0:01
There are plenty of problems that large language models are really not going to solve.
Search this video
0:06
This is called we actually think and it's like very messy and very probabilistic and very approximated in the way we approach problems.
0:13
If we can get a better system that can do coding for us, we don't want to just kind of probabilistic matching.
0:18
I think that could really turbocharge technology in general.
0:22
I've been in tech long enough to see that technological trends take a long time to spread.
0:28
If Thomas Edison were around today, we'd all be doing things by candlelight because they would make electricity illegal. Why?
0:34
Because electricity is incredibly dangerous.
0:39
If you build machine learning models for a while, there's one thing to actually think that something can be done versus like when they actually try to actually do it.
0:46
If you're Kaggle Grandmaster, you're Kaggle Grandmaster.
0:48
It's very hard to kind of fake.
0:49
Getting to that level is much harder than, you know, getting a PhD from some of the top universities. Well, hello everyone.
1:00
It's Jim O'Shaughnessy with yet another Infinite Loops.
1:06
Today I am super excited to have Boyan Tunguz, a person that I met on Twitter, completely on Twitter.
1:13
And why I met him on Twitter is an amazingly accomplished person, which we're going to get into in a minute, but the reason that I met him was because he's incredibly funny.
1:27
Uh I I started seeing his tweets going by and every time I I either had a smile or a laugh.
1:32
And then I And then I look into you, Boyan, and wow, like you are super impressive.
1:37
Born in Sarajevo, um came to the United States, you got multiple degrees in physics at Stanford, and then a PhD at the University of Illinois.
1:49
You are a Kaggle Grandmaster.
1:53
I got to ask you about that because that seems really, really cool. Um welcome. Thank you, Jim.
2:00
Thanks for having me on the show.
2:02
You know, you you and I interacted a lot for the last few months at least on Twitter and finally we get to chat with each other. Yeah, it's great.
2:09
And what's cool about it is what what I love about Twitter is I was originally attracted to you because you were funny.
2:16
And so I I started following you and then I started looking into you and I'm like holy this I mean, let's jump right into the old Kaggle Grandmaster thing.
2:31
I didn't know too much about it. And oh oh my god.
2:34
What you have to do to become a Grandmaster?
2:38
First off, there's only about 300, right? In the entire world.
2:43
Tell us about that and and what led you into wanting to do that given the fact that you moved from physics to machine learning, artificial intelligence. Uh you're at Nvidia.
2:54
I think I've heard of that company.
2:56
Uh they have a couple of things.
2:58
But but let's let's jump in on the Kaggle Grandmaster. Yeah.
3:03
Um so for those of you who don't know or don't heard of Kaggle and there's still a lot of people out there even in the tech world who haven't heard of it.
3:10
Uh Kaggle is data science competitions website that started like uh about I think 10 years ago like something like that.
3:20
It was uh machine learning competitions, that's how it became famous, but over the years tried to become much more comprehensive uh site for all sorts of data science related tasks.
3:30
Uh it uh uh kind of got me hooked when I was kind of switching my careers.
3:38
Uh that in itself is a long story how I end up in data science and machine learning.
3:41
Um you know, I I just today I had another conversation a little bit uh contentious online about like the future of education and in education in particular.
3:50
And uh you know, like I as as you mentioned, I do have a lot of pretty advanced degrees and I have you know, taught in the colleges and whatnot, but none of that's really kind of led to any consistent kind of career, especially in a field of like physics that are you know, I'm sad to say it's on the kind of it's dying legs.
4:10
So like I really needed to find something else to do with my life and I really didn't feel like going back to school.
4:16
So I tried a lot of online resources and you know, I and and machine learning community in general is very very generous with a lot of online teaching material.
4:27
Some of the first moves were in artificial intelligence.
4:29
I actually signed up for the first one.
4:31
I didn't finish, but I did sign up for it.
4:35
I did a lot of like well-known online courses like Andrew Ng's machine learning intro course, a few other Coursera, Udacity, but none of those courses were like really getting me anywhere professionally.
4:47
And like it's unfortunately, even though I think they've improved a lot over the years, but it's we still live in a world where like a college degree is sort of a gold standard that everything else seems like a much much less of an accomplishment.
5:02
So you know, I was looking around.
5:04
I heard of Kaggle ever since I joined data science, but I always thought it's impossible and don't have a you know, either background or like know-how to actually do anything there, but then one year I just decided okay, like I'm getting a lot of I'm just going to try something, see how far I get and and and kind of the rest is history.
5:23
I I joined the competitions.
5:25
I was just finding my way around and quickly discovered that um Kaggle kind of solves two big problems in online education, at least for people who are willing to kind of go through everything and you know, I got more and more into what takes to actually get to the top.
5:44
First of all, it's a very it has a very strong incentive structure to kind of keep you coming back for like learn.
5:51
So like a continuous learning like you can experiment and learn just by doing.
5:55
So that's a very very nice kind of setup that that they have over there.
6:00
And then I realized that people who do win these competitions, who like really are on top of the leaderboards, you know, you can't just get there like randomly.
6:08
You cannot just kind of it's not like lottery that you just kind of pick a lottery ticket and you kind of end up on the the top.
6:14
You really have to have a lot of good skills.
6:16
So people who are on top of the leaderboards and I've been blessed to kind of work with many of them over the years, you know, are very good smart people.
6:25
And like you these leaderboards really serve a purpose of very credible credentials.
6:29
So like something that a lot of other online resources don't quite get, you know, like don't you don't really you know, if you get some kind of online degrees, you know, from one of these online places like the question is like are you really as good as you know, someone else from some other online degree.
6:44
Like on Kaggle, at least people who are more honest about it.
6:46
You know, if you Kaggle Grandmaster, you Kaggle Grandmaster.
6:50
You know, like there's there's very hard to kind of fake, you know, and in the many ways, you know, getting to that level is much harder than, you know, getting a PhD from some of the top universities.
7:00
Yeah, and when I was reading about it, I I was kind of floored by first off, what you have to do to get it. Yeah.
7:08
Like the number of gold medals, the number of various projects, everything.
7:13
I really found fascinating.
7:13
Um but I also found it really interesting how hypercompetitive it is. Yeah. Yeah.
7:21
And and and then I saw a quote from you that was like, I love hypercompetitive learning. So so why?
7:28
Why do you love it to be hypercompetitive?
7:30
Um I don't know which word is that.
7:30
You know, I do and don't.
7:33
Um I I feel like uh you know, I and I've tried to reflect more about this over the years and I think it's one of the potentially more problematic sides of Kaggle because it doesn't really attract the smartest people, it also attract the most competitive people and a lot of them are not pleasant in the way they're competitive.
7:52
So you know, like I have a you know, war stories and war wounds over the years from from those kind of experiences.
7:59
Uh my own light side is like I I kind of don't personalize the leaderboards.
8:01
Like for me, leaderboards is like a kind of abstract thing, you know, and I kind of measure my own progress against the leaderboard, you know, in in abstract.
8:09
I don't think of these people who are ahead of me or below me like as human beings that I need to defeat or I'm better than or something like that.
8:18
So like I I I feel like a leaderboard is more like I'm being competitive with myself, you know, I try to be uh better version of myself and like I kind of use these leaderboards as a measuring tool for like a roughly gauge where I am in my own progress.
8:30
So like uh I am and I'm not competitive.
8:33
So like you know, I'm I'm not in competitive in the sense that I I uh want to be better than other people.
8:43
I want to be just better than my previous self.
8:44
So that's that's how I'm where I'm coming from.
8:48
I think that's actually the best kind of competitive.
8:50
Um I and so I totally agree with you, Boyan.
8:55
Uh it and what you're really saying is, "Okay, so here are these people above me.
9:00
Hmm, what do I need to do?
9:00
What do I need to do to to to get up there?"
9:06
Uh and then really not thinking other than, "Wow, okay, there's a whole bunch of people who aren't even where I am."
9:11
So that kind of gives you a dopamine boost there.
9:13
But but but as long as you're, you know, competing against yourself, I think that is the the best type of competition.
9:20
And like if you're going to be competitive, I I found that the best way to do that is to only compare yourself to your former self, right? Yeah.
9:29
Not not not to other people.
9:32
I you mentioned that at the beginning the the switch from physics to data science.
9:38
Uh talk to me about that.
9:38
I I I know that, you know, theoretical physics, not a lot of jobs outside of academia or certain hedge funds, I know.
9:46
Uh but but but what what uh tell that story.
9:53
So, you know, like I needed a job.
9:53
So, the question is like, what what can I do with I really didn't feel like going in a many of the options that were viable because, you know, I would still have to kind of go essentially go back to school, try the new things.
10:07
And you know, like I I was sort of fiddling online doing a lot of type blogging for a while that that was kind of keeping me occupied and I just accidentally stumbled upon the data science and and uh it really kind of jumped at me because it was just the right proportion of like technical skill and then analytical skills.
10:24
You know, like you know, it's you need to be more technical than you know, statisticians or someone who just does, you know, paper paper pencil or with small kind of uh work.
10:33
But, you know, you're not same thing as, you know, software engineers who are just kind of worrying over the code code all the time.
10:39
So, like, you know, it was right proportion.
10:44
It really kind of it played to the way my mind works.
10:46
So, like, it it really became like, "Aha, like I wish there was something like this when I was in school or something."
10:54
You know, that was that that kind of moment, you know, like I still love physics.
10:57
It it's really intellectually extremely gratifying uh thing, but it's very hard to kind of get that stuff and kind of apply to your work and life and and other things.
11:06
And but in data science, you know, like you can uh data science just your way through a lot of world you know, daily problems, you know, you can figure out how to optimize maybe your grocery list or like your your savings accounts or whatnot.
11:22
It it it it's really really fascinating once you you get into that whole field uh how much it can actually be applied.
11:29
And that that's one of the most wonderful things about data science.
11:35
So, you know, like I go you know, piled all these online resources.
11:38
It really clicked with me and you know, I'm I'm really happy to be that world.
11:44
Yeah, I my friend David Ha, who was briefly at Stability AI as our uh lead researcher.
11:50
He he and I were talking about it and he said what draws people to to the field in his opinion was just incredible curiosity, but they also love to tinker and they love to solve puzzles. Yeah.
12:06
Would would would would you agree that that's a good way of looking at it? Could be.
12:10
Like I'm more of a thinker than puzzle solver.
12:12
It's it's not really the same kind of I think again like the the whole world of data science and machine learning and now AI uh has many options for people with a different kind of sensibilities and skills, you know.
12:29
Like I I do know a lot of people who are really kind of hardcore puzzle solvers and you know, you can put them in a in a in a in a bookstore with one of those puzzle books and they be as happy as a as a clam.
12:43
Um it it's it's for me it's not so much the puzzle solving, but more about kind of big idea thinking.
12:48
You know, so like I like understanding the world.
12:50
So, for me understanding comes more gives you much more a deeper says personal satisfaction than just kind of solving a new puzzle.
12:59
You know, like I Puzzles are fine, but I find them a little bit dry kind of too abstract for my liking, you know, but you know, like being able to come into a new problem and then extract from it a new understanding about something about the world, that's the most gratifying thing that keeps me going.
13:15
Yeah, satisfying in terms of getting a better understanding of the underlying system and then letting that inform the way you approach it. I I I agree on that.
13:28
And you you said in 2021 that you've been floored by advances in natural language programming, but you also said, "Hey, it's only it's just getting started." Yeah.
13:38
Can you elaborate on that and what you see coming for our listeners and and viewers?
13:44
Yeah, um with years ago like you know where that was in Kaggle and uh like I think 5 and 1/2 years ago where that I really tried to learn more about natural language processing and really for the first time started applying myself to it and I came in the third in one of the competitions with my team.
14:04
Um I was I felt like, "Wow, you know, like I I finally understand this whole field and it's like very interesting and and it it has a lot of potential."
14:11
So, I thought like I I would get more involved with it.
14:16
But then things started moving really fast and really advanced and I'm like, "Okay, I I'm having replies here and I feel feels moving way faster than I kind of have the time or energy to kind of keep up with."
14:28
And like I'm like, "I'm going to kind of just sit on the sidelines, see how fast it goes and where it ends up and maybe in a couple years I'll I'll catch up and start like fiddling with it more."
14:38
And you know, there was no slowing down.
14:40
It just you know, went from you know, strength to strength and things are moving really really fast.
14:47
Uh and you know, now it it it's I think caught a lot of people by by surprise not only how fast it's moving, but how powerful these very conceptually straightforward approaches are.
14:59
You know, so that you know, the whole transformers and large language models that that can get you really you know, essentially you know, between the relationship between the words and you know, in the bigger and bigger context can actually produce some kind of emergent behaviors like pretty decent reasoning or thinking, you know, that that that can actually uh you know, by itself get you pretty far.
15:22
You know, it's still not perfect, far from it, but the fact that we even here, I I think it caught a lot of people by by surprise, even people who are uh who thought that were you know, the cutting edge of the field.
15:33
Uh in terms of where it's going, uh I think it's anyone's guess.
15:37
Uh I think we are still uh you know, you know, bigger and better models are being trained all the time.
15:44
And now that people know that this is the approach to uh better AI, I it's you know, everyone is in the field.
15:50
So, like I I think it until last year, I think there was still like fair amount of skepticism or people were not really impressed with that uh what's going on or like uh what we could have really accomplished and you know, how useful all this can be.
16:06
But I think now all those uh considerations are uh laid to rest and everyone who has any kind of uh interest in the field is is there.
16:13
Like we are just now just getting started with you know, full engine full throttle of all the resources that are being thrown at this and you know, you know better than I do like you know, what uh startups are working on.
16:25
So, you know, it's it's still very early because now we we finally have full attention of everyone who matters.
16:31
Uh you know, going all in on on this whole uh generative AI, I think. Yeah.
16:37
And uh uh another thing that I've seen you note, which I agree with, is that there are plenty of problems that uh large language models are really not good at solving.
16:50
And talk a little bit about that because I I res- that resonates very deeply with me.
16:58
Yeah, I I think the you know, it it's uh it became obvious uh with these large language models where people are you know, people making fun of them, but I think it's obvious now that the thinking and reasoning are two different things.
17:09
You know, that uh these uh it these large language models are sort of giving us the mirror to ourselves.
17:17
And what we see is like, okay, so this is how we actually think and it's like very messy and very realistic and very approximated in the way we approach problems and and uh you know, these large language models can get pretty far doing that and actually that's what we do.
17:33
You know, like you know, when we interacting with our friends or or co-workers or like in any kind of big interaction, we don't really try to really solve these deep reasoning kind of problems.
17:46
You know, that that we use a lot of heuristics, a lot of kind of approximate approaches.
17:50
So, you know, it can get us far, but if you really want to get to something that's really hard, you know, it's not going to get you there.
17:57
So, if you really want to solve some hard equations, just BSing through that is not going to get you there.
18:03
If you want to kind of create new knowledge, uh other approaches really would would kind of coming complimentary.
18:09
So, like our one big big area of research where I know some very famous AI labs are working on those automated math reasoning.
18:18
You know, it's a it's something that we still don't have any kind of uh real foundation for, but you know, mathematics is usually considered like it is the is the queen of all the intellectual disciplines as well as the and that's the most uh uh most productive in the terms of what we can kind of gain from from these systems.
18:42
So, like coming up with systems that can leverage some of these AI natural language processing things to get us to to sort of sort of like a place where we can actually apply other methods.
18:53
So, you know, like figure out figuring me out some kind of approximate solution that we can then uh solve exactly using different kind of methods.
19:03
So, you know, like get ideas of the outline of the proof for a math problems and then use these reasoning methods to kind of get us there.
19:10
I think we are that similar to the way uh chess was solved or Go was solved, you know, where where you you had these uh probabilistic models and then you kind of do exact calculations that as a part of the final process.
19:25
So, that'd be one of them.
19:25
I think coding would be the big thing that everyone is working on.
19:29
If we can get a better system that can do coding for us, uh beyond just kind of probabilistic matching, I I think that be extremely useful and that would really turbocharge technology in general.
19:45
Yeah, and you you make the point about Deep Blue and beating uh the chess grandmasters and then Go.
19:52
But, another point that you also make that I I that I like to highlight is the idea that after Kasparov lost, he figured, "What about man and machine?"
20:04
The so-called centaur model. Yeah.
20:09
And and what we found was not only did Deep Blue make human chess grandmasters much much better, it made just just people who weren't even at the grandmaster level much better chess players as well. What's going on there?
20:28
Uh in chess specifically or in general is this kind of model Let's let's let's let's start with chess, but then because I I hope that it it you could generalize it to other areas of learning as well.
20:41
Yeah, uh so my my understanding is that that now because you essentially everyone can have a mentor to teach you chess that's sort of beyond like top human player, it's much easier to kind of practice.
20:55
And I think that's one of the biggest opportunities for AI systems in general, not just for chess, but you know, having personal trainers that can really improve whatever you're trying to get good at.
21:06
Like, you know, imagine you know, having a better person who can read you for your math scores or or your drawing, for instance.
21:14
I want to get back to drawing and I want to figure out how to use generative AI to improve my own drawing skills.
21:21
Um and and I think that that uh that's really turbocharged a lot of amateurs to become even better players than otherwise would they would have been.
21:32
Because, you know, sparring against the best systems is always going to improve you, you know, like metal sharpens metal.
21:39
Um and then the whole centaur model that you mentioned was very viable with, you know, human-computer uh interactions where humans were providing kind of intuitive parts of the the puzzle with the computers, you know, working on the hard parts of things.
21:51
I think that was a very uh viable uh approach to, you know, being the best in chess for a while.
22:00
But, my understanding is like with the like latest systems, even centaurs really can't keep up with the uh best uh chess computer-only chess players.
22:08
So, like I I think that right now we are beyond centaurs.
22:13
So, I think something like that is going to happen with uh other fields where for a while a human plus AI is going to dominate just AI or just humans.
22:24
And then eventually AI is just going to kind of take off and just be in a league of its own.
22:30
One of the things that I that I believe pretty deeply in and and we've invested in companies.
22:35
So, for example, you're talking about drawing.
22:36
Uh one of our portfolio companies is called Wand, and it's basically for graphic artists, and it is literally a tool, but the the the wand attaches to the AI.
22:46
And then they essentially get to iterate against their own work.
22:52
It's really, really cool. Yeah.
22:55
And and but the part that I like about it is AI, at least for right now AI I see as an incredible tool.
23:03
Maybe one of the coolest tools we've ever invented. Yeah.
23:10
To to be able to allow for the flourishing and extension of human creativity and innovation. Yeah. Do you agree? Do you agree? Oh, yeah. Uh hopefully so.
23:20
Uh the problem is like uh all of a sudden I want to go back to all sorts of doodlings and and the little side projects I did decades ago because like hey, now I can do it even better than I ever you know, tried before.
23:33
But so it's a it's sort of a distraction from my day-to-day job because I feel like building AI versus you know, using AI to build something else it's like a huge trade-off, you know, like it did I I I wish I could like draw comics now or like make little you know, cartoons or whatnot.
23:51
Like I really wish I had like more time to kind of dedicate to some of these other activities.
23:57
So yeah, it's it's a very distracting uh for me at least.
24:01
It's incredibly enticing.
24:04
And I think a lot of people will be entering fields and and whatnot that otherwise would have probably not done so.
24:10
You know, like I'm I'm really excited to expose my own children to some of these tools.
24:14
So like I really want to see like what they can do like what what what what's their understand.
24:19
And they're still pretty young so they they're not at the stage where they can actually really take full advantage of these things.
24:23
But I I think it's going to be very exciting for a lot of creative pursuits.
24:28
And you know, I'm just giddy with excitement about it all. As am I.
24:31
You mentioned your kids and I saw somewhere that you are going to homeschool your kids. Yeah. Yeah.
24:38
I mean why are you going to take that path? Uh um many reasons.
24:45
Uh Part of it is you know, like I coming from outside of the US and being exposed partially to American public schools I was kind of really kind of let down.
24:58
You know, I I I felt like the schools were not quite as good as what I'm used to growing up.
25:03
Even, you know, some not so well-to-do countries.
25:05
You know, like they they they the schooling is much more rigorous and you really learn much more.
25:09
So, that was one of the uh kind of disappointments.
25:10
Um A- another big thing that, you know, became bigger and bigger over the years is, you know, what we were talking earlier.
25:18
Like the that uh uh I I'm not really happy with, you know, how educational system works in general.
25:25
And I feel like uh so much more so much of it is kind of wasteful and not really focused and not really tailored to individual people's needs.
25:35
You know, like I I have three kids.
25:36
Each one has a very different approach to learning, very different mindset.
25:40
And trying to kind of put them all through the same uh kind of one-size-fits-all process, I feel we very would be doing them a disservice.
25:50
So, like I I really uh want to do something that's much more tailored to each one of theirs needs and interests.
25:55
So, that's another reason.
25:55
Um and and another reason is like we like to travel.
26:00
So, we don't want to kind of be beholden to the school year and where it starts, where it ends.
26:03
You know, we can just pick up and go and uh you know, go across the country or across the world depending, you know, where it really suits our schedule, not really school years. Yeah.
26:14
Um I I completely concur about the antiquated nature of the United States school system.
26:20
Um I think actually AI tutoring, for example, is one of the best use cases.
26:25
Because uh we've seen at Stability AI, for example, we got a remit to educate using AI and tablets a third, I think it's a third, of the kids of the African state of Malawi. Wow.
26:41
And still very early days, but it appears that the kids who are have the AI tutors Yeah.
26:50
are like leaps and bounds ahead of the kids that are in the traditional uh classes.
26:55
And so um Like like using the technology with kids is very controversial uh subject and I have opposing views about it myself.
27:04
Um but I have seen it that that my kids have learned how to read and write on their own, like just by using these tablets.
27:12
So like I didn't even introduce some of the uh these things to them.
27:14
They kind of picked up by just playing these games, like putting letters around and you know, vocalizing what they're reading.
27:21
And you know, it can be a very powerful tool for like early uh literacy and you know, all sorts of things that were prior not possible.
27:27
And now especially with AI can really be huge boon to education across the board.
27:34
I agree and I I would take it a step further and say that children are designed evolutionarily to be learning machines. Yeah.
27:43
And uh one one of the big mistakes that we've made in the past in my opinion is that we spent way too much time trying to teach children what to think rather than how to think. Mhm.
27:55
And uh you know, we we give fellowships for our children's adventures and one of the first ones we gave was to a couple uh with five young kids and they they wanted to investigate and ultimately make a documentary about uh alternative ways to do it with children. Okay.
28:13
And and what we're seeing when you look into these methods is these methods by and large, not all, but by and large are are harnessing that child's innate desire to learn. Mhm.
28:28
And and then and then kind of letting them loose, so to Yeah. Yeah.
28:34
Um is is that something Have you designed a curriculum for your kids or are you not that far yet?
28:41
We we have sort of like a core curriculum that's actually very traditional.
28:44
It doesn't involve any technology.
28:46
So we go over, you know, reading, writing, math.
28:48
These are kind of core things that we do.
28:51
Um then I try to expose them to a little bit of history and geography through a lot of books that they they try to get them to read on their own. So, there's that.
29:01
Uh a lot of games are actually very educational.
29:03
So, like, you know, Minecraft, for instance, it can teach you a lot about mining and ores and all these things that I never myself knew about.
29:10
So, now kids might teaching me.
29:13
So, like, asking me like, "So, what's the bedrock?"
29:14
And I'm like, "Okay, I got to explain to you what bedrock is."
29:16
And I barely know it myself.
29:18
So, yes, I think um the core our curriculum is very traditional, just pencil and paper stuff, but the more of the additional things are uh that they kind of pick up on their own are a little bit more kind of technology-driven.
29:34
Um another thing is like, we're right now in the process of moving, so we're right now in a very small town in Indiana, and we're moving to Florida.
29:42
And one of the reasons why we wanted to move to a bigger place is exactly so we can actually have more uh um exposure to, you know, for our kids for more educational resources outside of the traditional system that we hope we can get in a bigger place. Yeah.
29:57
Another thing that that I know you're interested in and think a lot about, um especially with how AI might or machine learning might impact it, is is the future of work. Oh, yeah.
30:10
And and and and you and I share another thing, I think.
30:13
I don't want to speak for you.
30:14
You can correct me if if I'm wrong.
30:16
But, um I at least I, and I and I think you, maintain that, you know, human creativity is going to improve to a point where it would have been impossible to get it to that point without these new tools. Do you Do you agree? Yeah? Yeah. Yeah.
30:32
I I think it it uh people are saying that, you know, these these uh new generally AI systems are just uh uh uh stochastic parrots or something they're just kind of parenting back to us, but the kind of things I've seen are very original and very kind of unique.
30:49
And so, like it exposes you to much wider spectrum of things that you probably wouldn't have never thought about.
30:55
So, like, you know, like you can uh you know, give a prompt to one of these systems say like draw me a frog made out of cheese on the moon or something like that.
31:06
And like, you know, no one There was no prior, you know, image of that anywhere.
31:10
So, just kind of being able to kind of experience with these ideas or like what they would look like.
31:16
Um I think for me at least, and I think for a lot of people, the key to creativity is iterative experimentation.
31:24
So, like, you know, the more you can try more things you can try out, more things you can see how they work out.
31:30
And closer the the feedback loop is, the better you can get.
31:35
And like, every single thing that I became really good at in life, there was a very close iterative loop.
31:40
So, you know, the Kaggle, like, you know, you build a model, you submit it to the leaderboard, or like try to, you know, local validation, you see how it performs, go back to the drawing board.
31:50
So, like, it it was a very tight iterative loop that's really kind of helped me get ahead.
31:53
And I think generative tools as well, like, you know, like hm, you know, if you're artist or or you writing code and you can see like what certain function would look like, how it's do, it's will give you more options and more varied options to try out and kind of get better feedback from all of them.
32:11
So, I think that the the flywheel of this very tight iterative feedback loop is going to be amazing for a lot of lot of people in a lot of lot of fields.
32:22
I I agree on the the tightness of the feedback loop and the just the plethora of additional choices that generative AI makes available to you.
32:35
I love your example, like a mouse a on the moon made of cheese. And there it is, right?
32:42
And and then you you think, oh, well, and then that leads again to your tight iterative loop to ideas that you probably wouldn't have had.
32:52
When So, I'm 63 years old.
32:52
When I when I was younger, um pretty much people thought you were smart if you had a good memory. Yeah.
33:02
And and and now, I always thought that was a bad definition. Yeah.
33:06
And and cuz I always thought that uh the ability to think, to synthesize ideas, Yeah.
33:14
to uh iterate, as you say, was was superior.
33:17
And I definitely think that uh generative AI is going to help uh ex- a lot in this particular area.
33:24
If we're going to look at sectors, right?
33:27
Which sectors do you think are going to be able to use the tools of machine learning and uh generative and other AI as as the biggest lever?
33:37
And And which sectors would you say are probably not going to really be too terribly affected by it? Yeah.
33:46
I I I think the latter ones would actually be much wider uh number than than people people would realize.
33:51
Uh not only because of generative AI, because I I you know, I I've been in tech long enough to see that you know, technological trends take a long time to spread.
34:02
And there's a lot of inertia and there are a lot of barriers to entry for all sorts of reasons.
34:07
Some of them Some of them are legitimate.
34:08
Some of them are just kind of uh made up in order to kind of keep people in a in a certain uh positions.
34:16
And so, I think the whole process is going to affect you know, people on the ground much slower than I I had originally thought.
34:25
So, like, you know, when when all you know, ChatGPT came out in November last year, I was like, holy crap, you know, we're on the verge of just major like disruption across the economy, across everything.
34:36
This time next year, things the whole world is going to be completely different.
34:39
But not quite in didn't quite happen.
34:41
So, like I think on the low end, things will take a while to kind of percolate.
34:46
Um on high end, I I I I think people who already building AI tools are in the best position to actually build the tools for themselves that will make them even more uh effective.
35:01
So, people who know how to code, people who know how to kind of build better better models, they will be very efficient if they if they choose to uh and big big building even better tools and better models if if they choose to.
35:15
Um in between, I think uh our creative pursuits like uh uh literature writing, uh comic books uh could be affected. Uh the question is how? We really don't know.
35:29
Um I I feel like with every other technology, we will have uh new professions that will pop up that people never thought about before.
35:36
Uh this will be enabled and new skills that uh prior did not exist like this, you know, we all make fun of prompt engineering, but you know, it really is a very good skill.
35:46
And like people who know how to create very creative prompts and how to use these tools and how to chain them, you know, those are skills, you know, they are they are not, you know, maybe the same as software engineering skill, but you know, they're very life valuable skills that people are able to kind of prolong.
36:01
So, I I think you know, high end, people who know how to use AI will be better at getting even more out of AI.
36:11
At low end, you know, people who are I I don't know, um clerks at your local post office probably not going to be affected one way or the other for a long time.
36:19
So, I think there's a spectrum at least in the knowledge industry.
36:25
Yeah, and and one of the things that I worry about is um on the one hand I think it's it's vitally important for us to understand as best we can the technology um and it's it's uh benefits are obviously numerous, but we also need to spend time thinking about the the problems. Yeah.
36:50
But then but then I also see and it concerns me there's this kind of all in doomerism. Yeah.
37:00
And and and I look at a lot of the people who are like the most pessimistic. Yeah.
37:04
And I don't see too much domain knowledge in technology. Yeah. Yeah.
37:12
What What do you What are your thoughts on that?
37:14
Yeah, I mean I I I was recently uh interviewed for a blog post and that was exactly one of the points I was making that that it seems the most prominent doomers are really not technologists.
37:24
Not in the sense that you know someone who can actually build even decent machine learning model or something like that.
37:30
And uh one thing that you discovered you know, if you build machine learning models for a while is that there's one thing to actually think that something can be done versus like when they actually try to actually do it.
37:41
And you know, and that's one of the things that I loved about Kaggle.
37:46
It dispersed so many of my own intuitions and my own grand visions of what a great model would be like.
37:52
Like so many of them they have really panned out to be nothing or if anything were even worse than just very simple models.
37:58
So, you know, it's this you know, it happens everywhere where you know, outsiders have much different or more inflated ideas like what's possible than people who are actually trying to work in the field.
38:12
Yeah, but I think in in in the machine learning modeling it's particularly acute because uh uh you know, our models are great and not as easy to make as some people outside think they are, you know, and and it takes a lot it's still a lot of effort and a lot of uh uh you know, attention to details and things that people don't really pay attention to actually get some something even slightly better than what you've had before.
38:36
And then there are all sorts of unknown unknowns like that we may run into like like, "Okay, this approach and scaling laws work until maybe they don't."
38:45
So like, you know, it could be like we're thinking it's an exponential curve and maybe it's actually some kind of sigmoid and sigmoid starts flattening out maybe in a couple months, maybe next year or something and we've kind of reached saturation point that we did not anticipate with the current approaches.
39:00
And you know, but a lot of doomers just see exponential curve and you know, connect three dots for exponential and say like, "Okay, the next thing we know is these things are going to destroy our whole world."
39:10
Um I think what they also what misunderstand is how big a disruption rather than destruction is going to be.
39:20
So you know, we we did a talk a little bit right now about like which fields are going to be most affected, which ones not.
39:26
Uh there's some rough guesses what we can do and you know, I talk to people who are uh trying to estimate these things and uh and no one really knows for sure like which fields are going to be the most affected, which ones are not.
39:39
We we can sort of take a look and kind of take a stock of what specific tasks people in these fields are doing and they kind of try to estimate which ones are easier to automate, which ones are less.
39:48
But you know, it's still a very rough estimate.
39:53
So like we know that disruptions are coming but we just don't know how they're going to break.
39:57
And like to just go through these disruptions to full-scale doomerism, I think is a huge leap.
40:03
That's that's where I'm coming coming from.
40:06
I agree completely and the it's a challenge for me at least to try to walk the line where I want to be categorical with people where I say, I am not being Panglossian here.
40:18
I am not saying there are not going to be problems.
40:21
There are going to be problems.
40:25
And and and the way I look at it is much better to acknowledge that fact and then, you know, try to start thinking about countermeasures, try to start thinking about ways around them.
40:34
And it just concerns me because like I joke sometimes that if Thomas Edison were around today, we'd all be doing things by candlelight because they would make electricity illegal. Why?
40:46
Because electricity is incredibly dangerous. Yeah. Yeah. Yeah.
40:52
Yeah, that's that's a very good analogy.
40:54
Like I I I get that I think You know, when we are going we are in the early stages of like second industrial revolution.
41:05
So, like, you know, we know the first industrial revolution had an incredible impact on, you know, lives of people and society in very unpredicted predicted ways and there was a lot of the disruption.
41:16
Some of it was very violent.
41:18
You know, to this day we are, you know, struggling to kind of come to terms with a lot of disruption that the industrial revolution was.
41:25
And you know, I think similar things will happen with with this, you know, AI revolution.
41:29
I I'm not Panglossian either.
41:31
Like I I think we are in it for some very challenging times and, you know, I think overall, just like industrial revolution, overall lives are going to improve.
41:40
But on individual levels, a lot of people are will be winners and losers.
41:45
And you know, how do we deal as a society with vast numbers of, you know, losers in the sense that, you know, they're not on the winning side and, you know, what are the benefits AI brings.
41:54
And, you know, how to kind of deal with that is going to be a big challenge, you know.
41:58
Uh I don't have any answers.
42:00
I I'm just barely kind of speculating about what what's going to happen next.
42:06
Yeah, and one of the things that I think is kind of a a bug in human OS is first off we we tend to think that what we know now is kind of everything there is to know. Yeah. Yeah. Yeah.
42:18
Which of course is absolutely not true.
42:21
If we if we invented a if you invented a time machine and we went back 500 years and found the most brilliant human on the planet, we'd find that 90% 95% of what he or she believed was completely wrong. Yeah. Yeah. Yeah.
42:39
And and and so in the way I look at it is like it's we have no idea what professions are going to get created Yeah.
42:51
this new technology and and it's very easy for people to talk about professions that will be impacted or eliminated Yeah.
43:01
because because we know about that profession, right? Yeah. Yeah.
43:04
And and it's very difficult your your parades were dollar runs from felt space Yeah. known unknowns, right?
43:10
It's very difficult There's a great quote that you that I can't remember whose it is, but it goes along the lines of no matter how smart, no matter how creative, no matter how brilliant a human being, there's no way that you could ask them to make a list of things that would never occur to them. Yeah. Yeah. Yeah.
43:32
And and yet we we we kind of struggle with this.
43:36
We fall into this idea that we we can do that.
43:38
Um I I on the field with with a lot of the doomers, what I find is they go immediately to AGI, right? Like No. No.
43:50
And and that of course we have James Cameron's Terminator there.
43:54
By the way, another thing as I said at the introduction of our chat Your humor is great.
43:59
You you had on Twitter my favorite conference room and it's Skynet. Yeah. Yeah.
44:09
But but what what do you think?
44:10
Like uh do do you think AGI in in the um broader sense of like truly sentient machine intelligence?
44:21
Do Do you think that's either desirable or possible? Possible for sure.
44:29
And uh the question is like how we define AGI, what we think of AGI.
44:31
Uh I have a sort of a working uh definition of what I would consider AGI.
44:40
And this would be uh AI system that can pretty much replace any introductory knowledge worker in any any field that we can think of.
44:50
So like, you know, level three engineer or something in a software company or or an accountant or like paralegal person.
44:57
And how far are are we from that?
44:57
And that's the question that, you know, I going to spend a whole summer trying to answer.
45:02
Like I I was in the in Palo Alto in the Bay Area.
45:04
Uh talking to a lot of very smart people, people who, you know, affiliated with these things. And you know what?
45:11
There is this sense that those kind of systems are around the corner.
45:12
You know, like maybe not immediately within the next uh 6 to 12 months, but that's where we headed.
45:19
Like you you have agents that can perform any kind of knowledge work uh that any one of us could be doing like right out of college for any kind of company.
45:29
So I think that that that kind of AGI is coming.
45:30
Uh AGI that can pretty much do any human work uh is probably not that much farther down the road, either.
45:39
Uh I think again, like, you know, doomers you know, rely a lot of some science fiction and a lot of us have criticized them as as, you know, this is just a fan fiction that they come coming up with.
45:50
Um And I I I think in the classical movies we don't have many examples of benevolent AI that's not kind of embodied in these destructive robots.
46:01
You know, so like uh we immediately jump from like no AI to these perfectly anthropomorphic robots that can uh kill anyone.
46:10
And then like we kind of made it in the movies.
46:15
There are very few examples where uh you don't have anthropomorphic robots and yet you have kind of AGI agents like anyway.
46:23
Uh Her like housing is probably one of the best examples also that what happened in the movies probably one of the best example what could happen.
46:29
But you know, that's much more much closer to what we will first encounter over the next few years than the Terminators or like a Matrix or some of these other really kind of outlandish things or even the I, Robot or or like you know, Steven Spielberg's AI.
46:46
So these are not So we don't have many templates against which to think, you know, it's just as as a general population.
46:54
And I I I think uh we have to come up with a better Maybe we have to come up with a better fiction, you know, to kind of account for these things.
47:06
Uh maybe someone needs to make a better like a movie that that's not destructive but you know, it's disrupting in certain ways and like kind of challenge these things.
47:13
I think Her is was a good movie as well that kind of delves in that those issues and I think uh that else is just around the corner.
47:22
So you know, there are few examples of more realistic I think first encounters with advanced AI than the Terminator movies.
47:31
Yeah, that's an in fact that Oshana Ventures two of our verticals, Infinite Films and Infinite Media, are designed to actually as part of their uh remit do exactly that.
47:43
Come up with far more optimistic uh use cases, movie scenarios because like if you look at old science fiction like H. G.
47:54
Wells and others, it was pretty optimistic.
47:57
Not not the other way around. Yeah. Yeah.
48:01
Um one use case that as I was listening to you that I thought of because I've thought about it a lot is the potential great benefits that we could get from kind of a personalized mental health Yes. AI.
48:17
Yeah, talk talk to me a bit about that.
48:20
Yeah, I had a tweet about it too with the with a lot of interesting reactions.
48:24
Uh I think there's a still a lot of stigma associated with those seeking help for mental issues.
48:31
Uh The my wife, she's a psychologist and you know, not really psychiatrist or you know, even clinical psychologist, but still you know, like talking about mental states and uh uh mental health issues is you know, close to what we think about and you
48:47
know, how to you know, deal with what everyone's doing anxieties and you know, slight depressions or or uh uh some other like small scale issues like those could be handled I think very well with uh with some simple AI systems. Like right
49:00
Like right now we, you know, if we even talk about like an ideal constantly about men would rather act do acts than go to therapy.
49:09
Um you know, we we don't really necessarily need therapy, but we need like a close friends who can actually provide good kind of feedback.
49:16
And I think even that is like putting too much pressure on people, you know, lives that don't really need, you know, someone to kind of be mopping with them and trying to solve their lives issues with them.
49:29
Well, you know, AI would not have any of those issues.
49:32
Like you know, there's no stupid questions you can ask it all sorts of silly things. Not going to judge you.
49:38
You can actually kind of interact with it and it can if it can be very um effective, you know, then then I think it could be a very good tool.
49:47
One particular uh mental health practice that I I really consider very highly is called a cognitive behavioral therapy, uh, which which has a lot of applications, especially for people with mild depressions and uh, and uh, also saying anxiety disorders.
50:04
And I think teaching you how to think differently, you know, to rephrase your mindset, yeah, would be a very effective way of kind of helping a lot of people with their daily struggles that don't rise to the level of well, a psychia- psychiatric uh, issues.
50:22
Yeah, and uh, you one of the reasons why I'm so bullish on it is because on the one hand, you're absolutely right.
50:28
We we still have a bit of a stigma, although I I don't think we should, but we do. Yeah.
50:34
Uh, attached to people like men would rather fill in the blank than go to therapy, right?
50:38
It's It's funny because it's so true. Yeah.
50:41
Um, and and yet, um, we we also get the hook from our our tending to anthropomorphize tech, right? Yeah.
50:53
So, if we if let you brought up the movie Her, right?
50:55
He fell in love He fell in love.
50:57
Of course, maybe if it has Scarlett Johansson's boy, I'd fall in love, too. Who knows?
51:04
But but, um, so we we kind of have that going on, but we'd also intellectually know I'm not going to be judged because this is an AI that I'm talking to. Yeah. Yeah.
51:16
Uh, and and therefore, it's my thesis, I could be totally wrong, but my thesis is people would probably be much more willing to be really honest Yeah. with an AI, right?
51:28
Something they might they they might not even be willing to admit to a licensed psychiatrist. Definitely.
51:37
They would in fact admit to the AI, making it much more effective because maybe that's the real problem Yeah.
51:44
that they're that they're that they're dealing with.
51:45
Um Are Are Are there any other other than So, tutoring, mental health, another one that I put out there and and I got a lot of flak for this one. Yeah. Uh but was elder care. Yeah.
52:01
At least in the United States um we have a an aging population. Yeah.
52:08
And for the most part, even the families with the absolute best intentions Yeah.
52:14
if if they end up putting mom or dad or grandma or grandpa in a elder care facility, we're going to go all the time.
52:22
We're going to really go and visit them.
52:24
And then life interrupts and they don't. Yeah.
52:28
And so, I My idea was we should give them all a personalized AIs. Yeah.
52:33
That they can chat with, that they can reminisce with, that they, you know.
52:35
And and and boy, did a lot of people jump on me for that one.
52:40
They're like, "Oh, you'd rather give them a machine than, you know, Yeah. I'm like, "No.
52:47
The other Of course that's an ideal solution, but it doesn't exist. It doesn't happen." Yeah.
52:52
Well, I I I completely agree.
52:52
I mean, the you know, you know, I again, you know, I've been on Twitter and I've actually met a lot of wonderful people on Twitter and I'm really grateful to Twitter and some other social networks that they have been very instrumental for me.
53:05
Again, like I live in a small town far away from, you know, all the happening places in tech.
53:10
I probably would have gone bananas if I didn't have access to social media.
53:14
You know, like you know, like I you know, I have all these interesting things that I want to talk to people about, but there's no one around in my town who talks about I you know, I I often tell people that I'm only half joking that around here AI stands for artificial insemination. Yeah.
53:31
It's It's you know, it's more of a semi-rural place that people don't really care about tech much or if if they do, it's in a very kind of superficial way, which is fine.
53:38
But, you know, like I you know, having a place like Twitter or you know, some other places like very much help me keep my sanity.
53:46
You know, like I meet people like you and we have these wonderful talks and you know, I feel pretty fulfilled.
53:52
So, I think yeah, I think if social media can do this, I think other tools that that are semi-human like AI can really be helpful as well.
54:02
And you know, I see no no problem with that myself.
54:06
Yeah, and I I I feel strongly that Twitter despite I I think first off Twitter is pretty anti-fragile. Right?
54:13
I mean, if you you look at all of the either intentional or unintentional yeah, uh attempts to kill it, it it doesn't die. It it doesn't die.
54:26
And I think one of the reasons for that is what you just said.
54:31
Uh you know, despite all of the hassles and everything, I love it because I get to meet people like you.
54:36
And I I think that it's what part of my thesis with the new uh company is it's a lot easier to change your digital zip code than your postal zip code. Right? Yeah. Yeah. Yeah.
54:52
And and and if you can have a platform where where you can meet interesting people who share your interests and and then move it on to things like this where you actually speak with them and you actually uh actually become friends with them.
55:07
That's happened to me quite quite a lot. Yeah.
55:10
Um but but the other thing about it that's really interesting to me is there's a lot of people who do Twitter really badly. Yeah.
55:21
You're you're good at it.
55:21
Like so so what do you think is the reason if you if you do Sure.
55:29
that that many people do it so badly?
55:32
And and how would you give them some guidance on on how to level up a little bit.
55:40
I think, you know, I think that different people doing it different ways and different voices and I kind of appreciate that.
55:47
Uh Uh what I find like what I try to avoid as much as possible is becoming too confrontational.
55:55
Like even if I'm not really agree with you, I I don't want to kind of get in into arguments and you know, there are all sorts of jokes about it.
56:02
Uh how futile I arguing online is and it's 100% sure not true.
56:09
Uh So, you know, humor is one of one of my ways of really diffusing tension.
56:12
So, even if I'm bringing up some very contentious and very uh problematic issue, I try to do it in a light-hearted way so that even people who disagree with me don't feel kind of threatened or be be judged or like uh being uh attacked in some ways.
56:26
You know, people will, you know, disagree and I feel it that's fine and this is what the our marketplace of ideas is all about.
56:34
But, you know, try to become too argumentative or kind of jumping at people or like uh try to assert yourself in certain ways that are very confrontational.
56:42
I think those are uh things that should be avoided if possible.
56:48
Especially places like Twitter.
56:48
I think they're you know, if you have like small circle friends and you have certain kind of worldview that that kind of drives with those people, going to those like little forums and you know, by all means be as confrontational as you want to be.
57:01
But, you know, like this public square that that's Twitter, I I think it's best served if people are a little bit less toxic in the terms of like how they attack each other or kind of go about these things.
57:12
Um And like, you know, there are many many things many things, you know, we talked about, you know, close feedback loops.
57:18
That's I think one of the reasons why kind of got better and better Twitter because like I would post something, see people's reactions and kind of make make a mental note whether this kind of resonates with people or not and then kind of repeat that over, you know, many years with many, many little things.
57:32
So, I think a lot of people don't pay attention to what they're doing.
57:36
You know, they're just doing one thing all the time, you know, not really see like how it translates.
57:40
I'm not saying that you have to uh conform to people around you, but just like see like what what's really worth, what doesn't, and you know, kind of builds on that.
57:50
And that's so that those would be my my big messages.
57:51
And then depending on what you want to do with Twitter.
57:57
Like some people just want to kind of build like a particular product kind of identity and work around it. That's fine.
58:04
Uh you you want to have it a little bit more personal, that's another thing, that's fine.
58:10
Uh you know, it it truly depends whether you want to grow on Twitter or not.
58:14
If you don't want it to grow, that's fine, too.
58:17
Like you just want to kind of absorb information, that that's also perfectly okay with me.
58:23
Uh you I love the way you you praised the idea of the public square ideas.
58:27
Um and and then uh if So, what I heard, and I'm going to I'm going to repeat it uh because I think it's good advice.
58:37
Um if you if you treat the main timeline more as, "Look, you're in public here.
58:46
You're talking to a lot of different people.
58:47
Would you go to, you know, whatever, Union Square in Manhattan and start yelling at people? Probably not. Yeah. Yeah. Yeah.
58:56
Uh but then I also like your solution, which is if you find a group, then get into the DMs, take it somewhere else.
59:04
And and that way you can do it both ways.
59:06
So, I I think that's a really great idea.
59:08
And another thing we share is a love of reading.
59:10
And as I was researching you, like you don't do anything halfway, do you?
59:18
Uh you have you have uh you have reviewed hundreds of books.
59:20
You became a top 10 Amazon reviewer in 2011. Yeah.
59:27
Um so, how did you get so good at writing book reviews?
59:32
It wasn't just book reviews, it was product reviews as well.
59:33
So, like the So, it's started with books.
59:39
And like So, you know, Kaggle, Twitter, and Amazon reviews were like, you know, and by the way, I'm not on Amazon anymore, you know, that's we're going to take as a complete tangent why I'm not posting there anymore.
59:50
But, you know, like you post something and you get like really quick feedback from people.
59:54
So, like, you know, I did it the first book review I wrote on Amazon was a for a book that I absolutely hated.
1:00:01
So, like, I I just wanted to kind of write a review and tell people don't read this book, it's really awful.
1:00:06
Please, you know, don't go there.
1:00:08
So, I posted that and to my surprise, all of a sudden people are liking this review and all of a sudden I have a ranking there.
1:00:14
Like, okay, this is interesting.
1:00:14
I get feedback for my work.
1:00:17
Um, you know, that's a different from blogging where, you know, like a lot of times you're blogging and pretty much you never get any kind of feedback from people whether they like your post or not and and it's a it's a very different system.
1:00:27
And here I would read books and they're like, okay, I'm just going to tell people what I thought about it and was first out of books and then later on moved on to products and gadgets and that's that's kind of kind of got into tech blogging which was my second and a half career of between you know, physics and and being a machine learning modeler.
1:00:48
Um, but yeah, I I definitely read a lot of books.
1:00:54
I I I think my estimate is about 1,200 books that I've read in my life and I I don't read as often or as much as I used to, but they still read about uh, two or three books a month.
1:01:04
So, that's that's my uh, cadence these days. Um, uh, I don't know.
1:01:11
I just like it and and I think part of the reason I'm good at Twitter is because I I have appreciation for written word.
1:01:17
And like, you know, like when when I'm writing a tweet, a lot of times very short like three-word tweets would actually take me a long time to compose because I'm kind of weighing how much different words, how much context would resonate, you know, can I eliminate something?
1:01:33
So, I it it I I just like kind of written word in its own right and it it's it's a kind of naturally born to read and now write on Twitter. Yeah, for me, too.
1:01:45
Um, I and I I share your idea about Twitter and feedback.
1:01:48
That you know, cuz I would do I don't do them as much now as I used to, but I would do very long threads on Twitter.
1:01:57
And and and and people friends would say, "Why don't you just do a blog?" Yeah.
1:02:01
And I'm like I I'm like Twitter is the best microblogging site in the world cuz you get instant feedback. Yeah. Right?
1:02:08
And and you don't necessarily with blogs.
1:02:11
And you have your you have a website actually with your book reviews and gadget reviews, right? I used to. I think it's down.
1:02:17
I need to upgrade the WordPress for that. Okay. Okay.
1:02:21
Our researchers find everything. Yeah, they No, no, no.
1:02:26
Like if you if they were to let me ahead with upgrading that website again, I will.
1:02:30
I I mean, I've started a lot of kind of side projects over over the years.
1:02:34
I think if you're a techy person, you pretty much do it all the time.
1:02:38
And I needed to kind of upgrade my blog, but it was very low priority, so I never really went back to that.
1:02:45
I I I hope to do a little bit more of that in the future.
1:02:48
Um, I would like to kind of have a So, like again, one of the things that I'm known for is this machine learning for tabular data, which is sort of a uh unsexy cousin to natural language processing and image uh machine learning.
1:03:03
But, it's a kind of data that pretty much, you know, drives all the transactions that we engage in, you know, from banking, from you know, uh shopping online, everything is you know, recording some kind of transactional database and uh it's pretty much 90% of all the data that we have in the world.
1:03:21
And uh there are a lot of opportunities there, a lot of interesting problems to still to be solved.
1:03:26
And you know, I would like to kind of start a blog that really explores this whole topic a little bit more systematically at with some point when I have more time, which I never do, you know.
1:03:38
Yeah, well, there's the old maxim that if you want something to get really done fast, ask a busy person. Yeah.
1:03:46
That that's that's really resonates with me, yeah.
1:03:49
What what what are some of your favorite books?
1:03:52
So in fiction I would say Yeah, uh like Nabokov was one of my favorite writers, still is.
1:04:01
Um I think Lolita is absolute masterpiece, you know, it's very controversial even when it was published and even more so today.
1:04:09
Uh but in terms of like the writing it it's absolute masterpiece.
1:04:14
Um I of the contemporary writers my favorite one is Haruki Murakami. Yeah.
1:04:21
So he's he has a very quirky interesting way of uh looking at the world and uh it kind of really kind of caught me and I really enjoyed his writing.
1:04:32
Uh I when I was younger I really enjoyed Milan Kundera.
1:04:34
Uh he was a one of my writers, favorite writers.
1:04:36
Um These days uh You know, the books that I really enjoy uh was would be like Steven Pinker's Blank Slate. Mm. Love it. Yeah.
1:04:50
The bank debunking some of our misconceptions about, you know, human lives being just kind of blank slate that you can do everything about.
1:04:55
that you can do everything about. Um Right now reading a very interesting book called Thinking Like a State and like kind of looks at all these very misguided attempts by the states to kind of standardize everything from the
1:05:14
architecture to central planning of agriculture and whatnot and how it invariably leads to kind of disaster because there's so much information on a low level that these uh central planners don't really take into account and how it kind of leads to these uh these issues. And you know, I I
1:05:29
these uh these issues. And you know, I I can I also think there's a lot of uh problems in the way that the software projects get structured around these ideas that they won't have a you a lot of people won't have a very lean way of structuring the projects
1:05:48
where, you know, in the reality there's a lot of messiness that people learn how to work around and even uh uh leverage to to get ahead, but you know, a lot of central planners don't take these things into account and then you end up with a much more bloated and slower software than it really should be. So, you know,
1:06:03
So, you know, I think these are kind of interesting ideas to play with.
1:06:07
Yeah, I agree and you know, it's the problem with planning or top-down planning is complexity adapted systems emergence comes from the bottom for the most part and and trying to trying to invert that is not going to change that.
1:06:23
Which is why you get all of these famous debacles because because people just don't understand that that very, very simple concept.
1:06:30
What are you working on right now in the machine learning and AI space that you're most excited about?
1:06:37
Um right now we have a lot of projects within Nvidia that's building a new a gradient boosted trees library.
1:06:44
So, we you know, if again, if you follow me on Twitter, you know, XGBoost is one of my catchphrases and I'm promoting library a lot which is also maintained that uh supported by Nvidia.
1:06:58
Uh I think gradient boosted trees are very interesting algorithm in their own right and they've been very successful with the all the problems that they've been applied to.
1:07:08
And I think we can but but unfortunately they're not they've been kind of under-appreciated in other researched.
1:07:15
So, like you know, compared to neural network there's just tiny a fraction of people actually doing some kind of research on this stuff.
1:07:20
And and I feel like we can even leverage some of the learnings from our neural networks for gradient boosted trees.
1:07:26
And that's one of the things that I'm kind of working on.
1:07:29
Um I've also been very, you know, my social media presence has been appreciated at the many places at Nvidia.
1:07:37
So, I helped with the some product evangelization.
1:07:39
So, you know, like I I I got to play with H100s as soon as they became available.
1:07:44
Uh so, it's a it's a fun little kind of side not not core of my duties at Nvidia, but like in another project that I or another function that I like to perform all the time.
1:07:58
And and was also always a big fan of our Jetson little boards and I get to play with those. And all stuff.
1:08:07
Cute little videos or projects online.
1:08:09
So, these are kind of projects I work on. Terrific. And and lots of fun.
1:08:11
Well, well, Boyan, this has been absolutely fantastic.
1:08:16
Uh at the at the at the end of the each of our chats on Infinite Loops, uh we we uh we have a wand and we make you the emperor of the world.
1:08:28
Is that Uh you you you can't kill anybody.
1:08:28
You can't send anyone to a re-education camp.
1:08:33
But what you can do Yeah.
1:08:35
into a magic microphone and intercept two ideas, two thoughts into the entire world's population.
1:08:42
And whenever their morning is the next day, Yeah.
1:08:47
they're going to wake up and they're going to think that they came up with the two things that you intercepted them with and they're going to act on it immediately. Yeah.
1:08:57
What what what what are you going to intercept?
1:08:59
Um What would it be like to try to think more probabilistically?
1:09:06
Uh so, probabilistic mindset, I think it's it's a very effective and very I think you can diffuse a lot of people's anxieties just by adopting that kind of a mindset.
1:09:19
You know, instead of thinking in terms of black and white, you think about the spectrum of things, and I think probabilistic type mindset kind of helps you with that.
1:09:25
So, even if it's not really any kind of formal probabilistic things, just think about okay, you know, this is the worst that can happen, but you know, just probability of that is pretty low, and like there are all these other things that are more likely, and try to kind of cognitively kind of re-appreciate your own world.
1:09:43
Um um Another one of my mottos in life is is like uh you know, in theory, there's no difference between theory and practice, but in practice there is.
1:09:58
So, you know, it's it's one of the two or three mottos that I kind of repeat to myself all the time every once in a while.
1:10:04
So, you know, it's you know, all the art, the red color musings, and and high-level thinking that we kind of view ourselves with, it they're always going to be well short of like what practical considerations we have to have in our lives. I love both of those.
1:10:22
The second one reminded me of the old joke about the two old bankers talking to each other, and the one says to the other, "We know that it works in practice.
1:10:31
We're just worried it won't work in theory." Yeah, yeah, yeah.
1:10:37
I think I've seen that in New York a couple of times. Yeah.
1:10:38
And and uh the probabilistic thought has been one of the soap boxes that I've been on for a long, long time.
1:10:44
Uh because it would just help people so much, I think.
1:10:50
Because black, white, yes, no, zero, 100, Aristotelian, deterministic thought is can be really a nasty place to be if it's either got to be black or white when most things are gray.
1:11:02
And most things fall into maybe.
1:11:05
Boyan, where can other than Twitter, where can people find you?
1:11:09
Um, I also have a post a lot of stuff on the LinkedIn, which is a uh Mostly stuff that was there is similar to the what word for that I post on Twitter except all the jokes.
1:11:18
So, if I post something serious, you can find me there.
1:11:22
I do have a Substack blog, but I haven't posted there anything in months.
1:11:29
Hopefully to I I'll revive it.
1:11:31
Um I I think Twitter is the best place to find me.
1:11:34
I I'm a pretty I post pretty frequently although most of my posts are pre-scheduled.
1:11:40
So, you know, it's it's not a time I'm really posting all the time.
1:11:44
But, you know, I do try to interact with people there.
1:11:46
And if you you want to have any some kind of conversation about something I post there, like I'll be open to interact with you. Terrific.
1:11:53
Well, this has been a lot of fun and thank you so much for coming on. Thank you. I really enjoyed this. Yeah, so did I. Cheers.