Leaving Deepmind to start his own Startup! Aleksa Gordić

In this exciting podcast episode with Aleksa Gordić, a former research engineer at DeepMind who has now embarked on his own startup journey, we delve into a wide range of topics related to his experiences and insights. Throughout the interview, we explore his current priorities, his work at DeepMind, his decision to drop out of his master's program, and how he landed a software engineering role in machine learning at Microsoft and then Deepmind without an official degree.

Aleksa shares his journey of self-learning and building practical experience by participating in competitions, hackathons, and data science projects. He highlights the importance of gaining a solid foundation in mathematics and recommends resources like the "Cracking the Coding Interview" book and the FastAI courses for those looking to enter the AI field, which I also share in my guide for learning AI.

While discussing his time at DeepMind, Aleksa emphasizes the diversity of roles within the company, ranging from software engineering to research. He shares his experiences working on applied teams, collaborating on vision language models, and integrating data sets. He also highlights the dynamic nature of research and how large language model projects often involve both engineering and research breakthroughs.

Aleksa addresses the question of whether a master's or PhD is necessary to succeed in the industry, emphasizing that personal drive, self-discipline, and a strong portfolio of projects and practical experience can be equally valuable. He shares his own journey of transitioning from a hardware-focused background to software engineering and then machine learning.

Throughout the conversation, Aleksa provides practical advice for those aspiring to work in the AI field. He suggests engaging in hands-on projects, reconstructing papers, and becoming familiar with relevant APIs. For those interested in research roles, he emphasizes the importance of deepening knowledge in niche domains and standing out through open-source contributions, high-quality blog posts, or other unique perspectives.

The podcast episode covers a wide range of topics beyond Aleksa's journey, including his current startup, Ortus, which focuses on AI-driven content analysis (basically allowing you to discuss with YouTube videos, including mine!). We also discuss challenges, the future of AI in video creation, productivity tips, and more.

To fully explore these fascinating topics and gain deeper insights, I invite you to listen to the complete podcast episode on YouTube, Spotify, or Apple Podcasts:

Full episode transcript:

[00:00:00] This is an interview with Aleksa Gordić. Aleksa used to work at Microsoft and DeepMind as a research engineer in machine learning and has now started his known work with Ortus at Google Chrome extension to use on YouTube to answer all your questions about specific videos. It's super cool and using large language models.

Aleksa has lots of experience in contributing to the AI field, while also building his own stuff. He also has the YouTube channel AI epiphany that I am sure you are familiar with, and if not, you should definitely check it out. I hope you enjoy this interview.  

You are already managing a lot with YouTube and Ortus and just everything. Are you still planning to do videos or are you like going all in with Ortus, the company and your This project I am planning on, on, on, on doing the videos like still. The frequency is obviously reduced just because I, I don't have [00:01:00] that much bandwidth, but I, I am actually planning to, to have every, every two weeks at least something, maybe it's gonna be a bit more about the entrepreneurial story rather than like, hardcore ML .

Who knows? Maybe, I dunno, maybe I'll get back to papers as well. It's all tbd. Currently, I'm just focused on, on, on building a startup. That's kind of the number one priority, but I'm not gonna give up on, on the YouTube thing, so, yeah.  

Yeah, it's, it's really hard to manage everything. I don't know how you did with DeepMind and Microsoft to have a YouTube channel as well, just with the, the PhD, like my schedule is pretty, it's not free, but it's like I decide of my own schedule, so it's much easier to try to balance everything, but still it's quite hard to. Try to do videos as well as doing the thesis and everything. So I, I can't imagine with a real, a real job. I, I think it's similar, like you still have a bunch of flexibility when you're working for a good company. Ultimately what matters is you're [00:02:00] output like people, nobody's like tracking you and sitting down, oh, you didn't do like eight and a half hours today, or whatnot.

Like, you can, if you can do cool stuff and cool work in five, six hours, you're good. And then next time you can maybe compensate or like, ultimately the output is what matters, not, not the hours. And I, I, I hope that's, that's how many companies operate, right? Like ultimately you're not paying for people to spend their time because they're gonna, they're going to find a way to fill in the gaps and just do something else and not be productive.

So, yeah. Yeah. Unfortunately for most of my friends, I think, like, they go in person and they have to work eight hours per day no matter what. So it's, yeah, it, it is definitely better if you can. Just have like a, a specific workload. Workload. And if you are more efficient, well you just do it faster and you can do yeah, more other things. That's the, the ideal solution.  

And how was your, your work at DeepMind, what were you doing? Like, I [00:03:00] assume DeepMind is more like researchy, so were you more focused on the implementation of research or writing papers, code manage people? What were you doing there?  

It, it actually very much depends on which team are you in. Basically I was in applied team, so that means I was working more on the inference side of things. Initially when I just joined, I was basically collaborating with the folks from the Flamingo team, so like the, the vision language models. I did help integrate some data sets there. So you can consider that part of my job more researchy, like.

Ultimately, all of these bigger projects are now. Like boiling down to hardcore engineering. Like I depends how you define research, right? Like I would argue that all of these large language model projects, VLM projects are more like engineering efforts. And then research projects. Obviously a new paper comes out, somebody picks up an idea like flash attention or something.

Well, that's again, more of an engineering breakthrough, right? Like flash attention was just optimizing how the attention mechanism works and made it [00:04:00] more efficient. And so like, I think that mindset of al always trying to see what's going on in the field and staying at the edge and reading the papers and picking up ideas, even when the, those ideas are much more engineering ideas than, than research ideas.

I think that's what we can call as researchy nowadays, at least, at least on, on those types of projects. So I was, I was basically, as I said, I've done that. And then I maybe helped review the, the flamingo paper as well for that. I got the small like in the appendix or something like Grateful to Aleksa Gordić or whatnot because I, I joined like the project was already kinda wrapping down and so like, I was not there from, from, from the beginning.

And because of that, I couldn't be the co-author. Yeah. And so after that, like I was mostly focusing over the last maybe six, seven months, I was focusing on, on getting visual language models into production in various Google products. So that, that means I was mostly on the inference side running those big models, like up to 80 billion as you know, in the p that's kind of public information, obviously wouldn't share anything that's not public.

And [00:05:00] I was running those on like clusters of, of TPUs that, that, and now the thing is, Most of those things when you're at Google and, and these big tech companies, like all of these toolings are internal. And so a lot of the things in complexity is kinda abstracted away from you. So I didn't really have the gut feeling for what's going on all at all times.

And also some of the core people on Flamingo Project, I talked with them and I basically asked them some things like, Hey, do you understand how this charting thing works? And they were like, Nope, somebody else implemented that. I'm just focusing on this other part of the abstraction. So ultimately when you're in these bigger projects, you, you kind of narrow down, niche down in one particular, Yeah, sub area. And you don't necessarily see the bigger picture un unless you are maybe the team lead or whatnot. So I, I think coming into DeepMind, they definitely had different expectations. What it means, To be on some project. But again, that was very subjective because if you were on some smaller project where the output is paper, then the experience would be completely different.

And honestly, I I, I didn't have that experience. So...  

And [00:06:00] you are more of a generalist, right? You prefer to, to touch a bit everything and do everything or like, manage everything at least a hundred percent. Like I, I, I like to have a bigger picture and that, that's part of the reason why I never saw myself long term in, in a, in a, in a big tech.

Like you're always going to be like, I think it's over glorified. I think most people have not that great of experience, to be honest, because for imagine how many products are there in production, running at Google, running at Microsoft, somebody has to patch the bug. Somebody has to make those systems run.

And, and that's not the most exciting job in the world. But again, it just depends. Like some people who are. Passionate about certain part of the stack, and they end up doing that role and they're into optimization. And maybe like then for them it's a perfect fit. I, but I, I think on, on, on average, I think most people are probably not as super happy as if you were to start something on your own. Yeah. But then again, that requires a completely different spirit. You have to be [00:07:00] more entrepreneurial. It's, it's not for everyone. So just a personal preference, I guess.  

And would you say that it's also the same reason for why you dropped out of your masters? From what I've seen and also did not do a PhD, because from what I see, well, from what I know, the, the studies, the more you advance and the more like specific you become and the less generalist you can be.

For, for my masters I was in, I could say that I was working in computer vision and classification, like that's pretty general. Mm-hmm. And then for the, the, the PhD, I'm now doing medical imaging and segmentation. It's already, like much more. Not fine tune, but like much more specific. Specific specific. Like basically my coach says is, why did you drop out of the, the masters?

Okay. May maybe fir first to just address the whole PhD thing. I, I think, well, in my experience talking with many people who've done PhD, it varies a lot. Some people, like in like immediately over the first couple of months start digging into a certain[00:08:00] niche. Yeah. Some people spend 2, 3, 4 years trying to find their thing and then they end up publishing maybe one paper.

So for example, one, one guy I just recently talked with is literally in, in, in, in that class, like over the first three years, he was very broad and didn't niche down, even though it's a PhD. So I think PhD is also a spectrum. It's not one thing that, that, that's kind of uniform across all of the universities.

But back to your question on, on the Masters, why I dropped out is because I, I couldn't bear it. Like I, I was, it was not meant for like, I, I, I was not built to sit down in a classroom and, and, and listen to. Generic classes without any agency. Like I, I felt that okay, I know, I know what I wanna accomplish. And at that point, I already switched and decided I wanna switch to software engineering from electronics. So I, I knew I, I had to go more through algorithms, data structures, more software design, and the masters were very broad and still, I, I knew that's not gonna be the, the, the fastest path. To, to get where I wanna be.[00:09:00]  

And also like, what, what's the chance, what's the probability of, of your teachers at your faculty being the best experts in the world for the particular area you care about? That's more of a PhD thing if you're lucky. Right? If you, if you get the right advisor. So, so like, I, I, I'm not dismissing like in some alternative life, I, I might have taken the PhD route, but the master's part, I will skip it again.

Like, I, I, I, I just don't think that masters is very useful. I, many people do skip it, they just go directly to, to PhD. So Yeah, like, just, just because the way how I learn and, and, and the level of agency I have and the level of like self-discipline, I don't need this external structure to, to give me the motivation and in every other respect, I think is just inefficient.

You can do a much better job if you're self disciplined. If you can curate your own, personalize your own curriculum for yourself, there is no way that the generic curriculum is gonna cut it for you. So, because of that, for me, was a fairly, fairly easy decision to be honest.  

But sometimes for big companies like Microsoft, for example, you need the, [00:10:00] the credentials and just the, the title basically. You just mentioned that you did it, you, you, you learned by yourself. But how did you manage to end up at a big company like Microsoft, as a software engineer in, in machine learning without, without any official degree in, in this field? Yeah.  

Yeah. So the story is, I guess back in 2017, I was doing this android internship in Germany. And I reached out to my friend who was very who was great at, at like doing, he's done previously many software during internships at Facebook, Google, all of the FAANGs. Now the, the, the, the acronym is kind of different, but you, you get the point. And he told me to, to basically go through this cracking the coding interview book.

And so once I got back from, from that internship in Germany to, to Belgrade, I just. Double down on just learning algorithms as much as possible, going to hackathons, going to datathons, going to all of these competitive programming websites and, and trying to, to just execute on all of [00:11:00] those resources.

I, I knew, I knew are good, like the correcting the coding interview. And then I started applying. I got rejections, like I think anywhere, wherever I applied the first time. I was always rejected. It's almost like always, and I think, I think it, it's fairly common as well. And for some reason just because of lack of communication, I think most people think that all of these stellar engineers and people just, yeah, just first time you, you apply and you're there.

So I think that experience is probably romanticized a little bit as well. And so after having done all of those preparations, I ended up in this machine learning summer camp for which I found out completely accidentally. Like, I, I think that the deadline was like, like in three days. And I found out from someone, Hey, there is this ML summer camp, check it out.

And I was like, okay, let, let me, let me see what it, what it's about. And I immediately applied. And then later, because the people who were organizing that camp were from Microsoft, that helped me lend, lend the job at Microsoft initially as a software engineer. I was [00:12:00] applying also in parallel for Microsoft.

So those two combined kind of led me to, to, to lend the job there. And so I joined Microsoft and I joined the, the Microsoft HoloLens team, which was, I was super lucky. Like, in retrospect, joining that team was the best thing that could have happened. That was the best team, I would say at Microsoft at that point of time.

Probably even now, or maybe now, it's maybe now like working with OpenAI or like closer to OpenAI. That's probably the most exciting thing to be working at Microsoft. But back then, HoloLens, like super, like si like science fiction device. Like the, if you saw the Apple Vision Pro, like HoloLens too could do most of those things in 2018, probably not the same level of quality, but you already had instinctual interactions.

You could already use your hands for gestures. You could look at holograms and that the system knows that you're watching that holograms because you're basically, we are tracking your eye gaze vector. And I was actually working on the eye tracking myself and and so long story short amazing project a lot, lot of computer vision, obviously, it's very computer vision heavy.

[00:13:00] And because of that, you had to use a machine learning, right? Like the, the best way we know how to parse images nowadays is just machine learning. And so that naturally led me to ha to have to learn more about it. And then I started learning on, on, on the side, like Andrew Ang and, and reading papers. And then, yeah, that's, that's a whole different story, but that's kind of how it started.  

So it was always about being a software engineer, and then you ended up in machine learning as it was the, the main solution for the issue. You didn't enter the field because you wanted to do machine learning?  

I, I did because I, I entered the, the summer camp because I was interested in artificial intelligence. But like software engineering was also a transient role for me. Right. Because I was by, by my, like my bachelor was electronics heavy, I was focusing on digital electronics. I was even designing some high radio frequency electronic circuits, like on, on such a painful level of details. I was literally drawing the metal layouts and the polysilicon [00:14:00] layout of how certain, like up converter radio circuit should look like, or like there and the microcontrollers and embedded software and real-time operative systems.

So, so my background, like the official background comes from really, really deep hardware space. Yeah. And then I was slowly kinda raising up the levels of abstraction. All the way to applicative layer of software and then ultimately to machine learning, which is as Karpathy called it Software 2.0, which is even more like abstract and on, on a high level compared to what the, the folks be below us are doing. The compiler folks the, the embedded software folks. So yeah.  

Before that, that summer camp and, and Microsoft, you mentioned that you enter basically a lot of competitions and try to get various practical experience. If someone is just tired of college or some, or, or dropping out or finishing university and not wanting to do a Masters and PhD just like you, how would you recommend them to get into AI in 2023, [00:15:00] would you say that the best thing would be to go to Kaggle or to other competition's website? Or how would you recommend them to be able to finally end up at like DeepMind or OpenAI?  

It all depends what the ultimate role is. Do you want to be a software? You, you can be a software engineer at DeepMind, you can be HR at DeepMind. So like, it, it all depends what you wanna do, right? Like if you wanna be a research engineer, which is something I'm the most familiar with, I can probably give the, the, the best tips there. Yeah. But I also know how to go the other routes as well, like if you wanna be a research scientist et cetera, et cetera.

So you mentioned Kaggle. I think Kaggle is, is a separate community. If you go the Kaggle route, you're probably going more the data science route. Yeah. And, and, and, and you probably will not be dealing with unstructured data as much. Meaning text, meaning vision, meaning videos, more of like table tabular data.

Although I'm, I'm aware that Kaggle hosts like a slew of competition, but I think historically they've been more like, like I [00:16:00] know bu tous is the kind of one of the, the, the, the Kaggle grand masters and he's always XG Boost. XG Boost. You can just tell by the models they're using that they're very focused on tabular data.

Whereas I think everything exciting that's been happening in the field over the last many years has been more on the unstructured data front. So images, vision, all of that. So how to get to the role you want, basically, again, depends first on the role and then. There, there are some meta things that you should do no matter which of these roles you want, whether you want a machine learning engineering role or or scientist role.

You want to have some decent background mathematics. And I wrote the whole blog on this po on this, on this topic, by the way. So like I I'm gonna tell them as much as I as I like can right now. So I, I do encourage them to check it out. But do as many projects as as possible.

Maybe try and reconstruct a couple papers. Initially start with some of those high level courses. Like fast AI I think is probably the best one right now. Even, even more than Andrew Ng, like Andrew Ng just has the name. He's the Stanford guy, professor. [00:17:00] He's very like, big credentials and like Jeremy, who also has amazing credentials is just in that sense.

He, he, he's not like, As prominent using, like, he doesn't have the Stanford badge, basically. That's what I'm trying to say. But like I, I think that his courses are probably even better for many people, especially those who are coming from software engineering roles. He literally has the course that's meant for software engineers to transition into ML.

So that would be probably one of the first things I would do. And then immediately just start building, just start building. And and of course this is, these advices are more for engineering roles but like start building either reconstruct the paper, but even that is probably unnecessary. You probably wanna just get familiar with, with also prompting using the APIs.

And then, so that's kind of the, the top level abstraction. And then you can slowly start going deeper into reconstructing papers. That's kinda on the, on the, on the, on the bottom layer of, of of this abstraction pyramid of I'm kinda using better knowledge, but you know what I mean? So That's, [00:18:00] that's, that's some, some advice there.

For scientists, it'll be different. So for scientists, you, you would probably want to go with the piece. You're out you want to focus on reading and, and implementing papers as much as possible, getting familiar with the literature, everything that it takes to, to become a true expert in a very, very niche domain.

For example, if we take the, the research example, and the person is doing a PhD, how can you be sure that you, you will be able to end up doing research at the DeepMind if you are doing a PhD? Yeah. Like lots of people are doing PhDs and want to end up there. So how can you get in front of the competition or just build a better background?

That's a, that's a good point. I, I completely agree that just by going the standard route, even if that route is PhD, it, it is a fairly standard route, especially over the previous years. Maybe now people are a bit more reluctant with everything that's been going on to go the PhD route because it's, it's becoming increasingly.

Maybe it's a bad word, but irrelevant for I think, increasing percentage of, of the, of [00:19:00] the, of the tasks you'll be doing eventually. So you want to find some way to stand out. And there is, there is like a spectrum of ways you could do that. Like one of them would be create one open source repository that people find useful.

People are DeepMind people at OpenAI and just show that, you know, you're extremely good at this topic by doing that. That's one way. Second way is write a blog where again, you attack some previously unattached problem or just take this some existing problem, but have a completely new perspective and be sure to, to write a very high quality blog.

And if people notice you, that's again, one, one credential. I think a single blog post, a single good blog post can get you that role or at least opportunity to apply for, for those companies, because ultimately you will have to prepare to, to, to, to go through the interviews. So that's the second way.

And then for me personally, like YouTube definitely helped because I was covering very heavy papers and I had people before I [00:20:00] joined, way before I joined, I had many DeepMinders commenting and saying, you've explained this better than I could. And, and for me that was like such a huge recognition because h here am I like somebody who has zero background in all of these things, just like grinding and working and learning.

And all of a sudden, like I, I'm, I'm getting there like knowledge wise and also getting, getting those credentials. Also, if there is an open source project from your target company go and contribute. That's a perfect way to to, to get noticed. Maybe a small tangent. George Huts recently started his was it a tiny Corp company and he has this tiny GRA grad framework.

And he, the way he's basically recruiting people is by watching at the poor requests and finding high quality poor requests and, and, and basically interviewing those people. So I, I think that's the new way and that's how I will be recruiting for, for, for my company as well. Especially when you're a startup, you, you have that, that luxury of, of being much more, getting much more information than just four or [00:21:00] five hours of end-to-end interviews with someone that can be very easily tricked.

So, yeah. Yeah, and you have lots of choices when you want to contribute or just to, for example, if you have, you are doing a PhD or you have a full-time job, then if you want to excel or like be in front of the others, you basically need to do more than just the full-time job. So like contributing to, to GitHub repositories or building something or having a YouTube channel as you did.

And I wanted to ask, why did you start e explaining research papers on YouTube or just sharing on YouTube instead of directly building something or contributing to open source work? Like why, why for you, was it the YouTube thing?  

I've actually done all of those. I, I have a couple of very, very popular open source repositories. One of them is like graph attention networks from Petar Velickovic that had over 2000 stars. So that's, back then there was a huge number of stars. Like now you have all of [00:22:00] these GPT projects that, that get amassed, like tens of thousands of stars in a couple of days. But like, this is kinda over hype. But back then, yeah, that, that was, that was a big deal and people did notice all of the work I've done and that was not the only one.

I also implemented from all of the sub subfields. I was learning reinforcement learning or, or natural language processing or computer vision or, or, or a bit more niche areas like neural style transfer or deep dream. I've implemented all of them from scratch and that helped me learn how it works and also get, get a bit of. Exposure, like get, get those public artifacts out, out there.  

And would you say they were more useful than YouTube or just different, like what bring you the most opportunities?  

So it's, it's hard to estimate. Like ultimately the, the person I got referral from was Petar Velickovic. And, and I think already while I was doing, covering his videos sorry, not his videos, his papers like the graph attention number of paper, he did notice me and we started chatting casually.

And, and we are also [00:23:00] compatriots, so we're, we're coming from Serbia, both of us. And so that kind of connected us as well. And then later I also pointed the project. So he was very, very comfortable to refer me just because he knew me personally as a human being as well as like my technical ex expertise.

And so that, that was kind of my, my, my way in, into getting the referral. And literally how it happened is on Friday he pinged the recruiter. The recruiter replied, 30 minutes later and he said, oh, of course, I, I follow Aleksa's work already, like, over, over the past months. And I was like, oh my God. Like, DeepMind recruiters already know my, my, my my work?

And then literally on Monday I had like a call, Hey, let's, let's start the interview process. And I didn't start preparing at both. So it was very funny to, to, to, to, then the following weeks were very intense. Like, so like to, to be very, that's kinda understatement on the century. And yeah, that's, that's, that was kind of how it happened for me.

And was it a problem to not have the, the masters or PhD degree to, to get into [00:24:00] DeepMind? Was there anything you had to do to, to work around or?  

Nope. Nope. A hundred percent not, not like, as I said, for the research engineer role, you don't have to have a pg. So if I was applying for research engineer, this depends on whether OpenAIr or DeepMind or other companies.

But like, there are some companies, I think DeepMind is very strict about this. Don't quote me on this. Like, it's not, not the official stance of DeepMind, but like to the best of my knowledge, For the scientist role, you have to officially have PhD, even if you're better than 70% of PhD folks at DeepMind, which is hard.

But even, even if you're there, if you don't have that official diploma, I, I think you, you would not officially become a scientist. So that's, I think they're kinda rigid on that side. I think OpenAI is way, way more flexible and they don't even distinguish between scientists and engineers. I think they're, you're just like a member of technical staff, so that's kinda different.

And all of the other companies, maybe Mosaic ML and others, all of them have probably specific and, and different requirements for what it takes. But like ultimately [00:25:00] the, the same thing I mentioned before on, on, it doesn't matter how many hours you spend at your job, it shouldn't matter whether you went through the PhD route or master's route or if you just self-study.

Because we have amazing examples. Like Chris Olah is a good example. He, I think dropped out of Bachelor. He's probably one of the best scientists in the world for interpretability. Working at philanthropic as one of the founding, founding engineers. To the best of my knowledge, that might not be his role, but I think he's a good example.

There are, there are other examples as well. I like to think I'm, I'm one of them as well. Maybe not on that level of expertise in, in any particular topic, but like more general and, and yeah.  

Of course you are. I mean, and you at, at DeepMind you worked mainly on flamingo. You mentioned like, so visual language models and multimodal learning, which is definitely a trending topic and every, everyone is applying it or, or are doing that right now, just like you are with Ortus.

The main reason you [00:26:00] left the mine was probably because you couldn't, you, you were more attached to like something too specific, I assume. But would you, like, you, you could have asked to change teams or do something as you mentioned, that was about the paper or like more, more general towards a specific topic. Uhhuh. So why, why did you end up end up leaving and doing your own thing rather than trying to explore anything else at the moment where you, you already were.  

So the thing for me before I even joined, I already shut the clock ticking. I knew no matter how good it is that I'm not gonna stay for too long.

It was already a bigger company. It's like DeepMind is not, I was surprised by the way, I didn't know this, but like, DeepMind is now 1000- 600 people. Plus there was the recent merger with the brain. So like, who knows how big it is. So it's, it's not like a small syrup environment. Like, so, so because of that even before I joined, same for Microsoft.

Like I knew ultimately that I'm very, I'm, I'm too may, maybe too high agencies a bad word [00:27:00] because many of them are too high agency as well. But just like. The entrepreneurial side of, of, of me is, is too strong for me to have somebody above me, like a manager who's gonna tell me, Hey, do this task, do that task.

And, and regarding the switching teams, I actually wanted to switch a team and I was trying to switch a team. That's, that's actually building some of these larger scale models. That was, that was my, my, my goal. But like, it's, again, it's a pure company and it's not something that happens overnight.

Yeah. It's not that easy to, to actually switch teams depending on where you are. At the org it was very painful and, and with a lot of frictions. So like that, that's it. And I was already at, at year and a half almost. And I knew that even if I transitioned, like I, I would say for too short and like, I dunno.

So for me, this was the perfect moment. And then the day, like, I think the day I left, basically on Monday, GPT-4 came out. I was like a blast. I was Okay. This is, I'm, I'm, I'm living on, on the actual, on the top of, of the, of the, of the hype in the AI space. [00:28:00] Well that's, that's still to be seen, I think. I think the field is still kinda going upwards and I, I don't think that's gonna stop anytime soon.  

Would you say that the accessibility of those models to anyone able to, to build lots of incredible applications, is it like, is it more of a, a blessing that it is accessible or it it makes it so the competition is so high that it's even harder now?

Competition is always better for consumers and always not the best thing for those who are competing, right? Because it's kind of going, it's a race to the bottom as they say. Like, I think what happened is we lost this gatekeeping capability in the ML field. And now if you're a software or a web dev, you, you have a fairly good chance of competing with somebody who is like very proficient in machine learning, in my opinion.

If you're trying to build a product, If you're not trying to just innovate on the, on the model side, and that's empowering and annoying at the same time. You've had people like, I think his name is Pieter Levels. I might be mispronouncing his name, but [00:29:00] he's a very famous in entrepreneur who's built multiple companies own his own, literally like a solopreneur guy, like one guy, maybe a couple of people helping me him a little bit.

And he just like built these two companies called Avatar ai. I think he rebranded that one to photo AI or something. And then the second one into ai and both of the, both of those companies are, he took stable diffusion model and this became so easy that somebody like him, because like I I would bet like almost zero if not zero machine learning knowledge or expertise, was able to fine tune stable diffusion and, and get like the pictures in front of people inside of a web app.

And his current revenues are like his very public about this, by the way, like $100,000 every month. So, so the, the monthly recurring recurring revenue Yeah. Is $100,000. That's for many VC-backed companies, startups, YC startups as well. That's, they, they never achieved that. They, they failed before that.

So this is, here's a single guy with zero ML [00:30:00] knowledge all of a sudden building cool stuff in ML. So that's terrifying for those who are incu in the space and who thought that ML is their, like, defensible position. Yeah. But it's also very empowering because that means anybody can build cool stuff now.

And then of course, if you have deep expertise, and I, I, I like to think I fall into that category. I think it's much better to position yourself a bit lower in the infrastructure like stack as opposed to be on the applicative layer. And for me, that's mostly exploration at this point.

I'm just exploring and, and also just learning. Like I, I, I learned a ton because I had to build everything, the front end, the backend, the data pipelines. I, I'm currently the process of wrapping up the monetization parts. So I had to learn about how to deal with pedal with all these web hooks. And there's so many things and Google sign-ins.

And, and so just from the technical standpoint, you know, also just like building the whole thing. I think it's very empowering, but ultimately, I honestly think that's probably not the best place where I should compete. [00:31:00] Because like here, when you're in the, on this level of the, of the stack, you, you want to be fast and, and you want to be familiar with web technologies.

And that means that puts you in a pool with a lot of people because there is a lot of very hardworking people who know web. And so that's not kind of perfect position. But if you go deeper down the stack, maybe the, the something that like Mosaic ML folks have done, I think that's amazing. Like if you have deep expertise, You can start something like Mosaic ML, where they're basically allowing you to train your own models at lower cost.

You have bigger flexibility. Data stays on your side. You don't have to give your data away to like third party APIs. No matter what their privacy policies are. You, there is still the, the, the, the level of trust you have to have. Whereas here, you just train on your side. So I think that's kinda a good position to, to, to be in right now if you're, if you have the Deep ML expertise somewhere more deeper in this stack.

On your end, you had multimodal learning and vision language expertise, as well as lots of experience with YouTube. So [00:32:00] I assume the Ortus choice is also a great, a great choice for you where you have expertise. But before entering into that, could you cover or explain what is Ortus? How does it work, what does it do?  

So on a high level, what Ortus currently enables you, it's a Chrome extension. You open up a YouTube page and you watch video, and you can ask question about what's happened in this video, and you get like an immediate reply, and you also get the precise timestamp of where that relevant answer is in the video.

That's kind maybe the, the gist, but there is also the summaries. There are like the episode level summaries, the chapter level summaries and, and bunch more features. But the main gist is you want to find information in O one instead of in o n in sense, like when you're seeking some information, you usually have to go through the thumbnails and find details you care about.

Here, I'm just gonna ask and get immediate to reply. Like, so it's, it's in a way almost like a search engine, like, although the scope of the search engine is a single video. Yeah. And, and yeah, that's just kind of what it is on, [00:33:00] on the high level.  

It's funny because it, I, I feel like it's the same goal as the the shorts feature where we basically try to, for example, for a long, but yes, we try to find the, best bits of information and create short to reduce that surge that have, uh, complexity. And so like you are basically doing that with just a text query instead of manually trying to find the best sequence creating a short and sharing it, and so -it's, yeah, I, I feel like it's really cool. It's a really good timing where podcasts are becoming bigger and bigger. Like everyone is doing one and long form content is super popular, but what's not bad. But like the thing with long for content is that it's often not nearly edited.

It's not edited nearly as much, and so you need to, to watch it and wait for the information to come, but most times it's just to relax and like enjoy the conversation anyways. A hundred percent. A [00:34:00] hundred percent. If, if you want to gather. Insights and stuff like that. It's definitely not like you don't have a good bang for your buck listening to podcasts, it's like way too long to have information.

So it's a really good timing to, to build such tools. And have you seen a hundred percent, have you seen lots of competition in that? Because I, I, I haven't, like, I, I've seen some summaries obviously with like built with Whisper and, and stuff like that, that, that could make summaries of podcasts, but they are, they aren't really amazing.

But I've, I've, I have not seen things like, Ortus where you could have chapter summary as well as asking question to it. That's, that's pretty cool. And it's, I think it's pretty novel, but on your end, since you've, I I'm sure you, you did the market study, so did you see anything interesting and how did you end up doing better or trying to compete with them?

First of all, thanks for the, for the compliment. I, I think what I've seen is there is [00:35:00] a bunch of people who go on Twitter create a project using the link chain for like 30, 40 minutes, and then they say like, oh, I'm doing the same thing as you are doing. But like, it couldn't be further from reality because actually building something in production on a bigger scale that that works better is, is so much harder than, than you just connecting like a transcript of a video and then using three lines of flying chain to query it by using maybe some one of the vector databases like Chrome or whatnot.

So in that sense, there is a lot of competition, but I haven't seen any of them actually push it to make it useful to anyone on, on any serious scale or with any serious SLA or like, like we, yeah, we are actually available. You can go use us and, and, and, and get some value hopefully.  

And then I did see some of the. Extensions who are doing similar things. Chat Cube is definitely doing something super similar, if not the same at this point of time.[00:36:00] But as I said, we're kinda evolving, so for me, this is more exploration than, than the end goal. Like and, and then I think this cider uh, extension is doing similar things and not only, they're not kind of constrained to, to video formats on, on YouTube in particular.

They're for any blog or whatever you're doing, you can install them and, and, and have some type of summarization and question answering capability. So yeah, people are building some, some similar stuff. But I didn't see anyone who is super focused on just building this and, and, and accelerating.  

Are you looking to add articles and things like that as well? Because I, I assume you are working with transcripts, so it's, it should not technically difficult to. Add everything, but like in terms of the, the machine learning part, I assume it's pretty similar. So are you planning to, to add them?  

Not really. Not really because there, there will be now the spreading in a, in a direction that I'm not really interested in.

I, as I, as I said, like I, I, I don't wanna disclose too, [00:37:00] too much right now, but like I'm thinking of some other things of, of where I think I can have defensibility as opposed to, to going and attacking. Just, just going broad. I think it's much better to, to either be in a particular vertical, like currently for us, YouTube and doing amazingly good job at that, rather than spreading too thin across everything and getting some, a little bit of value across everything.

I think it's better to have a bunch of value in a, in a single thing. That, that's, that's my thinking there. And if you wanna go broad, for me personally, I, I wanna go deeper down the stack, not stay on the applicative till layer. So that's how I think about it.  

How did your more technical and and research, machine learning background help you compared to levels, for example, to build Ortus? Like, did, did it help you in any way or you, you just had to learn everything and just use APIs and connect stuff? Like how, how was, how helpful was it?  

Yeah, it definitely did help me to, to know machine learning, [00:38:00] especially in particular large language models and, and understanding tokenization and tokenization issues.

And thus you're better at prompting when you know some of those things. And, and also just the mindset of how you research and pick up stuff and where do you find exciting ideas to try out in your, in your system. And then like going through line chain and immediate understanding everything they're doing easily.

And, and so that's all of my previous background of understanding how ML works, the concepts, the terminology. So it was very easy for me to just pick up something and, and, and re-implement it on my side. When it comes to line chain, actually did. Create a whole custom logic of routing and, and, and, and bunch of things that I can't disclose because it's not open source.

But and then later I, I kind of put names on, on some of those techniques. I, I basically, you reinvented on my side just by going through line chain and understanding, Hey, they're calling this routing chains or they're calling this, I don't know whatever. And, and so yeah, it did help me. That's, that's at tldr.

I also had to learn a bunch of new stuff, especially Chrome extension. [00:39:00] Building a C Chrome extension is super painful. Google went through this transition from the manifest V2 to V3 version. 95% of the documentation online is completely obsolete and not working. It's, it's actually very painful to build a Chrome extension.

I, I didn't know that in, in advance. And also, there are many pities like. There is like this content script and, and, and there is this background script and there is the option and all of them have different access to different APIs. So sometimes you have to communicate from the content script, which is the thing you actually see on YouTube.

The, the, the widget. It has to send some data to some background script that's running in the background in the browser. And then that script is communicating further information and like there is just so many peculiarities and, and things I had to learn. Also, just the web design part, the, the, the web development, the backend part, like dealing building a nest, nest js framework on, on, on the backend.

I, I had a luck over the first weeks. I, I actually had a, a, a friend of mine who was, who was helping me out. I'm currently just solo in this with my girlfriend who, who is amazing at product design. She's a product designer by, [00:40:00] by, by profession. And so she's helping me on that front as well as on, on various other ideation fronts. But yeah, it was a lot of learning and machine learning did help.  

Awesome. I had a question about like, what was something that you didn't expect to be nearly as complex or difficult to do? So I assume that the Chrome extension was definitely one that you ordered, underestimated.  

But is there something, the top one thing there, there is top one thing. Top one thing is Google sign-in, you would not believe me, like Google sign-in was the most painful. You, it's supposed to be a single button that you click, you open a popup, you pick your email address, you sign in, yeah. And that's it. And you're all, all of a sudden you're using this so-called identity platform.

It was so painful. Like, again, the, the, the issue, the combination of using it to set up a Chrome extension added a additional complexity. So you couldn't use some of the existing React libraries. So you had to use directly this, [00:41:00] this Chrome API in combination with sound. Other APIs and it tur, it took me way longer than I thought.

I thought it was gonna be like, how hard may be every single website in Chrome extension in the world. Okay, maybe Chrome extension is a bit less are using that. It was so hard. Like, and, and so there was something I I've done next day, a week ago, so that was super painful. Other than that, let me think.

Technically, well, building this payment processing pipeline was, I wouldn't say difficult in the same sense. It's more you have to do this work, you have to figure out documentation and do stuff. Whereas with Google signing, it was literally every single code snippet I found online was broken. I had to go and fuse five, six different sources and, and get something to work.

And I'm, I'm going to actually push that into prod. On Mondays it's gonna be a bigger update. Monetization and Google signing are gonna be, Now a part of the app. And so it's gonna be fun. And that's something that people on Twitter with line chain definitely [00:42:00] don't see, or don't, don't hundred percent tackle yet.

Hundred percent. You just see a Google site. You, you, you just don't know the level of complexity that goes into making something work without issues handling all of the exceptions errors when you're dealing with, with with web. There is just so many things that could go wrong. You can have networking issues, you can have streaming issues.

OpenAI might, might be, for example in a, in a like search demand. And so the, the, you have to handle the errors there so that it's, it's a lot of engineering work that you never see as somebody who just has a widget in front of you and you're just dealing with. Asking questions. So it's a, it's a qualitatively different beast compared to just creating a MVP for a weekend and, and, and, and thinking that's, that's the same thing.

Exactly. Is there something on the opposite side that you expected it to be hard but was much easier?  

Not really. I think that as long as you put your mind onto something and you put your time and effort, and you give me enough [00:43:00] time, I'm going to. I'm going to solve it. Like most of the, of course, if it's blue sky research, there is a, there is an option that you will never get to the solution.

Like obviously AGI cannot be one of those tasks that they put in front of me. But like, because of that, I don't think it, it's usually just sometimes you just have to grind through it and, and, and learn. I didn't have anything that was surprisingly difficult to be honest.  

Like you, you mentioned that you may not be the right person to do everything. Like, for example, you, you may need someone that is an expert for front ends and for all the, the, the different parts. And so are you already looking for other people to help you?  

Yeah.  

And also, so that's great. If anyone is interested in, in helping you now you've built the, the whole foundation work.  

But having to start over, would you recommend someone to do as you did, and try to do everything by themselves or start right away with a small team of diverse backgrounds?

It all depends. I don't think of myself of as this omni potent [00:44:00] creature that can do everything. Like, first of all, I'm, I'm, I'm, I'm time constrained. If I'm doing backend, that means I'm not doing ML. If I'm doing ML, that means I'm not doing the signing. If I'm doing monetization, that means I'm not doing the front end.

So like, I can't do everything. So like, I'm, I'm very, I, I, I, I super appreciate m value, good teams, and, and I like working with people. So all of those combined make make it very, very like apparent that I should hire and help people work with me by, by this very nature of me being general, like I'm not the best at anything.

So that means I can hire a frontend engineer who's gonna be better than I am, even though I can probably pick it up in, I don't know, like 1.8% slower than that person who is like a professional or something. I actually became fairly efficient at this, at this point. But like, I still, I, I don't want to do it if nothing else.

So I wanna have somebody who is better doing the frontend. I wanna have somebody, somebody who is better during the backend. I, and, and, and then there's also the whole, the marketing, talking with VCs. Once we start raising money, that's a different beast. I was kind of so far rejecting the offers [00:45:00] because I wanna find my own conviction before I go there, but, Yeah, you definitely, if you can find a couple of friends and start building with them.

And that's how I started, like, don't, don't get me wrong, like I hopefully didn't get that impression, but as I said, my girlfriend helped me from the get go. Also, I had my ex co-founder now, so he was the friend I mentioned. But because of personal circumstances, he couldn't continue because bunch of stuff, financials and, and, and he's, he's currently working in a startup that he can't quit just like that.

And so, so a hundred percent if, if you un, un, unless you're somebody who doesn't know how to work with people, but then you'll never be able to build a big company, in my opinion, including levels. I, I think one of, one of the critiques I would give to him is, of course everybody has their own goals, but like, in my opinion, he's too, he's not too ambitious.

Even, even in the sense like he's never going to try and just focus on one single thing and make it huge. Huge, huge. Yeah. And, and also because he's very anti vc money. [00:46:00] And he just doesn't appreciate the fact that that can help you accelerate and build bigger teams. Because everything he's built so far has been a bunch of companies that are where you need a couple of people to run them.

And, and some of them are fairly successful, don't get me wrong. Like, like no, Nomad List is a fairly big website, but he's been working on that for seven years on his own, maybe with a couple of people helping every now and there, then as, as contractors. So I think ultimately if you wanna scale and build something that's huge and has a massive global impact and leaves a dent in the history of civilization, I'm, I'm becoming like, romanticizing this whole thing.

But I think then you, you have to build bigger teams. You have to raise money. Most of the time. Some, some people can be bootstrapped, but most of the time you have to accelerate because if there is competition and they raise money, how are you going to outcompete them? Yeah. So that's the thing.  

And you mentioned that you, you would replace yourself with exports in front end and, and all, all the, the, the specific areas like you are a generalist and I, I [00:47:00] think I am so too, that's why I'm on YouTube as well and covering all the d the, the different papers.

I just really love reading them and love learning about all the different things and touch to everything you said it yourself. You would hire people that are better than you at specific things, so like specialists and experts. But when would you hire another generalist? I,

I think that's one of those dark magic thingies that's based on intuition. There is, there is, if there was a simple rule heuristic I could give you, then everybody would be building startups. I think I, I, I recently listened to some podcasts where they were calling the generalist folks. Canons and then everybody else is, I, I dunno, like cannonball or something in the sense you sometimes need people who are more natural leaders and, and in general scope.

But like, when is that moment, when is the right moment is, is very too specific to every company and situation to know it. Obviously you don't want to start, like if you, if if you have a team of, of six people and [00:48:00] everybody's generalist, but not expert in anything I don't think that's as good as having six people who have expertise in one thing, but then general across everything else, like the, the, the sort, the, the T shape, the classic T shape.

So I think that's probably better. You wanna have people who compliment you across. Different areas. And, and then, but everybody, because those are early days of the startup and everybody, everybody will have to do everything. You want them to have the entrepreneurial spirit and generality and not be just like a cog that you can put on a single part of the system.

And then when everything else is burning, you cannot pull that person into just extinguishing the fires.  

So you definitely need some specialty, but it's still pretty much required to be interested and, and learn about the other things, at least to, to work with people efficiently. A hundred percent. You have to be willing to learn because you will have to learn a ton. There is just, everything is unknown.  

If we think of the mainstream researcher, like if we say Einstein [00:49:00] or whatever. Mm-hmm. We would think that they are like so much far into their, their own research and their own thing that they cannot even communicate to other people what they are doing since they are so experts in one thing, which we, I assume we need in, in, in some fields, but especially for the industry.

This, this cannot work. Or maybe it can work, but it's like very, it, it needs a good manager or something. I don't know if Yep. Like I, no. Would you say that someone that is like extremely good at something, doing a PhD in one very particular thing, super good, but has no interest in, in learning other thing and, and, and even working with people. Can, can this person be successful just because this person is so good?  

A hundred percent, but, but not, not in the entrepreneurial space. They are not going to build their own company. Well, Well, Larry Page and Serge Bryn would argue were arguably experts in the field and they were not really good at anything else.

So they early on hired Eric Schmidt to help them [00:50:00] be the CEO and build a company. You, you can even succeed in the startup world, although you'll never probably be the CEO that, that the prototypical CEO is, I dunno, like Elon Musk or, or whatnot. But like, if you're working as an embedded, like individual contributor, I think OpenAI and DeepMind is mostly consist consists of those people, right?

They're extremely good at something. Well, they also have to know how to communicate because ultimately if you a researcher, you know, you have to know how to present your paper, but only, only that much and not more, you don't need anymore. A good example would be this guy called, I think Alex GREs was his name.

He used to work at Deep Mine. He's one of the co-authors of the Neur neural altering machine paper, the NTM paper, which was very, very interesting. And he was an ic, so that means individual contributor. He was not a manager. He didn't have anyone. Below him, and he was like level eight at, at Google or something, which is like s stupendously high in the like VP level or whatnot.

And he, he didn't have, he had zero people underneath him. [00:51:00] So I think in the, in, in some of those value resource labs, you can definitely be completely, completely, like it's a derogatory word, but like, they call it as well, like in the sense you just know one thing, but you're so, so good at this one thing that, that you just compensate for, for the lack of generality. So there is space for everyone. It just depends how you, how you, how you position yourself and, and, and what's your goal.  

And if you are just an average person, how would you succeed In the, in the startup world or just in the industry? If you, if you cannot be the best, like there, there's always someone better than you.

I, I assume. And so how can you still succeed if you cannot be the expert? Is the, is the easier route to be more general, or should you still try to, to be like the best at something, even though you, you may not, is there something you would say is, is more easier or like, has more chance of success?  

Yeah. When you say [00:52:00] success, success at building a company, that's a particular goal we're optimizing for, right? Yeah. Have a successful career as in like more than average outcome, I'd say like of course. So build a company is, is definitely good. Or be like higher. Higher up in the. Google or, or big company.  

I mean, if you're, if you're average, then by definition you're going to probably be average in your outputs as well, unless you're super lucky. Right? So, so, so you, you do have to stand out in one way or the other. Now to your question of this spectrum between being broad in general versus being narrow expert, like, it, it, it's a spectrum between, like, you can also make, make it a spectrum between DFS and bfs. Like are you doing the depth first, exploration or breadth?

First exploration. And I think there are plenty of examples across that spectrum of people who've succeeded at whatever the final goal is. Be it the building, a startup, I mentioned Larry Page and surgery brain maybe being on the. One end of the spectrum, maybe on the other end, [00:53:00] I dunno if it would be a good example of super general in tech.

There's plenty. I think there are probably more general examples in the tech space than that, than super narrows examples. And, and the same goes for for career if you're trying to accomplish for engineers, a hundred percent. There are people who, who've been like very general like myself, who can make very nice progress in career.

And you also have people who are obviously super narrowed down, maybe compilers or whatnot, and they become amazing engineers. There, maybe Chris, Chris Lattner will be a good example of, of somebody who's like very super good expert or, or, or is it John Carmack? Like amazing expert and, and still they succeed.

So tldr it's, it, it doesn't really matter. It depends on, on your personal preference where you end up on the spectrum. But like all of them, if you're putting in enough effort and if you're above avarage, then. You have chances of getting there. Like I, I'm seeing you have chances because it, it's not guaranteed right there, there is no guarantee there, there is a factor of, there is ity in the system.

[00:54:00] So even if you're good you might not succeed as somebody who is much worse than you. And, but you can definitely increase your, your, your, your luck. You can, you can make your life Yeah, as they say[[. I definitely agree on what you said with the, the effort you put. And I disagree on, on what you said, on you need to be above average.

Well, I believe even if you are below average, if you put enough effort, you can definitely succeed much better than lots of average. Sure, sure. But then the definition, it depends on the definition. It's, you see, that means you're above average across the dimension of, of much, you can put much more effort.

So, so that says you're, you're not below average. If you're average, then by definition you cannot become the best unless there is some luck factor. But like I, I think that extreme effort is super underappreciated, especially like among intellectuals. I think that's probably one of the most defining traits that, that, that, that's kinda the best predictor for success anywhere across anything.

Because if you can, if you're tenacious, [00:55:00] if you can pull out, if you can work on something for 10 years in a focused way, you put a lot of work into it. It doesn't matter if you're below average on whatever metric intelligence or whatnot. I think you can compensate for many, many things. But if you have that extreme ability for, for, for that type of work, then you're definitely not average. There is no way you're average. Yeah, that's a, that's a super scarce trait to have. So in that sense,  

Ob obviously if you, if you are focused and you, you work more than other than than the others, you will basically have more luck just because you, yeah, you have more opportunities. But it, it's funny how like back in the days when we are younger or, or just when we are in school, the person that seems most intelligent or that, or that we, we are like impressed the most is the person that studies the least, which is like, it doesn't, it doesn't make sense.

It just, it feels like this person has talent by, by nature, but then [00:56:00] they, they end up like not studying enough or not working enough and they, they just sit on on what they have by nature, which is definitely not the best way to, to tackle like, yeah.  

But, but it also might be, again, overly romanticized, I think a lot of them. For example, you meet somebody in high school, you know, and, and all like, they're not putting in the effort or, or during your bachelor or whatnot, and they're not putting in the effort yet. They are, have these awesome grades. Well, you're forgetting that they had 12 years of education before they joined at least, or, or whatever the country is be before that.

And, and maybe they were like amazingly ambitious in high school and elementary school and they were putting in more effort than the other kids. Yeah. And then you see, because we are not at the same stage, like, and it's very hard to calibrate for where we come from. And so I think there is a lot of confusion because of this.

Ultimately, like talent does exist. I'm not denying talent. Like of course, somebody who is two meter plus [00:57:00] tall, of course they're gonna be better at slam dunking than somebody who is 160, like talking about centimeters. I'm not sure whether you guys use Yeah, yeah. The imperial system or, but but yeah.

Yeah, obviously I, I just want to come back, be before finishing the discussion, I, I just want to circle around to Ortus and if you can disclose it, what is the next step for Ortus? You mentioned the, the, the big update that you are going to do, but what is the next step related to machine learning or to any understanding task or, or videos itself? Like, is there anything you can say about what you are attacking next?  

I, I'm just going to say that I'm very passionate about training models and, and, and, and doing some deep tech work and some of the issues I've seen. Building the ortus gave me some inspiration to potentially start doing something additional. So yeah, that's as, as, as vague as I can get, but, At least, at least I'm saying that, that I'm going more deep [00:58:00] tech side.  

Awesome. I'm excited to see that. You mentioned you will stick with videos, and I know that you love videos and content creation in general. So how, how do you think AI will be changing videos and content creation over the \\next few years? Is there gonna be a big difference or will will we stick with listening to podcasts and watching shorts and TikTok?  

I think it's definitely changing and we already see companies. That are doing something along those areas. I think a good one would be, is it Descript? I think it's Descript. They're basically building this AI first video editor software Yeah.

That allows you to do stuff. Like, for example, let's say the two of us, we make a mistake now while we are speaking, you can post production. Basically what they do is they automatically transcribe your speech, which is something you usually don't get with the classical video editors. And so what you can do is you can edit this, the, the textual part of your video, and all of a sudden they synthesize the [00:59:00] voice using your voice.

You, you just have to kinda what's the word? Enroll beforehand with your, with your voice. And then you can literally edit out stuff without having to refill. Because historically we had to refilm right? Some, if you wanted. I, I'm usually super lazy and don't have time, so I just kinda crops trim stuff down or, or edit out.

But that, that's very empowering. That's one example of what people are doing. So I think both on the video editing side as well as on the stylization side. Like we've seen what's going on with the images now. Like I think the, the, the, the most recent one was drag GaN like the, the, the GaN that's allowing you to just, well drag points along the image and then thus edit the image.

And I think that's the type of, of, of powerful editing capabilities is definitely coming for videos as well. It's just a matter of time. And yeah, that's definitely gonna change a lot of, a lot of things.  

Do you think video generated content will be, will be there soon? We see a lot of [01:00:00] generated images, but they aren't really used that much. I feel like, for example, if we see presentations or, or advertisements or anything, If they are used, it's because people wanted to use AI and they are saying that this is generated. I, I feel like they, they aren't really used because of their quality and that's what they wanted. Do you think videos will get good in the next few years or will it take like extremely long to have? When do you think we can use AI to create videos and start creating just content, like the, the most difficult kind of content with ai?  

It is a hard, it is a difficult question. Like I think a straightforward way would be to say we're in the same moment of time with videos where we were with DC GaN with images like back in 2015 where things are not really looking great, but like, you know, that with a [01:01:00] couple of years of effort and research and engineering, you would get there.

Having said that, like videos are obviously so much more data intensive, like processing wise, you, you need much more flops. And I don't think the, the, the hardware is still there, especially not for five minute videos. Maybe for, if we're talking about shorts like super short, like five, yeah, yeah. Five second GIFs or ten second or up to 32 seconds.

I think that probably in the next year, two, three years, we'll we'll see a lot of progress there for longer form, like ]five minutes or, or more than that. I definitely think hardware needs to, to catch up as well. Like the, the just is, I think it's prohibitive from the, from the cost economic standpoint.

Although, Take everything I say with a grain of salt. I haven't been really following the latest and greatest papers from this particular subfield, although I, I do, I'm aware when something, something big comes out, so that, because of that, I'm fairly sure I'm, I'm, I'm fairly confident that my [01:02:00] answers are, are, are roughly okay.

Like, right. I like, we see a trend going to podcasts and just, for example, we, we like to go to shows and things that are really human first. And just like with podcasts, it's, it's a lot of unedited audio content and like, that's how we want to spend our time. Just listen and enjoy the content and like nothing has to be perfectly edited and, and perfect at all.

We just want human discussions. So I, I wonder if trying like, so hard to, to do video editing. Like I, I, I, I see the, the, the use for the script, for example, for correcting the audio and improving the content. But for like, do we really want something that generates content for people to consume or will we just end up with generating informative content with AI and then using Ortus to summarize it for us and [01:03:00] just like never consume content, basically just use AI to, to create it and AI to, to ingest it.

Do you see a use for video generated content? I mean, I mean definitely, especially if you're a creative, if you're a designer there, there is a lot of economical demand to produce such clips, right? So there is an industry for. For such things. I don't think it's mutually exclusive with longer form podcasts.

I think those are gonna probably remain for a foreseeable future in the, in the human like space. In the sense the shorter form videos will probably be more and more like AI generated, especially if your goal is to create a personalized ad that's, that's whole's purpose is to have higher conversion for whatever your business is.

So there, I'm fairly sure it's probably much better to do something like that video generated content and it, it'll probably become much [01:04:00] better than humans as well, because you can just set an objective that, that you want to maximize your conversion rate and you train your neural network using some RLHF for derivative or RL hf some ideas for lining.

To, to produce the, the, the most attractive type of short video, but for longer form like humans. Ultimately we are the best at being humans, and I don't think that's going to change. Although I, I might, I might stand correctly, who knows? Like, because ultimately neural networks are very good at mimicking humans and, and, and weathering mutation game is, is much different compared to what we already are.

Well, I think that debate is on since during times, but yeah, I, I do think there is, there is space for both and we will continue to coexist with, with the machines, with our AI overlords as they say.  

I also saw a study where, for, for beauty, like for facial beauty, basically the, the, the most average you are the, the more beautiful you are. Like if your eyes are centered [01:05:00] perfectly, if your nose is, is placed perfectly. And so. I feel like the, an AI is basically an average, like, it, it represents the population and so it might be able to mimic humans better than, than one individual can.  

Interesting, interesting thing you said about the human face. I, I, you, you probably know about this website, like this person does not exist. I'm not sure if it's still in production, but when you try and generate human faces, they usually seem very beautiful. Like at least, at least when you correctly generate like a woman face or whatnot. And when you now said that it's hemorrhaging, it kind of makes sense, right?

Like all of those like kind of generic faces of beautiful women and, and, and men and, and so, and it makes sense. I mean just the statistical nature of, of how these systems are trained. Yeah.  

But I assume it's also trained on like, I don't remember the name, but celebrities data set, celebrity something, so Yeah, yeah, yeah. It [01:06:00] biased towards, towards beauty. And yeah. I have one more question for you that is not related at all to everything we discussed, and it's, it's related to something that, that you mentioned that you always try to learn and you, you, you try to be the most productive possible. You like you, you like sports, you like health and, and, and how to improve learning.

And so I, I wonder if you have any recent discoveries or, or book you've read or anything that could help the listener to in improve their life in any way, like either health or, or just learning efficiency or,  

I think I ingest most of my information ;;;nowadays through podcasts, to be honest. Maybe blog as well, which are condensed versions. So I don't have any book recommendation per se. I would definitely recommend Andrew Huberman podcast for anyone who wants to do these types of like live hacks, optimizations sleep nutrition. [01:07:00] Supplements everything else. So for me, I, I think the main thing, again, Bo, it boils down to tenacity. Just create some plan, some structure in your day-to-day and try and keep it up for many months or years.

And like you will undoubtedly alert a ton along the way. You'll tweak, you will, you will evolve I wake up, I'm going to drink that glass of water with like maybe one pill of omega three amino acids and I'm gonna, I'm gonna stick with that.

And then after three months, that becomes kinda second nature. Or maybe you have, like for example, I do have alarms on my mobile phone. Yeah. That remind me to do some stuff. And then you can introduce the second thing, like maybe, okay, now every day or, or every second day or whatever the curriculum creates, I'm going to go out for a run.

For like 15, 20, 30 minutes. And I think everyone in the world, like no matter how busy you are, you have 15, 20, 30 minutes to do some physical [01:08:00] workout and everything else is just a cop. Like you're just lying to yourself, obviously. So that's just, and, and then you, you, you, you keep it up for like three, six months or whatnot, and then you enroll into the future.

If you start younger, like in five, 10 years, you'll have so many habits that you've grown over the years that people look at you like you're some type of a, like a wonder can, when, when in reality it was a painful, slow process that took years and years. And it's the same thing for me, like I think when I was 12 years old, I'm 29 now.

Since recently, like I started doing some of these types of curriculums and that's, that's like 17 years of me exploring and sticking. And I actually never stopped. Like I, like I, my, my biggest break from a physical routine was maybe three weeks or the span of like 70 years. It was like the biggest break where it was not working out, and so then you just accumulate so much knowledge.

It takes time. Yeah.  

Have you read or listened to Atomic Habits?

[01:09:00] Atomic Habits, I know about the book and I think I'm probably applying everything they're saying there already. Yeah. But I never read it. I might hear words like a video summary. That's what I usually love to do with those books. Like I, I really don't have time.

There's too many books. You just go and watch like a five, six minute like a summary. And if I see that's roughly in the space of everything I already knew, I'm like, I don't have the time to read another book. But the, the, the audio books are, are really amazing to me. Like it's just podcasts for me. I, I listen to them when training or running or, or biking or anything.

And I feel like it's really, I, I prefer audio books than podcasts, but I also watch off podcasts. And I think Atomic Habits is a really good one to create, to, to learn how to create habits as you described. Just one thing, for example, is like if you are trying to start being more active, like running, instead of trying to go to, to say to yourself that, that like, I need to run twice a week for 30 minutes.

Instead, just try to every day like, put your shoes and go outside and just, just do [01:10:00] that. Like, if you don't feel like running, come back inside and that's all. But when you do that step first, it gets easier to, to put your shoes and, and just go for run. But then once you are out, you, you often will be just like, okay, what, whatever.

Just I, I will start running like it. There's no way I will get back. Right, right now. And yeah, it's, it's a good thing to just. Like diminish the, the, the first entry step for creating a habit like ma make, don't already aim for the ideal habit that you want to build. Just start, start slow and easy. Yeah.

And another great thing that you mentioned is when you want to create a habit, it's much easier to link an habit to another one. If you, if we have your, your example of drinking a glass of water. Every day when you wake up, you can start doing that, and then you add your omega 3s and then you add your, whatever I have, like calcium or something else that, that I take in the [01:11:00] beginning.

But you, you don't have to start with the perfection. You can just start with something. And, and.

Great point. And this is something I, I, I, I've, I've learned like naturally, like throughout my life, and that's, that's one of my best habits. Like even if I have 15, 20 minutes to spare, I will open up a paper or I will open up a podcast and I will just like, I think this is probably not like the same for everyone, but I can super quickly get into the focus mode.

I, I need 30 seconds. Like I, of course you need the spectrum again. Like you need 30 minutes. I'll maybe be the, be in a deeper flow, but I can already. With 15 minutes, I'm going to use it up instead of waiting for the perfect time slot. Yeah. For the perfect opportunity where I finally have four hours of free time.

Yeah. And then four hours comes and you're like, oh my God, that four hours looks intimidating. Like, I'm not gonna start. And most people just keep on procrastinating. Yeah. Like, like that because oh, I never have the time. I'm working full time, et cetera, et cetera. And then weekend comes, and then you just try and make [01:12:00] yourself busy, which is just like mass procrastination basically.

Yeah. I think for people like you and I that, that you, you and me that are doing, for example, when you were at Microsoft, you started a YouTube channel and did other stuff. But what's difficult then is that you, you have a full-time job. Like you have a, a, an eight, an eight hour occupation in your life and you need to, to fill in the, the little 15, 30 minutes with things that will actually produce something.

Which is something that like we are managing to do, but I, I believe is hard for many people just because for example, if I, if I take my, my girlfriend as an example, she, she often says like, I, I just have 20 minutes. I cannot do that. Or like, there, there's not enough time. And do, do you have anything, any recommendation?

Like is that genetic? Why, how, how can you [01:13:00] go around that? Yeah. Yeah. It's, first of all, it's hard. It's definitely hard. It's downstream from mindset. And I know this is a cliche thing to say, but if you have this mindset of urgency, I personally, like, I, I, I, I feel the sense of urgency throughout my day, to be honest.

And so if I'm not working, if I don't feel productive, I like, it sucks. I don't feel good and I don't feel I'm like, long term I will reap the, all of the benefits of, of like compounding interest of my time that I put it in. So like, Because of that urgency. I, I almost like, it's not panic, but like, it's almost like alertness starts rising in my body and I, I have to go and do something like I have to do something productive.

Of course, I can always, I, I do know how to, to, to relax as well, and that's a skill you learn because initially it's very easy once you become super productive, to become overly productive and, and then you're very close and sometimes you go into burnout. And so you also want to know how to relax and like you have to know how to switch to, to make the switch from [01:14:00] both modes.

Most people are very good at just switching to just chill mode. But once you build the up the muscles to build to, to, to switch into the alert mode, then you have to additionally build the muscle for going from the alert mode back to the, to the relaxed mode, that makes sense. Yeah. You have to have high pain tolerance.

Like sometimes it's literally physical pain, like you can feel like in your, I like sore plexus, I think is the, this part, like you can literally feel pain there from like, oh my God, I have to not do this. I would rather go outside or whatnot or just like, lay down on the bed or sleep. But you push through it.

I, it's, it's in that respect, I think it's very similar to physical workout. Like of course it's hard, your muscles are burning. You would probably rather drop the bar from your shoulder down to, to the floor, but instead you keep 'em doing the squats or whatever you're doing. So I think I, I learned a lot about going through that, grinding it through, through my calisthenics and, and physical exercise, a routine that I've developed, as I said, since I was 12 years old or something.

Yeah. Like that. There, there is no. [01:15:00] Magical pill. You, you have to push to it and it gets easier with time. Definitely gets easier. Yeah. It's all in the painter, and I think the physical training definitely helps with everything. O on my end, when I was in a high school, I was just like gaming and not really studying.

And then a few years into college, I started training, felt much better. And then I started trying, try trying a bit harder in school, and then I created the YouTube channel and everything. And I like nice. It all started when I started training. Like it's, I think it's, it's definitely necessary in anyone's life.

Yeah. For many people, I think that might be the thing that gets the ball rolling. You, you, you, it's a prerequisite that you have energy, right? And usually the physical, that's the easiest part to tackle if you have more physical energy. You will probably, but not, not necessarily have more energy for mental work.

I think that's, again, a [01:16:00] different step. Most people never go from one to to the other one. They don't learn how to channel the, the same type of energy into the intellectual endeavors. And so they just stay at, at some something that's more physical. But yeah, a hundred percent agree with, with what you said.

Yeah. It's just like feeling bad or partly for depression, like when you are not physically active, you are, you end up not being mentally active, which ends up like, it, it is just a vicious circle. And if you, if you get one thing right, it should help all the others. And also just hundred percent in general manifests itself in health.

So I assume that's the same thing for, for just being physically in shape. You, it will make you mentally in shape as well. Just like ready to Yeah. Pull new challenges and just like, try to think and, and brainstorm and whatever. It's, it's really linked.  

Just a quick remark. I I, I definitely don't think it's, it's necessary condition though. Like, I, I think there are [01:17:00] plenty of examples of, of entrepreneurs who are chubby. Yeah, definitely not in physical shape. Like Reid Hoffman is probably a good example. The, the, the, the ex founder of, well, the founder of, of LinkedIN then if you take a look at Elon Musk, he definitely does not look like somebody who takes super much care about his, his body.

Well, recently that this, there's this whole thing with between Zuckerberg and, and Elon, so they're kind of competing or whatnot, but like historically, I think he, he didn't do that much workouts, if at all. I'm very curious. What was his. Routine if, if any, if anything really. So I think you can ultimately push it through.

Yeah. Like if you just hear about intellect, there is plenty of people who just do that. And, and I mean, I mean it's, it's actually, that's the stereotypical mental image people have. When you see scientists or somebody doing intellectual works, somebody who, who has their, their head can like stick like on top of their body, but the body is, there is, is a, is a, almost like a vehicle to just transport the body from point A to point B.

So, yeah. Yeah. But I would say that [01:18:00] the, I don't know if there are any studies on that, but they, they may be bar more prompt to burnouts without physical activities. I don't know. I, I, I just, I feel like it's way more healthy to just be overall healthy, but Yeah. Hundred percent that's right. That lots of people are prioritizing one or the other.

And I feel like both are equally important. I, I want to say yeah. To anyone to go try out. ortusbuddy.ai or on a lot of YouTube channel. Not all of them now, but I believe you are announcing them on your LinkedIn and elsewhere. Yes. So, and on Monday I'll have many more, many more channels on Monday coming up.

Awesome. And I know you already have Lex Friedman and Hubeman, so that's if, if you are, if you, the listener are not familiar with Huberman, you can definitely use Ortus and just summarize some of his, his styles because it's, they are really long but super interesting podcasts and yeah, I definitely recommend using Ortus, but also listening to [01:19:00] Huberman and to that I would add, I don't know if you've added it to Ortus, but the Tim Ferriss podcast is also pretty useful and interesting, I feel like.

Yeah, so that's, I I have many trouble. I'm, I'm planning, but like, it's, it's yeah, we're, we're trying to prioritize the quality as well, and so we're not expanding as much, but we'll definitely, yeah. Tim Ferris is on, on the list. Definitely. Awesome. As well as your channel. Thank you very much, and I'm excited to see that and try that.

Yeah. Thank you. Thank you so much for taking the time. Thank you for sharing all your insights and your, just your life in general. It was really cool to talk with you and yeah, thanks a lot.  

Thanks for, for the invite, Luis. It was a pleasure.[01:20:00]