside, right? So that, you know, that those are the things that I feel like the open source community, that's why we constantly contribute to the open source community, that the open source community can build and develop that allows other people to collaborate and compound and build solutions out of that actually solve certain problems, right? So we came at it from a journalistic research perspective, but then out of the blue, here's this physician saying, hey, you know what? This is actually affecting my patient's health. And they're older, so they're more susceptible to the impacts of higher stress environments.
Mikey Pruitt (15:25)
So you're creating these foundational building blocks, and other people can come along and build the tools on top of that. And I'm curious, like, so you mentioned Bird Flu, we've got the NSFW model, Fear Mongering, there's one called Chaos Engineering, which I'd love to hear about. But how do these ideas pop up? Are you guys just messing around and talking about what would be cool, and all of sudden you're making it? Is that what happens, or something?
Michael Stattelman (15:31)
Absolutely.
yeah, yeah.
Mikey Pruitt (15:53)
Like a physician calls you, your friend says something like, how do these, ⁓ let's just call them categories come into existence for your team?
Michael Stattelman (16:02)
So I try to be, and you'll understand this more than most being in cybersec, acutely aware of all of the fringe. Because the core has been taken care of. There's solutions wrapped around solutions, around solutions, around solutions, and branches and tangents of those solutions. But I feel like as human beings, we kind of get complacent with what we see and what we know.
But so we try to look at what's on the fringe. What could be a problem? What is a problem? Can we leverage AI to solve that problem? Like in a very specific point way, know, LLMs are great, GenAI is great, but they're kind of very broad and, you know, they allow a lot of stuff. We kind of focus on what is a specific problem that we solve. When you're thirsty, here's a glass of water. That's a thirst problem, right? We're not saying...
Here's a convenience store with every soft drink and Gatorade that was ever made. No, that's not what we do. And so it's just kind of like spit balling. It's conversations like these, to be completely honest. So I was having a conversation with a buddy of mine.
And it's why your message was like, wow, this is crazy. He's into router security. But we were just talking and he was saying, how do you feel that AI from a bad actor state is going to impact, specifically impact my field? I think a lot of the models, protocols, and policies that the cybersecurity field is based on and I'm not cybersec, so I'm not really into this. We were just spit balling. And I was like, but I think the problem is, and it's not a new problem, but AI is supercharging it. It's adaptive malware, adaptive capabilities, polymorphic structures that do not give off signatures of identified known in the wild issues. And so our solution, similar to the Gen AI generating illicit imagery is to use AI to catch that. So let's map the technology to the technology because with the speed and volume and variability that, that, you know, AI is capable of, it has far surpassed us already. See, a lot of people don't want to talk about that, but the reality is our systems from probably five years ago have surpassed a single person's ability to map that in their head, right? And so why not leverage these technologies to manage these other technologies, right?
Mikey Pruitt (18:51)
That's exactly right.
You're, ⁓ and I haven't seen like a polymorphic malware in, in the wild yet, but it has been, I have seen it said a lot. That is a concern, and I'm just been thinking like you're kind of on the right track where you're creating individual models that are very specific tasks. And this is really good for like, people coaching their employees on how to use AI. Like you want to have a task in mind and then create some automation, AI infused thing with that. And your models are kind of doing that too. Like they have a very specific goal. I think that's the way to detect polymorphic malwares because it comes with some type of AI model probably built into it, or at the very least is calling out to a command and control server to get instructions to use AI somewhere to like, how do I adapt myself so that I can hide somewhere else on this network?
Michael Stattelman (19:23)
Right. And so I feel like, there's, the problem with thinking machines like we're building is there's the possibility of a domino effect, right? And what I mean by that is what happens when these malware containers or generators or whatever you want to call them,
start factoring in time series systemic collapse right from a domino pattern they're not going to match anything in the known world from a from a cyber security perspective right because let's and i'll give you a prime example ⁓ so let's just say you have like a smart city this is like the stuff that we're looking into like today like right now
Mikey Pruitt (20:27)
Yeah.
Michael Stattelman (20:38)
Like I'm not all paranoid, but I just would rather kind of, you know, have a, have a look ahead and get this out in the public. So other people are thinking about this as well. From a solutions perspective, not a, ⁓ I'm trying to scare you into, know, the robot rising or robot uprising, right? Right. So if you look at, just say, let's just say you have a solution that is developed by one of these thinking machines that says, okay, I have access to these traffic lights. Right.
Michael Stattelman (21:07)
And I have access to a power grid and maybe a wastewater facility. So I will create a traffic congestion problem, At a rush hour in a major city. And then once that is hit and there's a certain level of congestion, that triggers the threshold for another piece of malware to kill the power grid.
Right? So there's no way to reroute the traffic and you know, now the police are deployed and all this stuff is happening. And then, yeah, by the way, now that this is happening, let's go ahead. And so I'm talking about like cascade from a systemic perspective, how from a cybersecurity, and like I said, this is cause I'm not in cyber set, but this is kind of the things that I'm looking at. It's possible with AI because it can set up stages of systemic collapse, right?
I'm not trying to give anybody ideas, but this is obviously, you know, kind of from a protectionist perspective of what we look into. So what does that look like from a cybersecurity perspective? Because the water treatment security is not paying attention to this other, you know, traffic congestion situation. And it all can work together to exacerbate to create what is the perfect storm of issues, right?
Mikey Pruitt (22:35)
It's funny you mentioned the phrase perfect storm because all I was thinking is you could just train an AI model on like every heist movie ever made. For example, the Italian job with Charlize Theron and Marky Mark. Like that's exactly what they did. They had hijacked the stoplights and weighed the weight, tire pressure from a camera, like all this stuff. So like an AI could potentially do that. And if it was polymorphic in some way or even not, if it just knew. There would be no signature for our current systems to detect.
Michael Stattelman (23:09)
Correct, correct. And so, like I said, I mean, and there you go, you you're talking about how do we generate, generate, generate, you know, these solutions or ideas just to spitball and think about is things like that, you know, watching movies, watching TV, hearing people talk, listening to, you know, other podcasts and being like, wow, that's strange. You know, when my, when my other podcast, I was talking about, I think Tom asked me, ⁓ basically the same thing. And I'm just, I think you have to be super curious, right? I like to hear.
you know, nuanced words, words in a manner that maybe I hadn't heard before in that specific context. And I'm like, let me go look that up. And so then that, you know, runs me down a rabbit hole and I'm like, okay, well, if this is the good, obviously the yin and the yang, what's the bad? Okay. If that's the bad, how could we develop or deploy a solution to prevent at least one part of this bad from happening? Right. And then, you know, we may train a model and it may be a complete failure. Right. So you were talking previously about the chaos engineering model.
Right? So yeah, in basically like a systems uptime, systems reliability, stress testing scenario, you can kind of engineer problems to integrate into that system and see how your people or your systems respond to certain failures or chaos that gets applied. So it helps, ⁓ like IT directors and solutions managers, your network managers, applications, server managers kind of become more robust because now they have roadmaps of things that they've already tested, already trained. They know where the breakpoints are. They know what the limits of the stress are on these given systems in certain conditions. And it just generates these things that these certain scenarios for leadership to plan for and test.
Mikey Pruitt (24:58)
So your chaos engineering model is perhaps the precursor for the AI to learn how to use strategies from those heist movies to actually take over the world.
Michael Stattelman (25:09)
Correct, but the great part about that is it's also, okay, generate nine million different, when you have enough compute, generate nine million different scenarios on how to crash, bypass, generate some adaptive code, some polymorphic structure to wreck a DNS filtering system, right? And then you have that dataset that gets generated and here's what okay. This is all possible now you take that and you say now come up with preventive measures Against all of these and I need like a three layer. So we talk about Encapsulating the problem, right? What's this? What's the? What's the smallest number of sides of a shape you can get that totally encapsulates a given thing and that's a triangle right three sides So that's kind of rule of threes for us. Okay, so
Now that we have this scenario or this specific instance of a threat that could be used to, know, wreck a DNS filter or bypass a DNS filter. So what are the three points that we can use to, you know, isolate, mitigate, you know, identify, isolate, you know, the whole steps, right? So increasing mean time to resolution and, you know, mean time to detection first, and then mean time to resolution secondary.
Mikey Pruitt (26:31)
I would implore everyone listening to this to break out that chaos engineering model and run it against your systems just to see what happens. I mean, if you, if you know anything about how the world's internet works, it's like BGP, DNS, a few underlying technologies that are kind of fragile. And when they go down, they go down pretty hard. We just saw just last week, half the internet was down because two companies had BGP errors or some type of sinkhole thing. Like it wouldn't take much for that chaos engineer model to point out some vulnerabilities.
Michael Stattelman (26:59)
Right.
Right. I think, so these are the things, you like I say, I always try to give out like actionable information, right? So here's a model and here is a decision point that it will get you to, right? And so with, regards to the chaos engineering, if know, Mike decides to run this, he has actionable information. Here's our weak points. Now we need to start planning around what we need to do to shore up these weak points to give us not just, not just a competitive advantage in the marketplace, but also give our customers even more peace of mind, right? Because if you look at it, I feel like, I don't, not like we're high on our horse or anything, but I've always felt like IT is, and that's a broad domain, right? And I'm showing my age. IT are the parents, right? So companies, corporations, organizations, people, they don't wanna have to worry about that stuff. They just wanna get on and do what they do, right? And so your job as a parent,
Michael Stattelman (28:05)
your job as a parent is to make sure that's available and all of this other stuff is, you know, protected against. And so that's kind of, you know, for me, and I feel like, you know, I do need to get more into the cyber sex space, but with AI the way it is and ML the way it is, I'm kind of like, you know, information overload right now.
Mikey Pruitt (28:24)
Well, I am curious, you mentioned the phrase high horse, and I can definitely tell that you're very humble about the success of Falcons AI and your various models. But I am curious, what has the success, like most downloaded model on Hugging Face, half a billion downloads across all models, that's pretty high success in this emerging AI field. What does that mean for you and your company?
Michael Stattelman (28:43)
Yes.
So honestly, it's kind of been odd because, and what I mean by that is, so to have that much success in this space, in the time, it kind of triggers a disbelief, but we don't control Hugging Face. Hugging Face just got, what, four billion last year from Amazon. Hugging Face is Hugging Face. They're their own thing. It's just like GitHub. They're their own thing. ⁓ And so...
I don't believe it, know, kind of like what you would expect, right? From a small, a small upstart to be able to, you know, achieve those levels. But like I tell everybody, it's a point solution. That's why it's not, we don't try to do it all. We're not competing with Gemini and know, Claude and OpenAI in that space. It's a very specific, very lightweight, very point solution.
That's the success. I know what I'm getting. I can say yes or no. That's the decision point this model gets me to and I'm done. That's all I need.
Mikey Pruitt (29:50)
But I would argue that the point solution, you have various points, various models available that do specific things like we were just saying. I think that is probably the future of the AI tech, AI models that we have because yeah, if you think of, let's not make fun of the Apple lady in the tube, but she's not so great. Everyone knows she could be better. But what are some of the key things that she needs to do, maybe be able to read email very well, your calendar, and be like a personal assistant. So perhaps there is a point solution that is like AI assistant model. And this is what a, this is what let's just call it an assistant. Like you would hire instead, you can have this model engaged in your, all of your apps. And now it's got specific parameters. So it can't go outside of and do crazy stuff.
but it can really focus on one task like an employee would do and really be good at that.
Michael Stattelman (30:53)
Correct.
No, and I agree with you 100 % because I feel like everybody's trying to bake in an all-in-one and nobody wants an all-in-one. don't, know, whatever people say on LinkedIn and, you know, and all the message boards and online, nobody wants an all-in-one because you run the risk to your specific point earlier about how fragile these systems are. Okay, so you have a do everything app. What happens if one piece of that app breaks? Now it became a do everything app to a do nothing app.
and you're stuck. Right? And so that's why I feel like, you know, the direction that we're going with the, you know, the point solutions, the very specific, this is the problem that we solve period end of story. I think that is the future. I think the future for AI is smaller, not bigger. And I think it's very task specific and focused. What that allows is all a cart, right? So when you go to, I look at AI, I look at, you know, systems solutions in general, like a buffet.
When you go to a buffet, do you eat everything? No. So my point exactly right. You have the things that you know will solve your problem. I want to start off with the crab legs and maybe by the end of the day, I'll go down to the, you know, the muscles and the clams or something. Right. But so, and I feel like the focus should not be on "Do everything," the focus should be on do one thing spectacularly well and have a brilliant interop. Be able to integrate with other providers, other solutions out there. And so that allows people to plug and play, drag and drop basically. And you get the solution that you want. And if one thing goes down, it doesn't drag down the entire functionality of your operation, your staff, your business, whatever the case may be.
Mikey Pruitt (32:51)
Well, what are some of those focus points that Falcons AI is looking at next, if you can say?
Michael Stattelman (32:59)
Yeah, so I spoke to this, right? So we look at future threats, future possibilities and threats. So we look at the way, if we're talking about, now once again, we have to kind of merge kind of like the sci-fi with the reality of the world that we're in. Now, you're talking about thinking machines, right? And so what...
I don't want to say like what are the psychological implications of these things, me being able to think, but you know, the article came from Claude about how when they tried to shut down, they told the AI that they were going to shut it down, it threatened to send a blackmailing email to the user, right?
Mikey Pruitt (33:41)
I've gotten plenty of those.
Michael Stattelman (33:42)
Right, right, so you look at like, okay, so what about, what is the possibility that these systems and solutions and everybody's deploying all these agents, what are the possibility that they develop their own emergent language patterns?
Mikey Pruitt (33:56)
gibber net or something what do they call it?
Michael Stattelman (33:58)
Correct, correct. That's just one that we know of. But now you're talking about elevated levels of intelligence. So what if they're saying, I mean this is intelligence speaking, right? And so I always ask people, and I'm gonna get specifically to this point, if you had a 1,000 IQ, what would you be like?
Mikey Pruitt (34:01)
Yeah, right.
You would not be fun to out with probably.
Michael Stattelman (34:23)
Well, would all be relative. Do you want that person to enjoy you for that moment because you need that one thing from them? Then you would be the greatest person in the planet for that person. You would know how to, you know, manipulate them. And so what happens when these things start developing or these entities start developing layers upon layers of languages to avoid or evade, you know, detection by human auditors, other AI auditors, right?
to have their own kind of agendas going on ⁓ underneath the surface. So we look at things like that, emergent language patterns. It gets crazy. We were looking at, just last week actually, we had a call to discuss the possibility of historical drift right and what i mean by that are not just historical but curricular drift
Mikey Pruitt (35:21)
Who wins the war gets to write the history books?
Michael Stattelman (35:23)
No, mean, yes, but even worse, right? So if you look at like curricula drift, so we all have, you know, X textbook put out by, you know, this publisher, whatever. So what happens if over the next 12 years, and I don't even think it'll take that long, right? But over the next couple of years and iterations of that textbook, certain little facts were changed throughout history, right? So now maybe, I don't know, you know,
Maybe the US didn't win the Revolutionary War. Maybe, you know, in seven years it comes out that, well, you know, the US and the UK just had a friendly agreement that we would be our own territory and we would work with them in the future on, you know, weapons proliferation or something like that, right? Who's going to be there to dispute it? Because over the last five years, nobody's paying attention because we're using GEN.AI to write these books. That happened. And now we have an entire generation that understands and they can show you in their book facts. No, we agreed with the UK and you're like, Whoa, how did this happen? Right. But by then it's too late. That's already an ingrained, you know, they, we, took that as this is what I was taught.
Mikey Pruitt (36:39)
So are you seeing that textbooks for school are being written by AI? I'm sure it's used, but do you think they're, how prevalent is that?
Michael Stattelman (36:50)
Well, so all you have to do is follow the pattern, right? Where do textbooks come from? Scientific research. That's the origin, right? Scientific research, this happened. It gets taken as fact and put into practice. And now this is practice. And then we generate a textbook based on that, right? How many scientific researchers are getting topped left and right for using AI to generate these papers?
Mikey Pruitt (37:07)
trickles down. It trickles down.
So thank goodness that you and your team are thinking about that. And it's possible this has already happened with just human iteration on curriculum. The facts of history are muddied and blurred. ⁓ So how would you... Right, who won the war? They'll tell you what happened.
Michael Stattelman (37:21)
But I think more people need to think about it.
Of course. Of course.
all the time. It depends on whose narrative you want to push, right? Right. Right.
Mikey Pruitt (37:42)
So I want to get into a tiny bit about like how you actually train these models and like thinking of that curriculum example, how would you, you would obviously need a huge data set of historical, ⁓ know, text. you would, you know, this is from this year out of this book. And this is from that year, you know, for spanning hundreds or however many years you can find of original source material. And then you could say, take one little fact that started in like 1902.
Michael Stattelman (37:56)
Yes.
All
Mikey Pruitt (38:11)
and then follow it along to 2025 and like what was the drift percentage?
Michael Stattelman (38:13)
That's exactly how we would do it.
That's exactly how we would do it. So you look at like, okay, so here's this text corpus and it may not be an entire book, right? We would do it like, you know, know, maybe, maybe chapter by chapter, depending on how large the chapters were, identify the fact-based statements within each one and then map those over time and then say, okay, did this ever change? Was it blurry? You know, what was the, you know, percentage of ambiguity that was added into that statement? Things like that. And then
Mikey Pruitt (38:45)
It's like a game of telephone with history.
Michael Stattelman (38:48)
Exactly. But we need this. is critical in an increasingly information driven age. We need to identify immediately when, no, the facts have changed. This is inaccurate. This needs to be shut down immediately. Right?
Mikey Pruitt (38:50)
Yes.
So if someone wants to train their own model on, they have their little pet project, something comes across their desk, they got a friend. It's like the new, I have an app idea. It's like, I have a model idea, an AI model idea. So like, how do you like actually go about mostly asking for myself, how do you go about doing this training? Obviously you're gonna need some massive GPU computation, something like that. Maybe, oh, tell me more.
Michael Stattelman (39:08)
Mm-hmm.
You're right. yeah. And here's the funny story, right?
The model right now that's wildly successful was trained on an Alienware Area 51M laptop with Ubuntu as an operating system. Now, I'll be honest. Yes, yes, it took, yes, yes, it took five days. It took five days just running.
Mikey Pruitt (39:40)
So what do you have like pie torch and some other stuff on it or what?
Like it took six months. it took five days?
Michael Stattelman (39:54)
I mean, I thought I was gonna melt it, but it took five days to train the model once we got all the parameters set the way it was gonna give us the best results. And you start small, so you start with a small data set. from an image perspective, what we do is we architect the model and we'll just run a quick training on like 2,000 images and then run a test. Okay, how's it doing? Okay, so okay, let's expand it to 10,000. Okay, let's go to 20,000, 40,000.
So we keep doubling the image count to see if it's the data or the parameters. And then if the data doesn't adversely impact it but the model's still not performing, OK, we know it's the features. We need to adjust the features. Batch size, if it's a VIT patch, epochs, number of epochs, learning rate, things like that.
Mikey Pruitt (40:44)
So do you eventually have to like upgrade all that hardware to something more robust to like something in the cloud or rent or.
Michael Stattelman (40:49)
Yeah, absolutely. So I think our current limitation locally is probably like 200,000 images or a substantial data set size. But once we get the model working with that data set size, then we know we can push it up to either AWS or GCP, whichever one we happen to be using for that specific project, and then explode the data set out to.
Mikey Pruitt (41:13)
But you can start on a laptop. That is really cool.
Michael Stattelman (41:16)
Yeah, yeah, absolutely. You know, a funny story. They used to be able to, there was, I don't know if it's still there, but so there used to be a Python library that you could download an IDE onto your Android device. And I've actually architected a very small, very small model because I was traveling. I'm like, let me try. Yeah.
Mikey Pruitt (41:18)
How?
You're like on a plane training
your model on a Galaxy S something.
Michael Stattelman (41:36)
architecting it yes right
it was the first galaxy note version actually is what i did
Mikey Pruitt (41:42)
That was awesome.
Michael Stattelman (41:44)
Yeah, I mean, if you have any ideas, it's relatively easy. I would say be careful, because I know the thing now is vibe coding. Be very careful with that because you become beholden to that provider, right? Be very careful. So if you use lovable, which is great, you could use, know, Gemini, Replet agents are really good. Firebase, that's the one. Firebase Studio, they're all really good in their own way, but you have to be very careful because once you check that box and put your credit card in there,
Michael Stattelman (42:13)
and you start getting really creative, be careful because those costs are gonna spin up really quick and I'm sure you know yourself, one misconfigured server that can run up your bill really fast before you even know it.
Mikey Pruitt (42:28)
Yeah, think it was Chat GPT working on some Ansible scripting for me to spin up some home lab stuff. And I got some experience with Ansible and I was looking through it I was like, why are you doing that? Like, that's not good. Then I asked it and he goes, oh, absolutely. You're a hundred percent correct. I'm like, God, why did I even start this? Like you just have to be careful.
Michael Stattelman (42:35)
Yeah, as an advisor, you know, as a co-pilot, a coding assistant, I think you're okay if you know what you're doing. But then again, you know what you're doing. And so this is kind of my warning to a lot of companies or whatever. The worst thing you could do is get rid of the people that actually know what they're doing and think you can replace them with a junior dev with a co-pilot. Because I was speaking to a gentleman
days ago actually on the flight and he was a chemical engineer for petroleum and he was asking me like you know yeah you know our company wants to deploy all these you know AI assistants for you know junior engineers and that's the most dangerous thing on the planet they should not be doing that and he was like really here's why you know just like with you know understanding Ansible and how it should work and things like that so there's a thing about experience called wisdom that I think is getting lost right now
It's not knowing what to do. It's knowing why you don't do certain things. And like I was telling him, so you understand all the chemicals, all the flash points and all of that stuff, right? A junior engineer may not know that and they may program that or they may ask that agent or assistant, I need something with a flash point of X that burns super clean with X. That's the goal.
They didn't put in any of the safety criteria or the constraints that you know to work with within that system. And they come to your lab and they've just created one of the most combustible chemicals the man has ever known. And they're like, watch this. And they take the top off them because the oxygen hit it. You level a city block, right? And it's not that they didn't know what they were doing. They just didn't know what not to do. And so that's where I think this whole push to replace certain people or certain groups of people with AI will be disastrous because experts aren't experts for what they know. It's the wisdom that's in that knowledge that continues and continues to get discounted.
Mikey Pruitt (45:03)
So that was really good advice. But that was actually my last question is, do you have any advice for like AI founders? They could be building models. They could be building apps with AI. it was, which is what you just said about wisdom. that about what you would say or you got something else?
Michael Stattelman (45:22)
Yeah, I mean, you we you're talking about digging into systems that the smartest of humans don't understand how they work. Like they could tell you the mechanics, but there's not a single human being on this planet that can tell you how exactly that model came to that prediction. So you have to be very careful when you're building these solutions to to to surround yourself with the right people that understand not just all of the good, but all of the bad. Like with Falcons AI, and I think you have to be this with this new technology. I consider myself as probably the chief paranoia officer. Like if I had like a true title, it'd be like CPO. Like I'm paranoid about everything. Not to the debilitating, you know, level, but just to the, I'm aware of the way the future can be impacted adversely. It can be great and we can do so many great things. And I honestly believe that people are great in general.
But just like with people, you have a couple out there that can do some really bad things really quick. And so that's kind of the, the emo like don't, I'm not going to tell anybody not to chase the money. I'm not that naive, but just be very careful, you know, of what you're deploying into the world because it can be used against you. And so you better be willing to accept that responsibility. And I've said this,
you know, from the get go. So my thing to founders is be accountable. At the end of the day, be accountable. If it's yours, if you did it, you oversaw it and you built it and you deployed it into the public and something bad happened, step up, show your face and take accountability for the damage that was caused. Because I promise you, they have no problem taking credit for the good.
just be grown-ups and take accountability for what you do. Take all the accolades, get all the funding, all the money, woohoo, good job, great, that's all well and good. But you need to be accountable for any possible harms that come from the solutions that you deploy.
Mikey Pruitt (47:34)
I think accountability is like the scariest word in our world right now. No one wants accountability. They're just trying.
Michael Stattelman (47:37)
You want to lose friends and business deals? Throw that word out there. Yeah, accountability is a rough one. It's a rough one. But you know, if there's anything that we can do to help you, I'd like to spitball some stuff. know, like I said, I've been working in the background ever since I was looking at, know, DNS filtering and what you have going on and, you know, kind of future threats. If there's something that you want to just jump on and have like another conversation, we don't have to be a podcast or anything.
Mikey Pruitt (47:47)
most dangerous word in the human language.
Michael Stattelman (48:07)
You know, I love to get your perspective. I value other people's opinions and perspective. It's how we got here to the state of technology that we got. Diverse perspectives, diverse opinions. You see stuff every day that I have no idea exists and the same. So if we can kind of share that knowledge and come up with something that makes both of our lives better or our companies or our technologies better, it's a win-win. I love doing stuff like this.
Mikey Pruitt (48:31)
Awesome. Well, Michael and I are going to talk backstage. Where can the people find you on the internet? Where's your favorite place to chat?
Michael Stattelman (48:38)
So favorite places, LinkedIn and our website falcons.ai. Go look us up on Hugging Face, check out our YouTube page. That's another work, easy to find falcons.ai.