Actually, the reason I picked the same name. So I started, you're right, in 99, I started a company called X -Plane out of Aachen, Germany. Aachen's name was Ex -le -Chapelle, as a French name, Ex -le -Chapelle. It had something to do with the water there, being basically a chapel at the water and there was the...
And that's why I chose Xplained at that time. In 2004, I sold the assets out of Xplained to a company in the US, which is why I moved actually from Germany to the US in 2004. And I had basically then an empty shell and I decided in 2020 when I started the company again, not knowing what name I should take, I just picked an empty shell and
filled it with the company. And then the customers liked it so much because Xplain has something to do with AI, has something to do with explanation, has something with, I mean, all this basically fits nicely with what we do. The customers were saying, no, keep the name, don't change the name. And hence it is what it is today. Yeah, I started in the mid 90s actually in NLP. As you said, I mean,
did language models. My first language model was from 1995 actually. And machine translation and speech recognition were my work in my master's thesis, in my PhD work, etc. So it was all about speech translation, speech to speech translation to be more precise. So basically,
what you can do now with many applications where you basically speak one language in, in real time it translates something back. And initially it was a government funded project. It was hard. I mean, it still is not easy, but it was really hard then because computer was not there, data was not there.
Hassan Sawaf3:50And in 2000, I decided to basically start a company out of the research work, which I was doing in academia and build basically a company to offer machine learning services for everyone who wants to integrate it. That was before cloud. So delivery of that technology was not trivial.
So you had to basically always sell some server with connection into the companies or you had to build machine like speech recognition, which is very, very small to fit in a car, for example, et cetera. So there was, there was challenges then which we did. We don't have today. In 20, in 2004, I sold it basically to a government contractor in, in the East coast in DC and
went into working with DARPA projects, government entities, larger corporations as well, and so forth. So what made me then, so there I was actually, I mean, it was beautiful. I had large resources at SAIC, later at eBay, at Amazon, and then Facebook, big resources, a lot of support from leadership.
But I wanted to start something where I could build basically a platform where everyone can utilize AI, not only the engineer. And that's what made me start Explain. So basically speaking to the non -engineer was the time where I was at EWS, not necessarily a focus point, for example.
AWS is built for the engineering community in particular, first and foremost. And I decided I wanted to build a platform which basically speaks to the entrepreneur, someone who is non -technical, to the artist maybe, and to people who are lesser technical so that they can actually be part of the game building AI. And that's what Xplained is building.
Hassan Sawaf7:34So actually, I didn't get, I didn't choose to go into that field. The field shows me to some extent. So what happened is I was actually working at Daimler before I actually went into academia. I was working at Daimler to build basically automation for airlines and airports to communicate with each other, inform each other about the plane, the distribution of
like the passengers, weight distribution, is there children on there, how many wheelchairs do you need? I mean, all the logistics in and around the planes were communicated in natural language between the airlines and the airports. And Daimler basically put me on a project to read those messages and do automatic actions to optimize the process.
And that was 92, 93. So that's quite a long time ago. And in 1995, actually, I was presenting the work. And then one person out of the audience came to me and said, Hassan, what you are building here for this, we need for a speech to speech translation project, which is funded by the German government, et cetera, et cetera. And then basically,
Being bilingual myself, so I have Syrian origins, was born and raised in Germany, so I speak German and Arabic fluently. It piqued my interest, of course. I mean, this is a cool thing that you can have at some point a device which can in real time translate what you're saying in the other language. So yeah, that was basically where this is coming from. And actually for me at that time, it was very interesting.
and quite unique at that time to basically think about the speech to speech translation process as a single process instead of basically saying you do the one thing speech recognition and then you do a machine translation and then you do text to speech. How can basically the models be more integrated? They're still different models but they're deeper integrated with each other and that was very fascinating and a little bit before its time actually.
Alp Uguray11:15and seen how the big companies are working, they adopted the technology, it explained what are, so you started in 2020, so it's been four years, what are some of the trends that you see that surprised you, that was unexpected? There are the parts that are expected, but like what was...
Unexpected that that I was surprising
Hassan Sawaf11:48So what I actually was surprised about is when you are giving basically AI which is easy to use into the hands of people who are outside that field, how much innovation comes to the table. You give the right people the right tools into their hands and
the innovation, the level of innovation explodes, which is beautiful. So this is something which I was, I mean, I knew it, but the scale which I'm seeing is bigger than I have expected, which is really good. And that aligns very nicely with what Xplain is doing. We want to build tools like a kit. We are building a kit.
for people to basically take data they have, for example, put it on their fine -tune a system, fine -tune an open source language model, for example, or open source speech recognition, or whatever it is, and then basically build their own derivative which is unique to them. We're also giving them tools, the people tools, to build agents in a very seamless and easy manner.
where you basically anything you onboard on Xplained, let's say from hugging face or your own models or your own data with an existing model somewhere, et cetera. Everything basically turns into an agent when you onboard it on Xplained. And you are making then on Xplained a little bit self -aware, meaning every model will know that what it can do and what it cannot do. If you allow it, if you basically
do some benchmarking on that model. It will also say, where am I strong? Where am I weak? Now, as a developer, you actually like to see that so that you know basically what to do to improve so that your model is stronger in the next iteration. For the people who are integrating into a solution, into a more complex agent, for example, they will know basically...
Hassan Sawaf14:08where the limitations are or where the strengths and weaknesses are. And then they can combine it with maybe other models as well, where basically the agent has multiple tools to solve bigger problems. So in essence, we're giving the user the tools into their hands to basically recruit the right agents into a group together and make them solve a bigger problem.
together as a team. It's like we are basically changing the context or concept of building solutions from an engineering only task to entrepreneurial task. You are an entrepreneur, you hire the right people, put them in the right set up together, describe the problem in a...
smart way in an educated way so that all the agents inside your new team, so to say, can actually solve the problem the way you want. And then from there, you can actually see how they collaborate, maybe help them a little bit here and there, but they can collaborate to build the solution for you. After some time, you have done that solution. That time,
When you have only humans in the loop and only humans basically working together might take days, weeks, months. With agents, it might actually be seconds, milliseconds and whatnot. If you have the human in the loop, it might still take minutes and so forth, but it's much faster today where you can actually mix and match the human muscle with the AI muscle together into like one of these agent groups, so to say.
You're building incredible solutions with relatively small efforts and not as much technological expertise needed. You need to understand the problem. It's more important for you now to understand actually what is the domain? What is the problem you want to solve? What are the constraints? What are the policies you need to be aware of, et cetera? And basically,
Alp Uguray16:51That sounds very fascinating. Like when I'm thinking about it, then like I'm thinking how, when it scales that like in a company, there are 200 employees or whatnot. And then there are, they are working with like 500 AI agents and like each AI agent is like, has a specialized model that certain like permission settings, prompt templates, and then they talk to each other to, to
Hassan Sawaf18:21So I think we need to learn to lead. I mean, think about leadership skills, really. I mean, either it is leadership skills between ourselves, humans. That's something which we definitely should need to learn. And whether you... And basically translating...
problems into like, let's say, contextual information, which is interpretable by AI and by the human. In essence, you need to be able to write good business plans. That's something which is going to be very helpful. Or what we did at Amazon, like write basically document -centric development. Everything inside the company was very
document centric, we were working on PR FAQs and six pages and two pages and all of these things. In essence, if you learn on how to do that, that's actually very beneficial to you. It was always beneficial, even for the scientists at Amazon when basically they were reporting to me when they come to me with a six page document describing their problem or describing basically the solution they want to implement. And it was an easy to
for us all to be on the same page within like 10 minutes reading, everyone was on the same page, we were communicating, converging super fast, being able to ask the right questions and then executing fast. So this experience now with the agents can actually be used again with the same kind of like vision here. And the other thing which we need to be able to do, which aligns actually nicely with this.
is that we need to start thinking on a larger scale. We need to think about problems, not necessarily... I mean, it's similar. I mean, every phase shift which we had in computing, we had to do that. When we started computing in early days, and you used to need to take like punch cards and understand basically how those can be coded and whatnot down to...
Hassan Sawaf20:45the terminal where you program assembly maybe to Fortran, to Cobalt, to other things. And then at some point you have MS -DOS and UNIX and whatnot, CPM at that time, where it was simplified and more human to basically communicate to the machine, graphic user interfaces, and then the iPad and whatnot. And now with AI, I mean, we're basically making things
easier so that you can think on higher level, more complex problems to solve. I mean, imagine we needed to do something like drug discovery and we can only program computers with assembly. It's going to be very, very hard to just even think about how to program it, right? Now we... Exactly, exactly.
Hassan Sawaf21:43Now we can actually with AI, we can think on a higher level with bigger blocks, so to say, and allow the AI basically to do some of the simple work. In essence, I think that goes nicely together with the concept of thinking in descriptions and policies and so forth because
When you think about programming on the very high level, that's what we need to do. I mean, how do you program? You program by basically defining what is your environment? What is the context? What are the policies? What are the constraints you have to basically set? And then basically, and then start basically programming top down like you do when you're programming with object -oriented programming languages, for example. Yeah.
Hassan Sawaf24:05Mm -hmm.
So everyone who is basically married to what they're building for too much is gonna actually suffer eventually. I think an important problem, an important thing to do is always think about your users challenges and how you can actually solve that.
and think about the problems in how to break them down even into individual sub -tasks, for example. I mean, I think that is important. I think the company should not be afraid. Someone who is building technology should not be afraid, not companies only, even individuals, even scientists.
should not be afraid to basically throw out something and say, okay, there is a better way to do it. I mean, if you think about my field, if I think about my own field, machine translation, when I started, it was more rule -based and prologue and Lisp and things like that. And then basically statistical learning was happening. And then suddenly you have to basically throw the code out and basically
redo, rethink on how basically translation can be done using statistical learning methods and then machine learning learning methods and the neural approaches and then LLMs now. I mean, it's changing all the time and we should be allowing ourselves to rethink, reinvent ourselves all the time. This is very important. To be able to do that in a larger corporation, you have
Hassan Sawaf25:58big systems and if the system is a monolith, is this individual thing and basically you can only update the whole thing together, then there is a challenge. So my suggestion is if that is a company which basically has that, how can actually be this system be taken?
and re -engineered into basically modular components. Most of the systems today are built on microservices, right? I mean, this is a good thing. So it is already broken up. You can innovate already. Problem is that some of these big systems are so many microservices that it is getting, you're losing.
you're losing oversight about what is there and how do these microservices interact with each other. Every microservice has to be maintained by an engineer at least, if not multiple engineers, et cetera. So the big companies have large armies of people which basically are necessary so that they can actually maintain the system. And I think now we are going into the next iteration to some extent.
instead of thinking in microservices where you have to actually still feed each microservice, what is input, what is output, what's the business logic, how do I change the business logic, maybe you can actually start thinking about them as agents and build basically some smarts into them. So when you basically onboard a model onto Xpane or you build an agent on Xpane and you design it that it basically can...
can work with XML files. Then you send some JSON into it. This is a simple task for us engineers, but the system, if the API is not designed for that, it's going to break. Something is not going to work, and then basically you have to have an engineer come back, do some work, and then basically get an update. Now agents can actually be smart to say, okay, you're bringing a new thing in there, the content I have.
Hassan Sawaf28:19I have seen something similar to that. I know what you want, what your intent is. I can basically reformulate and basically change the pipeline inside that agent to solve that problem. And suddenly it evolves into something which is already stronger and you don't have an engineer involved in it. Engineers of course will get informed that something is happening there, but they don't need to like get into each of the hundreds of agents and do every step.
They can focus on the things which are relevant instead of basically focusing on all of these things. So I think this is a new concept of how to build systems. And with this kind of technology in mind, you can actually deal with larger systems with a smaller team. Meaning we can actually build big systems again and again much faster.
We need less resources to build these systems. Startups can actually, more startups can actually afford to think big and build things and all of that. And we see that. We see that on XPlane. We have some startups which are building really cool things which they're gonna announce in the near future. They would be needing, I mean, I don't know, 10 times more people if they was not basically like the agent.
the Injit concept and not only 10 times more people, but also tens of millions of dollars more than what they already raised. So this is going to be super cool to see more of that.
Alp Uguray30:57When it comes to training the model and then as the model gets better, of course, there's the designing of the model, the parameters that go into it. But from your perspective, especially open source, like having an open source model, how can we make sure, and then this maybe ties to more of the...
the regulation part of it, how can we make sure that the scraping of the data part and the usage is well tracked and also somehow profitable maybe to the person or the company that generates that, creates that content? So that's the first part. And then second part is, as the closed source accelerates, right? And then they have all the compute to scrape the world.
Hassan Sawaf32:21All right, let's talk about the first one first. So yeah, I think the regulation is going to basically at some point control what kind of content we are able to take and feed our models with, which is good. We need to be aware of where data is coming from and whatnot. And by the way,
Doing that does not necessarily mean that the model is going to decrease in quality if you're constraining it. I mean, if you look at Lama 2 and Lama 3, they put some thought into Lama 3 on where is the data coming from, what data am I going to use, et cetera, et cetera. And if you look at the quality of Lama 3 and compare it to Lama 2, there is worlds between them.
even though the size of the model is basically the same, right? So quality does not necessarily need to degrade when you basically constrain or control where the data is coming from. But there is concerns. We need regulations basically that everyone knows where the data is coming from. Copyright and other things have to be taken into account.
To some extent, I mean, this question was actually one of the questions which came to my mind when I actually started Xplained. I wanted to build a marketplace. And that marketplace, basically, four kind of assets are basically going to be part of it, or are part of it today. Datasets. So if I have a dataset, if I have data,
which I don't know what to do with it, but I know that someone will probably benefit from it. I want to find a place where I can put it in there. And then basically someone who is going to use it for training, they basically are going to share the value they're generating with me. That's something you can do with Xplained. The other thing is when that person who is taking your data and training their system on your data,
Hassan Sawaf34:40The data, the model is hosted on Xplained. It's not hosted somewhere else, meaning you cannot reverse engineer the data set from the model you basically, the weights you basically are downloading. So I, as a data owner, feel comfortable that actually when I basically share my data through Xplained, it's kept safe. And if there is a second generation of that model, I'm going to benefit again, et cetera, et cetera.
So this is one thing. So this is datasets. Models, algorithms are free today. I mean, basically, you can go into GitHub and find basically almost any kind of model you need in some way, in some form, et cetera, et cetera. So they need to be made, they need to be made findable or searchable. I mean, that you can find them. And then you have basically
trained models, like not only baseline models, but foundation models, maybe optimized models for certain domains, et cetera. You need to be able to find those. And then also pipelines, which means a combination of models. For example, speech recognition is usually something like segmentation, speaker diarization, the actual core speech recognition, blah, blah, et cetera.
So there's pipelines in there which you can actually also put into the platform. And then of course, people. People need to be findable somehow and collaborative with these other agents. So this is what Xplained is offering. In essence, so this is to your first question. I mean, if I know that I can find a place where I can put my data and whoever is using it is using it in a responsible way and
I benefit with them about this one, I might actually feel comfortable doing that. And we see that. I mean, we have customers who are using these models, building kick -ass models with them, et cetera. So this is great. Then to your second question about what do we do with all these closed source models and open source models? Is there...
Hassan Sawaf37:02a winner takes all kind of situation eventually that basically all the open source things are going to fall sideways. And my belief is it's not going to happen. There's going to be a world where you see them side by side. It's actually similar to the operating systems. There is something called Linux. Everyone uses Linux when they want to scale, when to do enterprise, et cetera, et cetera, and so forth.
And everyone uses either Mac OS or Windows or something like that, sometimes Linux as well, when they want to basically have like a personal machine do things, they don't need to worry about how to set up, et cetera, et cetera. Then they basically go and get the Windows machine or the Mac OS or something, device, et cetera. But that doesn't mean that...
Linux is dying. Even companies which were long time against using Linux are using Linux. Look at Microsoft. I mean, so IBM for that matter, right? I mean, the world is not as like black and white anymore. So I think there is value in both. And in essence, I believe in the second model in the basically let's have
people focus on building like specialists agents and put them together into like a super agent which can basically like do the same thing like the monolithic models can do. Right. I mean you can't you have the monolithic models they're offered by a handful maybe two handful companies and so forth and there is value in using them for someone who doesn't need to worry about
building agents and whatnot. And then basically you have, if you want to build enterprise solutions, oftentimes you can use them, but you use them maybe with other things together. Maybe you don't need to use the monolithic models and just focus on like a set of specialist models and generate better results. There is so much research which is being done today to prove that for specialized solutions.
Hassan Sawaf39:19it actually might make much more sense to take a 3 billion parameter model, have it being trained and optimized towards your problem versus taking a 500 billion parameter model, which takes so much more resources, takes so much longer to run, et cetera. I mean, imagine I'm not going to be able to see that a thermostat in a house, for example, is going to have a 500 billion parameter model built in.
But I can imagine there's going to be a 1 billion parameter model or like a 200 million parameter model which can be put in the thermostat. And that's all the knowledge it probably needs to solve a certain problem, right? So I think there is, it's not a one or the other. It's going to be the together. And the market is so big and the applications which you can think about today are so many.
that it's not going to be open AI taking everything. I'm pretty sure of that.
Alp Uguray40:25Yeah, that's the exciting times coming ahead. And it's very good to hear that specialized AI models for certain tasks that are getting triggered by maybe getting orchestrated by a different agent, then it's going to be potentially how enterprises work or orchestrate their work.
I got two more questions. I know we're almost out of time, but in the best case, I would have keep you for hours, but I don't want to keep you away from dinner. When it comes to the language, and I think language is how humans interact with each other, how animals interact with one another. And then there's the element of
Hassan Sawaf42:12All right, so I'm actually excited about this topic quite a bit. If I know that the agent which is solving the problem is understanding me to deeper and deeper way, and it's a machine, and it's not like stumbling over things which it was supposed to be doing and it's not doing, like in the past when...
we had voice user interfaces, right? I can actually access, I can actually finally access someone to help me, something to help me in a continuous way, which was always a dream offer for us. Right now we're doing it, I mean, before the latest AI solutions, we're doing it by basically trying to make a
to use call centers and then basically use overseas call center and so forth. And then there is a communication issue sometimes and then there is cultural clashes. And I mean, all these things are happening, were happening. And I think those things are gonna be lesser. Now, does it mean that we don't need the human? And that's actually gonna be an interesting question. I believe there's going to be the human
The human is going to bring value, but I as a human, as a human agent serving a customer, am not under time pressure necessarily like I was before. I can take my time and wherever the human element, the you and I, right? I mean, not the transactional things. Those can actually focus on those things.
And because the human is not under pressure to like solve and finish and to the next customer and so forth, because AI is basically taking the, the grant of the work and I'm basically taking, paying attention to the very sensitive pieces, the human to human interaction is going to increase in quality, which is right. I mean, I'm going to give you an example in the healthcare. If you're talking to the doctor.
Hassan Sawaf44:38Today, often many doctors are under such pressure that they have only got so much time with you. They can ask you questions, talk to you, but they don't have really much time for you. They have to do everything in like five minutes or 10 minutes and they need to be done to the next patient, otherwise they're gonna lose money. When AI can take much of that work, the doctor can focus on the things which...
are the human aspect can be given some more focus. This is one component. The other component is even the human, when I'm talking to any customer, even if I'm a doctor and paying attention to my patient, having AI assistance with me, giving me all kinds of information to the...
to my attention versus only focusing on certain things which my mind is able to focus on gives me a better foundation to communicate with my customer, to communicate with my patient, et cetera, et cetera. So in essence, I don't believe that there is going to be a shift away.
into basically using that we are just using AIs. I think we're going to use AIs more, but the human is going to have the same amount. We are going to still need basically the human in many cases. So the transaction is not going to be lesser. It's going to be more quality though, instead of quantity. And we have more time and whatnot. So I think that's going to be very, very important and very good for us. Quality.
Hassan Sawaf47:31I mean, if it was a business, I mean, for a business, I would actually ask, are you prepared for, how are you preparing your company for AI in terms of data, in terms of like structure and so forth? I think that's going to be important for people to think about.
that and I would like to hear from companies how they are thinking about this one.
Founder, Alp Uguray