Back to OmniBlog

Podcast: Marin Smiljanić talks Machine Learning on AI Searched


May 7, 2024


Founder story

Marin Smiljanić and Tanya Bendis, the co-hosts of AI searched discuss the start of Machine Learning and the progression of AI today.

Episode 101 of the AI searched Podcast brings you Marin Smiljanić, ex-amazon engineer, who discusses the challenges of starting his own company Omnisearch and the history of Machine Learning and how the resources today allow us to push forward with AI innovation. Watch the podcast or read the transcript below.

Follow along the transcript:

Tanya Bendiš:  Welcome to the first episode of the podcast AI Searched, where we talk about all aspects of artificial intelligence and machine learning across multiple industries.

Tanya Bendiš:  I'm so excited to be here today with my co-host, Marin Smiljanić, a colleague of mine, who is the co-founder and CEO of OmniSearch.

Tanya Bendiš:  Welcome, Marin, thanks for joining me today.

Marin Smiljanić: Well, thank you very much for having me.

Marin Smiljanić: I'm excited about this.

Tanya Bendiš:  Cool, well, let's start from your past, I guess.

Tanya Bendiš:  So when did you start getting into machine learning and your experience at Amazon, and how did you come about to found a company working in AI and machine learning?

Marin Smiljanić: Yeah, so I think that if you really think about it chronologically, I didn't start off as any kind of an AI person.

Marin Smiljanić: Strictly speaking, I was a distributed systems person.

Marin Smiljanić: So through both what I did in my first job, which was at a startup called Single Store, and afterwards at Amazon, my first couple of jobs were definitely all distributed systems things, meaning, you know, make a system scale to hundreds of thousands of machines.

Marin Smiljanić: So that was my bread and butter for the earlier part of my career.

Marin Smiljanić: I guess the turning point sort of where I got more interested in machine learning and started following it more was when I switched teams at Amazon.

Marin Smiljanić: So first, I was on AWS, and then afterwards, I moved to the Alexa team.

Marin Smiljanić: And the Alexa team, like you can imagine this being, you know, a very important exponent of AI as you call it.

Marin Smiljanić: Generally, the methods there were quite primitive at that point.

Marin Smiljanić: So they didn't integrate all the cool new LLM things and all the new advances that you had in the past couple of years.

Tanya Bendiš:  Yeah, well, yeah, I can imagine that big companies, a little bit slower, functioning systems.

Tanya Bendiš:  What was the major turning point at Amazon that made you kind of turn on that light bulb to start OmniSearch and kind of take the risk of being a founder and working with your co-founder Matej and just starting your own thing, even though you didn't know exactly where it was going at the beginning.

Tanya Bendiš:  And you have to pay your bills, you have to survive somehow. What was the choice? Where was it?

Marin Smiljanić: Do not fear death.

Marin Smiljanić: And so I would say, like, it actually has a fairly prosaic motivation, because there was actually a problem that I saw at Amazon that really annoyed me. So think of both my teams, S3 and BLXA team, both are fairly technically complex.

Marin Smiljanić: You can't just get somewhere out of the blue and start working on those products.

Marin Smiljanić: There is actually a lot of information you need to absorb about the systems, about the entire philosophy behind them and so forth.

Marin Smiljanić: And so Amazon had this massive repository of internal video training materials, lectures.

Marin Smiljanić: And these would be like the legends who built AWS would be giving these hour-long lectures, going into algorithms, like really laying it on thick.

Marin Smiljanić: And you couldn't really make any use of those kinds of materials.

Marin Smiljanić: Like you could find a video by title, by description, by something like that, but to actually find that little nugget of information that you're looking for in those VAS databases, that was a no-go. That was just not supported.

Marin Smiljanić: And that's actually the spark.

Tanya Bendiš:  Yeah, incredible.

Tanya Bendiš:  Well, I mean, now you're very much experienced in machine learning and artificial intelligence, and you're extremely active in the community, both in Silicon Valley and Seattle, as well as in Europe and Croatia.

Tanya Bendiš:  Where did machine learning come from? This podcast is about AI and machine learning, so just as the first episode, let's kick it off with a little history lesson, you know?

Tanya Bendiš:  Let the audience know, like, how did this all start? And why is it such a buzzword today when actually it's been happening for so long?

Marin Smiljanić: And that is a very good question.

Marin Smiljanić: So I would say you always had two different philosophies for achieving artificial intelligences, as it were.

Marin Smiljanić: So the first one that you, I think even predates the other one, is the sort of approach where a system learns from data. And these would be the very early neural networks, like perceptrons and stuff like this. So this was, I think, all the way in the 50s, 60s even, like barely after Alan Turing.

Marin Smiljanić: The other philosophy, I think, appeared in probably the 60s, maybe late 50s, but probably 60s, which is the symbolic part of AI.

Marin Smiljanić: And so what this means is you basically have systems that are fairly rule-based, they use logical reasoning, they attempt to sort of codify human knowledge and build systems based off that.

Marin Smiljanić: And actually, for historical reasons, due to a very clever guy writing a very hatchet job book about perceptrons, the ideas from the first branch, which is learning from data, became defunct for a fairly long period of time.

Marin Smiljanić: And the symbolic AI was dominant.

Marin Smiljanić: That changed, you always had like a little group of people that were working in the background on the defunct versions and ideas in AI.

Marin Smiljanić: And in fact, I think two things happened in this latest wave of machine learning, deep learning, AI, like whatever you want to call it. One is that the hardware finally became good enough.

Tanya Bendiš:  Exactly.

Marin Smiljanić: Yes.

Marin Smiljanić: So the humble graphics card actually became an incredibly powerful part of machine learning infrastructure.

Marin Smiljanić: And so the stuff, even though you had the ideas in the 80s and even before that, you couldn't make them work sufficiently well in the real world.

Marin Smiljanić: With this stuff, you could really train some good models, very powerful models.

Tanya Bendiš:  Well, it makes you think, today, I would just say, in general, consumerism is something that we're so used to.

Tanya Bendiš:  And since the millennium, pretty much, I sure I own more things than my parents did and my grandparents, of course. So in the 50s, my grandparents were around my age.

Tanya Bendiš:  And now you think about it, we put so many resources and we test so many things on these models, machine learning, AI models, chat, GPT. And realistically, not all of it is useful, and it's very interesting to see how back in the day, when we didn't have all this GPU, right, to kind of let this work, they were very limited with the resources and how precise they were versus now.

Tanya Bendiš:  I can ask ChatGPT what the weather is gonna be like, or some silly little prompts. And what's your feeling on that? How sustainable do you think that is?

Tanya Bendiš:  What are we looking for in the future with that?

Marin Smiljanić: So, I think it's always a matter of having a really significant technical breakthrough, but then that thing really reaching wide adoption.

Marin Smiljanić: I don't mean the superficial level, which is like ChatGPT has this many daily active users, because it has a lot of really, really good uses.

Marin Smiljanić: But then if you go into more specific use cases that actually are the rest of the iceberg from specific industries to medicine, whatever it is, you simply need to make use of a whole lot of data and to make use of it precisely.

Marin Smiljanić: So you can't just leave everything up to the model to just conjure things up.

Marin Smiljanić: So you need to have some sort of a database in the background.

Marin Smiljanić: So I think that that's going to be just in the next couple of years, in the next decade, really, it's just going to be making sure that these things that make for cool demos can actually be deployed at large.

Tanya Bendiš:  The models are efficient and not using as many resources because we just won't be able to handle that.

Marin Smiljanić: And hallucinations, So yeah, we need to make sure that these great new advances are applied in the right way too.

Tanya Bendiš:  Well, on the search, working in multimodal search and just being able to analyze different types of data is just taking a huge section here of all the possibilities of what we can do with AI.

Tanya Bendiš:  As you said, in Amazon, there was just so much going on. You weren't even able to filter through what was actually necessary materials.

Tanya Bendiš:  And I think in a lot of industries, we see these problems where you can't exactly pinpoint what you need just because it's just too overwhelming.

Tanya Bendiš:  What are your next steps with your startup and what are you looking forward to changing in this world, in this technology?

Marin Smiljanić: I think any founder in this space will always tell you, I think the first thing is make heavy investments into the algorithms, make heavy investments into the models and make those more powerful, more efficient, more flexible, more able to interact semantically with the end user and with the customers.

Marin Smiljanić: And then the other thing, of course, is just commercialization, meaning apart from the education and media industries where we're most active, opening up a couple others and simply growing the business and bringing these benefits of multimodal search to as many people as we can.

Tanya Bendiš:  Yeah, and I mean, for example, Omnisearch is incredible because it has such a technical team.

Tanya Bendiš:  There's just so many award-winning math competitors and incredible people, compromising this team.

Tanya Bendiš:  And I just think, like, what would you say? Like, what would be your inspiration for somebody working in AI and machine learning? You know, maybe they don't have as big as a marketing department. Maybe, you know, they're not doing so much for sales and growth, but they have a technical team.

Tanya Bendiš:  Is there hope for these people?

Tanya Bendiš:  Or is AI too strong of a buzzword that, you know, you just kind of hop on the wagon? Or how much is, like, high quality work, you know, important at this moment? And how do you see that going into the future?

Marin Smiljanić: It'll end up being very, like, the fundamentals will end up being extremely important.

Marin Smiljanić: I think that we're seeing this now, like, Mate, our co-founder, and myself, we're just seeing it all over the place where there's people who make nice demos, but they can't follow it through with a proper enterprise use case.

Marin Smiljanić: And so I think that having the fundamentals nailed down right is going to be really critical in the next couple of years, and not just building wrappers for XYZ.

Tanya Bendiš:  Yeah, exactly.

Tanya Bendiš:  For example, in marketing, within three months of ChatGPT being open to the general public, we saw PDFs for sale about ChatGPT prompts, how to use stuff, and just so many people profiting on very basic things, and good for them.

Tanya Bendiš:  They jumped on that wagon, so it's going to be very interesting to see more technical teams now who are, let's say, a little bit slower in their development, obviously, because algorithms are a little bit harder to create than a marketing PDF.

Tanya Bendiš:  But it's going to be very interesting to see how this technical side of machine learning grows.

Marin Smiljanić: You've got to also be careful if you're a technical founder with a heavily technical team.

Marin Smiljanić: You've got to focus on the customer.

Marin Smiljanić: You've got to close the deals.

Marin Smiljanić: You've got to actually provide support.

Marin Smiljanić: You can't try to build castles in the sky that nobody actually needs. So you've always got to focus on the customer and on the user.

Tanya Bendiš: Yeah, user experience is important. Making sure it's worth it. Yeah, absolutely.

Tanya Bendiš:  What is your favorite thing about... Well, you hop around the world. You've lived in Canada. You've lived in the US for a little bit. You've grown up in Croatia. You've worked in a lot of different places and a lot of different large name companies.

Tanya Bendiš: What's your favorite thing about being, let's say, both a polyglot and a citizen of the world?
That's how it's said.

Tanya Bendiš:  It's like traveling around and seeing life in both places.

Tanya Bendiš:  What's your favorite thing about being in the Silicon Valley versus your favorite thing about being here in Zagreb?

Marin Smiljanić: I think it keeps you honest in some sense.

Marin Smiljanić: It makes you not drink the Kool-Aid as much usually.

Marin Smiljanić: When you're talking about somebody being in one ecosystem their entire life, it makes you, I think, see things a bit more clearly.

Marin Smiljanić: It makes you compare things, lifestyle versus business considerations.

Marin Smiljanić: Or it just makes you see clearly some of the patterns, so I think it really helps. I would always advise people, young people, that are wrapping up college or during college to give it a shot.

Tanya Bendiš:  Yeah, give it a shot. Travel, see the world a little bit, work in different places. Always a good experience, right?

Tanya Bendiš:  Yeah, incredible.

Tanya Bendiš:  Okay, last question here.

Marin Smiljanić: That's good.

Tanya Bendiš:  Lots of them.

Tanya Bendiš:  So there's just been so much talk about legislation and privacy in this whole AI and machine learning space.

Tanya Bendiš:  I'm really interested to hear your opinion of, you know, just kind of like the space race, like how much money can be invested and also how much is legislation going to kind of slow down the innovative process?

Tanya Bendiš:  What do you think about that?

Tanya Bendiš:  And who do you think is going to come out here stronger?

Tanya Bendiš:  And what is your proposal and idea for that?

Marin Smiljanić: Proposals and ideally, it's always harder to be constructive.

Marin Smiljanić: No, but I, as a general rule, I find that you end up stifling innovation if you try to expose, to impose too much of a regulatory burden.

Marin Smiljanić: I think that the regulation can all, can in many different instances come from a good place, a place of good intentions, but still make it more onerous for, you know, developers who are actually trying to push the boundaries.

Marin Smiljanić: I think that it's very welcome, for instance, in the EU AI Act, trying to regulate, say, facial recognition in public spaces.

Marin Smiljanić: That's a wonderful thing.

Marin Smiljanić: And one thing that my distinguished former employers at Amazon don't know about that much, I think.

Marin Smiljanić: But still, it seems to me like that they are placing limits on, you know, model sizes and, you know, the data that you train it on.

Marin Smiljanić: And I think it'll simply be difficult for smaller players to comply, while at the same time you've got, you know, OpenAI and the guys that, through the past couple of years, have had this explosive growth, who are in fact now lobbying for this kind of legislation, which, of course, suits them quite well.

Marin Smiljanić: So my recommendation is always, in such new industries, don't stifle the innovation.

Marin Smiljanić: So I would always recommend, you know, going low on the regulatory burden.

Tanya Bendiš:  How do you think that these regulations are going to change the public perception of AI?

Tanya Bendiš:  Because right now, you know, as far as we've done some interviews on the street, it's been very interesting to see how the general public actually reacts to AI, because they kind of either shun it or they love it.

Tanya Bendiš:  You know, they're using ChatGPT.

Tanya Bendiš:  They're creating content.

Tanya Bendiš:  They're making cool pictures of themselves.

Tanya Bendiš:  They're super happy, you know, very, like, top of the iceberg AI and not really understanding, like, the deeper algorithms about it.

Tanya Bendiš:  Or they're just so scared that, like, robots are going to take over the world and eat, you know, eat up their jobs and everything.

Tanya Bendiš:  Like, do you think that through these regulations and through time, people are going to be able to accept that, you know, we kind of live together with machine learning?

Marin Smiljanić: I think eventually, yes.

Marin Smiljanić: Because it seems to me you always have the concept of a hype cycle, and it usually ends up, you know, it starts off very high.

Marin Smiljanić: And at the peak of the hype cycle, you know, you have these hysterical media narratives of, you know, XYZ is going to take your job or whatever.

Marin Smiljanić: But then through time, you see, and in AI, this has been particularly brutal, where the expectations were not met by reality at the points where it was expected to.

Marin Smiljanić: And so you then have a bit of disillusionment, but then eventually it rises to a sustainable point.

Marin Smiljanić: So I think that, yeah, I think that at the end of the day, people will realize what the benefits are.

Marin Smiljanić: And I think eventually they'll be convinced that, you know, adopting this technology is a good way forward.

Tanya Bendiš:  Exactly, and that it can actually improve your work and improve your life in general.

Tanya Bendiš:  I think this is really some feedback we were beginning from a lot of people is just having their everyday lives be easier, even if it's for private tasks, not only for business.

Tanya Bendiš:  Well, thank you so much for joining me today.

Tanya Bendiš:  And I'm excited to continue our podcast, you hosting people over in the United States, me hosting people in Europe, and see where this is going to go.

Tanya Bendiš:  You can catch AI Searched on Spotify, on Apple, YouTube, and Google Podcasts.

Tanya Bendiš:  We look forward to connecting with you and hearing from you, and let us know in the comments if you have any questions about AI and machine learning.

You can watch the AI searched podcast on:

Apple Podcasts



Subscribe to our newsletter

Sign up to our newsletter and receive the latest updates!