Looking Toward The Future: AI Innovation with Michael Abramov Of Keymakr & Keylabs.ai
26:52 Share

Forget frequently asked questions. Common sense, common knowledge, or Google. How about advice from a real genius? 95% of people in any profession are good enough to be qualified and licensed. 5% go above and beyond. They become very good at what they do, but only 0.1%.

Hello, this is Richard Jacobs with the Finding Genius Podcast. My guest today is Michael Abramo. He's the CEO and co-founder of Keymaker.

or the web site at elab.ai. It's a data annotation platform that allows people to annotate and label data. So my guess is if you have someone's name and phone number, maybe you get their email as well. Or if you have name and email, you get their phone number. That's about all I know about data annotation, but we'll go into that. So welcome, Michael. Thanks for coming.

Hi, thank you. Thank you for introducing me. Yeah, tell me a bit about your background and then, you know, we'll talk about the company. Like, background. So, my name is Michael Abramov. I am a software engineer. Last five, six years, I'm a CEO of Keymaker and Keylabs. It's actually two different companies under one roof. Keymaker is a data annotation service provider and Keylabs is data annotation platform. Yeah, and I wouldn't say it's about emails and phones. It's more about...

We deal a lot with computer vision, actually. So what does data annotation mean in the computer vision sense? Yeah, so imagine you drive, I don't know, Tesla or any other car.

car that has self-driving capabilities. The car has cameras or any other sensors and it has to understand what it actually sees. Like, is it a person on the road? Is it a red light or green light on the tri-foot light? So the camera has to recognize it. And in order to recognize it, it uses some kind of AI, kind of computer vision models that, you know, they see it and they are

understand what they see. So we at Keymaker, we prepare this data, we actually take all these images, videos, photos, whatever, visual data it is, and we actually label objects so that computer vision development companies could take our labelings, the objects on the images, and could train their models and they could say, hey, this is a pedestrian or this is, you know, a dog, a dog, stuff like that. But more complex than I explained, yeah. Yeah, as...

you know, the past couple of years now, Google on my phone, you know, from the pictures, it'll, it'll show me like a slide show of, you know, one of my dogs throughout time, all the pictures I've taken or my son or that kind of stuff. So it's getting better and better. I see. Yeah. That,

That's what we do. I mean, our training helped Google and others to do that. How do you... Are humans doing this? Or you have AI that... Yeah, so like six years ago, it was humans only when we began. But now it's hybrid. And I would tell the different companies do it in different ways. So some companies would do like 90% automatic and then 10% human moderation. We are still old school because our clients coming to us with very...

specific very boutique request so i would say we do 90 manual and 10 automatic things but it's mostly manual process there is a lot of tech behind that because we have to provide manual labelers really really high efficient tools but yeah it's done it's done i know how do you identify i don't know what you'd call this but let's say you're you're trying to identify cats or dogs and picture a

I'm sure there's obvious easy ones and there's ones where it's like, hmm, I don't know. Like, what do you call the obvious easy versus the hard ones? And how do you know what is hard for the AI? How do you establish the boundaries of what it gets right and wrong? Right. That's a very good question. And that's what we're trying to come up with our clients. And as you said, the cat or dog is really easy or like car or motorcycle. Those are, I don't remember when we last did that because

because that is fully automated by our clients and then you know the world can automate today all of these simple obvious object detections we work more on like you know deeper understanding of what's happening on the image or in the video let's say behavioral analysis so you can see a person with a knife so what's your what's your judgment person with a knife cutting things is is it a you

You know, criminal. Maybe it's a chef in a restaurant, right? So you don't know. Maybe it's a... All right. Now, maybe it's a print of chef with knife on a bus and the bus is driving by. Not even like a person with a knife in a kitchen is contact, but a picture of a person with a knife on a bus in the middle of a shooting zone. Okay. Yeah. Yeah. So we have to take, you know, consideration of the...

like bigger context and the smaller context and what's happening and analyze. In many cases, we should analyze the behavior. Sometimes we should analyze the intent. I mean, there is still no behavior, but we know that there will be intent. Let's say you gave me a video where a person walks in a mall and

And after, I would say, two minutes, the person does something really dangerous. I don't know, hits somebody or something else. Now, when we go two minutes before that, we can see that the person's walk or the person's behavior is a little bit aggressive. So we can be more precise on that. We can say, yeah, I mean, his attitude was aggressive because we know the consequences, right? So it's kind of going to the larger... Yeah, so there's the...

If you have images of this person before, you look for that. What was their previous behavior if you have it? What is their current behavior? Does the behavior change the speed of it? Yeah, there's so many factors. Very interesting. Not only that, we can... It's funny because I'm right now at CVPR, it's Computer Vision Conference in Nashville and Tennessee. And I'm...

I just met a person who asked us to do annotations for like data labeling for his casino project. So, and the funny thing is that they want to recognize same person over different cameras and in different times as well. So imagine one casino might have like 3,000 people

like three, 400 cameras and it's small, medium-sized casino. And what we have to do is not only define the person, the object or subject in this case, and define the subject's behavior, we should also relate that same person through all of the different cameras and say, this is the same person that we see on the other camera.

And it might even be in different times. So it's not only like, you know, two cameras at the same time might be like he, you know, the person just walk around and

And then there's also too, like you guys are a hybrid. I'm sure they have these systems look for aberrant behavior. And then there's a threshold. You can set the sensitivity up or down on various parameters and it gets flagged and probably goes to a human for secondary review. And you don't want too much of that. You don't want too little. So I'm guessing you're probably balancing all these factors.

Yeah, this is the difficulty because, yeah, I mean, it's our client's headache, not ours. I mean, they develop the computer vision model and they develop the outcomes of the model, you know, what model predicts the prediction. But it heavily depends on how we did, how well we

we did our job for them, like how precisely we did our job for them to give them the training data set so they could train their model. Yeah, sure. Do you have another example? You know, I know you don't want to violate client privilege, but, you know, what can you say? Like, what other examples do you have that are very interesting? Yeah, oh, so I can tell you my best, my favorite example and it's environmental. It's not from the P2P

people kind of example. So it's from agriculture. Now imagine, do you know what is the most important animal in the world, like for the world to be alive? Like an earthworm? The what? Animal? Like an earthworm or like a phytoplankton? No. Okay, let's say insect. It's a bee. Oh,

Oh, a honeybee. Yeah, yeah. I mean the bees. Yeah, yeah. So if there will be no bees, there will be no plants because the plants grow because of bees. Yeah, thank you. Thank you a lot. Yeah. Now imagine when, you know, when you're agricultural, a big agricultural company and you have your fields of whatever, I don't know, crops or whatever, and you have some kinds of diseases or some kind of weeds or some kind of bad bugs that you want to...

determinate, what you're going to do is you're going to bring lots of chemicals and distribute them over the field and that usually kills all the good things as well. It kills the bees, it kills other insects that are not harmful, it pollutes the environment really badly, etc. So there is technology, new technology right now

It's a big machine with lasers. And actually what it does is it has cameras and it goes on the field, drives on the field, and it recognizes with the cameras, it recognizes bad bugs or

weeds and it just burns them down with laser, but it doesn't really pinpoint it. Okay, so the whole other environment, the animals that live on the ground, snakes, mice, and the bees and all the others that should not be killed are not being killed. I remember I saw this years ago at CES Keto's and I thought it was stupid, but then now it seems like it actually could work.

I don't know why, I just thought it was like so crazy 10 years ago, but now it's not. It might seem stupid, but in my opinion, this is one of the biggest inventions and one of the most important technological advancements in agriculture because you can save so many lives of different species and you don't pollute the land, the ground. I don't think it's a flaw.

anymore. I think it's great. But I remember seeing something like that. Yeah, that's great. Yeah. So like you, you'll burn off their wing or something and you know, let's say there's a particular fly that bothers a plant and then it, so there's the wing gets burned, it falls to the ground and then it's natural. There's no chemical. Yeah. Very cool. Wow.

So what are some of the technical things that you have to consider? Again, it looks like context, movement pattern, I guess, change in shape, change in color, other objects associated with it, maybe its interaction with other living creatures in its vicinity. I mean, what are some of the factors that you'd use to make a model like this? So, yeah, I think you have lots of other examples from defense, which I...

don't like speaking about. Oh, why? Because they're really... They might be harmful to some people to hear about, yeah. But I mean... What are some examples? There's lots of things to reveal what's hidden. Like, I think about it, you, like the person is walking in the...

public space and the person has a gun hidden under their like clothes under t-shirt or even in the bag and there are different features that might hint us about the person having a gun or like some kind of arms and actually we have lots of things like that to reveal the hidden and

And in many cases, more sensors are being used, temperature sensors, sometimes x-rays and sometimes magnetic sensors, etc. And they work all together in some kind of fusion.

And we collect the fusion data and not only the camera data. And we try to reveal all different kinds of things like that as well. So I think that many startups today, our clients mostly are startups. Even if it's corporates, it's some kind of small startup in Qualcomm or in Intel or in AMD. They are still startups because they kind of have, you know, combined capabilities.

combinators inside their companies. So many startups today are trying to work on non-obvious use cases. So cat and dog is obvious or like the person is killing another person it's also obvious or like aggressive behavior also became something obvious for our computer vision to reveal. But now many many many startups try to focus on non-obvious things and they try to focus on preventing things

And sometimes the prevention goes like they can understand, they can prevent things days or months before they happen. And you could ask me how. So there is one startup that focuses on understanding emotions and psychological state of the person by face, by emotions, by gestures, by behavior, et cetera. - You could try using infrared too to see the changes in blood flow, what parts of them get hot.

As well, as well, yes. Yes, definitely that as well. And how much the person is doing, yeah, how much the person is lying or saying, you know, things that...

They shouldn't say it in different places. And then they try to, you know, they try to pinpoint that person and say this person is dangerous by like 60 or 65%, you know. So they try to predict. You remember that movie with Tom Cruise? I forgot the name. Yeah, Minority Report. Minority Report, yeah. So it's not a...

a fantasy anymore. I actually know startups that do that, that try to do that, that try to, you know, they raise money for that. I think that's terrible with the pre-crime stuff. It's just, I don't know, I just personally wouldn't, I wouldn't want to be surveilled. I mean, I understand if you're going into a controlled environment like a casino or I'll play three and I'll put that weapons, whatever, but...

Even this surveilled everywhere I go. Of course, I agree with you. And that's a very philosophical question. And I think many people are, I mean, people that I meet, some of them say, I don't mind. Let them collect my data. They can't do anything to me because I'm clean. I'm not doing all the...

Right, I have nothing to hide, stuff like that. But then I'm telling them, hey, okay, you have nothing to hide, but imagine that in 10 years or maybe in 50 years, your grandchildren or your children will do something. Maybe unintentionally, but they will do something. And by that time, the technology will be sitting in the courts or in all these police departments, enforcement departments, etc.,

and their behavior will be analyzed and your behavior or your data will be also used for context for their behavior. It can be done, right? A police station has this. They're not going to have the surveillance from the police officer, they're going to have it on you. So that people will game the system and get you all rattled or cause you to evoke behaviors that you wouldn't normally evoke. They say, "Oh look, they were sweating" or "They're lying".

This is not... People are foolish when they say, oh, I have nothing to hide because people can twist this tremendously to cause a lot of harm. And there is another behavior... There is another risk here that many people don't think about is impostering. Like, you know, let's say I'm a hacker, I'm a social engineer or something, and I want to behave like... I want to make some harm to you for some reason. Or maybe not harm, but I want to steal something from you and I want to manipulate. Now, how can I do that? I can...

I have your photo. I can print your face and I can bring it

you know, to different places and I can do different strange things while some cameras are looking at me even in casino, etc. So I know that I have injected, now I have injected your face, your ID with some kind of weird behavior into the system. Now, if I know that the system is global and all these cameras, all these casinos are sharing this data and like, you know, collecting them in one place and processing them in one place, training one big model,

So I already have injected your identity with bad behavior into the system without you doing that, without you knowing about that. Which afterwards can cause some problems to you that I can take advantage of these problems that you have. I can call you later and say, hey,

I'm from police and I know that you did this and that and you will notice that I'm not lying because you already have got some kind of invitation to the court or something. So you're going to talk to me. What about data poisoning? I've heard that term. That's exactly the example I gave. Data poisoning is when

It's one of the examples. So data poisoning means I will intentionally place wrong ideas, the wrong data into the processor of the data. Right. So we have to understand that the data is not important without the processor of the data. So there is one big processor somewhere in the cloud or, you know, let's say ChatGPT, let's call it a processor or any other big model.

And when you give any input to that, it takes the input and learns something. Now, as a bad guy, you can give bad input, but it's not just curse words or something. Bad input might be some very smart input that you put there in order to manipulate the whole system. You can teach it to be, I don't know, Nazi or to have some kind of ideology.

ideas that our social environment is not accepting. And that's how you can manipulate the masses, I would say. You want people to vote for some candidate and you have enough capacity to poison the big model so you can write lots of articles and create lots of materials with some idea.

And while knowing that this model will process this and it will see that there is a lot of content with the same idea, so it's going to think that this idea is true. True, yeah. What other, I don't know, what else comes up that makes it difficult? So, yeah, as I said, data injection, data poisoning is something that we have to struggle with. And I

I think that the other big thing that the world will have problems with is, do you know, did you hear about concept save now, use later? Yeah, okay. Oh, you can, yeah, you save data sometimes for years and then it could be used later. Yeah, so it's, the term is called SNUL, like SNUL, S-N-U-L, and it refers to, so we have too much data today and let's say all this video data in 4K and, you know, it's very heavy and without

We don't have enough processing power today or enough electricity to process it all. It's very hard. So we choose to process only significant or meaningful data for us, right? And I would say that it's less than one thousandth of percent of data which is going to the processor. And not all of that is being processed today. We don't have enough power. But what

But what we can do is we can save all this data somewhere because we do have, say, like the storage capacity. And in 10 years, in 20 years, when the hardware...

is going to be much more powerful and it becomes powerful every year. You can see that, you can see that, you know, what Nvidia does and all these big corps create. You can actually take all this stored data and reprocess it and take out some information out of it. So let's say if today I maybe have no enough data to be imposter, like to

look like you while having one photo of yourself or one or some you know your voice etc and it's not enough for me while storing this data in five ten years i can i can do much more with that be much more dangerous right right that makes sense well yeah i mean you'd also have the problem right you you can't create too much data you have to be very lean about what you do you don't want to throw away important stuff but you can't keep everything yeah i mean even within even within a

frame, a video frame, and you designate areas that are important to look at and just throw away the rest? Like how much, you know, for, let's say, if it's going to be used in a court case, I don't know, you know, how much do you keep? How much do you throw away? How do you make sure that your systems are very lean so there's not tons of extra junk in there? Yeah, I think we don't have a, there's no exact answer to that question today. Everyone will find

find answers in their algorithms. And some algorithms are better, some are worse. Some need more data to, you know, have to make conclusions. Some need less data to make conclusions. I think today it's a little bit early to say because all these algorithms, they don't work really like 100% or even 90%. But in five years, I believe we're going to be

in the much more dangerous world and we should be very cautious. I don't know how to be cautious to be honest. So some of the reference that can, for example, I saw this lady made it out of like 100 different license plates and the automatic license plate reader, I couldn't read it. It looked like there were like 30 different cars there, you know, and it,

it wouldn't work. And she made another piece of clothing with like partial faces of people all over it and the facial recognition wouldn't work either. Yeah. So it was like an anti-surveillance device. Do you do any anti-surveillance? Yeah, yeah. There is a lot of development in that as well. So you can, I saw some guys printing some, you know,

you know, some patterns on the t-shirts. So the whole surveillance system doesn't see him as a person. It just ignores him. I mean, he goes in a crowd of people. It can recognize all of the people except him, even though he looks like, you know, two hands, two legs, a head, a face, everything is... Nothing's hidden. But because of this pattern printed on his t-shirt...

The system doesn't recognize him. All of the things, they will fade as systems become smarter and smarter. Yeah, and I think that it's not going to help us. We can't hide, but we can't hide from this technology. And we are continuing building it. And the only thing that, in my opinion, we can do is we should be...

aware of its capabilities in not today but in years we should always think okay today I'm giving away my voice I'm giving away my face I'm giving away my other things what can be done in 20 years with that what can be done after the laws change because today the laws say you know the court cannot use any of these things they have to you know look at me and ask me questions and speak to me before they accuse me in you know crimes or something but you don't know that

what's going to be in 50 years? Maybe in 50 years you're going to be automatically judged by AI even without knowing that, right? So we always have to think ahead of time and try to understand what's going to be then. Try to simulate that situation. But all this, I would tell you another thing. All these corporations, all these startups, all these people, they don't care about it at the moment. They are, you know, we're engineers, we're scientists,

scientists, we just, our biggest goal is to come up with something, you know, to invent something, to find some formula, to solve some problem. We don't mind how it's going to be used. It's like Oppenheimer, he knew he's doing, I mean, at some point he knew he's doing the nuclear bomb and other people they knew, like before him, they knew that, you know, you might create some reaction that will destroy a lot. But they kind of, they said, okay, if I don't invent it, somebody else is going to invent it.

Why would I stop? It's my ego. It's my publicity that I want to get when I invent this big thing. The same happens to startups today and the companies. They don't want to think a lot about where it can be taken to the bad place. They just want to develop new technology, which is amazing in my opinion. Well, I mean, it's short, silent, and foolish. Yeah.

I mean, we people are hard to be, you know, destroyed. So we will have to somehow struggle our own technologies and somehow to make sure they don't destroy us. It's a feature. But I believe we're going to survive somehow. All right. Well, what's the best way for people to find out more and follow up? Where can they go to find out more about your company? Oh, is it about the company? Yeah. I just wanted to... We're, you know, getting close to the end of time. So I wanted to find out again what was the... So if...

If you want to learn about our company, please visit our website. And yeah, I mean, Killops AI or Keymaker.com. We do data annotation, we do labeling, we train, we help training models, computer vision models, LLM, like large language models, everything. So I don't know, reach out to me in LinkedIn. Well, I do want to ask you something. We have been asking all the AI interviewees, what has changed in the past couple of years

to suddenly allow for LLM to work and allow for machine vision to take off and all that. It seems like AI has been around for a long time, but now all of a sudden it's just so much better. So is it more computing? Is it more layers? What is it? You know, I think all the answers are correct, but I have one answer that is more correct than others. It's democratization. We democratize that. We gave it to the people.

Okay, we gave it to everyone and people are amazing. People are smart. Even those who are not PhD, even those, I mean, you don't have to be a super intelligent PhD professor or Nobel Prize scientist in order to, you know, come up with great idea. And at the moment, the AI was democratized and everybody has access to it and everybody can come up with doing what they enjoy.

you know, their idea, like implementing their ideas, I think that that's the biggest boom in AI. That's the biggest explosion. That's the biggest reason for explosion of AI. Yeah. Yeah, like 3D printing and you will make all kinds of models. Yeah, that makes sense. Everything. It's like industrial revolution. When everybody had access to it, it just blew up. Because industrial revolution, you know, it began tens of years before it blew up, exploded, I mean. Yeah.

Gotcha, right. And the same with internet. Once it was democratized, you know, it just got other colors and other, you know, distribution level. Well, very good. Well, Michael, thank you for coming on the podcasting. It's very interesting. And I could see why you laughed when I said emails. And I was thinking of like data annotation at the lowest level. But, you know, now it's a whole nother world. Yeah. Very good. Yeah.

If you like this podcast, please click the link in the description to subscribe and review us on iTunes. You've been listening to the Finding Genius Podcast with Richard Jacobs. If you like what you hear, be sure to review and subscribe to the Finding Genius Podcast on iTunes or wherever you listen to podcasts. And want to be smarter than everybody else? Become a premium member at FindingGeniusPodcast.com.

This podcast is for information only. No advice of any kind is being given. Any action you take or don't take as a result of listening is your sole responsibility. Consult professionals when advice is needed.