AI After Dark

00;00;00;15 - 00;00;26;02
Unknown
So. So we were just talking about before we started this was, was, that your career didn't start anywhere from, like, a traditional path of data. ML now transitioning to AI because there's big things like nobody has AI experience, you know, in their hiring, right? Now, building an AI native product in the recruiting space. So what? What convinced you that you were the right person to do this?

00;00;26;04 - 00;00;53;26
Unknown
Yeah, well, I don't with my wife. We have this conversation where you think she thinks that I'm no romantic? Because I said, well, you could have married at least ten guys that I know, and you would have been fine. And she's like, well, that's very romantic, but but this is kind of the same thing. The way I look at things is I have a problem in front of me, and then I look at what tools are available to solve the problem, and then it just the tools for the problem.

00;00;53;26 - 00;01;14;03
Unknown
And if there is a new tool that I don't know how to use, just learn to use the tool. And then you solve the problem. So I'm driven in that way, in whatever I have in front of me is like, okay, what is the problem in this particular scenario? This is the problem. Okay. Let's do some research about the problem, what the tools are or what things are out there that can help us to solve this problem.

00;01;14;05 - 00;01;34;16
Unknown
And then you you get on the tools actually first we we Santiago, which is my partner in this in this company we he's been doing recruitment for more than 20 years. So he really knows recruitment really well. I was his client before we became friends. He was doing recruitment for me when I was a CTO in one company in New Zealand.

00;01;34;18 - 00;02;03;12
Unknown
And we been having conversations for years about all the things that are broken in the recruitment industry. And a few years ago, in one of these conversations, we were like, well, that is the problem. This is the way we could solve it, but the way to solve it required AI. So I just went and did a post grade, studies in University of Texas just to get that post grade so I can understand AI so I can solve this problem.

00;02;03;12 - 00;02;19;21
Unknown
So it's really kind of the problem comes first. Then you understand what the tools you need to solve that problem. And then you either master tools or you learn to master those tools. So you solve the problem. That's how I ended up there. But right now, I mean, the tools are evolving so fast. Oh, yeah. So quickly.

00;02;19;21 - 00;02;48;23
Unknown
How do you. I mean, I feel like that's a very nice, structured way. Yeah, of being able to solve a problem. But do you have the time to do all that or how do you keep up? Well yes. There are like two different levels in that. One is the tool is a. Yeah. An AI is basically, it trick is a mathematical trick using statistics that can give you the illusion of intelligence is really artificial intelligence is a misnomer because it's a glorified averaging.

00;02;48;23 - 00;03;11;22
Unknown
And people that know about AI, they probably going to be catching the pearls on my comment right now, but I see that in a context. So you have a lens which is large language models. You have all this corpus of text, and they are really good. And in kind of predicting what the right word will be to follow the word that they just use or you're using to ask the question or whatever, that is the tool.

00;03;11;24 - 00;03;39;24
Unknown
Now, all the other things that are evolving really fast and they're changing all the time. It's Citra is how you use that tool. So as an example, in the field of graph databases that they are very closely use with models for certain applications, you have new graph databases popping up and closing all the time, and every one that comes could be the unicorn that solves this, because the ones that are well-established are really slow.

00;03;39;24 - 00;04;05;04
Unknown
Like neo for J, for instance, is really well established, solid. You can always go for new for J, but it's relatively slow. So for some applications they don't give you what you need. There are a lot of attempts to solve the problem and give you the ultimate graph database. And they pop up and I so quickly that when you need to choose one for your application is a challenge because it's like, okay, what do I do?

00;04;05;04 - 00;04;28;08
Unknown
I go for neo for J, and I pay the penalty of slow or whatever, or B is heavy Java, or I go for the latest and greatest. I promise to be the solution of the problem, but in three months they may not even exist anymore. So that is a problem. But it's a different label. When you're thinking about the solution, you're thinking, okay, I'm going to use an L model to solve this part, and then I'm going to use a graph database to solve this part.

00;04;28;08 - 00;05;02;12
Unknown
And I'm going to embed the vectors here to do retrieval amended or whatever. And then when you go okay, which database am I going to use for the graphs. Which database I going to use for the embeddings. What framework am I going to use to handle the alarm and all the agents that I need to create? That is the part that you need to balance between being conservative enough that you don't go on the bleeding edge because you're going to bleed, or being a little bit more edgy and saying, okay, you're going to just try this because I know that I can replace it, for instance, relatively painlessly, if I have 2 or 3 things

00;05;02;12 - 00;05;22;21
Unknown
that are if it is open source, you may always keep it, even if they don't support it anymore. If it is something that is well designed, you can have an API that is relatively cheap and easy to replace. The the tool behind that API. So there are some techniques at the implementation level that can solve that. The problem.

00;05;22;24 - 00;05;51;01
Unknown
My personal approach to those things is I will not use the latest and greatest and most edgy, because the risk is too high, especially for a small startup that you don't have the bandwidth to say, oh, I guess we are going to have to re-implement all of that. But I tried to not go with the too old kind of, tool because it's slow sometimes, or I mean, usually if it is old and people are trying to replace it, there is a reason for that.

00;05;51;03 - 00;06;10;06
Unknown
So probably not the best tool, but you don't use the latest and greatest right to go in the middle is like, okay, this one has been around for a while. Probably they're going to stay. They already have big customers. They seem to be fine and it's good enough for what I want, even though it's not the shiniest. Now, the other problem with those is some of them are really expensive.

00;06;10;08 - 00;06;31;13
Unknown
So you also need to. So give me give me a real example. I mean, so I guess as context, your CTO of a company called Nomi. Yeah. Nomi is building an AI native recruiting platform. Yeah. Give me an example of a situation where you have transitioned like the it almost feels like a hub, you know, the centralized piece that goes and calls all the different elements that it needs to be able to go execute?

00;06;31;16 - 00;06;54;16
Unknown
Yeah. So for instance, when we started, to see the implementation phase of the company. Right. Basically what do we do? So we understand the context is we take long format interviews with people without any particular job position in mind. So if we interview you, for instance, we're not going to be interviewing you for this particular position in this particular company.

00;06;54;19 - 00;07;17;06
Unknown
It will be, hey, we interviewing you, Alex. We want to know your career, your trajectory, your experiences, what you've done. What are the different challenges that you had in your professional career, how you overcome those challenges that that's what we do is we not focusing on this or that job description with all that information, plus, of course, your LinkedIn profile and your resume and whatnot.

00;07;17;09 - 00;07;42;25
Unknown
Then we ingest all of that and we use that as a as a representation, virtual representation of you. So we can do a few things like generating a particular resume for a particular job description. Now. So that's where possible possible job description comes into play. And we ask the platform, hey, given this job description and all you know about Alex, how will you write a resume for to apply to that job position?

00;07;42;28 - 00;08;08;05
Unknown
We can generate that. We can also give the potential employer a chat box that you can ask questions. And of course it's not you. Is really the AI answering the questions on your behalf, but using all the information that we have. So that's in a nutshell what the platform does. Also, employers can search candidates using natural language, and the platform can go on all the embeddings and etc. etc. to find those candidates.

00;08;08;05 - 00;08;27;04
Unknown
So that's in a in a nutshell, what it does. When we started with implementation, one of the first question was, okay, cool. We know when to do. Yeah, because it's basically a huge corpus of text that we want to make some decisions based on that. And that is the realm of AI. Fine. Now do we use it from Google?

00;08;27;04 - 00;08;49;02
Unknown
Do we use long chain? We do. We use llama index. Now there is a new one from Google called 8-K. A genetic development kit that is really doing the rounds. That's now. But back in the day, those are the three that we looked and considered. We tried the three of them. We did proof of concept with those and we ended up using llama index.

00;08;49;05 - 00;09;12;04
Unknown
The main reason is llama index is one of together with long chain is one of the two most well-established, platforms that are out there. The frontend we were building was, of course in react and that is TypeScript. So a llama index had the library in TypeScript and we thought, okay, we can keep it simple and everything in one language, the easier the better.

00;09;12;06 - 00;09;29;19
Unknown
Which is again, another thing that I like to do is what is the simplest possible solution for this? Not the shiniest, not the most complicated, nor the one that I can show how crazy I am about. Technology is just what is the simplest way to solve this? But it's going to work. Yeah, we were talking about that, five minutes ago.

00;09;29;20 - 00;09;50;00
Unknown
Boring is good. Yeah. So lemma index had. Well, Lang chan also has a TypeScript, implementation, but my index had a UI as well. Is my index UI, which gives you a chat box already built for you. So we were like, cool, that's the one. Why? Well, because he's been there for a while. He's not going anywhere. Clearly is not going anywhere.

00;09;50;02 - 00;10;14;16
Unknown
It is actively maintained. So new things are being incorporated all the time. Now, we started implementation in TypeScript and then I realized very soon, okay. The TypeScript version of my index is like a reduced version of the Python one, and I started to have some problems interacting with the database. And then I went to discord, which they have a discord channel.

00;10;14;16 - 00;10;41;13
Unknown
So they're really, really responsive with, with the community. And I was like, hey, I'm having this problem interacting with this database. And the answer from their CTO is like, oh, we haven't tested that. So if you find the problem and you can fix it, please do. The give us the PR and I'm like, cool, I'm not that good to be undertaken in your platform to mess around with your platform to fix it.

00;10;41;16 - 00;11;03;02
Unknown
And I don't have the time, honestly. So we decided to move everything to Python. Why? Well, the penalty is now you need to do TypeScript for the frontend, and you need to do Python for all the AI backend. But the library in Python is way more stable, mature, well maintained, widely. Use all the new features that they released.

00;11;03;02 - 00;11;23;02
Unknown
They released them first in Python, and then if you're lucky, they're going to port them to TypeScript. So it's like TypeScript is like the second class citizen in the ecosystem. So we ended up choosing Lambda index because it came with a UI. Admittedly, I could have used the UI with any other framework, but again, why you're going to mix and match if you can just keep it simple.

00;11;23;02 - 00;11;46;12
Unknown
Use Lambda index. All this stack, use it in Python, which is what clearly they, they prefer. And yeah, that's how we ended up choosing those technologies. Now 80 K from Google is really promising. I've been playing around a little bit with it, but again, you need to now justify why are you going to migrate your entire application or you're going to have two frameworks.

00;11;46;12 - 00;12;09;13
Unknown
You may have some features supported in lambda index and some others supported in 80 K. Why? I mean, simplicity for me is key because otherwise it's really a nightmare to maintain. When you hire somebody to work with you, they need to be familiar or learn more things, especially in this environment, which is not like you can go and say, hey, I want somebody with five years of experience.

00;12;09;13 - 00;12;35;17
Unknown
In 80 k K was released. I don't know how long, but not not long ago. So whoever comes along, it doesn't have much experience or haven't played with the tool at all and they need to learn it. So keeping it simple, I think that is also very important, especially in the AI environment where there are a lot of new frameworks, new tools, everything is moving all the time.

00;12;35;19 - 00;12;53;09
Unknown
And the other thing to your point of how do you look at an environment that is so, changing some of the change in SA, for instance, a year ago, everything was about the prompt prompt engineering, which is, it's funny for me that we put the name engineering to something. We make a thing out of it.

00;12;53;10 - 00;13;18;27
Unknown
It's like, write a good prompt. That is not engineering. It's just do a good job of writing. Prompt. But it was like, yeah, prompt engineering is going to do the trick. And then now we are like, no, no, no, no prompt is a thing of the past. Now we're working on context engineering, which is I understand and I appreciate the difference, but it's like a difference without a distinction because the name will receive both prompt and context.

00;13;18;27 - 00;13;38;23
Unknown
And context is kind of part of the prompt, because it's basically all the information that you give the limb to point it in the right direction. So when the limb starts to make educated guesses about what you want, it has a good vector behind that is pointing in the right direction. Basically. And you can do that with prompt or with context.

00;13;38;25 - 00;14;00;10
Unknown
But they are not two fundamental different things. It's different how you construct them, but they are not different once you put them in front of them. Is this is this how I guess you would describe a reinforcement learning then is is like you say, okay, AI is basically just an educated average of what you what this typical next thing should be agreed.

00;14;00;12 - 00;14;20;18
Unknown
But then you've got now the ability to like have the system learn on its own what you when you say to learning is actually just like contextualizing within a particular channel or niche. Is that is that the idea or how would you define reinforcement learning? Yeah. So in reality, the proper reinforcement learning happens before you create or when you create the model.

00;14;20;21 - 00;14;46;06
Unknown
So if you're running your own models, you can do learning. If you're not running your own models, then technically you're not doing learning. What you're doing is in any case, context engineering. What is the difference? The difference is when you build a model, you are training that model. That's when you train it. You don't training afterwards, once it's built, is built, you can train it, but you have a new version of the model.

00;14;46;09 - 00;15;05;04
Unknown
And you need to own the model to do that. Training the model is super expensive. You can train a model if you want that model to perform a very specific thing. You could potentially train your own model. You have a hugging face, which is a platform that has thousands of models you can go and use.

00;15;05;06 - 00;15;30;22
Unknown
Those models are good if you want to perform one specific task, and you could, if you wanted, go and train the model in a particular, field. Let's say. So that will be like I grab let's say let's start from scratch. I grab an empty, neural network, and I start feeding the neural network with, let's say, contracts in the oil and gas industry.

00;15;30;25 - 00;15;56;18
Unknown
So that neural network is doing all the all the parameters for contracts in the oil and gas industry. And then the the feedback that you're talking about is me as a human reinforcing the right things. And getting rid of the wrong thing. So let's say that the model is learning. It doesn't understand what is right, what is wrong, what is proper, is not a lawyer is just learning correlations between words.

00;15;56;21 - 00;16;24;00
Unknown
And some correlations are useful and some correlations are not useful. So you have a human in the loop saying okay this is useful. This is not this is useful. This is not. And the model is going to penalize the relationships that are not useful. And will reward the ones that are useful. You training the model. Once you train the model with this corpus of text, now you have a model that is really, really, really good at working with contracts in the oil industry.

00;16;24;03 - 00;16;44;22
Unknown
So once you have that model, then you can use it for, for instance, generate contracts, new contracts. You can give the model instructions. Now the problem is that model is going to be really going to suck at many old things. It's going to be really good at that and probably what it gives you. You're going to have to for the process that to get something meaningful.

00;16;44;24 - 00;17;10;29
Unknown
That's where the the generic large language models come along, right? You have things like, grok or OpenAI or whatever, Gemini from Google. Those are generic, like large language models. They've been trained with everything that is extremely expensive to do. And those things are good for everything, but they need a context because otherwise they know too much.

00;17;10;29 - 00;17;33;05
Unknown
Let's say, and that's where the context engineering comes along. Now, you can call that training. The model, while technically is not really training the model, but is giving the model the boundaries and giving more context about what you want. That's why, for instance, the prompts typically are you are a Java developer with ten years of experience, you're really good at Spring Boot and blah blah, blah.

00;17;33;12 - 00;18;07;16
Unknown
What you're doing is saying, okay. Forget about everything that doesn't include that word in that context. So the M is focusing on all the content that it has, that it knows Cervantes and he know Shakespeare. It knows all sorts of things. But now it's saying, okay, cool. All of that doesn't is not really relevant. You're going to focus on the text that I have in my, in my parameters that are related to Java, Spring Boot, Kubernetes or whatever you are putting in your context.

00;18;07;19 - 00;18;38;26
Unknown
And that's how you help it to be so specific. So you can call that training is not really training. And some people call it training, but in my opinion it's not. Yeah. Yeah. And so if I think about intelligence to your point is like it's not really intelligence, intelligence is the ability to take, a lot of experiences or different, data points prove points or whatever, and cultivate something, you know, net new, right, where you have an assessment that you say, okay, I see a trend that no one else has seen, or I see an ability to, synthesize this in a different way.

00;18;38;28 - 00;18;58;27
Unknown
Is that why you think elms are never going to be intelligent? Because when you're doing this, you're basically telling it, forget whatever it is you think doesn't apply to Java in this case, when the reality is like, I would like to know if you see something new that I haven't seen yet, that should be related to Java and isn't today, is that how do we get there?

00;18;59;00 - 00;19;18;18
Unknown
Yeah, there is a there is a component in human intelligence, which is recognition of patterns that let them do really, really well. And that's why there is people. And this is an opinion. People may have different opinions, but these people that have the opinion that they are genuinely intelligent and it's just a matter of degree, right. They not as good as we are at recognizing patterns.

00;19;18;25 - 00;19;46;24
Unknown
And that's why they are not yet on the general AI, which they're looking for. So it's a matter of degree. I don't hold to that opinion. I think that pattern recognition is part of what makes us intelligent. But there is something else. There is a component that in philosophy goes into qualia, which is we can talk philosophy if you want, but that is a different field, which is beyond pattern recognition is something else that we can put into it, that it goes beyond pattern recognition.

00;19;46;27 - 00;20;09;27
Unknown
The eyes are always going to be restricted to pattern recognition. They can't do anything else because that's what they do. That's all they do. That's the logic. But they are so good at it that they really, really look like if they were intelligent. Now there are concepts that don't belong to models like right or wrong or good or bad that doesn't belong to models.

00;20;10;00 - 00;20;30;29
Unknown
That's why when they train the models, going back to training models, part of the human loop that is involved in training, the model is giving the model a certain sense of appropriateness or good or bad, if you will. Right? And depends on how you do that, you're going to get a model that is a little bit opinionated and you can get political.

00;20;30;29 - 00;20;52;25
Unknown
But, Google was famous when they released the first version of Gemini because if you ask Google to give you the Pope, he was giving you female popes, all sorts of or or, I don't know, a German soul during World War two. And they were coming with black soldiers in Germany in World War two. There were no black soldiers because they train it to be, not white.

00;20;52;25 - 00;21;13;28
Unknown
I mean, they kind of they want to have it in this concept of having a variety of colors. Some people and the model is like, okay, you want variety of colors. Some people like, and I give you variety. So all of that is given in the human in the loop training that we give it afterwards. So all those things don't belong to a model.

00;21;13;28 - 00;21;40;22
Unknown
The model, the only thing it does is pattern recognition. But he's so good at it that he really looks like he's intelligent and he's superhuman in pattern Recognition. They really, really can see patterns that we cannot see because of the amazing computing power they have. One of my, partners, Adrian Faustino, he he he theorizes that in five, maybe ten years, the most important studies will be actually philosophical, right?

00;21;40;24 - 00;22;02;22
Unknown
Because to your point, it goes beyond the capability of of what is commoditized now, which is pattern recognition. So I do actually want to talk philosophy. You mentioned qualia. What what what elements of philosophy do you see, separates human from AI? Yeah. So if we're going to get into that field, which is for me, it's very interesting.

00;22;02;22 - 00;22;24;29
Unknown
The first thing is I don't have a degree in philosophy. So whatever you sound like an LLM, you're like, yeah, I'm just I'm just making sure that we put this in context. I'm not a philosopher. I'm not. I don't have any of that. But, it's interesting to see the evolution of ideas. I saw a recent, but not recently, but, there is a conference that is called ideas have Consequences.

00;22;25;02 - 00;22;48;06
Unknown
The place with the phrase election has consequences, but this is talking about ideas. And in that presentation, which for me was an eye opener, it's like, okay, in this particular year, some philosophers came up with this idea, and that was the training idea. And everybody started exploring that idea. And 50 or 100 years later, society was shaped based on those ideas.

00;22;48;08 - 00;23;09;23
Unknown
But then some other philosopher came up with a different idea, and we started exploring that. And 50 or 100 years later, society was organized on those ideas. And then you realize that philosophy is always ahead of us in the sense of what you take for granted. What you take for common sense today is really what the philosophers were theorizing if you decades back.

00;23;09;25 - 00;23;38;24
Unknown
So if you look at the world with that, which is not just AI, it has to do with everything we are living. And of course, that process is accelerating because of course, the capacity we have to transmit, transmit ideas, transmit information is way faster than before. So these phenomena happens faster than before. So if you start to look backwards today, we live in in the philosophy of the 1950s, 19 early 19th century.

00;23;38;26 - 00;24;05;25
Unknown
So starting when Nietzsche and all those guys actually, both fascism and Nazism are somehow derived from from that philosophy, which is, there is no God and therefore there is nothing transcendent, which means that humans are the top of the chain in any case, and whatever I can get away with is the supreme good. Because if there is nothing above, then whatever I can get away with is the supreme good.

00;24;05;25 - 00;24;36;25
Unknown
And this fine. So when you put that into practice, you get the 20th century. And now we are more into the, French philosophers of mid to 1920 century, which you have all the post-modernism and the construction of reality, et cetera, etc.. So we live in a very deconstructed society where you have conversations where people will say naturally, well, that's your truth, but it's not my truth, which in reality doesn't make sense because truth can cannot be more than one.

00;24;36;27 - 00;24;56;12
Unknown
We can have different opinions what the truth is. And that's where philosophy comes in. And it's like, okay, let's have a conversation and let's see if we together we can discover what truth is. And that creates a society that is a little bit more united. Because even if we disagree, we are both trying to find the truth. But as soon as you like.

00;24;56;12 - 00;25;18;23
Unknown
No, no, you can have your truth and I can have my truth because we are now deconstructing reality, because we are postmodernists. Then suddenly we don't have anything to talk about. You can have your truth. I have my truth and what is left? Well, whatever I can get away with is what? So I'm going to try to impose my truth over yours, not because it's better, it's just because it's mine.

00;25;18;25 - 00;25;33;25
Unknown
And now we have a society that works like that. Everybody have their idea of what the right thing is, and nobody wants to talk to their side and say, hey, let's sit down and see if we can arrange. No, no, no, they don't need to talk to you. I just need to push for mine. That's what it is.

00;25;33;27 - 00;25;58;14
Unknown
So what has that to do with I. Well I that's pattern recognition people training those eyes and doing the the human in the loop is people that we are all embedded in that philosophy right now. So you have models that are more inclined to one truth in current terms, and some other models that are more inclined to different truth.

00;25;58;16 - 00;26;18;07
Unknown
And when you what do we tell them? You need to take that into account because it's going to react to that. You can ask the general aim of your preference and start asking uncomfortable questions and you see what you get will be really different. But he's not because he is. Recognizing those patterns is because the Liliom was training.

00;26;18;07 - 00;26;41;00
Unknown
Read it. And then for the trained by somebody to have a particular moral view of reality. Like if you think that diversity is good, I'm not saying that if it is or if it isn't, but if that's what you think, then you're going to train your them to have diversity as a value proposition, and they will respond accordingly because it's just following their training.

00;26;41;03 - 00;27;03;00
Unknown
Right. So all of that is philosophy, but is it's pattern recognition. Yeah. That's wild. I mean to think to think. You're right. We're in a we're in a day and age where we all feel because ideas are so easy to disseminate and distribute. We there's a lot of power in that. And so and sometimes that will corrupt us and think into thinking, oh, man, what I think is the right way.

00;27;03;00 - 00;27;26;17
Unknown
And everybody should follow that. As, as AI becomes more powerful, more expansive, faster, everybody's going to be a philosopher, you know, it's like I say influencer, but really it's just a philosophy, right? Of sorts. And and then how does that impact society in a world where you're like people are constantly espousing what they believe is like core core truths, right?

00;27;26;17 - 00;27;52;06
Unknown
Like there's, the dude, is it a balloon? You know, like this? It is. When does it pop? And we realize, like, oh, no, this was dumb. I don't know, really. I mean, I can see what you're saying because we need to compose that with the capacity we have nowadays of create a bubble around ourselves. You go to any social platform with your preference and that social platform, we're going to start learning from your preferences what you like, what you don't like, what you react to.

00;27;52;12 - 00;28;18;09
Unknown
And he's going to start creating content tailored to you. So you end up more and more separated in that bubble where your truth, again, truth is one. But we believe that it's not that you have my truth and everything that the internet is showing me seems to agree with me, so that's really cool. Therefore, the guy that have a different idea he must be an idiot, or must be dishonest or must have an ulterior motive.

00;28;18;11 - 00;28;38;02
Unknown
Because clearly I'm right. Everything, everywhere I look, it tells me that I'm right. And that is really, really dangerous. I think that we if we just know if we don't snap out of that, that is really a one way road, because we're going to end up in a solipsistic society where on top of that, we now more and more are working from home.

00;28;38;02 - 00;29;10;08
Unknown
So we don't get to the office and have to deal with the guy in the idea that they have different ideas and having coffee with him, you just in your house working on your own. So that's a really dangerous, tendency. I think that eventually we're going to have to snap out of it because otherwise it's it's getting to be really so to contextualize it for the recruiting space, what I've, what I've observed is a lot of times where it falls apart or the biggest problem is the employer doesn't really know what they want.

00;29;10;08 - 00;29;35;19
Unknown
Right. Oh yeah. And, and and then to, to kind of put it in the context of what you're saying, like I'm sure the employer believes that they could achieve this unicorn being for, you know, only paying them 50 K a year, whatever it is. How do you, how does how does know me or the platform you guys are building approach the employer engagement side of this to help, to help them see, you know, truth set that they are not surrounding themselves with.

00;29;35;23 - 00;29;55;28
Unknown
Yeah. Yeah. So that's a really good question because there are two big problems. One is the potential candidate doesn't know what he wants, and the potential employer also doesn't know what he wants. And there is a process of discovery there. And people that work in recruitment, they solve it by throwing CV, seduce résumés. Adieu, until something sticks.

00;29;56;01 - 00;30;16;08
Unknown
So it's like, I know you don't know what you want, so I'm going to show you everything I have is like a bazaar that they you work in and you, like, have all of these choose whatever you want. That is not, of course, is not useful because the employer has to review it, create number of résumés. If is really résumé is a poor representation of the person.

00;30;16;10 - 00;30;39;11
Unknown
If you find out too late that some people embellish the resume, to put it politely. So there's all sorts of problems with that approach, right? But at the same time, we recognize that the employer needs to go through that process of discovery. I need I think that I know what I want, which is the worst thing, is they engage thinking they know what they want, but they really don't.

00;30;39;14 - 00;30;59;04
Unknown
So you start the process of recruitment and you can see that they're refining that, thinking about what I want as they go in part is because they don't know what is available. So it's like if you can buy a shirt, you can walk in, and you may have a shirt in your head that you imagine how you want it.

00;30;59;06 - 00;31;12;23
Unknown
But then at the end of the day, you need to see what is in the rack, right? And you're going to have to choose what is in the rack, and you're going to choose whatever is closer to what you had in your head. Or you may even see a shirt that has nothing to do with it. And I don't know, I like I actually like that better.

00;31;13;00 - 00;31;32;15
Unknown
So that's exactly what happens when people is recruiting. They come with an idea that sometimes is vague, and then they start to refine that idea. Sometimes they come with an idea, but as they see candidates, suddenly they hire somebody completely different that when they say at the beginning and that happens, is normal. And that's the way it works.

00;31;32;17 - 00;31;58;23
Unknown
So you can't complain about how it is. That's the way it is. And if you want to change it, good luck. But that's not what we want to do. So the way we approach that is to say, okay, we are going to give the employer the tools to run that process as fast and efficient as possible, because at the end of the day, the process has to happen, but it doesn't need to take 42 days, which is the standard in the industry right now.

00;31;58;26 - 00;32;18;14
Unknown
Because what happens is you get a job description, you start finding résumés. If you're a traditional recruiter, you get a job description, start finding résumés of candidates that have the same keywords that the job description. And then you send off resumes to the employer. The employer is strolling through hundreds of resumes and telling you this one. Yes. This one.

00;32;18;14 - 00;32;46;18
Unknown
No, this one year is this one. No. You start to arrange first interviews with those people, then they get interviewed the next week. Some of them and some others the week after. Some of those interviews need to be reprogramed because the guy has a day job currently. So whatever. When you put all of that on the table, you end up spending more than a month close to 42 days in just going through that process of revising resumes, interviewing candidates and whatnot.

00;32;46;21 - 00;33;07;10
Unknown
At the end of that process, the employer, the potential employer, has a better idea of what they want because they've seen the rock a little bit, and they kind of have an idea of what what is in store. And now you can really start searching, right? But you wasted 42 days or so. So what we think is, is, okay, how can we speed up the process that has to happen?

00;33;07;10 - 00;33;32;14
Unknown
There is no way an employer are going to come right off the bat with the correct job description of what they want, right? So what we did is do these long format interviews, store all of that information, and made that available to the potential employer so they can interact with the platform and search. And by the way, it solves a related problem, which is if I want people for, say, customer support.

00;33;32;14 - 00;33;46;28
Unknown
And I know the people when I call and I'm going to be really angry when they call, then probably I would prefer them to be women, because you tend to chill out when there is a women on the other end. If you disagree, you're going to shout at him, you're going to insult him. You don't feel guilty about that.

00;33;47;01 - 00;34;05;17
Unknown
But what do we mean? You to do is more polite statistically. Now, of course, statistically. So sometimes there are requirements that are to written requirements that are, in my opinion, genuine requirements. But they are unfreedom requirements or if they're going to interact over the phone, you want them to have a clear accent because they need to interact over the phone.

00;34;05;19 - 00;34;36;18
Unknown
So it can be a super capable candidate. But if their English is a little bit English or Spanish or German or whatever, if it is not a clear accent that you can communicate easily over the phone, then maybe that candidate is not the right one. It's a threat to try. There are a number of unwritten, requirements that are it not polite to put in a job description, so the potential employer can interact with the platform and start searching in natural language saying, hey, I'm looking for customer support representatives.

00;34;36;18 - 00;34;56;06
Unknown
They need to have a clear English spoken language, and I can even say I prefer them to be women or whatever, because it's not a legal process yet, and they can start to see what candidates pop up and they can say, oh, that one looks promising because they the system is giving you a little bit of a view of the candidate, let's say.

00;34;56;08 - 00;35;15;08
Unknown
And based on the requirements that you are putting forward. Right. And so you can click on that candidate and you can interview the candidate right then and there. Hey Johnny, how much experience are you having customer support and the are you going to reply based on the interview? I have 15 years of customer support. What companies have you work for?

00;35;15;08 - 00;35;46;04
Unknown
I work for these companies. What responsibilities have you had or I been in charge of my team of customer support. So. And then I can answer those questions quickly for you so you can get a sense of this is the right person or not. So something that before was like, okay, I review the résumé. Now I need you to go and find these people and see if you can arrange a meeting for me so I can go and have a half an hour conversation or 15 minutes conversation and try to figure out if this guy is the right guy.

00;35;46;06 - 00;36;03;04
Unknown
That can be done in an afternoon with the other benefit that because we interview them for one, two, three hours, there are a few things that we already figured out, which is they are not lying, right? If they lie, you can keep up a lie for 1 hour or 2 hours. And they're not lying. They are real people.

00;36;03;04 - 00;36;25;26
Unknown
They are not a bot, which is part of a problem right now with with online applications. So all of that is already, sorted for you. And, the only thing that is left is you do a quick check of the main characteristics you expect in a candidate. If they satisfy your expectations, then you can say, okay, now I want to have an interview with this person, but not just to check the basics.

00;36;25;26 - 00;36;44;19
Unknown
Now I want to sit down with that person because maybe I'm going to hire him. So something that before took 40 days or so. Now I can take an afternoon and you can interview, I don't know, 15, 20 candidates in an afternoon because you ask two, three questions. If the candidate is not interesting, you can move on. You don't need to be polite and hold this video meeting anyway.

00;36;44;21 - 00;37;02;03
Unknown
You can just move on. Yeah, he's not going to be offended. So that I guess is one of the main reasons why we build this is because you need to speed up the process. And the way to do it, in our opinion, is to put in front of the potential employer. Yeah, yeah, all the information. Ask whatever you want.

00;37;02;03 - 00;37;24;11
Unknown
You can ask, have you traveled? Yes. I been in different countries because of work or because I like traveling and because all those things in an hour, two hours conversation come out and it can be important for you because maybe you want somebody that has some experience traveling, I don't know. Yeah. I, I think right now there's, there's at least in my world, I found that we've, we look for a lot of companies that are like the cursor of whatever.

00;37;24;11 - 00;37;43;04
Unknown
It's like, hey, how do we make it so that our customer is, is building their own applications within our platform? It feels a little bit like that. Maybe it's it's not like they're building applications, but it is having this education that you were saying takes 42 days on average in a traditional environment is now happening much quicker.

00;37;43;07 - 00;38;02;08
Unknown
And the customer is doing it all completely on their own. Yeah. They're starting to, to see all these things on their own that this because I was thinking as you were talking, I was like, if there's thousands of companies out there that say, you know, we have the promise of digitizing the recruiting, you know, industry. And this has been happening for decades, right?

00;38;02;08 - 00;38;26;26
Unknown
Recruiting. I mean, that is the age old problem, right? How do I find the best people? And and I was thinking, okay, so what's different about Nomi? Why now? Like, why why is why is I helping this happen now? And and I hadn't thought about it through the lens of, oh, you're you're what? You're truncating isn't necessarily the actual work of finding the person and interviewing the person, whatever.

00;38;26;26 - 00;38;54;04
Unknown
It's the educational piece of teaching both the employer and the employee how to present the best opportunity to create that connection. Yeah, yeah, because he's the part that nobody wants to do. So the reason why the recruitment industry people have been trying to automate and do whatever to make it fast, to make it painless, and it doesn't work and it still doesn't work, is because nobody wants to do the hard part, which is sit down with the person and talk to them.

00;38;54;07 - 00;39;16;05
Unknown
That's the part that everybody wants to avoid. They send you an AI to do the interview. They want to get your résumé or your LinkedIn profile and go through that filtering by keyword and whatnot. But it's like, okay, can you sit down with the no, no, no, we don't want to. That's exactly the point. The point of we got kind of one of the problems we have is market fit, because the industry is actually going in the opposite direction.

00;39;16;05 - 00;39;36;27
Unknown
We're going the industry is going in the direction of let's put the person behind the system. I don't want to deal with the person. I don't want to interview them. I don't want to talk to them. How can we make this automated? And we're like, no, you need to sit down and talk to the person, because if you don't have that, then the information you have is crap and crap crap out.

00;39;37;02 - 00;39;57;22
Unknown
So if you have a resumé that doesn't really say much about the person, well, good luck with that. If you're getting a Lims, if you train the ILM with, I don't know, I want to find an example that is not offensive to anyone. But let's say that you grab the worst possible books you've ever read. I don't know, some flat Earther book or whatever.

00;39;57;29 - 00;40;20;18
Unknown
Right. And you train dealing with that. What do you think you're going to get when you ask questions? Right. So this is what is happening to the recruitment industry at large is like we just trying to make it so we not need to deal with the person we send. We publish an advert online, people apply online. We send them an online I interview where they do the interview online.

00;40;20;18 - 00;40;39;26
Unknown
I don't need to deal with that. That's all hidden behind the automation. And then eventually at the other end, he pops some candidate and he said, well, he's not the right one, of course is not right. You need to sit down. So the challenge for us was, okay, we know that somebody has to do that and we are willing to somebody has to sit down and talk.

00;40;39;28 - 00;41;09;19
Unknown
Part of it is because Santiago enjoys doing that. But even if you enjoy doing that, it's not scalable. You can sit down with for two hours with every candidate, for every job proposal. Because of course is is unsustainable. But we were like, okay, we willing to do that and we're willing to even train people to do that. So we can have which is part of the the roadmap is eventually to get the position of interviewer as a, as a, as a position that you can hire somebody or sign a contract with somebody, train them to do that, and they do interviews.

00;41;09;19 - 00;41;30;09
Unknown
That's all. That's what they do. Fine, we can do that. But now we need to scale it because of course, otherwise not going to scale. So the way to scale it is once you have that is all these little things, it's like, okay, I interviewed you once and now we can use that to apply for as many jobs as you want in the rest of your life, because we have mechanisms to keep that up to date.

00;41;30;09 - 00;41;49;17
Unknown
Of course. So that's where the scale is. The scale comes along. It's like the employer can do this process instead of 42 ways to figure it out. What I really want, I can do that in an afternoon because I interview 20 guys in a row and like realize, you know, okay, cool. So this idea I had is probably not realistic or.

00;41;49;20 - 00;42;10;25
Unknown
No. Now that you mentioned that you really, really need to have experience with it, I didn't thought about that. Now I'm going to put it on my requirements. You can do it in an afternoon. So that gives you scale. And from the candidate point of view is the end of first interviews. It's like the 15 minute interview with whoever to see if they want to offer you a job.

00;42;10;27 - 00;42;33;12
Unknown
You need to organize for that. You need to take time out of your job somehow, go to some dark corner in your office. Or if you're working from home, you need to disconnect for 15 minutes so you can have this interview. And that's tiresome because you have one after the other after the other, and it's always the same questions and it's like, oh my God, they never reply to you because nine out of ten they don't even reply to you.

00;42;33;14 - 00;42;51;20
Unknown
So all of that is like gone. You can interview me as many times as you want, go and have at it. You can get a report after work, see how many people interview you. You may not see who was that interview you, but you can see that you were interviewed and that can we can analyze that.

00;42;51;25 - 00;43;12;08
Unknown
We're not doing that right now, but we can analyze that conversation to understand where the conversation stopped. So if we find that everybody is asking you to have experience on whatever and you don't, and that's where the conversation stops, maybe it's a good idea for you to go and learn it, right? So there are a few things that we can do with that, that we're not doing at the moment, but we can do so.

00;43;12;08 - 00;43;29;02
Unknown
We can help you to improve your professional career so we can have scale after the fact, but somebody has to sit down and talk with the person. That's the part that everybody's trying to avoid. And we're like, no, we're willing to do it. Yeah, well, for as much as for as much as we this is another thing we were talking about.

00;43;29;04 - 00;43;57;26
Unknown
I has made it so easy to produce at least prototypes. Right. Or, you know, something that we can engage with and say, oh, this is pretty neat. There's. Yeah, there's something here. You know, I could build this more, and so that's created a lot more noise. And, but when you do find the signal in my, in my opinion, and in what I've seen, that signal is so much stronger because you find organizations like Know Me, which are doing the unscalable things so that there's a differentiated edge amongst all the noise.

00;43;58;03 - 00;44;13;07
Unknown
Yeah. And so I think I think it's beautiful. I think I think it's awesome. And candidly, there's a part of me that hopes you never scale. You know, you don't get this hyperscale because you're doing, you know, the core human things that nobody else will do. Yeah. And that is a good point. I haven't thought about it.

00;44;13;07 - 00;44;33;02
Unknown
I'm going to have to think about it a little bit more. But I think that you strike something there when the only things are. And this is the thing that I need to think about. The only things that are going to remain are those that are not scalable by AI. Yeah. So whatever you doing, if it is scalable by AI, you may be replaced by.

00;44;33;04 - 00;44;51;01
Unknown
But yeah, by definition is commoditized. Yeah. But but I need to think about that. But that sounds like something that we should meditate on. And and that's again that's why we're like, okay, nobody wants to interview people. Let's do what everybody else doesn't want to do because somebody have to do it. And just because people don't want to do it doesn't mean that is valuable.

00;44;51;07 - 00;45;12;23
Unknown
That's something that is, I guess, important to say. The reality rewards results, not good intentions or anything like that. So time will tell if this idea of let's interview people, let's sit down and talk to them. Let's make sure that we capture who they are. What experience do they have? Let's filter the noise out of the signal in the sense of or you embellish your CV.

00;45;12;23 - 00;45;29;05
Unknown
Good luck trying to keep up with the conversation with me for two hours about what you're doing in your job. If you made it up this morning, I interview a guy and he was telling me that he's been exploring AI because he's super interested in AI, and I'm like, okay, cool. Which frameworks have you used? And he goes, oh, I'm using long chain, blah blah blah.

00;45;29;13 - 00;45;48;23
Unknown
Oh, so you're doing Python? No, no, no, I'm do it in a Java like okay. Lang chain does have Python and TypeScript doesn't have Java. So it's really you can put lang chain in your CV is going to pop up in the text search. You're going to call the guy, book an interview for next week, meet with him 15 minutes.

00;45;48;23 - 00;46;05;22
Unknown
In 15 minutes. There is no chance you're going to figure that out, and then you're going to go in the next day or whatever. And for us is like, no, no, that doesn't float. Yeah, yeah. But you do have you do have a generative. Well someone was going to ask you about it. You have a generative, part of the platform.

00;46;05;22 - 00;46;25;11
Unknown
You have an agenda, part of the platform. Right, with with the generation of the different documents and, and submission of that. Yep. On the generative side in particular with this use, with this case, how do you avoid it embellishing on behalf of the candidate. Yeah, right. That's a really good question. So again everything comes back to there are kind of two things there.

00;46;25;11 - 00;46;47;29
Unknown
But everything complex comes back to the interview. The two hour long interview. One thing that is a pain that we didn't thought about that when we started, but now we're dealing with it and we thinking about a few ways of making it less painful is that when you finish the interview, you have to sit down and go through that interview yourself.

00;46;48;03 - 00;47;08;19
Unknown
Whoever did the interview and kind of cleared it up. And the reason for that is twofold. One is you want the person to feel comfortable, and you want this to be a conversation. You don't want this to be a questionnaire or you want to put a light on their face. And that's the question is, has to be more of a conversation where they tell you an anecdote, they tell you stories and whatnot.

00;47;08;22 - 00;47;25;27
Unknown
In the course of that, it's very common that people start telling you, yeah. So what happened is I just, break up with my girlfriend. We were three years in and I was planning to marry, but she didn't want to marry. So we broke up. And then I decided to travel because I was like, there is a lot of stories that you like.

00;47;26;00 - 00;47;51;15
Unknown
Maybe I shouldn't leave that in the in the, Transcript. Although there are some applications for this in dating, but that's different story. We're talking about recruitment now, so you should do that one in the same. I feel like every year recruiting and recruiting a way for recruiting. So yeah. Yeah. If you if you're talking about careers and actually there is an application of the same pattern to the platforms of, matchmaking.

00;47;51;17 - 00;48;12;17
Unknown
Right. But let's put that on the side for a second. There are a lot of conversations that happen in this interview that you may not want to keep. There are parts where the interviewer is talking that you may not want to keep. So we go and edit all of that, and we make sure that whatever we have is, is representative of what the person decided to share with you.

00;48;12;23 - 00;48;33;13
Unknown
Right? You're not breaking their trust when they told you something personal and you clear it up. All the things that are noise, right? Because you may spend, I don't know, ten minutes, complaining about the technology that you hate. Now they can understand the context a little bit. But if you search that technology, that guy going to come up because he talk about it, right?

00;48;33;20 - 00;48;58;07
Unknown
But he can't talk about it because he hates it. So probably not what you want. So you need to deal with all of that is let's say you don't have to, but it's easier if you deal with all of that in the transcript before you read it, or embedded in the system. Once the transcript is embedded in the system, then you have the retrieval part of it, and that's the part that you want to avoid hallucinations.

00;48;58;07 - 00;49;26;10
Unknown
You want to make sure that it's not making up, etc.. And there are already well known techniques. And so if somebody is working in AI they problem familiar with it. So and I'm going to go into detail on that. But you have retrieval of it. So what you do is you tell them go and search in the you have the transcript, are stored in embeddings, which is basically pre-calculated the words so the computer can do faster.

00;49;26;12 - 00;49;48;07
Unknown
The proximity is like the relationship between words. But you have the chunk of text there that the person talks about. You chunked the transcript, you convert it into numbers, and then you put that in a database. And then you can tell the okay, here you have a corpus of text, go and find their whatever is relevant to answer this question that has been asked to you.

00;49;48;10 - 00;50;17;01
Unknown
And you need to then write an answer on the behalf of the person being asked. So the M will go find the chunks that are relevant. Say that you are asking, do you have experience with Java? So is going to go and you're interviewing, say Johnny. So you're going to go to all the chunks of Johnny, whether it is the transcript or the resume, and we'll find everything that mentions Java in some way or anything related to Java.

00;50;17;01 - 00;50;34;26
Unknown
So if you're talking a spring boot, spring boot, spring boot, that will come out because the M understands relationship between words and with those chunks is going to build an answer. And then you can have an agent going back and saying, hey, I want to make sure that you didn't introduce anything that is not there because the NLM actually knows everything.

00;50;34;26 - 00;50;55;29
Unknown
So they have read Wikipedia and the manual of filenames and the manual of Java. They know everything. So you want to make sure that whatever the answer is, the M has done the semantic construction of the response. So it sounds natural English, but the actual content is coming from the chunks and not from the original training, let's say.

00;50;56;02 - 00;51;14;01
Unknown
So that's the hallucinations part. So you can have an agent that take care of that. Even in retrieval you have teams of agents working together. One is going and searching and finding the relevant parts. Another one is validating, and there are a few things that you can do to make sure that the answer represents the candidate is funny.

00;51;14;01 - 00;51;39;22
Unknown
I even played with answer in the same style and tone. Then the candidate and we were trying with a few different styles of interviews, and I realized that when we interview people that we know and we have some some personal relation, those people were way more informal in their answers, in the way they answer. So, as you know, in Argentinian Spanish, you use curse words as normal language.

00;51;39;25 - 00;52;02;29
Unknown
So when you tell the M reply as if you were this person using the same domain language, you get really funny interviews. So you need to account for that and make sure the response is professional. You need to make sure that, the user question is not appropriate. You don't provide an answer. So there are a few things there, but you can solve that with a genetic application.

00;52;02;29 - 00;52;31;16
Unknown
So I guess that is interesting. I because I was listening to, I was trying to learn how to like, speak less honestly. Yeah. You know, hear, listen more, speak less. So, you know, it's because you didn't get rid of that. Now, this is great for me. I and it was talking about, like, a neuro mirroring in terms of, like how, you know, just when your CEO speaks, they tend to speak much more calmly, direct, short, because that helps calm the other person in the room and you know, where they're in front of.

00;52;31;19 - 00;52;50;10
Unknown
And and you're right. I hadn't thought about that. And if we if you are going to do a lot more human to human interaction that that feeds that becomes the data pipeline for the Nomi platform. There is that dynamic that the the sentiment of the the candidate is going to differ depending on the sentiment of the interviewer. Oh, yeah.

00;52;50;12 - 00;53;15;14
Unknown
And how you get that consistency, or at the very least, you get an opportunity to, to, to at least stabilize within a certain range. Do you find that like again, by being a two, three hour long interview, the interviewer is able to kind of stabilize amongst a norm? Or do you have moments where it's like, I'm having a bad day, this is going to be bad versus I just, you know, I did really well.

00;53;15;16 - 00;53;41;28
Unknown
We're going to have a great conversation. Yeah. So part of the roadmap eventually plans to convert the recruiter. Sorry. The sorry the interviewer position. The plan is to convert that into an actual job. Right. What do we hire people to do? We do apply for them. Whatever. Some, license so you can become an interviewer. However we do it, that is it's a job in itself.

00;53;42;01 - 00;54;04;22
Unknown
And one thing that is really valuable for me and for San Diego, but especially for me, is doing the interview. So you understand the dynamic, right? Because I'm the CTO. Fine. In theory, I shouldn't be doing interviews in theory, but it's really, really useful that I've been doing interviews because it helps you to understand things like that is like the way I behave here really drive the way the interviewer is going to behave.

00;54;04;24 - 00;54;26;18
Unknown
So if I grab a tablet and I start asking questions with the pain in my hand and ask question and he answer and I take or do something that, he will behave in a particular way because he's been basically interrogated. But if I foster, an environment where he can just tell me a story and I'm listening, I'm laughing when he tells me something funny, and they make it very informal and very relaxed.

00;54;26;21 - 00;54;43;22
Unknown
The other person engage at that level. So you can really drive that. And it's something that when we start training interviewers, we're going to have to account for that and teach them how to strike the right note. Let's say that is not too informal, but not too formal, because if it is too informal, then it becomes very personal.

00;54;43;22 - 00;55;02;16
Unknown
And most of what people share with you is not relevant anyway. And you can use it. But if it is too formal, the person is like in a situation where he's being interrogated and the idea is in two hours, you relax, you sit down, and you are yourself. So you need to strike the note that allows the other person to be themselves.

00;55;02;16 - 00;55;23;27
Unknown
And then he a there is a lot of empathy that you need to in a way, understand who is the person that is with you and try to help him to be himself and not the other way around. So it has happened to me that you have people, that you ask a question and the answer your question, and then you ask another question and the questions are open because you want them to talk.

00;55;24;03 - 00;55;46;11
Unknown
You don't want them to tell you yes or no, but still you ask open questions is like, all right, am how long you been working with, I don't know, AWS oh, five years. Yeah. Okay. What you've done in areas of all sorts of projects is like, oh, right. But that's why the interview is long, because that is at the beginning.

00;55;46;11 - 00;56;03;03
Unknown
But as soon as you start to realize that, you really want to hear what he has to say and that you keep kind of driving that tone of conversation, oh, that's really interesting. In five years, you must have done a lot of things. So in this job that you did, what was the situation? Well, well, in that job.

00;56;03;03 - 00;56;20;28
Unknown
And then the person start to lose a little bit. There are other cases where the person is absolutely talkative and you can ask a question and they go crazy in different directions. You need to bring them back. So it's really a dynamic is really it's a job in itself. So doing podcasts and stuff like that helps. That's fascinating.

00;56;21;00 - 00;56;40;11
Unknown
So on the same I guess along the same lines, you're building generative tools. You're building a genetic tools. Yeah. Do you find that like this? You need a specialization in how you approach building a genetic versus generative or the same. Are we getting to a point where, like, the same person should be able to build generative, a genetic analytical of all the different forms of AI that that are going to be?

00;56;40;11 - 00;56;58;04
Unknown
Yeah, I think that there are there are things that are common and there are things that are specific. But what I find is you need to understand the genetic, you need to understand AI in general, how an alarm works. And you need to be realistic about what the Lim can do. I always say that it's like a three year old AI knows everything.

00;56;58;07 - 00;57;30;20
Unknown
So yes, he knows everything. But still three years old. He doesn't have a criteria, right? So you need to kind of make sure that you get them there and that you can see that even if you're using, for instance, I'm using cloud AI to generate code. So it's really herding cats. I need to be if I keep Claude in a short leash, is like, I want you to do this, but before you jump into coding, you're going to think it through, you know, make sure you apply good practices and you can show me the solution you want to apply and make sure that you don't go crazy and start implementing features that I didn't ask.

00;57;30;23 - 00;57;52;26
Unknown
I don't say that, but that's basically what I'm checking for because AI is really powerful. Claude can generate really huge amount of code. He knows a lot about good practice and whatever, but you need to kind of keep it short. So there are a few dimensions on that. One is for me, there is no difference between generative AI or a genetic AI.

00;57;52;27 - 00;58;15;21
Unknown
You use a genetic AI to do generative AI. Is more about two things one understand the limitations of the tool. Understand how to put the guardrails for the tool, what it is prompt. What is context? The just the guardrails that you need. And one thing that is, in my opinion, becoming more and more important is you can always overcomplicate the tool.

00;58;15;27 - 00;58;41;28
Unknown
And that will be super expensive because it goes crazy really fast in terms of how many tokens are you spending on solving a problem. So a combination of deterministic and a genetic is nowadays is necessary. You need to think then I solve this without AI. And if you can then solve it that way. Because otherwise you using too many tokens and it's becoming too expensive.

00;58;42;00 - 00;59;07;25
Unknown
So for instance, one thing that I'm solving now with the yeah is the, the agent, the agent team that is supporting the, company in the search for candidates. Right. So you start interacting with that agents and then you go, I'm looking for, I don't know, an oil and gas engineer find the AI realizes that you're working looking for oil and gas engineer.

00;59;07;27 - 00;59;35;01
Unknown
That's a requirement. Has to be an oil and gas engineer. Fine. But you haven't said enough for the search to be meaningful. So what you're looking for is what is the oil and gas is in Texas, is in Pakistan, is in Kazakhstan. Whatever you is so having a solution that makes sure that you ask enough questions and remember enough questions can be Java developer.

00;59;35;01 - 00;59;56;14
Unknown
It can be an oil and gas engineer, can be an accountant, can be a neurosurgeon. You don't know beforehand what they're going to be looking for. It can be a CTO, can be a CEO. So how do you figure that out? Well, easy, right. You send it to an l, m and you say, hey, if they looking for oil and gas engineers, what are the requirements that typically are required for an oil and gas engineer job description.

00;59;56;16 - 01;00;13;29
Unknown
And I'm going to answer that. It's going to be a really good answer probably. So you can use that to prompt the user right in the loop. You prompt the user saying, hey, tell me more about these. For instance, where are they located? How much experience do they need to have? Do they need to have this or that certification that is proper to the industry?

01;00;14;01 - 01;00;36;04
Unknown
Really good. Super expensive because you're going to be doing that again and again and again and again. So you try to do that deterministic if you can. And if you really, really, really can't solve it, then you go a little M and you come back. So all those type of techniques are across either see if you generating or you are doing any other sort of AI.

01;00;36;07 - 01;00;58;01
Unknown
Those are the things that for me are the skills that we need to, understand and develop is be creativity is about when do you see when not to use AI and what is the cheapest way to build this because otherwise it gets expensive really fast. Yeah, I love, I love that like even just a year ago it was like, hey you, you need to show me that you can do this with AI first before we can make any decisions.

01;00;58;07 - 01;01;17;03
Unknown
And now it's like, no, no, no, we're going back to the, you know, back to the original thesis, which is, I think, smarter. Show me that you can do this without AI or really, what are you genuinely good at? What is your special talent? And then let's leverage AI to augment that. Yes. Right. Versus like you know, no, but aim for AI first, right?

01;01;17;05 - 01;01;41;17
Unknown
I don't know. I think it's funny. Ironically, AI is bringing us back to our humanity in terms of like, how do we how do we figure out what makes us us? Yeah. And is not artificial anymore. Maybe that's how we avoid, you know, this this balloon pops. Yeah, maybe. I think that you say something every year that I think the year on the money, that all the things that I can do are not the human things.

01;01;41;20 - 01;01;59;27
Unknown
And we're going to have to focus. We can all be philosophers, you said. But basically what you're saying is we're all going to have to focus more in our human, really human talents and capacities and characteristics and not so much about what are, you know, how to write code in Python or whether you know how to the knowing how to do stuff.

01;02;00;02 - 01;02;12;02
Unknown
I can do that for us. The physical stuff is is beyond risk. Yeah, yeah, we still are the CTO of Know Me. Thank you. This is an awesome conversation. I really enjoyed it. Yeah, yeah. Me too.