WEBVTT
00:00:00.160 -->
00:00:03.680Okay, so you're listening to a very special episode of members of technical staff.
00:00:03.759 -->
00:00:05.759I'm here with Steven from Lambda Labs.
00:00:05.919 -->
00:00:14.480This is how we introduce people like you know, this is not like your your typical podcast of like name, last name, title, company.
00:00:14.720 -->
00:00:16.160You are Steven from Lambda.
00:00:16.640 -->
00:00:17.760Well, thank you, Jaden.
00:00:18.719 -->
00:00:19.679I've been introduced like that.
00:00:19.839 -->
00:00:26.079Oh, and I must I must say, we were just talking about this before, but um I must say that uh Lambda is one of the great overnight successes.
00:00:26.160 -->
00:00:27.120So uh you would have noticed.
00:00:27.600 -->
00:00:33.520It was instant, it was basically I just kind of you know started the company and it just instantly became a success.
00:00:33.840 -->
00:00:35.920It was it was basically no work at all.
00:00:37.600 -->
00:00:43.119Yeah, exactly, exactly, you know, and so and so and so and naturally it's just a behemoth in the space, right?
00:00:43.200 -->
00:00:46.079Like the AI thing happened and you were there at the right time.
00:00:46.240 -->
00:00:54.799Uh I should I should actually say uh for for people that cannot pick up on any sort of Australian sarcasm, which is a fair few people I've noticed over the years.
00:00:55.039 -->
00:01:01.280Like the amount of times that I've just been absolutely dragged on the timeline because it's like, oh, well, clearly my quip is not.
00:01:01.439 -->
00:01:06.400Some would say that, okay, you just have a bad sense of humor if you can't if people can't pick up on the sarcasm.
00:01:06.640 -->
00:01:08.480Others would say autism snipes.
00:01:08.719 -->
00:01:22.400I think that one of the interesting things about discourse on X in particular is just what every time I read a post, I might as well look at it and say, all right, how many myriad ways can this be interpreted by somebody?
00:01:22.560 -->
00:01:22.719Yeah.
00:01:22.879 -->
00:01:24.560You know, is it is it sarcasm?
00:01:24.640 -->
00:01:26.159Is it rage baiting?
00:01:26.400 -->
00:01:31.120Are they uh are they uh, you know, uh vague posting?
00:01:31.599 -->
00:01:36.000Are they are they trying to concrete post but are actually vague posting?
00:01:36.560 -->
00:01:45.680Um and you know, I feel like you can look at any sentence and be like, well, is it do they mean what they mean, or do they mean the exact opposite of what they mean?
00:01:45.840 -->
00:01:48.000And it's just very, it's very interesting.
00:01:48.400 -->
00:01:49.599Exactly right, exactly right.
00:01:49.840 -->
00:01:53.519And naturally that's why the posters become great comms experts.
00:01:53.920 -->
00:01:54.159You know?
00:01:55.840 -->
00:01:58.879You have to have a sixth sense as to how something's gonna be taken.
00:01:58.959 -->
00:01:59.040Yeah.
00:01:59.280 -->
00:02:04.480And you also kind of need to have a sixth sense as to what is cringe and what is not.
00:02:04.640 -->
00:02:06.239Because if you're cringe, it's over, right?
00:02:06.400 -->
00:02:07.599Everybody remembers cringe.
00:02:08.080 -->
00:02:13.039Um but if you're funny, you know, that's cool, people will remember you for 24 hours.
00:02:13.759 -->
00:02:15.520And so you need to keep doing it over and over again.
00:02:16.000 -->
00:02:19.039And so, and so, and naturally, that is that's just that's just how it is.
00:02:19.199 -->
00:02:21.360I spend a lot of my days thinking about this.
00:02:21.520 -->
00:02:26.240Um But you know, to segue into something that you think about a lot, well, uh Lambda.
00:02:26.639 -->
00:02:38.080If I were to characterize uh Lambda as a company, my sort of laymans, I say laymans, I'm a I'm a somewhat technical person, converted to non-technical recently, so that's kind of how I think I identify now.
00:02:38.240 -->
00:02:47.680But Lambda, would it be safe to say that for for those that have no idea what Lambda is, would it be safe to say it's like uh like AWS but for AI specifically?
00:02:48.000 -->
00:02:50.639I think that that's a great way of describing what we do.
00:02:50.800 -->
00:03:04.080I mean, we've always um since entering into the compute market been uh a place where we we just want to provide sort of computation to people who are doing this kind of work.
00:03:04.560 -->
00:03:06.319Yeah, and it's kind of a big deal at the moment.
00:03:06.479 -->
00:03:07.280Uh the AI thing.
00:03:07.599 -->
00:03:08.479Yeah, the AI thing.
00:03:08.639 -->
00:03:10.879Well, you know, it's a controversial uh segment.
00:03:11.280 -->
00:03:48.080Going back to how easy it was to start the company um you know in 2012, uh when when all of this was uh just getting started, uh I think that very few people could have just really predicted that that we would actually get to see that the the the ramp up that we've been been seeing in our in our lifetime because because I mean you know one of the first things you learn if you like ever study about AI either in college or in your own spare time is the AI winter.
00:03:48.400 -->
00:03:54.800And right, it's so it's almost like you're you're kind of traumatized from your very first exposure.
00:03:54.960 -->
00:04:00.080Oh, there there was once this big huge buildup and then the AI winter came.
00:04:00.400 -->
00:04:06.639And um, you know, that's never stopped people who are really passionate about the space from getting into it.
00:04:06.879 -->
00:04:07.680Of course, of course.
00:04:08.080 -->
00:04:15.120But uh uh it was is kind of interesting to see uh you know how far we've come from those beginning and early days.
00:04:15.439 -->
00:04:20.240I was gonna say, because like I mean, what, 2012, you're doing neural network stuff.
00:04:20.319 -->
00:04:20.480Yeah.
00:04:20.879 -->
00:04:25.040Uh I think back to two 2012, I think about Tumblr.
00:04:25.519 -->
00:04:25.920Tumblr.
00:04:26.240 -->
00:04:40.560You know, it's like it's like like um I I think of sort of, you know, social networks taking off, you know, uh f Facebook is like relatively mature by that hour, but I've been thinking wasn't the Instagram acquisition like 2013 or something?
00:04:41.120 -->
00:04:44.240And so it was sort of like end of 2012, 2013.
00:04:44.480 -->
00:04:44.639Yeah.
00:04:44.800 -->
00:04:48.240And did you have like like were you were you just getting like thrown into lockers?
00:04:48.319 -->
00:04:50.720Like like were you just bullied, you know, like neural network guy.
00:04:55.199 -->
00:05:03.920Um, and I was the first product that we ever made at Lambda was this thing called Heads Up, and it was a facial recognition contact book.
00:05:04.160 -->
00:05:10.800You know, so like in the normal sort of explanation in a short form interview, yeah.
00:05:10.959 -->
00:05:17.519I'll usually say, well, Lambda started off as a face recognition company and XYZ, and then we get get get into selling compute to people.
00:05:18.160 -->
00:05:23.759But the first product was actually a face recognition contact book called Heads Up.
00:05:23.839 -->
00:05:25.199And it was an iOS app.
00:05:25.279 -->
00:05:25.439Yep.
00:05:25.680 -->
00:05:28.240And what it did is, you know, the idea would be we'd meet.
00:05:28.399 -->
00:05:30.240I'd be like, hey Jaden, nice to meet you.
00:05:30.319 -->
00:05:31.279Let's take a picture together.
00:05:31.360 -->
00:05:33.040We'd take a picture together, yeah, yeah, yeah.
00:05:33.199 -->
00:05:41.680And it would scan the cloud to see if it could recognize your face and automatically download your V card or your contact book basically into my um into my phone.
00:05:41.839 -->
00:05:42.160Yeah.
00:05:42.480 -->
00:05:53.360And so there's a video of me and I'm posting on um I'm posting on the something awful forums, which for a lot of the people here, you know, have no idea what that means.
00:05:53.600 -->
00:05:53.839Yeah, yeah.
00:05:54.079 -->
00:06:03.279But it's basically, you know, I was posting on the something awful forms to commission a um to commission a logo, an app logo.
00:06:03.439 -->
00:06:05.360And at the time, like dribble was around it.
00:06:05.519 -->
00:06:14.560You know, you remember that sort of old dribble style 2012 of like the very like detailed, hyper detailed, like kind of lickable design of the app logos.
00:06:14.800 -->
00:06:15.759Does that ring a bell at all?
00:06:16.000 -->
00:06:20.000I mean, I mean you also you also I mean just like I'm I'm thinking of even just the name Jibble, right?
00:06:20.079 -->
00:06:23.120Or like like that kind of like class of name, the like like Jibble.
00:06:23.680 -->
00:06:26.319Dribble with three Bs, Tumblr with an R.
00:06:26.639 -->
00:06:31.439Yeah, I mean uh Delicious being like was it Josh Schachter's company, Delicious?
00:06:31.680 -->
00:06:31.759Sure.
00:06:32.000 -->
00:06:36.399Uh with the domain hacking D E L I, you know, O U.
00:06:37.120 -->
00:06:39.199We don't we don't name companies like that anymore, you know?
00:06:39.279 -->
00:06:41.439We don't like we don't have we don't have names like that.
00:06:41.519 -->
00:06:45.839It's just it's just like the something, you know, the San Francisco Company of America.
00:06:46.319 -->
00:06:47.680Like that's how we're naming companies now.
00:06:48.879 -->
00:07:00.399It's funny because it got back to like sort of this, there was this this like people yearn for the the the the early 1900s, late 1800s heights of of unfettered capitals.
00:07:01.600 -->
00:07:03.519Oh, it is what is it, the the the British India Company.
00:07:03.680 -->
00:07:04.720What's what's the very famous one?
00:07:05.279 -->
00:07:06.959Dutch East India Company, of course, of course.
00:07:09.120 -->
00:07:11.279So no, it's funny because we haven't gone that far back yet.
00:07:11.360 -->
00:07:11.600No, no.
00:07:11.839 -->
00:07:19.680It's really was more like standard oils, you know, so standard cognition, standard, actually, you know, standard something, American something, but Dutch East India.
00:07:19.920 -->
00:07:27.759So that would have to be like America, and then you have to add some sort of like some old uh an old uh sort of colonial names.
00:07:28.720 -->
00:07:31.360We we're returning, we're returning very quickly.
00:07:31.600 -->
00:07:32.480We ran out of names.
00:07:32.560 -->
00:07:33.279We ran out of names.
00:07:33.519 -->
00:07:35.360But no, I mean, yeah, even just even just a Taiwan.
00:07:35.680 -->
00:07:37.279American East Indochina Company.
00:07:37.519 -->
00:07:42.560Like a that would be if you saw that as like a new AI company is the American East Indochina.
00:07:43.120 -->
00:07:46.639It needs to go up against the uh the Taiwan semiconductor manufacturing company.
00:07:47.279 -->
00:07:47.680Yeah, wow.
00:07:47.839 -->
00:07:52.480So we've I mean we we we we have we have one of those behemoths around us right now.
00:07:52.800 -->
00:08:03.360It's like descriptive, descriptive um if you ever you know read about trademark law, it's very descriptive names, as we as we call them.
00:08:03.920 -->
00:08:04.560Exactly.
00:08:04.879 -->
00:08:06.959I mean, you know, and and what, members of technical staff?
00:08:07.040 -->
00:08:08.000I mean, who knows what that means?
00:08:08.160 -->
00:08:13.519I mean, we were just saying before, I mean, like like the different the different names that companies have internally.
00:08:13.759 -->
00:08:21.040I my the my play on members of technical staff is always that it's just like I think that everyone calls themselves a member of technical staff now.
00:08:21.120 -->
00:08:22.480So no, so why not me?
00:08:22.639 -->
00:08:23.120Why not me?
00:08:23.439 -->
00:08:23.839Why not you?
00:08:24.000 -->
00:08:35.039But here's the the real question is this when members of technical staff become when members of technical staff become a sort of media conglomerate, what will be the demonym of members of technical staff?
00:08:35.360 -->
00:08:40.559Well, I'm just gonna call everyone a MOTS, you know, because because obviously the you know then there's been some confusion recently, right?
00:08:40.639 -->
00:08:44.080Because it's like, oh you're called MOTS, like like why are you not called MTS, right?
00:08:44.159 -->
00:08:47.039Which is the abbreviate like you know, why not members of the members of technical?
00:08:50.000 -->
00:08:54.240One of my friends um is an account called Squirtle says, naturally, like all of my friends are weird, weird names.
00:08:55.759 -->
00:08:56.639Weird names online.
00:08:56.960 -->
00:08:58.320Um They're a real person.
00:08:58.480 -->
00:08:59.440Yeah, exactly.
00:09:00.480 -->
00:09:01.039You have no idea.
00:09:01.120 -->
00:09:02.799It's like, oh yeah, yeah, Squirtle says he's actually.
00:09:03.200 -->
00:09:10.240Squirtle says invent the squirtle dancing posting, or is that like He did not, but he just he just did it, he just posted every Sunday for a while.
00:09:10.559 -->
00:09:11.200I remember that.
00:09:11.279 -->
00:09:15.120I just remember following just because he'd post the Squirtle Sacks.
00:09:15.519 -->
00:09:15.840Exactly.
00:09:16.080 -->
00:09:22.799What a great, like, what a really interesting example of sort of niches in media and niches in posting.
00:09:22.960 -->
00:09:23.120Yep.
00:09:23.279 -->
00:09:28.799Is that you know you you can kind of it's like uh, you know, you can do the squirtle sacks thing.
00:09:29.200 -->
00:09:30.960You can do loan for anything.
00:09:33.759 -->
00:09:35.919There's all these other people, like pilgrim.
00:09:36.240 -->
00:09:38.159Yes, I've got there's many, there's many.
00:09:38.320 -->
00:09:40.320I mean, even him, you know, and it's uh it's so funny, right?
00:09:40.399 -->
00:09:43.120It's like I always picture the scenario of like explaining to my mom.
00:09:43.360 -->
00:09:49.360It's like, no, what you need to understand is that uh Squirtle says, like, he's actually he's actually an incredibly talented engineer.
00:09:49.519 -->
00:09:54.480It's like, oh, like, you know, here's my friend uh, you know, P Pee Poo-poo online, but he's like a quant.
00:09:54.559 -->
00:09:54.879Yeah, he's got a lot of things.
00:09:54.960 -->
00:09:57.519He's like, oh, trust me, he's a really high-powered uh.
00:09:58.000 -->
00:09:59.840Yes, yes, these are influential people.
00:10:00.000 -->
00:10:08.960But where is uh the reason I bring up Squirtle is because he had he had this thing, he's like, oh, you know, um uh it's like oh they're called members of technical staff, but yet there's only like where where where are all the other members?
00:10:09.200 -->
00:10:10.000It's just me.
00:10:10.159 -->
00:10:10.480Yeah.
00:10:12.399 -->
00:10:19.200Okay, so I just woke up here, and so this is gonna be my day in the life uh ad reading through the day.
00:10:19.519 -->
00:10:27.919You can actually see that I've um I've not even like shaved my face yet and wearing a hat uh because you know I've still got to look semi-presentable on camera.
00:10:28.080 -->
00:10:29.919Um, but I want to start it start started.
00:10:30.159 -->
00:10:31.759I can't even speak, not even had coffee yet.
00:10:31.840 -->
00:10:35.759I wanted to start this day off by mentioning uh Cappy.
00:10:36.399 -->
00:10:38.480They got me this cool uh vinyl.
00:10:39.279 -->
00:10:42.240They pressed my jazz recital into a vinyl.
00:10:42.639 -->
00:10:49.200So naturally they are the nicest guys ever, uh, which is exactly how you sell software because it matters how nice the people are.
00:10:49.360 -->
00:10:52.480Uh but basically they also released Cappy V2.
00:10:53.200 -->
00:10:54.960I think it was last week at this point.
00:10:55.120 -->
00:10:57.039Gosh, it feels a while ago.
00:10:57.200 -->
00:11:06.559Uh but of course, your AI software engineer, of course, Cappy works in the cloud for your enterprise needs.
00:11:06.799 -->
00:11:11.919And uh essentially what they can actually do is basically they can actually route for you.
00:11:12.159 -->
00:11:19.360So, depending on your code base, they can actually have their router choose the best model for your code base, which is actually pretty cool.
00:11:19.440 -->
00:11:22.639You can go ahead and code on your phone, you can go ahead and sell pull requests on Slack.
00:11:22.799 -->
00:11:26.960All of these things that you can do using uh Capy VMs, right?
00:11:27.039 -->
00:11:28.000And Cap E V2.
00:11:28.559 -->
00:11:30.159So definitely go ahead and check them out.
00:11:30.320 -->
00:11:35.360I'm gonna go ahead and have some coffee and um get into this day.
00:11:35.440 -->
00:11:40.080I have no idea where we're where we'll be doing ad reads from today, but we'll see.
00:11:40.240 -->
00:11:42.399As long as I have a phone, I can do an ad read.
00:11:42.720 -->
00:11:44.799But yeah, I've I've tangented quite a bit.
00:11:44.879 -->
00:11:49.360I mean, you know, going back to the sort of like, you know, the the 20 the 2012 days.
00:11:49.440 -->
00:11:54.399I mean, you know, how what like how does it feel to be so early?
00:11:55.360 -->
00:11:59.600Like or so so direct directionally correct as people are calling it now.
00:11:59.919 -->
00:12:08.080It drives you really crazy because you're um you're so early.
00:12:08.320 -->
00:12:19.919I mean, so I'll I'll tell a story I've never told before, yeah, which is uh I'll tell you this is a story of of when I was um pitching heads up to Y Combinator.
00:12:20.000 -->
00:12:31.200Yes and I got an interview because so before Lambda, I was um working at this uh Y Combinator company, Hyperink, which I had acquired my previous company.
00:12:31.279 -->
00:12:31.440Yeah.
00:12:31.679 -->
00:12:36.080And you know, it was kind of winding down my company, decided to join another shop.
00:12:36.559 -->
00:12:38.639Um and that's a story for another day.
00:12:38.799 -->
00:12:40.399Yes, but it's also a very interesting one.
00:12:40.559 -->
00:13:00.159But you know, I um worked with um uh long story short is that I got a recommendation, got an interview, and um I think on my interview panel it was Sam Alman and um uh Jeff Ralston.
00:13:00.639 -->
00:13:05.440And I want to say Michael Siebel, but I'm not exactly I I I don't exactly remember, I think it was the three of them.
00:13:05.600 -->
00:13:05.919Yeah.
00:13:06.960 -->
00:13:27.120And I remember I was explaining how convolutional neural networks are really changing the way that people are doing image recognition, and that uh you know I wanted to make this application that was recognizing faces, and that I was very excited about AI, and that in general I I just knew Lambda was gonna do something in this AI space.
00:13:27.200 -->
00:13:27.519Yeah.
00:13:27.840 -->
00:13:36.320And um Sam was like kind of sitting there on the chair like this, and like I think he was like sitting on his leg in a funny way.
00:13:36.399 -->
00:13:40.240Oh yeah, and it was just like really he was paying attention to the thing, yeah.
00:13:40.559 -->
00:13:46.000And and and um and I kept mispronouncing um Jeff Ralston's name.
00:13:46.080 -->
00:13:48.480I call him Joff because it's spelled that way.
00:13:48.799 -->
00:14:05.440It happens, but um and and but I was like testing out this thing, and they thought it was interesting, but like it it just it was it it felt at the time a lot more like technology in search of a uh of a of a solution.
00:14:05.600 -->
00:14:06.159Yeah, right.
00:14:06.240 -->
00:14:10.559And and and I'll like admit at that time like that product really was a hundred percent.
00:14:11.039 -->
00:14:27.279But like there was this huge passion that I had for AI and for computer vision and for for this stuff that um that didn't die even despite not getting you know, I I Lambda was rejected from YC from that, you know, in that interview, and it's fine.
00:14:27.519 -->
00:14:43.120Um but but you know, I think that there's this feeling when you're so early to something, and it and especially if an industry is still so early, because like at that time, really deep learning at the time was still very much an academic field.
00:14:43.440 -->
00:14:43.600Of course.
00:14:43.840 -->
00:14:44.000Right?
00:14:44.080 -->
00:14:51.759Like and and it wasn't even within academia, you would still be a little bit kind of like uh you're a little bit on the fringes.
00:14:51.840 -->
00:14:53.120You're not really in this core.
00:14:53.200 -->
00:15:08.320I mean, at the time the core was like kind of more traditional things like uh Fisher faces or local binary patterns or um, you know, uh um SIFT features and support vector machines.
00:15:08.960 -->
00:15:18.240And right, uh and it was uh the the fully uh learned features thing was was still a little bit avant-garde until the Alexnet paper.
00:15:19.519 -->
00:15:24.480And so I think the thing that it you know, how does it feel to be so early?
00:15:24.639 -->
00:15:29.360You know what's really funny because I I remember seeing um her when it came out.
00:15:29.440 -->
00:15:29.679Okay.
00:15:29.919 -->
00:15:41.679The same, the same same it kind of her came out around the same time as AlexNet and also one of the papers I really love, which is Alex Graves handwritten handwriting um generation paper.
00:15:41.840 -->
00:15:42.080Yeah.
00:15:42.240 -->
00:15:45.759And I thought to myself, oh man, Hollywood's already onto this whole AI thing.
00:15:45.840 -->
00:15:46.720Like, you know what I mean?
00:15:46.879 -->
00:15:49.200Like, oh, it's like it's already too late.
00:15:49.440 -->
00:15:50.000Right, right.
00:15:50.159 -->
00:15:51.840But I mean I mean I'm I mean science fiction.
00:15:52.879 -->
00:15:55.279For those listening that might not know, like what around what year was that?
00:15:55.440 -->
00:15:57.039You know, like it like what what time period?
00:15:57.120 -->
00:15:59.360So you're thinking because again, Lambda's 2012.
00:15:59.679 -->
00:16:04.000Yeah, her must have been, I think, 2013 or early 2013.
00:16:04.399 -->
00:16:08.320Because for I know for me personally, it's basically the transformer paper and everything after that.
00:16:08.559 -->
00:16:10.240Like everything before before that for me.
00:16:22.720 -->
00:16:26.639And that was and that's for those again, those listening, that's 2017, the transformer paper.
00:16:26.960 -->
00:16:28.960The transformer paper was 2017.
00:16:29.279 -->
00:16:48.159Um, you know, I think that some of the interesting things that like really kind of got us towards that that thinking, if you, you know, is uh Ilya's um I think the the neural GPU paper, neural Turing machine papers, the the uh I think it was Alex Gray's paper, is the differential neural computer, the DNC paper.
00:16:48.639 -->
00:16:56.480And these are all examples of like some of the kind of thinking that I I was following and is really fascinated by.
00:16:57.039 -->
00:17:25.119Like I I I just was so into this idea that um you could um feed in sequences of like these bit streams of information, learn something about it, and you know, accurately predict it and um accurately uh there was what was the old embedding, there was this old embedding paper.
00:17:25.279 -->
00:17:32.400It was one of the first ones um where they were demonstrating the sort of vector math that you could do inside of the latent space.
00:17:32.640 -->
00:17:33.200Oh, I don't know.
00:17:34.079 -->
00:17:35.279Does this ring a bell though?
00:17:35.440 -->
00:17:35.759No, I don't know.
00:17:35.920 -->
00:17:42.559I forget that um in I remember seeing Chris Ola present this at like NUREPS 2015 or something.
00:17:42.720 -->
00:17:44.960Yeah, and it was like 2015.
00:17:45.279 -->
00:17:48.160I was grade I was grade 11 and I was playing saxophone all day.
00:17:48.400 -->
00:17:48.880Were you really?
00:17:49.119 -->
00:17:51.119You were you were in grade 11 at that point in time?
00:17:51.519 -->
00:17:52.720Absolutely, absolutely.
00:17:52.960 -->
00:17:56.319So I I can look back on when the Transformer paper was done.
00:17:56.480 -->
00:17:59.759I remember um making simple little programs using GPT-3.
00:18:00.079 -->
00:18:07.680Of course, I think you know one of my one of one of the favorite things of a of a of a terminally online poster was to make like green text generators.
00:18:07.759 -->
00:18:08.000Right?
00:18:08.160 -->
00:18:09.200Did that with GPT-3.
00:18:09.599 -->
00:18:10.880It was like document vectors.
00:18:11.119 -->
00:18:15.440So this was a paper called Document Embeddings with Paragraph Vectors.
00:18:15.759 -->
00:18:20.720And it was with um Andrew Dye and and and uh Quakla.
00:18:21.039 -->
00:18:22.640Is it Quakla is La or Lee?
00:18:22.799 -->
00:18:23.519I think it's Quakla.
00:18:24.480 -->
00:18:28.559Um but anyways, that document embedding with paragraph vectors paper.
00:18:28.960 -->
00:18:29.839I remember seeing it.
00:18:29.920 -->
00:18:46.400It was like Chris was present presented this, he goes, Yeah, you can take a vector and you can embed king and then you can subtract out the vector embedding of man, and you can add in the vector embedding of woman, and you can get queen.
00:18:46.640 -->
00:18:58.640And that's next to that's very that that resultant linear algebra in vector space is very close by the embedding of straight, you know, the straight embedding of queen.
00:18:58.880 -->
00:19:01.440And I was just like, oh my gosh, this is unbelievable.
00:19:01.759 -->
00:19:04.319I mean, I felt I felt that way when I did my first linear regression.
00:19:04.559 -->
00:19:04.880Yeah.
00:19:05.200 -->
00:19:08.640I was like, because I I worked, I mean, people don't actually know this, but I did a year at a hedge fund.
00:19:08.799 -->
00:19:09.680So Oh, did you really?
00:19:09.759 -->
00:19:09.920Okay.
00:19:10.160 -->
00:19:10.480Which one?
00:19:10.720 -->
00:19:11.680Uh Geneva Trading.
00:19:11.920 -->
00:19:12.480Okay, that's cool.
00:19:12.720 -->
00:19:13.200Yeah, yeah, yeah.
00:19:13.519 -->
00:19:15.359Very, very random, because I went to music school, right?
00:19:15.440 -->
00:19:22.079So it's very much sort of like who who I who I knew, but it's like, you know, again, I've got to I've always got to try and beat the beat the non-technical allegations.
00:19:22.400 -->
00:19:24.079Hedgy for a little bit, at what did you do?
00:19:24.160 -->
00:19:29.680Like what like what like what like what does that actually what does it mean to work at a hedge fund?
00:19:29.839 -->
00:19:31.279Like what do you actually do with the code?
00:19:31.440 -->
00:19:35.680Oh, so I I was I I I was um I was working with an X Tower researcher.
00:19:35.920 -->
00:19:39.440Um and you know, we just had like a very, very tiny uh prop allocation.
00:19:39.599 -->
00:19:44.559It was literally like, for me it was literally luck of the draw because I knew this guy, I knew how his strategy worked, and I knew how to program it.
00:19:44.880 -->
00:19:44.960Okay.
00:19:45.119 -->
00:19:48.400So you know, like like so you're like a research programmer a little bit.
00:19:48.559 -->
00:19:50.000When you read the paper and then write the code?
00:19:50.319 -->
00:19:53.119Like if I put it on my LinkedIn, I'm just a software engineer as far as I'm concerned.
00:19:53.440 -->
00:19:55.839Yeah, it's not ourselves research engineers.
00:19:56.000 -->
00:19:57.279It sounds so much better.
00:19:59.039 -->
00:20:05.599I always called myself a research engineer before, you know, uh, but but uh, you know, there's a lot a lot of back testing, right?
00:20:05.759 -->
00:20:12.000There's like a lot of a lot of backtesting, and then you know, tested in prod and then like we were like like in the prop world, we had a tiny, tiny allocation.
00:20:12.079 -->
00:20:17.680I think we only had like eight mil, which for anybody who's worked at a fund, that is they they know that is a tiny drop in the ocean.
00:20:17.759 -->
00:20:20.319But you know, got to fling some SP 500 futures for a year.
00:20:20.640 -->
00:20:22.720And you don't need to worry too much about moving the market.
00:20:23.279 -->
00:20:23.519Not at all.
00:20:24.319 -->
00:20:26.480The market moves on the pools of liquidity.
00:20:26.960 -->
00:20:30.480Well the good the good thing is, I mean, you know, like it was like one of those funny things, right?
00:20:30.559 -->
00:20:34.079And you know, um am I still under NDA?
00:20:34.160 -->
00:20:34.960I guess so, technically.
00:20:35.200 -->
00:20:43.759I mean, I you know, I don't have to like like r release the actual details of the strategy itself, but I can say is you know, it's one of those things in which um scaling strategies like that very, very fascinating.
00:20:43.920 -->
00:20:44.720That's all I'll really say.
00:20:44.960 -->
00:21:04.720It's very fascinating, and you kind of when you today you look on the AI side of things, you kind of wonder, wow, you know, how much AI is being used, to what extent, how successfully, like it's pretty clear if you look at um a lot of the the folks who are out there that there's a lot of AI being used in the markets these days.
00:21:05.200 -->
00:21:07.599Um Yeah, it's like it's so it's so hard to do that.
00:21:07.680 -->
00:21:14.880But it's such a black art because of just like the degree of the amount of money involved, the degree of confidentiality and everything.
00:21:14.960 -->
00:21:17.039But it's it's it's it's certainly fascinating.
00:21:17.200 -->
00:21:44.799You know, I remember so speaking of like what it was like in like the 2012-2013 era, I remember talking to somebody who I used to um sort of live with was uh in a hacker house in uh in San Francisco, uh this guy Alex Peake, and he said, you know, some people hypothesize that the first AIs to ever the tr first true AGIs will be stock market trading AIs.
00:21:45.200 -->
00:21:45.279Right.
00:21:45.440 -->
00:21:52.640Yeah, it's like the the joke where it's like, oh, you know, the the uh ASI will not be born out of an uh open AI or an anthropic, it'll be born out of Jane Street.
00:21:53.039 -->
00:21:53.920Yeah, yeah, right?
00:21:54.240 -->
00:21:57.519No, it's like hey, we again I keep saying it directionally correct.
00:21:57.759 -->
00:21:58.559Could be directionally correct.
00:21:58.799 -->
00:22:07.680I mean like like as far As I'm concerned, I always get um, you know, I mean, like when when you're on the West Coast, you get the talk of like, oh well, like um LLMs can just trade it for you.
00:22:07.759 -->
00:22:11.680And you know, having been on the inside before, you know, I I always balk at the idea.
00:22:11.839 -->
00:22:15.359You know, I'm I'm sure there's always going to be some sort of particular time period.
00:22:15.519 -->
00:22:18.079If you overfit enough, an LLM is gonna look amazing.
00:22:18.160 -->
00:22:20.720Uh like as if you overfit anything.
00:22:20.960 -->
00:22:23.039Like you overfit literally anything.
00:22:23.359 -->
00:22:28.480You back test anything on a particular time frame, and you know, you can put your sharp ratio through the roof.
00:22:28.640 -->
00:22:41.759Um, but with that sort of thing, like I I always find it hard to believe that um, you know, those that are sort of like, okay, well cool, we we got we got LLMs trading our portfolio that it's like you know, to to to go up against Jane Street, like what do you have that Jane Street does not have?
00:22:41.920 -->
00:22:43.839What do you have that Citadel does not have?
00:22:44.079 -->
00:22:48.240Like this is like like the kind of engineering that a that a Jane Street will have.
00:22:48.319 -->
00:22:55.440It's like they they are not simply like they they have stuff in which you can't even you're not even running on a CPU, like you're running on an FPGA.
00:22:56.240 -->
00:22:58.079Like and you're plugged into the network.
00:22:58.240 -->
00:23:01.200Like it's like so so yeah, right.
00:23:01.279 -->
00:23:08.000It's I mean that's like without even getting into the sort of the high frequency aspects of it.
00:23:08.240 -->
00:23:08.400Yes.
00:23:08.880 -->
00:23:30.720Which which totally different strategies still you know one of the interesting things is uh about you know being in in this market for for as long is that when you're so uh early in it, sometimes you'll even get people who will say, Oh, well, when is this bubble gonna pop?
00:23:30.960 -->
00:23:33.279Like in 2018 and 2017.
00:23:33.680 -->
00:23:34.160Seriously.
00:23:34.640 -->
00:23:48.160Because because because people, if you're an academic and you're attending NERUPS, you're looking at the basically excuse me, the uh the NERUPS attendees are doubling every single year.
00:23:48.799 -->
00:24:18.079And uh and you're kind of wondering, you know, and and and and and and around the time when when when when TensorFlow and Cafe started to become really popular, which was kind of really, in a sense, the democratization of of these deep learning libraries and frameworks, because it used to be back in the old days that we would train on, you know, we would use either theano, you either wrote bear CUDA, you did theano, or you did torch, meaning Lua torch.
00:24:18.240 -->
00:24:18.640Yeah.
00:24:18.880 -->
00:24:21.359And uh those they were all three good options.
00:24:21.759 -->
00:24:24.559But Lua torch is in Lua, the programming language.
00:24:24.880 -->
00:24:30.480As in the original Torch before PyTorch was PyTorch was built on Lua.
00:24:30.799 -->
00:24:31.119Yes.
00:24:31.359 -->
00:24:36.559And and and it had implementations for all the kernels in both CPU LAN and GPU LAN with CUDA.
00:24:36.880 -->
00:24:38.960I use Lua for for NeoVim.
00:24:39.359 -->
00:24:39.519Yeah.
00:24:40.559 -->
00:24:44.480Um but yeah, that we have to used to program our neural networks in Lua.
00:24:44.720 -->
00:24:45.039Yeah.
00:24:45.279 -->
00:24:47.359And um using Torch.
00:24:47.519 -->
00:24:47.680Yeah.
00:24:48.000 -->
00:25:08.400Um and and then you know, you I guess your bare CUDA or a some sort of uh templated auto, what is it called, you know, an auto-tuned uh kernel, which I think I would imagine people still to do today, although I haven't really stayed up on top of the like GPU kernel, jockey literature or otherwise work.
00:25:08.559 -->
00:25:12.880But um, I got a much larger company.
00:25:13.279 -->
00:25:28.480But those are the days that uh those the those were the good old days, and then when everybody started flooding into the market after that, and everyone's like using TensorFlow and Cafe, and those made it significantly easier, I think.
00:25:28.640 -->
00:25:34.960Uh mostly like certainly on the Theano side, this graph-based programming model was like very complicated and confusing.
00:25:35.119 -->
00:25:35.440Yep.
00:25:35.680 -->
00:25:42.160Um and PyTorch just really fully, fully democratized and also brought all the power, I think, of Torch.
00:25:42.400 -->
00:25:42.559Right.
00:25:42.640 -->
00:25:44.319Um, I mean I learned on ScikitLearn.
00:25:44.559 -->
00:25:45.519Was it on Scikit Learn?
00:25:45.680 -->
00:25:54.240Oh yeah, like man, the you ever do you ever do any of the Android uh uh deep learning tutorials?
00:25:54.559 -->
00:25:58.720No, I was I was very much like, you know, MATLAB.
00:25:58.960 -->
00:26:01.440Yeah, I was I was very much like Python for finance.
00:26:03.920 -->
00:26:04.799A lot of those were MATLAB.
00:26:05.279 -->
00:26:07.279A lot of people learn their neural networks on MATLAB.
00:26:07.519 -->
00:26:08.000That's so good.
00:26:08.319 -->
00:26:10.640And you know, that's that's how you know.
00:26:10.880 -->
00:26:13.519Um that was not even ten years ago, you know what I mean?
00:26:13.839 -->
00:26:14.480Right, no, right.
00:26:14.559 -->
00:26:24.720And and and so so what's interesting, the effect is is that people sort of who are early see like see this huge influx of people and they're like, all right, is this a bubble?
00:26:24.880 -->
00:26:25.039Right?
00:26:25.200 -->
00:26:31.599Like, and it's funny because people have just been calling this non-stop for for for a very long time.
00:26:32.000 -->
00:26:42.640Um, but what's been I think phenomenal to see is just the extent to which the capabilities of the models continues to expand with every generation.
00:26:42.960 -->
00:27:02.480So you've got um, you know, just uh something that's trained on a black wall class model is gonna just be so much better than the Haber one, and then the the Vera Rubin class models are just so much better than the and I I don't see that stopping any time student as a society, we're continuing to plow more and more capital into this.
00:27:02.720 -->
00:27:07.279So we're gonna continue to get bigger, you know, be able to expand the model capacity so dramatically.
00:27:07.599 -->
00:27:07.839Yep.
00:27:08.000 -->
00:27:11.119So I mean, you know, like like Moore's Law continues.
00:27:12.559 -->
00:27:35.440Well, you know, I mean, not Moore's Law, but but certainly the uh sort of it you know, uh the the capital formation seems to continue to be growing at uh on a nice uh clip and the uh the compute uh you know NVIDIA continues to do a phenomenal job of releasing new architectures.
00:27:35.759 -->
00:27:36.480Yeah, absolutely.
00:27:36.640 -->
00:27:36.880Absolutely.
00:27:37.039 -->
00:27:42.799I mean, even uh uh as of as of day of recording, OpenAI's jalapeno come comes out.
00:27:42.960 -->
00:27:43.119Yeah.
00:27:43.519 -->
00:27:44.160I mean, you know.
00:27:44.400 -->
00:27:45.759Did I look into it much this morning?
00:27:45.839 -->
00:27:46.559No, I was busy.
00:27:46.799 -->
00:27:50.400I ha yeah, I haven't I haven't I haven't caught up on it myself either.
00:27:50.559 -->
00:27:55.200Although uh cool, we can leave that bit in and you know it's just like this is what it is.
00:27:55.279 -->
00:27:56.319Anyway, thanks for coming, Stephen.
00:27:57.200 -->
00:27:58.079We'll wrap it up.
00:27:58.400 -->
00:27:58.880Exactly.
00:27:59.359 -->
00:28:00.720Yeah, cool place to wrap up.
00:28:00.799 -->
00:28:04.240Uh, you know, naturally as a podcaster, I can just comment like, oh, it seems cool.
00:28:04.400 -->
00:28:14.640Like and and Do you think do you think that the jalapeno color is a like a little bit of a uh because it's green, do you think that's a little bit of a a nod to to NVIDIA?
00:28:15.039 -->
00:28:15.359Oh.
00:28:15.759 -->
00:28:16.400It's gotta be.
00:28:16.559 -->
00:28:16.960It's gotta be.
00:28:17.039 -->
00:28:25.119I mean, this is like this is like going back to our earlier conversation around like you can interpret a tweet any number of ways and like people can interpret names and how much you read into it.
00:28:25.279 -->
00:28:28.000Wasn't the OpenAI logo, wasn't that originally green?
00:28:28.240 -->
00:28:28.720Was it?
00:28:28.960 -->
00:28:29.519I don't think so.
00:28:29.599 -->
00:28:31.599I can I can check the I think it was always black and white.
00:28:31.759 -->
00:28:32.480It's always black and white?
00:28:32.880 -->
00:28:34.880Yeah, I could be I just could be making stuff up.
00:28:35.039 -->
00:28:39.279Hey, which is my right to do as a as a as a as a poster online, naturally, naturally.
00:28:41.119 -->
00:28:44.400Okay, so I just finished meeting one of the day and they gave me a liquid death.
00:28:44.480 -->
00:28:45.519That's very, very nice of them.
00:28:45.680 -->
00:28:47.680Shout out to the um doing the guides.
00:28:47.759 -->
00:28:54.480But I'm just gonna do the dad read in the elevator real quick and be kind of redo in front of them circle back when I have your company brain making meetings memorable again.
00:28:54.640 -->
00:29:03.440Of course, because members of your meetings have to be able to do stuff with that information, like follow-up email, they can actually do it for you connected by connected linear, all of these sorts of things.
00:29:03.599 -->
00:29:08.079Hopefully, somebody's not out of the elevator games that be extremely embarrassing.
00:29:08.400 -->
00:29:10.079Um wearing the click.
00:29:11.119 -->
00:29:11.920That's fucked up.
00:29:12.160 -->
00:29:13.440This is why I couldn't do the vlogging thing.
00:29:13.519 -->
00:29:17.839I'm just so self-conscious and I don't want to be rude, but circleback.ai, go ahead and check them out.
00:29:17.920 -->
00:29:23.680Um check out circleback.ai slash mods first 60 days free for you and your team.
00:29:23.920 -->
00:29:24.799Check them out now.
00:29:25.599 -->
00:29:29.119Um, I think um one thing, you know, because we've kind of kind of gone through the evolution.
00:29:29.200 -->
00:29:42.720I know if if we if we just jump, if we just skip everything and we jump right to now, I know the the interesting sort of discourse right now is along the lines of like just like open source versus closed source, because we, you know, you're sort of mentioning every era, right?
00:29:42.880 -->
00:29:48.799I mean, everybody's always worried about quote unquote AI winter, every era, there's like there's like bubble talk, whatever.
00:29:49.039 -->
00:29:57.839Um I think the the the the most interesting thing over the last couple of months, I mean, I know there was a lot a lot discourse when Kimi K3 came out, for example.
00:29:58.079 -->
00:30:04.400And I feel like, and again, it's all feels and vibes based for me, which is great, because you know, I can never be wrong that way.
00:30:04.480 -->
00:30:31.599Uh but when it when when when Kimi K3 came out, a lot of discourse around like, oh, well, uh if if open source models can even harness like a like a decent fraction of the capability of that which is on the frontier, say I don't know, fable, or you know, enterprise level fable or whatever, then enterprises may actually just switch to them and save a bunch of money, you know, uh uh for the same amount of capability.
00:30:32.640 -->
00:30:36.720Do you do you think that that is sort of, again, I'm gonna say it again, directionally correct?
00:30:36.960 -->
00:30:44.000Or, you know, do you do you believe that enterprises are probably always gonna want that top-tier, frontier level intelligence and we'll pay whatever?
00:30:44.240 -->
00:30:55.200Because the reason I ask that is because it does seem to me that if the former happens, then Enthropic's top line revenue maybe, maybe doesn't, maybe doesn't go down, but maybe it slows down.
00:30:55.680 -->
00:30:57.680Maybe the AI lab's revenue slows down.
00:30:57.839 -->
00:31:00.960I know that's just one thing that I've been thinking about a lot as a podcaster.
00:31:01.440 -->
00:31:12.319So, so um the first thing I'd say is that at the very frontier of value creation, people are always gonna want to the very frontier of models.
00:31:12.480 -->
00:31:12.799Yeah.
00:31:13.279 -->
00:31:37.039Because, you know, especially if if you think about it where you're dedicating your life to try to start a company and you're using a model to write the software for your company, and let's say it's a software company or otherwise the software is a really important part of your company, like it is at Lambda, like it is at any other, you know, venture-backed company.
00:31:37.839 -->
00:31:40.000You are gonna use the frontier of models.
00:31:40.640 -->
00:31:40.799Right?
00:31:40.880 -->
00:31:59.599Like if if these are the models that are guiding, for example, whether it's solving a math proof or doing some sort of new metallurgy work or doing some sort of as those features get unlocked, because just to be clear here, we're only at the very edge of being able to start to address difficult math proofs.
00:31:59.920 -->
00:32:00.240Yes.
00:32:00.720 -->
00:32:00.960Okay.
00:32:02.400 -->
00:32:09.440The current addressable market with this current set of frontier models is really more on the software engineering side of things.
00:32:09.680 -->
00:32:09.839Yep.
00:32:10.240 -->
00:32:20.799But like in one or two model generations from now, it's gonna go from, well, they can program well to maybe they can do a little bit of guided scientific research well.
00:32:20.960 -->
00:32:26.799Maybe they can do material science a little bit better, maybe they can do some chemistry, some industrial process engineering better.
00:32:26.960 -->
00:32:48.160And if you're talking about something like that, right, and if you're a global Fortune 500 company that's let's let's say, because I I really do believe that today the models aren't quite there yet to be able to, like, for example, you know, come up with an amazing new alloy that's going to um totally change the way that we make natural gas turbines, for example.
00:32:48.400 -->
00:32:49.039Sure, sure.
00:32:49.279 -->
00:33:01.359Um, but when the models are there and and that sort of feature, if you will, becomes unlocked in terms of the model capacity being big enough to conduct scientific research, the what is the value of that?
00:33:01.519 -->
00:33:08.799Well, and you you you you might you you you you better be paying the full token price because you want you want your genius scientific researcher.
00:33:08.880 -->
00:33:11.680You know, you want to have the John von Neumann on your team.
00:33:11.839 -->
00:33:20.720You don't want um, you know, I won't say any names as like what a second-rate scientist might be, but like certainly many are second rate towards you know computer.
00:33:20.880 -->
00:33:23.119You want the John Neumannus level and the country of genius.
00:33:23.519 -->
00:33:23.920Exactly.
00:33:24.160 -->
00:33:50.640And so like what a a big part of this too, by the way, is you know, it comes down to price signals, which is to say that um the frontier models being more expensive, you know, when compute is scarce, frontier model companies absolutely need to charge a higher premium so as to better reallocate that compute towards those with the highest willingness to pay.
00:33:50.960 -->
00:33:51.200Of course.
00:33:51.519 -->
00:33:51.759Right?
00:33:51.839 -->
00:34:06.240And and uh who thus and the who has the highest willingness to pay, those who are creating the most value with their stuff, who have the highest spread between you know the cost of uh of using AI and the value that they're creating.
00:34:06.480 -->
00:34:15.199And so I I I think that that's why I think that you're gonna continue to see the Frontier Labs rocking it.
00:34:15.280 -->
00:34:15.519Yeah.
00:34:15.840 -->
00:34:18.480I think that it's really important to have open models.
00:34:18.719 -->
00:34:19.280Yes.
00:34:19.599 -->
00:34:26.800And the the the the the I'm I'm I'm not trying to like kind of split the baby and get like a milk toast answer on this.
00:34:26.960 -->
00:34:37.519I I you know being the I'll take you back to 2013 where where I was working with um my uh some two guys I joined their their company.
00:34:37.599 -->
00:34:43.039I was working at Lambda and and and after um after a little bit of time I wanted to learn from them.
00:34:43.119 -->
00:34:44.559They had finished their PhDs.
00:34:45.039 -->
00:34:50.559And um so I was like the first employee at their company, it was called Perceptio, which eventually got acquired by Apple.
00:34:51.119 -->
00:34:54.320And we were running these models locally on the iPhone.
00:34:54.400 -->
00:34:54.639Yep.
00:34:54.880 -->
00:34:55.199Okay.
00:34:56.000 -->
00:34:58.960You need to have, and and it was it was so cool.
00:34:59.119 -->
00:35:11.119Nico is this rock star GPU programmer, and he was writing all of the kernels in uh OpenGL ES, and it was like the coolest, just the we were so far ahead.
00:35:11.440 -->
00:35:20.639I mean, we were just doing it was so ironically, if you know this Apple's thing eventually ended up being called metal, their like local CUDA thing.
00:35:20.800 -->
00:35:22.239And this is way before metal.
00:35:22.320 -->
00:35:24.239Yeah, and it was really it was metal.
00:35:24.639 -->
00:35:27.760It was metal before metal, is is the joke I would make.
00:35:28.000 -->
00:35:33.440And um uh you realize that so at that time it that was a proprietary model.
00:35:33.519 -->
00:35:36.800We we had just trained a model from he he had trained a model from scratch.
00:35:36.960 -->
00:35:42.960My job was to like write the harness for it and to you know to bullet everything and mem map everything into memory.
00:35:43.039 -->
00:35:43.119Yes.
00:35:43.280 -->
00:35:52.000And I'd mem mapped it into the memory, and then the DSP would do the you know the convolutions uh using this library called GPU image at the time.
00:35:52.320 -->
00:35:52.960GPU image.
00:35:53.280 -->
00:35:59.039Yeah, and it's I think it's still probably an iOS library for convolutions and different stuff on the uh on on mobile.
00:35:59.360 -->
00:35:59.760Sure.
00:36:00.639 -->
00:36:09.760And um you need open models and open weights to be able to do stuff like that.
00:36:09.920 -->
00:36:10.320Yep.
00:36:10.719 -->
00:36:11.039Yep.
00:36:11.360 -->
00:36:21.599Anthropic is not, you know, like it at least I've yet to hear from either Sam or Dario, you know, for them when they pu to talk.
00:36:21.840 -->
00:36:35.280I have not seen them mention that they want to come up with something where they're going to, you know, have a proprietary like binary blob that gets loaded up and decrypted and then run locally, and then they get some sort of licensing fee off of that.
00:36:36.079 -->
00:36:36.639Right.
00:36:37.039 -->
00:36:40.400That's I think a religio relatively untenable model.
00:36:40.559 -->
00:36:41.039Yep.
00:36:41.360 -->
00:36:44.719And so that's why you really need open models.
00:36:44.880 -->
00:36:53.360You need open models because there's so many applications that are going to be benefiting from running locally, running in real time.
00:36:53.599 -->
00:37:00.159I mean, just go and look at um, you know, go and look at uh how self-driving works on the Tesla.
00:37:00.639 -->
00:37:02.480You can't have it running in a data center.
00:37:02.639 -->
00:37:02.880Yes.
00:37:03.039 -->
00:37:03.840It's unsafe.
00:37:04.320 -->
00:37:05.440It's irresponsible.
00:37:05.679 -->
00:37:05.920Yeah.
00:37:06.079 -->
00:37:08.719Um and and that's an that's on the extreme end of things.
00:37:08.960 -->
00:37:09.280Of course.
00:37:09.679 -->
00:37:15.679But um, you know, the the the the there's so much advantage to having local privacy preserving models.
00:37:15.760 -->
00:37:28.719And so though that's the reason why I'm still very much like a anthropic and open AI are gonna continue to do exceptionally well in terms of the frontier, and there's huge amounts of money there because people are just ripping tokens, generating stuff.
00:37:28.960 -->
00:37:29.360Yes.
00:37:29.599 -->
00:37:40.320But also there's gonna be a vast ecosystem of people on the uh open open model world that are gonna do great too because people need to run stuff locally.
00:37:40.639 -->
00:37:52.960Yeah, I mean I I speak to even pre-seed stage startups that are already thinking about, well, you know, my my information, or you know, I feel like my information's pretty damn valuable, so therefore I want to run a local model either in the cloud or on-prem, right?
00:37:53.039 -->
00:37:55.039When they're when they're when they're a bit further on.
00:37:55.280 -->
00:38:07.039Other startups, I mean, especially those that are a little bit more further on, they might be, and a lot of them are, just totally fine connecting to all the leading providers and being like, cool, you guys have got zero data retention, that's good enough for me.
00:38:07.199 -->
00:38:07.679Yeah, right.
00:38:07.760 -->
00:38:09.119We got an enterprise contract.
00:38:09.280 -->
00:38:09.360Yeah.
00:38:09.519 -->
00:38:17.519Let's I I think generally speaking, everybody should feel comfortable uploading their code base and everything into OpenAI and Enthropic with no problems.
00:38:17.599 -->
00:38:18.800Like, you know, that's that's my opinion.
00:38:19.039 -->
00:38:19.519Yeah, yeah, yeah.
00:38:19.599 -->
00:38:21.440I mean, I know, I know personally, I don't care.
00:38:22.480 -->
00:38:23.519I'm pretty I'm a pretty big deal.
00:38:23.599 -->
00:38:39.920So you know, yeah, yeah, no, and and I I mean it's the the value you get on the other side of it is just so immense that uh you know and I I think that they they take their data retention and privacy so so seriously, so deathly seriously that that it's uh they're trustworthy.
00:38:40.239 -->
00:38:40.960I think that makes sense.
00:38:41.199 -->
00:38:45.599So I don't know if you can see right now, but I'm actually at a San Francisco staple.
00:38:45.920 -->
00:38:47.840Um that is the black rock department.
00:38:48.079 -->
00:38:53.039I just paid$15 for a burrito bowl the size of my hand.
00:38:53.280 -->
00:39:05.199Now my hands are pretty big, but if you actually know of the size of a human hand, they're not big enough to provide adequate calories for 5.765 calories over.
00:39:05.440 -->
00:39:13.440What I digress is supposed to be an average for length.com, your number one LLM platform.
00:39:15.119 -->
00:39:20.559Check out your LLM observability platform, if you've got observability, you can uh do tracing.
00:39:20.800 -->
00:39:27.920Do all of these things to help improve your applications and it help them become self-improving.
00:39:28.159 -->
00:39:29.599That's the cool thing about them.
00:39:29.920 -->
00:39:32.400Actually met the guys around the corner at the Salesforce Part.
00:39:32.800 -->
00:39:35.119I'm living the San Francisco dream, damn it.
00:39:35.599 -->
00:39:40.079You know, I mean, obviously, like even even with this wave of the open source models, right?
00:39:40.159 -->
00:39:46.800I mean, again, the the the leading open source models, like I mean, of course, you know, Alex Wang talked about open sourcing new Spark, right?
00:39:46.960 -->
00:39:51.039Uh and well, and and Zuck started tweeting again, so it must be a pretty big deal.
00:39:51.119 -->
00:39:56.000Uh but the but the Chinese models, I mean, gosh, like, as far as I know, they're still not making any money.
00:39:56.079 -->
00:39:57.599Like, what happens if they just start making money?
00:39:57.760 -->
00:39:58.880They just go closed again?
00:39:59.360 -->
00:40:15.679Well, I mean, I think that there's plenty of of routes that they have within that uh that uh governmental system to provide for a subsidized uh open model, especially if they feel like that somehow is you know giving China some sort of competitive advantage.
00:40:15.920 -->
00:40:16.159Yep.
00:40:16.480 -->
00:40:42.239Um but yeah, I mean, you know, it's really interesting to think as um I was actually polling people within within Lambda today and just speaking, you know, you're talking about Chinese models and I'm talking about sort of government-backed, you know, whether it's a subsidize or whatever, to to to generate these models, uh, because maybe they think it confers to them some some sort of geopolitical advantage.
00:40:42.880 -->
00:40:51.119But you know, I was polling people as of what they thought was gonna be the third largest consumer of compute five years out from now.
00:40:51.440 -->
00:40:59.760So let's say we've got the two frontier leaders today, um Anthropic, OpenAI, you know, you've got XAI and Google as well.
00:41:00.000 -->
00:41:00.079Yes.
00:41:00.400 -->
00:41:04.400But just holding aside any these are these are the known four players.
00:41:04.559 -->
00:41:04.880Yeah, yeah.
00:41:05.199 -->
00:41:06.480You can add meta in there as well.
00:41:06.800 -->
00:41:06.880Sure.
00:41:07.360 -->
00:41:09.599But these are the known four to five players.
00:41:10.159 -->
00:41:19.760What do you think is gonna be the like that sort of where if we were to run this poll in 2022 or 2023, some people would have said anthropic?
00:41:19.920 -->
00:41:22.400Some people would have said, oh, it's gonna be, you know.
00:41:23.199 -->
00:41:25.360I would have said I would have said one of the fan companies back then, probably.
00:41:25.679 -->
00:41:25.920Right.
00:41:26.239 -->
00:41:26.400Yeah.
00:41:27.039 -->
00:41:27.519Yeah, exactly.
00:41:27.599 -->
00:41:30.800You would have been like, oh, maybe it's gonna be Microsoft with Copilot.
00:41:30.960 -->
00:41:31.039Yeah.
00:41:31.280 -->
00:41:40.639What do you think is gonna be the like maybe the dark horse or like the the thing that comes out of nowhere in the same way that anthropic probably came out of nowhere relative to 2023?
00:41:40.960 -->
00:41:41.760Uh Chinese government?
00:41:42.079 -->
00:41:42.559Chinese government.
00:41:42.719 -->
00:41:43.199So government.
00:41:43.599 -->
00:41:50.400So this is it's funny you say that because that was the exact response that I mean that was the mode response within the company.
00:41:50.639 -->
00:41:50.719Yeah.
00:41:50.960 -->
00:41:58.719Was that government was going to be uh you know, this this huge consumer and and and I think there's a bunch of good reasons for that.
00:41:59.039 -->
00:42:02.639So and so to be clearly this is government generally or or Chinese government specifically?
00:42:02.960 -->
00:42:08.079Um people were just saying gov, which I'm assuming, you know, could could be interpreted.
00:42:08.239 -->
00:42:13.679Yeah, some people were saying gov, some people were saying sovereign AI, which is kind of I would buck it into the same thing, you know.
00:42:13.840 -->
00:42:17.360Um in your opinion, is it US government or Chinese government?
00:42:17.440 -->
00:42:18.800You think you you were saying Chinese government specifically.
00:42:18.960 -->
00:42:20.559I'd say Chinese government specifically, actually.
00:42:20.880 -->
00:42:24.480And and so I don't think that I feel like yeah, US has has the two leading labs.
00:42:24.719 -->
00:42:26.239Like why I'm yeah, yeah, exactly.
00:42:26.320 -->
00:42:28.159There's there's there's there's not the same incentive.
00:42:28.719 -->
00:42:49.119Um but but just like government, uh I think government's realizing that um their and and and Jensen was so early on this in terms of the sovereign AI push, which is like government's realizing that um your country being run, your country's economy is increasingly being run by these models.
00:42:49.280 -->
00:42:49.440Yeah.
00:42:49.599 -->
00:42:55.119And increasingly decisions within your country are being guided by these models.
00:42:55.280 -->
00:42:55.599Okay.
00:42:55.840 -->
00:43:01.440Whether you know today it's code choices, tomorrow it's maybe managing other more important things.
00:43:01.599 -->
00:43:09.119Eventually it will be managing and optimizing infrastructure and other things that are of pretty critical national security importance.
00:43:09.280 -->
00:43:09.760Yes.
00:43:10.079 -->
00:43:16.480And having a model that is patriotic to your country is an important thing.
00:43:16.639 -->
00:43:17.039Yep.
00:43:17.519 -->
00:43:23.679And I I I really do think people are gonna start to realize that making sure your model's a patriot is important.
00:43:24.000 -->
00:43:24.159Yeah.
00:43:24.400 -->
00:43:27.920I mean, well, I mean, we we kind of already saw this with the initial DeepSeek models, no?
00:43:28.079 -->
00:43:28.239Yeah.
00:43:28.480 -->
00:43:33.119I mean, you know, like regardless of whatever how however thinks of you know a particular country's politics.
00:43:33.679 -->
00:43:37.679So no matter what, you can you can't doubt that that that that that Deep Seek's a patriot.
00:43:37.920 -->
00:43:39.280Yeah, of course, of course.
00:43:39.360 -->
00:43:42.719It's literally like you can you can try the prompts for yourself.
00:43:42.880 -->
00:43:45.519That's that's all I'll say, you know, just to not get any money in trouble.
00:43:46.079 -->
00:43:46.960But but and people will.
00:43:47.039 -->
00:43:47.440And people will.
00:43:48.239 -->
00:43:52.719And I I think that I I think that we ought to see more American models that are American patriots.
00:43:52.800 -->
00:43:52.880Yeah.
00:43:53.119 -->
00:43:53.679That's what I think.
00:43:53.840 -->
00:43:54.000Yeah.
00:43:54.239 -->
00:44:20.639And I and and I do I do I d I do think that there's aspects where um Where you know you kind of look at how everybody is deciding to do their stuff and without getting into everybody's thing, but one of the things I think is quite admirable about what Elon's approach has been to Grok is just how much he defines that as truth seeking as being the highest goal.
00:44:20.880 -->
00:44:21.199Yes.
00:44:21.440 -->
00:44:34.000And um, you know, because you do want your models to not be sycophants, and you do want your models to push back if you have a fake view of reality or a view of reality which is somehow mistaken.
00:44:34.159 -->
00:44:34.400Yeah.
00:44:34.559 -->
00:44:34.719Yes.
00:44:34.960 -->
00:44:44.400Um even if you don't like it, even if it is considered some sort of a faux pas or otherwise, you want your models to push back and say, you know, it's sort of, does it pass the Galileo test?
00:44:44.480 -->
00:44:58.320I think as he says it, which is, you know, um, you know, given all the information, the scientific information, even if it's, you know, a goes against the orthodoxy, does it still say and yet it moves?
00:44:58.559 -->
00:44:58.880Yep.
00:44:59.119 -->
00:44:59.360Yep.
00:44:59.599 -->
00:45:00.719No, it totally makes sense.
00:45:00.880 -->
00:45:01.440Totally makes sense.
00:45:01.599 -->
00:45:04.239I mean, you know, I think again, we've we've seen it over and over again.
00:45:04.320 -->
00:45:06.320And you know, people will continue to try and jailbreak this stuff.
00:45:06.480 -->
00:45:07.840You know, that's that's uh that's the thing, right?
00:45:07.920 -->
00:45:16.960It's like I I still remember when when DeepSeek first came out, it's like the first thing that was going viral is like those that are like, you know, pushing up against uh, you know, CCP proclivities.
00:45:17.039 -->
00:45:18.079Let's just put it that way.
00:45:18.239 -->
00:45:25.519Uh and you know, I think, yeah, we we we we are in an era of like AI psychosis, you know, however you want to put it, right?
00:45:25.599 -->
00:45:39.519It's like all the all the uh all the all the Instagram reels or the memes online where it's like, you know, oh yeah, I got um it's like me using Chat GPT to get uh gym advice, and it's like you know, it's like a video of somebody who's like dead on the ground, and it's like, oh, my mistake, I thought you meant, etc.
00:45:39.599 -->
00:45:40.559You know, it keeps going on.
00:45:40.719 -->
00:45:54.559And so, you know, to Elon's point, it's like, yeah, maybe, maybe have uh maybe have the LLM pushback uh if you are new to the gym and you're saying, oh, like I am this weight, should I put 400 pounds in the bar?
00:45:54.800 -->
00:45:55.280You know?
00:45:55.920 -->
00:45:56.719Seems like a good thing.
00:45:56.880 -->
00:45:58.320Yeah, exactly, exactly right.
00:45:59.039 -->
00:46:00.159Exactly right.
00:46:00.719 -->
00:46:02.880Okay, so now I'm walking through the mission district.
00:46:02.960 -->
00:46:07.199I just came from four-barrel coffee from two back-to-back meetings.
00:46:07.280 -->
00:46:11.760Um, but nobody gives a fuck because what we really want to hear about is black box.
00:46:13.039 -->
00:46:18.400Enough oil code plates by over 30 million box developers worldwide.
00:46:18.719 -->
00:46:28.079Use them for different uh redo around the world.
00:46:29.440 -->
00:46:30.239Redo truck.
00:46:30.480 -->
00:46:37.519Um encrypted inference enterprise that's a lovely sort of thing.
00:46:37.679 -->
00:46:38.639Check out Blackbox.
00:46:40.159 -->
00:46:50.079Yeah, I think um I I think something to to wrap on really briefly, and again, it's another thing, uh, as much as I'm a media guy and just kind of uh say a bunch of stuff online, I do like to think about these things.
00:46:50.239 -->
00:46:58.800And one of the uh another thing that I'm noticing is, again, you can sort of correct me if I'm wrong, if I, you know, if I think it's a trend and it's maybe not a trend.
00:46:58.960 -->
00:47:06.000But it seems to me that even higher up the stack, the higher up the stack that we get, you know, even I'm even thinking uh inference providers.
00:47:06.079 -->
00:47:14.400Um it seems to me that there is more of an emphasis on companies even higher up the stack starting to own their own hardware, right?
00:47:14.559 -->
00:47:18.000Buy GPUs, even thinking about data centers.
00:47:18.079 -->
00:47:18.880I won't say who.
00:47:19.039 -->
00:47:21.760But you know, like it does it does feel like it is is a trend, right?
00:47:21.840 -->
00:47:25.360Like the the hardware, the need for hardware is is moving up the stack.
00:47:25.679 -->
00:47:32.559Do you think uh, you know, a few years out from here, do you think that that will be an absolute necessity in in order to be a player in the space?
00:47:32.800 -->
00:47:35.039Like maybe you're maybe you're reselling tokens.
00:47:35.119 -->
00:47:37.679Do you need will you need hardware this far out?
00:47:39.840 -->
00:47:43.760Well, if you want to use AI, you need compute.
00:47:43.920 -->
00:47:44.239Yeah.
00:47:44.719 -->
00:48:00.079And you need to build an AI factory or at least have either ownership rights of an AI factory, right of use on an AI factory, meaning a lease or otherwise some sort of compute offtake agreement that you've uh uh uh signed uh and have a right of use.
00:48:00.159 -->
00:48:11.199And so, you know, how you decide to do that, I think is more of a financial question as you know, building it yourself, leasing it from somebody, um, you know, doing spot uh across a variety of clouds with no commitments.
00:48:11.360 -->
00:48:16.639You know, these are all I think equally valid ways to get access to compute that companies need.
00:48:16.880 -->
00:48:22.239Um I mean, I guess is your question like, you know, uh, am I seeing that as well?
00:48:22.719 -->
00:48:26.800Like, like, like Yeah, you know, I I I think that people are happy either way, to be perfectly honest.
00:48:27.039 -->
00:48:34.800And I think that they have to kind of look at, you know, they just have to look at the market and say, all right, well, how mission critical is this to my business?
00:48:35.039 -->
00:48:43.280Obviously for the frontier labs, it's pretty mission critical, and thus they need to lock it down for longer periods of time.
00:48:43.840 -->
00:48:52.000Um but uh, you know, I I think we'll we'll we'll see.
00:48:52.159 -->
00:49:05.199I mean, I think that that that you're gonna see the development of a computational grid in America that's gonna look, you know, it there's gonna be some shared things, I should say, yes, with our electrical grid in the United States.
00:49:05.519 -->
00:49:06.960And so this is exactly where I was going, actually.
00:49:07.679 -->
00:49:14.480The electrical grid is very much defined by the regulatory frameworks that the electrical grid seem to have formed around.
00:49:14.559 -->
00:49:20.400And I don't even know enough about the history of it to to really speak uh intelligently about it, so I won't.
00:49:20.639 -->
00:49:37.599But but you know, the the there's there's aspects where many of the utilities are sort of granted these sort of franchise monopolies on their um in their in their areas because they have to lay down wire and you don't want to double lay down wire, et cetera.
00:49:38.000 -->
00:49:43.280But you know, that doesn't at all apply to building an AI factory.
00:49:43.440 -->
00:49:43.599Right.
00:49:43.920 -->
00:50:04.079But I do think that you will see just sort of a lot of different companies participate in the building out of the American grid and building out of the American computational grid to support the massive, uh, as Elon put it, supersonic tidal wave of AI demand.
00:50:04.480 -->
00:50:04.960Totally, totally.
00:50:05.119 -->
00:50:10.719Now it's a it's a perfect thing because I wanted to set I wanted to set that up with like hardware now, power next, right?
00:50:10.960 -->
00:50:13.199And and so and and and we we can talk about power.
00:50:13.280 -->
00:50:26.239I think one of the other things I would definitely want to talk about is sort of like pushback uh within communities on AI data centers and also just like some of the reasons why actually having an AI data center in your backyard is a good thing.
00:50:26.320 -->
00:50:26.480Yeah.
00:50:26.800 -->
00:50:31.039Because there's so much disinformation in the world about this today.
00:50:31.199 -->
00:50:32.079It's unbelievable.
00:50:32.320 -->
00:50:32.480Right?
00:50:32.559 -->
00:50:38.960You know, people say, oh, it's gonna, you know, uh, you know, AOC's up there shaking dirty water.
00:50:39.199 -->
00:50:40.400It doesn't even make any sense.
00:50:40.480 -->
00:50:45.679That's not that's not what that's not what living next to a data center is like at all.
00:50:45.840 -->
00:50:51.199Um, you know, may maybe is I I I what is even the origin of the dirty water thing?
00:50:51.360 -->
00:50:54.960Like is it, you know, is is it if somebody's doing blasting and it breaks a pipe?
00:50:55.199 -->
00:50:55.519I don't know.
00:50:55.760 -->
00:51:00.639Is that any different than building a building or building doing any other construction project in the world?
00:51:01.039 -->
00:51:01.199Totally.
00:51:01.519 -->
00:51:05.840I mean, you know, the if it's it i the the AIs themselves actually disprove the whole water consumption thing anyway.
00:51:06.480 -->
00:51:13.360Because because they were they were talking about evaporative cooling versus you know, the vast majority of AI data centers are closed loop cooling.
00:51:13.840 -->
00:51:23.119So for like approximately a 10,000 GPU cluster, you might use between 20 and 30,000 gallons of water, which is not very much water.
00:51:23.199 -->
00:51:23.360Right.
00:51:23.519 -->
00:51:26.400It's about 20 to 30,000 almonds.
00:51:26.559 -->
00:51:26.719Yeah.
00:51:26.960 -->
00:51:27.199Okay.
00:51:27.519 -->
00:51:29.760What is the caloric value of an almond?
00:51:30.639 -->
00:51:30.960You must know.
00:51:31.199 -->
00:51:31.840I'd have to know.
00:51:31.920 -->
00:51:33.519I mean, well, I don't actually eat much almonds.
00:51:33.599 -->
00:51:36.320Uh you know, it's a dozen it's like it's like fine.
00:51:36.480 -->
00:51:37.519I mean five, I don't know.
00:51:37.599 -->
00:51:39.760You know, yeah, five, five units, I don't know.
00:51:40.159 -->
00:51:41.360Five, five, uh, five.
00:51:41.920 -->
00:51:45.760It's like maybe like uh like uh like an almond half a gram of protein, I don't know.
00:51:46.079 -->
00:51:46.880Right, right, right.
00:51:47.119 -->
00:51:50.960So so the point is that like it's 20 to 30,000 almonds.
00:51:51.199 -->
00:51:51.519Of course.
00:51:51.840 -->
00:51:53.920You know, um of water that's used.
00:51:54.079 -->
00:51:54.159Yes.
00:51:54.400 -->
00:51:58.960So it's this complete misinformation that that that just gets cycled through.
00:51:59.119 -->
00:52:04.159There's a bunch of uninformed people posting on Facebook uh the noise of the data center.
00:52:04.320 -->
00:52:05.760Have you ever walked by a data center?
00:52:06.079 -->
00:52:06.800No, it's pretty quiet.
00:52:07.119 -->
00:52:12.480It's it's it's like a quiet, it's a quiet, friendly residential neighborhood.
00:52:14.880 -->
00:52:16.480Oh I've walked past some pretty much.
00:52:16.639 -->
00:52:28.239I I don't know the exact decibel level, so I couldn't be able to say, but but but it's only when you know you're doing maybe uh if you have like a behind-the-meter gas turbine that's there that's running.
00:52:28.800 -->
00:52:32.960Um and and and that's the vast majority of data centers are not that.
00:52:33.280 -->
00:52:33.440Yes.
00:52:33.679 -->
00:52:34.400They're not that.
00:52:34.559 -->
00:52:40.079And um, I mean not to mention not to mention the um, you know, the the opportunity for a blue-collar workforce.
00:52:40.159 -->
00:52:49.599I mean, there's sort of been this narrative over the last few years of like, oh well, you know, i i if white-collar work is gone, I mean, I don't actually believe this, um, but you know, uh like like okay.
00:52:50.400 -->
00:52:53.119Like there's such a saturation in the colleges, I should say.
00:52:53.360 -->
00:52:56.719Therefore, why don't we send people to to the trades or whatever?
00:52:56.960 -->
00:53:00.639Well, here's your opportunity for the trades to go make decent six figures.
00:53:01.920 -->
00:53:02.719You know, on this build out.
00:53:03.519 -->
00:53:08.159I don't think that we're gonna see some big influx of white-collar workers going to the trades.
00:53:08.239 -->
00:53:12.159I think that most white-collar workers are not gonna survive going to the gas station.
00:53:14.480 -->
00:53:25.760Going to the gas station, getting a Zinn, a 24-ounce monster, and a roller hot dog in the morning because you know they don't have the constitution that most of the folks have.
00:53:26.079 -->
00:53:26.639Of course, yes.
00:53:26.960 -->
00:53:46.320Um, and the the reality is that the jobs that these are bringing to people who are working great jobs in the electrical, plumbing, mechanical industries, construction is so first of all, that's just like the during construction part.
00:53:46.719 -->
00:53:50.719During the construction part, hundreds of jobs get created.
00:53:51.679 -->
00:54:00.960During the maintenance part, you have amazing HVAC contracts, you've got amazing security contracts, you've got great electrical maintenance contracts.
00:54:01.360 -->
00:54:04.960And what you that's just the jobs for the community.
00:54:05.039 -->
00:54:05.360Yeah.
00:54:05.679 -->
00:54:12.000The tax income can be so massive, and people say, oh, well, you know, they're getting a tax incentive.
00:54:12.159 -->
00:54:17.440You know, many of these states, some of them is a full tax incentive, but many of them is only a partial tax incentive.
00:54:17.599 -->
00:54:17.760Right.
00:54:18.079 -->
00:54:25.440And uh and and and these are tax bills that could be equal to the budget of the entire county sometimes.
00:54:25.599 -->
00:54:25.920Yes.
00:54:26.159 -->
00:54:31.360Um, and so they're they're immensely good for the county's like financial standing.
00:54:31.599 -->
00:54:34.639Um, they actually are really quite quiet.
00:54:34.880 -->
00:54:47.119Like that, like there's just so much misinformation around um uh around this stuff that I just really wish people would just go and go and visit, go find a data center and go travel to the data center.
00:54:47.199 -->
00:54:47.519Yeah.
00:54:47.679 -->
00:54:52.159And you'll see that it's actually, you know, um it's gonna chew.
00:54:52.400 -->
00:54:55.760It's it's a pretty everything depends on it.
00:54:55.840 -->
00:54:57.840And then also here's the other fundamental thing.
00:54:58.079 -->
00:55:22.880There is still a responsibility that data center manufacturers and data center builders have towards, and AI factory builders have towards making um making them engineered so that they do consume a low amount of water, do not uh disrupt power prices in their communities and are quiet.
00:55:23.360 -->
00:55:26.400You can spend a little bit of money to do extra baffling.
00:55:26.639 -->
00:55:43.199You can you can um you can you can stand up behind the meter power and you can you can um you can you can provide um help to grid operators and to utilities to make it so that there is not an impact on the community.
00:55:43.440 -->
00:55:43.840Yes.
00:55:44.159 -->
00:55:49.840And that's what that's what that's what we're all more than happy and willing to do and are already doing.
00:55:50.079 -->
00:55:50.400Yes.
00:55:50.880 -->
00:56:05.039And so it's been really interesting to see it because there's just so much mis misinformation, and you kind of wonder maybe this is just an artifact of an election cycle where where maybe it's uh kind of ramps up during that because it's like a contentious topic that everyone likes to do.
00:56:05.360 -->
00:56:05.840It started it.
00:56:06.000 -->
00:56:10.320I mean, I you know, I like it started, as far as I'm concerned, it started even late last year.
00:56:10.480 -->
00:56:13.199You know, like like you had a bunch of message testing on either side.
00:56:13.280 -->
00:56:15.360So it definitely definitely makes sense to me.
00:56:15.519 -->
00:56:19.599Um gosh, I mean I think not to not to uh you know award myself a medal.
00:56:19.679 -->
00:56:29.519I think uh I was on core memory right at the beginning of the year saying that this this I I think this would become this would become the trend, and you know, uh them their messaging probably needs to change in order to accommodate.
00:56:29.599 -->
00:56:30.559But uh that was the thing.
00:56:30.719 -->
00:56:38.159I mean, as far as I'm concerned, uh my thing has always been, you know, this is the the largest upgrade to the American energy infrastructure in over a hundred years.
00:56:38.400 -->
00:56:41.440Yeah, well and and and think about the prosperity that's gonna come from that.
00:56:41.599 -->
00:56:41.840Yes.
00:56:42.239 -->
00:56:58.880You know, I I I talk mul multiple times, so I think that it's it's hard to disentangle sometimes how much of this pushback is really just like there's major fundamental and need to be addressed sort of issues around people's perception of inequality in the country.
00:56:58.960 -->
00:56:59.039Yeah.
00:56:59.360 -->
00:57:04.239Because I think people just don't generally understand that it's not a fixed pie.
00:57:04.559 -->
00:57:20.159People don't understand how wealth creation works, and so they kind of hold up people like Mark Zuckerberg or Elon Musk as like villains or something when the reality is that the world is made a much better place because they exist and they've they've built these businesses.
00:57:20.639 -->
00:57:25.840You kind of that's like an immovable object that's like impossible to address, I think.
00:57:26.480 -->
00:57:56.480Um but but going back to like what the benefits are to the overall society here is that like AI data centers are going to reduce our power cost and dramatically increase the United States energy production such that I think we might be able to start to see the return of certain types of industries that are core national security industries come back to the United States.
00:57:57.599 -->
00:58:06.320Whether it's certain types of um, you know, metal refining and smelting or certain types of manufacturing jobs.
00:58:06.800 -->
00:58:27.519I think that, you know, both because you're gonna lower that sort of hurdle rate, it's sort of like I I I've talked on a couple uh of places about this hurdle rate uh idea around energy, which is you know, in an economy okay, so in in project finance, if you've got a hurdle rate, which is your minimum financial return that you're gonna allow in the project, right?
00:58:28.159 -->
00:58:32.880A bunch of projects will get proposed and those with the highest hurdle rates will sort of get selected.
00:58:33.039 -->
00:58:33.119Yes.
00:58:33.440 -->
00:58:39.119And and and let's say you know your your your hurdle rate in an economy from a financial perspective is treasury.
00:58:39.440 -->
00:58:46.960Because you know, if it doesn't make more than if it doesn't return more than treasury, it's a risky project, and you might as well everyone's gonna put their money towards treasury.
00:58:47.119 -->
00:58:55.039So any any any anything that has a hurdle rate lower than treasury, or a return rate, I should say, lower than treasury is not gonna get done, right?
00:58:57.760 -->
00:58:59.119That's that's the exactly what I mean.
00:58:59.280 -->
00:59:07.440And then with activities that are energy intensive, which every activity is, there is a hurdle rate on the energy.
00:59:07.519 -->
00:59:27.199If your energy cost across the United States is 12 cents a kilowatt hour, like it is in some states, especially like I, you know, um you know, with within some states, you don't get certain economic activity, some sort of, you know, so you're not gonna have steel smelting in California, right?
00:59:27.280 -->
00:59:32.480You know, um, and so that's gonna go somewhere else with a lower cost of energy, like Texas or whatever.
00:59:33.039 -->
00:59:59.840Um and I think that's like a really positive thing that AI is gonna enable this power grid build-out that might result in a lot more data or excuse me, a lot more power plants being built that is gonna lower the hurdle rate for great job creating economic activity, like some of the basics of on material science, on metallurgy, on manufacturing.
01:00:00.559 -->
01:00:03.119And this is such a good thing.
01:00:03.440 -->
01:00:03.679Of course.
01:00:04.239 -->
01:00:19.519And and and that I think is like one of the biggest reasons why you should support and everybody should support things like nuclear reactors, data centers, and massive solar and battery storage projects.
01:00:19.760 -->
01:00:20.239Absolutely.
01:00:20.639 -->
01:00:27.199Because that is going to you know unlock so much wealth creation for everybody to enjoy.
01:00:27.280 -->
01:00:27.440Yep.
01:00:27.599 -->
01:00:30.480I mean, nobody does not benefit from having lower cost of energy.
01:00:30.719 -->
01:00:31.119Of course.
01:00:31.679 -->
01:00:39.840You get lower gas prices, you get lower uh metal prices, you get lower cost of goods for cars, every single thing.
01:00:39.920 -->
01:00:41.360And you see it in China.
01:00:41.440 -->
01:00:43.280Yep and why can't we have that in the United States?
01:00:43.440 -->
01:00:43.920I think we should.
01:00:44.159 -->
01:00:44.480Totally.
01:00:44.639 -->
01:00:45.039Absolutely.
01:00:45.199 -->
01:00:46.800Well, I think it's a great place to wrap.
01:00:46.960 -->
01:00:52.159Uh we've wrapped on a white pill, you know, as we' as as uh as people like me on the line.
01:00:52.480 -->
01:00:58.239We gotta end with a white pill because the world is already full of too many black pills and uh never take them.
01:00:58.639 -->
01:00:58.800Right.
01:00:59.039 -->
01:00:59.920Exactly right.
01:01:00.079 -->
01:01:01.920Well, Stephen, thanks for joining me here.
01:01:02.000 -->
01:01:04.239You've been listening to another episode of Members of Technical Staff.
01:01:04.320 -->
01:01:07.039We'll go ahead and see you on the timeline again, real soon.
01:01:07.840 -->
01:01:20.480Okay, so it is the evening, and I'm about to sit down and actually edit this video, but there is one, excuse me, more company that I need to talk about here, and that is Mots's newest sponsor, Chatbase.
01:01:20.559 -->
01:01:27.360Now, I'm actually gonna read this because I still haven't memorized the damn thing yet, and that's the kind of value you get from MOTS.
01:01:27.440 -->
01:01:30.880So apologies for the value uh before I actually talk about this.
01:01:31.039 -->
01:01:43.519So, real quick here, real quick, Chatbase is the complete platform for building AI customer support agents that can answer questions, automate workflows, and resolve customer issues across chat, email, and voice with powerful integrations, real-time knowledge, and enterprise grade security.
01:01:43.679 -->
01:01:51.920Chatbase helps businesses deliver fast, accurate support while turning every customer conversation into better outcomes.
01:01:52.159 -->
01:01:52.880Better outcomes.
01:01:53.119 -->
01:01:56.159Cool thing with chatbase is that they're also bootstrapped.
01:01:56.320 -->
01:01:59.599And uh Yasta the founder, he lives around the corner from me.
01:01:59.679 -->
01:02:03.039So if I ever do a poor job, he can come over and beat my ass.
01:02:03.760 -->
01:02:04.239Or something.
01:02:04.320 -->
01:02:05.280He wouldn't actually do that.
01:02:05.440 -->
01:02:09.519Is uh quite nice because I actually partner with the nicest and greatest.
01:02:09.599 -->
01:02:10.960Uh and that's that's that.
01:02:11.039 -->
01:02:13.039I mean I think that's that's enough value for now.
01:02:13.119 -->
01:02:16.000Uh let's go ahead and wrap this up, shall we?
