See https://www.youtube.com/watch?v=0KQdivzfXdk

Transcript:

Chapter 1: Introduction to New AI Model

Hey everyone. We just finished training our brand new AI model and I'm really excited to share uh what we've been working on with you. Uh this new model isn't just a small tweak. It's a complete rebuild of the AI model or AI analysis that we do. So let me take a step back. Um, what I'm talking about here is when we scan the screenshots trying to find bad content, the the tool or the thing that we use to do that is called an AI model, right? So, we're using an artificial intelligence model to do that. Now, we did that previously as well. This new one is just an updated model with a completely new foundation built from the ground up by us and it gives us a lot of flexibility. Um, but we trained this thing on over 15,000 hand annotated images. Um, and the results are looking really promising so far. Um, so what does this actually mean for you? Well, first uh you're going to

Chapter 2: Model Performance and Benefits

see significantly better accuracy uh in detecting like bad stuff in the screenshots essentially. Um, so we're talking about more reliable detection across all different types of images, all different types of apps and browsers and and uh things of that nature. Um, it is also going to mean that there's going to be fewer false negatives and fewer false positives as well. At least that's our hope. Uh, based on our testing, that seems to be true. But we won't know for sure until we get it in your hands. Um, but the really cool thing is, uh, with this new foundation, we can continually improve this model over time as well. Uh so as you guys provide feedback about the areas in which you're finding it's working really well or not working well we can actually uh make adjustments and improvements there which that is something that previously we weren't really able to do. Now one thing that I want to point out uh is we have always

Chapter 3: Detection Capabilities

prided ourselves on our AI algorithm including the old one uh its ability to detect small details and screenshots when we've done our testing and compared uh with other products which we haven't done for several years but uh customers reach out to us and let us know like when they test these models when they test these accountability or parental control apps they find that ours tends to catch more things um especially in the small det details section. Well, it was true for our old model, but with this new model, it's even more impressive, right? Um, so we've anyway, we're just really excited about that. Um, and the the reason why that's important is because we don't want there to be a simple loophole where it's like, yeah, you just make the image small, you know? So, so this new model will continue to be really exceptional um in in that area. So, I'm going to touch base here just if you're a little bit

Chapter 4: How AI Models Work

curious um on how we built it. So basically the way that these models work uh the the way that all artificial intelligence uh works if it's called an artificial intelligence algorithm essentially what it is is it's not somebody writing specific code for looking at specific things if that makes sense. Essentially what it is is you are giving it uh data and you're telling it showing it examples of what to detect and what not to detect. And through doing that the algorithm actually teaches itself how to detect or not detect things, right? And this has really improved um you know like back in the day in order to detect bad content like say a nude photo or something um like a lot of the algorithms were just like looking for skin color and doing really basic rudimentary stuff. Um these new AI models in the last decade or so have really improved it significantly. Um and the reason why I want to point this out though is that it's not all positive.

Chapter 5: Model Limitations

So, so what I'm going to say here is uh sometimes we'll have customers reach out and they're like hey like why didn't it detect this thing and the honest answer is like well we don't know right I mean we'll keep working at it but the reason why is because like we just feed it like I said 15,000 images and these images show it examples of things to flag and things not to flag right now if that thing that you're showing us is an example where you But we like it's just not represented in those 15,000 images, right? Well, it the answer's kind of easy. We just need to go and get images that look like that and add it to our data set and then train the model again and then it'll be improved, right? But if it's, you know, an image where it's uh already kind of represented in the data set, it gets a little bit tricky to just tweak the model to now handle that image better essentially. Um anyway, so I know we've had the occasional customer reach out just kind of disappointed that we can't just like fix it and improve it. And it's because it's uh it's a little bit of a I I don't know how to I I don't really know how to describe it other than it's just like it's a little bit messy and a little bit complicated. Um but I did want to point out uh we are looking for beta testers to help test these models out. So, if you would like

Chapter 6: Beta Testing Request

to help us out, if you would like to help Truple out, if you'd like to help all the other customers out there, especially, you know, there are some customers who aren't very tech-savvy, um, and they rely on our products, right? And when we've got people willing to beta test, it allows us to kind of prove the model, prove the new feature update with customers who are willing to, you know, put a little bit more effort in, right? Because so beta software generally speaking, if you don't know, beta just means like we think it's good. We've gone through our alpha testing. Alpha meaning like AB uh you know beta being the B. We've gone through our alpha testing internally. It looks good. We're ready to put it in the hands of people to let them test it, but uh you know, it might still have some rough edges. Right now, we've done a lot of testing on this model. We hope that those rough edges are smoothed out, right? But uh you know, you never know until you actually get it in the hands of thousands of customers and then and then we get feedback. So we are looking for beta testers. If you would like to beta test this new model, please reach out and let us know and we will be thrilled to have you help us. Um not every platform is ready for it quite yet. Um iPhone, Android, Kindle, and Mac will likely be the first platforms, but we'll get the other ones added pretty quickly. That's it for this video. Hope you guys are doing well. Have a good day.

To beta test, follow these steps: https://support.truple.io/articles/macos/macos-migration