Marc Andreessen’s 2026 Outlook: AI Timelines, US vs. China, and The Price of AI (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) this new wave of AI companies is is (00:00:02) growing revenue like just like actual (00:00:04) customer revenue, actual demand (00:00:05) translated through to dollars showing up (00:00:07) in bank accounts at like an absolutely (00:00:09) unprecedented takeoff rate. We're seeing (00:00:10) companies grow much faster. I'm very (00:00:12) skeptical that the form and shape of the (00:00:14) products that people are using today is (00:00:16) what they're going to be using in 5 or (00:00:17) 10 years. I think things are going to (00:00:18) get much more sophisticated from here. (00:00:19) And so I think we probably have a long (00:00:20) way to go. These are trillion dollar (00:00:22) questions, not answers. But once (00:00:23) somebody proves that it's capable, it (00:00:25) seems to not be that hard for other (00:00:27) people to be able to catch up, even (00:00:28) people with far less resources. When a (00:00:30) company is confronted with fundamentally (00:00:32) open strategic or economic questions, (00:00:34) it's often a big problem. Companies like (00:00:37) need to answer these questions and if (00:00:38) they get the answers wrong, they're (00:00:39) really in trouble. Venture, we can bet (00:00:41) on multiple strategies at the same time. (00:00:43) We are aggressively investing behind (00:00:44) every strategy that we've identified (00:00:46) that we think has a plausible chance of (00:00:48) working. If you want to understand (00:00:49) people, there's basically two ways to (00:00:51) understand what people are doing and (00:00:52) thinking. One is to ask them and then (00:00:54) the other is to watch them. And what you (00:00:56) often see in many areas of human (00:00:57) activity, including politics and many (00:00:59) different aspects of society, the (00:01:01) answers that you get when you ask people (00:01:02) are very different than the answers that (00:01:03) you get when you watch them. If you run (00:01:05) a survey or a poll of what, for example, (00:01:07) American voters think about AI, it's (00:01:09) just like they're all in a total panic. (00:01:10) It's like, "Oh my god, this is terrible. (00:01:12) This is awful. It's going to kill all (00:01:13) the jobs. It's going to ruin (00:01:13) everything." If you watch the revealed (00:01:15) preferences, they're all using AI. (00:01:22) A lot of folks have sent questions ahead (00:01:24) of time and and what I what I've done is (00:01:25) kind of curated into a few different (00:01:27) sections uh in in an AMA this morning (00:01:29) with uh with Mark. So, what we thought (00:01:31) we'd do is cover uh four big topics. So, (00:01:33) AI and what's happening in the markets, (00:01:35) policy and regulation, um all things (00:01:37) 816Z, and then we've got a a fun (00:01:39) catchall which we're we're calling (00:01:41) sandbox of things if we get to it. So, (00:01:43) starting first maybe with uh with the (00:01:45) biggest question. We're sitting in the (00:01:46) middle of the AI revolution. Mark, what (00:01:48) inning do you think we're in and and (00:01:50) what are you most excited about? (00:01:51) >> First of all, I I would say this is the (00:01:53) biggest tech technological revolution of (00:01:54) my life. Um and you know, hopefully I'll (00:01:57) see more like this in the next whatever (00:01:59) 30 years, but I I mean this is the big (00:02:01) one. Um and just in terms of order of (00:02:03) magnitude, like this is clearly bigger (00:02:04) than the internet. Um like the the the (00:02:07) comps on this are things like the (00:02:08) microprocessor and the steam engine and (00:02:10) electricity. So that this is a really (00:02:12) this is a really big one. um the wheel. (00:02:15) Um the reason this is so big, I mean (00:02:18) maybe obvious to folks at this point, (00:02:19) but I'll just go through it quickly. So (00:02:21) um if you kind of trace all the way back (00:02:23) to the 1930s, uh there's a great book (00:02:25) called Rise of the Machines that kind of (00:02:26) goes through this. Um if you trace all (00:02:28) the way back to the 1930s, there was (00:02:29) actually a debate among the people who (00:02:30) actually invented the computer. Um and (00:02:33) it was this this sort of debate between (00:02:34) whether computer they kind of understood (00:02:36) the theory of computation before they (00:02:37) before they they actually built the (00:02:39) things. Um and um they they had this big (00:02:41) debate over whether the computer should (00:02:43) be basically built in the image of what (00:02:44) at the time were called adding machines (00:02:46) or calculating machines where you know (00:02:48) think of sort of essentially cash (00:02:49) registers. Um IBM is actually the (00:02:52) successor company to the national cash (00:02:53) register company uh of America. Um and (00:02:55) so like and and and that was of course (00:02:57) the path that the industry took which (00:02:59) was building these kind of hyper literal (00:03:00) you know mathematical machines you know (00:03:03) that could execute mathematical (00:03:04) operations billions of times per second (00:03:05) but of course had no ability to kind of (00:03:07) deal with human beings the way humans (00:03:09) like to be dealt with and so you know (00:03:10) couldn't understand you know human (00:03:11) speech human language um and so forth (00:03:14) and and that's the computer industry (00:03:15) that got built over the last 80 years (00:03:16) and that's the computer industry that's (00:03:18) built all the wealth of uh uh and and (00:03:20) financial returns of the computer (00:03:21) industry uh over the last 80 years you (00:03:23) know across all the generations of (00:03:24) computers from mainframes through to (00:03:26) smartphones. Um but but they knew at the (00:03:28) time they knew in the 30s actually they (00:03:29) understood the basic structure of the (00:03:31) human brain. They understood they had a (00:03:32) theory of sort of human cognition and (00:03:34) and and actually they had the theory of (00:03:35) neural networks. Um and so they they had (00:03:37) this theory that um the there's actually (00:03:40) the first neural network uh paper (00:03:42) academic paper was published in 1943 you (00:03:44) know which was over 80 years ago which (00:03:46) is extremely amazing. Um there's an (00:03:49) interview you can read an interview or (00:03:50) you can watch an interview on YouTube (00:03:51) with uh these two authors Makulla and (00:03:54) Pitts and you can watch an interview I (00:03:55) think with Makulla on YouTube from like (00:03:57) I don't know 1946 or something. He was (00:03:59) like on TV you know in the in the (00:04:01) ancient past and it's literally like (00:04:03) it's amazing interview because it's like (00:04:04) him in his beach house and for some (00:04:05) reason he's not wearing a shirt um and (00:04:07) he's like you know talking about like (00:04:08) this future in which computers are going (00:04:10) to be you know built on on the model of (00:04:11) a human brain through through neural (00:04:12) networks. Um and and that was the path (00:04:15) not taken. And basically what happened (00:04:16) was right the computer industry got (00:04:18) built in the in the image of of like the (00:04:20) adding machine. Um but and the neural (00:04:22) network basically didn't happen but the (00:04:23) neural network as an idea continued to (00:04:25) be explored in academia um and sort of (00:04:27) advanced research by sort of a rump you (00:04:29) know movement that was originally called (00:04:31) cybernetics and then became known as as (00:04:33) artificial intelligence uh basically for (00:04:35) the last 80 years and and and (00:04:36) essentially it didn't work like (00:04:38) essentially it was basically decade (00:04:39) after decade after decade of excessive (00:04:41) optimism uh followed by disappointment. (00:04:43) When I was in college in the 80s, there (00:04:45) had been a famous kind of AI boom bust (00:04:47) uh cycle in the 80s in venture and in (00:04:49) Silicon Valley. Um I mean it was tiny by (00:04:52) by modern standards, but it it at the (00:04:53) time was a big deal. Um and um you know (00:04:56) by the time I got to college in '89 um (00:04:58) in computer science departments, AI was (00:04:59) kind of a backwater field and everybody (00:05:01) kind of assumed that it was never going (00:05:02) to happen. But the scientists kept (00:05:04) working on it to their credit and they (00:05:05) they they built up this kind of enormous (00:05:06) reservoir of of concepts and ideas and (00:05:08) then basically we all saw what happened (00:05:10) with the CHIGPT uh moment. all of a (00:05:13) sudden it it sort of crystallized. It's (00:05:14) like oh my god, right? It turns out it (00:05:16) works. Um and and so you know that (00:05:18) that's the moment we're in now. And then (00:05:19) you know really significantly that was (00:05:21) what you that was less than three years (00:05:22) ago, right? That was the summer of 20 it (00:05:24) was the the Christmas of 22. So, we're (00:05:26) sort of three year we're we're sort of (00:05:28) three years in um to, you know, (00:05:31) basically what is effectively (00:05:32) effectively an 80-year revolution um of (00:05:35) actually being able to deliver on all (00:05:36) the promise that the that the people on (00:05:38) the all the on the alternate path, the (00:05:40) sort of human cognition model path, you (00:05:42) know, kind of saw from the very (00:05:43) beginning and and then, you know, the (00:05:44) great news with this technology is it's (00:05:46) already it's kind of ultra democratized. (00:05:48) you know, the best AI in the world is (00:05:49) available. Launch at GPD or Grock or (00:05:51) Gemini or or um you know, these other (00:05:53) you know, these other products that you (00:05:54) can just use um and you can just kind of (00:05:56) see how they work and you know, same (00:05:58) thing for video, you can see with Sora (00:05:59) and VO kind of state-of-the-art uh with (00:06:01) that you can see with music, you can see (00:06:03) you know uh Suno and IDO and so forth. (00:06:05) Um and so like you know we're basically (00:06:08) seeing that happen and now and now (00:06:10) Silicon Valley is responding with this (00:06:11) just like incredible rush of enthusiasm. (00:06:14) And you know, really critically this (00:06:16) gets to the magic of Silicon Valley, (00:06:17) which is, you know, Silicon Valley long (00:06:19) since has ceased to be a place where (00:06:21) people make silicon that, you know, (00:06:22) that's that not long ago moved out out (00:06:24) of the out out of California and then (00:06:25) ultimately out of the US, although we're (00:06:27) trying to bring it back now. Um but but (00:06:29) but the great kind of virtue of Silicon (00:06:31) Valley o over the last you know over the (00:06:32) last you know 80 years of its existence (00:06:34) is its ability to kind of uh recycle (00:06:36) talent from previous waves of technology (00:06:38) and new waves of technology uh and then (00:06:40) inspire an entire new generation of (00:06:42) talent you know to basically come join (00:06:43) the you know join the project. Um and so (00:06:45) Silicon Valley has this recurring (00:06:47) pattern of being able to reallocate (00:06:48) capital and talent and build enthusiasm (00:06:50) and build critical mass and build (00:06:51) funding support and build you know human (00:06:53) capital and build you know everything (00:06:54) enthusiasm um you know for each new wave (00:06:57) of technology. So, so that's what's (00:06:59) happening with AI. Um, you know, I I (00:07:02) think probably the biggest thing I could (00:07:03) just say is like I'm surprised I think (00:07:05) essentially on a daily basis of what I'm (00:07:06) seeing. Um, uh, and and you know, we (00:07:09) we're we're in the fortunate position to (00:07:10) kind of get to see it from from two (00:07:12) angles. Uh, you know, one one is we (00:07:14) track the underlying science and and, (00:07:16) uh, and kind of research work very (00:07:17) carefully. And so I would say like every (00:07:18) day I see a new AI research paper that (00:07:20) just like completely floores me um of (00:07:22) some new capability um or some new (00:07:24) discovery uh or some new development (00:07:26) that I that I would have never (00:07:27) anticipated that I I'm just like wow I (00:07:29) you know I can't believe this is (00:07:30) happening. And then um on the other side (00:07:32) of course you know we see the flow of (00:07:33) all of the new uh products uh and all (00:07:35) the new startups. Um and you know I (00:07:37) would say we're routinely um you know (00:07:40) kind of seeing things and again kind of (00:07:41) have my my jaw on the floor. Um, and so, (00:07:43) you know, it feels like we we we've (00:07:45) unlocked this giant vista. Um, I do (00:07:48) think it's going to kind of come in fits (00:07:49) and starts. Um, you know, the these (00:07:51) things are messy processes. Um, you (00:07:53) know, you know, this is an industry that (00:07:55) kind of routinely gets out over risks (00:07:56) and overpromises. Um, and and so, you (00:07:59) know, there, you know, there will (00:08:00) certainly be points where it's like, (00:08:01) wow, you know, this isn't working as (00:08:02) well as people thought, or you know, (00:08:03) wow, this turns out to be too expensive (00:08:04) and the economics don't work or (00:08:06) whatever. But, you know, against that, I (00:08:07) would just say the capabilities are (00:08:08) truly magical. And and by the way, I (00:08:10) think that's the experience that (00:08:11) consumers are having when they use it. (00:08:13) And I think that's the experience that (00:08:14) businesses are having for the most part (00:08:15) when they uh you know, when when they're (00:08:17) working on their pilots and and looking (00:08:18) at adoption and and and then and then it (00:08:21) translates to the underlying numbers. I (00:08:22) mean, we're we're just seeing a this new (00:08:23) wave of AI companies is is growing (00:08:25) revenue like just like actual customer (00:08:27) revenue, actual demand uh translated (00:08:29) through to dollars showing up in bank (00:08:31) accounts. Um you know, at like an (00:08:33) absolutely unprecedented takeoff rate. (00:08:34) We're seeing companies grow much faster. (00:08:36) um uh you the the key leading AI (00:08:39) companies and the companies that have (00:08:40) real breakthroughs um and have real have (00:08:42) very compelling products are growing (00:08:43) revenues that you know kind of faster (00:08:44) than any any way I've certainly ever (00:08:46) seen before. Um and so like just just (00:08:48) from all that it kind of feels like it (00:08:49) has to be early. Like it it's kind of (00:08:51) hard to imagine that we've like we we've (00:08:53) topped out in any way. It feels like (00:08:55) everything is still developing. I mean (00:08:56) quite frankly it feels like the products (00:08:58) to me it feels like the products are (00:08:59) still super early. Like I'm I'm I'm very (00:09:01) skeptical that the form and shape of the (00:09:04) products that people are using today is (00:09:05) what they're going to be using in five (00:09:06) or 10 years. I think I think things are (00:09:07) going to get much more sophisticated (00:09:08) from here. Um and so I think we probably (00:09:10) have a long way to go. (00:09:11) >> Maybe on that that topic. So one of the (00:09:13) big knocks is yes the revenue is immense (00:09:15) but the expenses seem to also be keeping (00:09:18) pace. So like what are people missing as (00:09:20) a part of that discussion and topic? (00:09:22) >> Yeah. So just start with just like core (00:09:24) business models, right? Um and so you're (00:09:26) right. There's basically this industry (00:09:27) basically has two two core business (00:09:28) models. consumer business model and the (00:09:30) quote unquote enterprise uh or (00:09:32) infrastructure business model. Um you (00:09:33) know look on the on the consumer side we (00:09:35) we just live in a very interesting world (00:09:36) now where where the internet exists and (00:09:38) is fully deployed right. Um, and so I'll (00:09:41) give you an example. Sometimes people (00:09:42) ask us like, "Is AI like the internet (00:09:44) revolution?" It's like, well, a little (00:09:45) bit, but like the thing with the (00:09:46) internet was we had to build the (00:09:47) internet. Like we we like we had we had (00:09:49) to actually build the network and we (00:09:51) actually had to, you know, and (00:09:53) ultimately it involved enormous amount (00:09:54) of fiber in the ground and it involved (00:09:55) enormous numbers of like mobile cell (00:09:57) towers and, you know, enormous number (00:09:58) of, you know, shipments of of of (00:10:00) smartphones and tablets and and and (00:10:02) laptops in order to get people on the (00:10:03) internet. Like there was this like just (00:10:05) like incredible physical lift um, you (00:10:07) know, to do that. And and by the way, (00:10:08) people forget how long that took. Uh (00:10:10) right, the the the you know, the (00:10:11) internet itself is a invention of the (00:10:13) 1960s, 1970s. Um the consumer internet, (00:10:16) you know, was a new phenomenon in the (00:10:18) early '90s. Um but, you know, we didn't (00:10:20) really get broadband to the home until (00:10:21) the 2000s. You know, that really didn't (00:10:23) start rolling out actually until after (00:10:24) the com crash, which is fairly amazing. (00:10:26) And then we didn't get mobile broadband (00:10:28) until like 2010. And and people actually (00:10:30) forget the original iPhone dropped in (00:10:32) 2007. It didn't have broadband. it was (00:10:35) on a it was on a narrowband 2G network. (00:10:38) Um it did not have high speed like it (00:10:40) did not have anything resembling (00:10:41) high-speed data. Um and so it wasn't (00:10:43) really until you know really about 15 (00:10:45) years ago that we even had mobile (00:10:46) broadband. So so the internet was this (00:10:48) massive lift but but the internet got (00:10:50) built right and smartphones (00:10:51) proliferated. And so the point is now (00:10:53) you have 5 billion people on planet (00:10:54) earth that are on some version of you (00:10:56) know mobile broadband internet right um (00:10:58) and you know smartphones all over the (00:11:00) world are selling for you know as little (00:11:01) as like 10 bucks. Um and you know you (00:11:03) have these you know amazing projects (00:11:04) like geo and India that are bringing you (00:11:06) know you know the sort of the remaining (00:11:08) you know kind of the remaining (00:11:09) population of of planet earth that (00:11:10) hasn't been online until now is coming (00:11:11) online. So, you know, so we're talking (00:11:13) five billion, six billion, you know, (00:11:15) people and and then the consumer, the (00:11:17) reason I go through that is the consumer (00:11:18) AI products could basically deploy to (00:11:20) all of those people basically as quickly (00:11:22) as they want to adopt, right? Um, and so (00:11:24) sort of the internet's the carrier wave (00:11:26) for AI to be able to proliferate at kind (00:11:28) of light speed uh uh into the broad base (00:11:30) of the global population. And and that's (00:11:32) a let's just say that's a potential rate (00:11:34) of proliferation of a new technology (00:11:35) that's just far faster than has ever (00:11:37) been possible before. Like what you (00:11:38) know, like you couldn't download (00:11:40) electricity, right? you you couldn't (00:11:42) download, you know, you couldn't (00:11:43) download indoor plumbing. Um, you know, (00:11:46) you couldn't download television, but (00:11:47) you can download AI. Um, and and and (00:11:49) this is what we're seeing, which is the (00:11:50) AI consumer, you know, the AI consumer (00:11:52) killer applications are growing at at at (00:11:54) an incredible rate. Um, and then and (00:11:56) then they're monetizing really well. Um, (00:11:58) and and again, you know, we we I (00:12:00) mentioned this already, but like (00:12:01) generally speaking, the monetization is (00:12:02) is very good. Um, by the way, including (00:12:04) at higher price points. Um, one of the (00:12:06) things I like about the um, you know, (00:12:08) about watching the AI wave is the AI (00:12:10) companies I think are are more creative (00:12:11) on pricing than the SAS companies and (00:12:13) the consumer internet companies were. (00:12:14) And so it's it's for example now (00:12:15) becoming routine to have $200 or $300 t (00:12:17) per month tiers uh, for consumer AI (00:12:19) which I which I think is very positive (00:12:20) because I I think the I think a lot of (00:12:23) companies cap their kind of opportunity (00:12:25) by by capping their pricing uh, kind of (00:12:27) too low and I think the AI companies are (00:12:28) more willing to push that which I think (00:12:29) is good. So anyway, so that you know I (00:12:32) think that's reason for like I would say (00:12:33) you know considerable rational optimism (00:12:35) for the scope of of consumer revenues (00:12:37) that we're going to be talking about (00:12:38) here. And then on the enterprise side, (00:12:40) you know, there the question is (00:12:41) basically just, you know, what is (00:12:43) intelligence worth, right? Um, and you (00:12:46) know, if you have the ability to like (00:12:47) inject more intelligence into your (00:12:48) business and you have the ability to do, (00:12:50) you know, even the most prosaic things (00:12:52) like raise your customer service scores, (00:12:53) uh, you know, increase upsells, um, uh, (00:12:56) you know, or reduce churn or if you have (00:12:57) the ability to, um, you know, run (00:12:58) marketing campaigns more effectively, (00:13:00) um, you know, all of which AI is (00:13:02) directly relevant to, like, you know, (00:13:03) these are like direct business payoffs, (00:13:05) um, you know, that people are seeing (00:13:06) already. Um, and then if you have the (00:13:08) opportunity to infuse AI into new (00:13:09) products and all of a sudden, you know, (00:13:11) all of a sudden your car talks to you, (00:13:13) um, and everything in the world kind of (00:13:14) lights up and starts to get really (00:13:15) smart. Um, you know, you know, what's (00:13:17) that worth? And and again there you just (00:13:19) you you kind of observe it and you're (00:13:20) like, wow, the the leading AI (00:13:22) infrastructure companies are growing (00:13:23) revenues incredibly quickly. Um, you (00:13:25) know, the pull is really tremendous. Um, (00:13:27) and so, you know, again there it's just (00:13:29) it feels like this just like incredible (00:13:31) uh, you know, product market fit. Um and (00:13:33) and and the core business model, right, (00:13:35) is is is actually quite quite (00:13:36) interesting. The core business model is (00:13:38) is is basically is basically tokens by (00:13:39) the drink, right? And so it's it's sort (00:13:41) of tokens of intelligence uh you know, (00:13:43) per dollar. And oh, and then by the way, (00:13:45) this is the other fun thing is if you (00:13:46) look at what's happening with uh the (00:13:48) price of AI, the price of AI is falling (00:13:51) much faster than Moore's law. And when I (00:13:54) could go through that in great detail, (00:13:55) but basically like all of the inputs (00:13:56) into AI on a perunit basis, the costs (00:13:59) are collapsing. Um and and and and then (00:14:01) as a consequence there's kind of this (00:14:03) hyperdelation of per unit cost and then (00:14:04) that is like driving you know just like (00:14:07) you know a more than corresponding level (00:14:08) of demand growth you know with with with (00:14:10) elasticity. Um and so you know even (00:14:13) there we're like it feels like we're (00:14:14) just at the very beginning of kind of (00:14:16) you know figuring out exactly how you (00:14:18) know expensive or cheap this stuff is (00:14:19) getting. I mean look there's just no (00:14:20) question tokens by the drink are going (00:14:21) to get a lot cheaper from here. Um (00:14:23) that's just going to drive I think (00:14:24) enormous demand. Um and then everything (00:14:26) in the cost structure is going to get (00:14:28) optimized right? Um and so you know when (00:14:30) when people talk about like you know the (00:14:32) chips or you know whatever you know kind (00:14:34) of the unit input costs for building AI (00:14:36) you know you now have these like m the (00:14:39) losses of blind demand are are going to (00:14:40) are going to kick in right um what's the (00:14:43) you know in any market that has sort of (00:14:44) commodity like characteristics you know (00:14:45) the number one cause of a of a of of a (00:14:47) glut is a shortage and the number one (00:14:49) cause of a shortage is the glut right um (00:14:51) and so you have you know to the extent (00:14:53) you have like shortage of GPUs or (00:14:55) shortage of whatever infest chips or (00:14:56) shortage of you know whatever data (00:14:58) center case, you know, if you look at (00:14:59) just the history of humanity building (00:15:01) things in response to demand, you know, (00:15:03) if there's a shortage of something that (00:15:05) can be physically replicated, it it does (00:15:07) get replicated. Um, and so there's going (00:15:09) to be like just enormous build out of (00:15:10) all I mean there is there's just (00:15:11) hundreds of billions or at this point (00:15:13) trillions of dollars maybe going into (00:15:14) the ground um in all these things. And (00:15:16) so the the per unit cost of the AI (00:15:18) companies are going to drop like a rock (00:15:20) um you know over the course of the next (00:15:21) decade. Um and so like I yeah I mean the (00:15:25) economic questions of course are very (00:15:26) real and of course there's you know (00:15:27) microeconomic questions around around (00:15:29) all these businesses but the the sort of (00:15:31) macro forces have been at least here I (00:15:32) think are very strong um and and yeah I (00:15:35) I just given the underlying value of the (00:15:37) of of this technology to both the (00:15:39) consumers the enterprise users. Um, and (00:15:42) given the just like incredibly (00:15:44) aggressive discovery that's happening of (00:15:45) of all the ways that people can use this (00:15:46) in their lives and in their businesses, (00:15:48) like it's just it's really hard for me (00:15:49) to see how it both doesn't grow a lot (00:15:50) and generate just enormous revenue. (00:15:52) >> Yeah. And actually, I think it was like (00:15:54) two or three weeks ago where AWS was (00:15:55) saying like the the GPUs that they've (00:15:57) been using, they've been able to extend (00:15:58) back to even like seven plus years. So (00:16:00) like the shelf life also of the GPUs (00:16:02) that they're using is now extending in (00:16:05) ways of which they can optimize better (00:16:07) than maybe perhaps the last couple of of (00:16:10) cycles. as well. Is that the right way (00:16:11) to think about it as well? (00:16:13) >> Yeah, that's right. And then and then (00:16:14) that's one that's that's one really (00:16:15) important question and observation and (00:16:17) and then by the way that also gets to (00:16:18) this other kind of question um where (00:16:20) there's different theories on it. Um (00:16:22) which is basically big models versus (00:16:23) small models. (00:16:24) >> Um and so a a lot of the data a lot of (00:16:26) the data center build is oriented around (00:16:28) hosting um training and and and and (00:16:30) serving the the big the big models, you (00:16:32) know, for for all the obvious reasons. (00:16:33) Um but there's also the small the small (00:16:36) model revolution is happening at the (00:16:37) same time and and and and if you just (00:16:38) kind of track you know you can get get (00:16:40) the various research firms have these (00:16:41) charts you can get um but if you just (00:16:43) kind of track the if you track the (00:16:44) capability of the leading edge models (00:16:46) over time what you find is after 6 or 12 (00:16:47) months there's a small model that's just (00:16:48) as capable um and so there there there's (00:16:51) this kind of chase function that's (00:16:53) happening which is the capabilities of (00:16:54) the big models are basically being (00:16:56) shrunk shrunk down uh and provided at at (00:16:58) at at smaller size and then therefore (00:17:00) smaller cost you know quite quickly. So, (00:17:02) I I'll just give you the the most recent (00:17:03) example that just got hit over the last (00:17:05) two weeks. And again, this is a thing (00:17:06) that's just kind of shocking. Um is (00:17:08) there's this Chinese company that has a (00:17:10) um well, I forget the name of the (00:17:11) company, but it's it's uh the company (00:17:12) that produces the model called Kimmy, (00:17:14) which is spelled Kim Mi, which is one of (00:17:16) the leading open source models out of (00:17:17) China. Um and uh the new version of (00:17:20) Kimmy is a reasoning model that is at (00:17:22) least according to the benchmark so far (00:17:23) is basically a replication of the (00:17:25) reasoning capabilities of GPT5, right? (00:17:28) and and and these new models of GPT5 (00:17:29) were a big advance over GPT4 and of (00:17:31) course GPT5 costs a tremendous amount of (00:17:33) money to to develop and to serve and all (00:17:35) of a sudden you know here we are (00:17:36) whatever 6 months later and you have an (00:17:37) open source model called Kimmy and I (00:17:39) think I don't know if they had it's (00:17:40) either shrunk down to be able to run on (00:17:42) either it's like one MacBook or two (00:17:43) MacBooks um right um and so you can all (00:17:46) of a sudden if you have like an applica (00:17:48) you if you're a business and you want to (00:17:49) have a reasoning model that's GPT5 (00:17:51) capable um but you you know you're (00:17:53) whatever you're not going to pay the (00:17:54) whatever GPT5 cost or you're not going (00:17:56) to want to have it be hosted and you (00:17:57) want to run it locally, um, you know, (00:17:59) you can do that. Um, and and and again, (00:18:01) that's just like another just it's just (00:18:03) like another, you know, it's another (00:18:04) breakthrough. Like it's just it's (00:18:05) another another Tuesday, another huge (00:18:07) advance. It's like, oh my god. And then (00:18:08) of course, it's like, all right, well, (00:18:09) what is OpenAI going to do? Well, (00:18:10) obviously they're going to go to GPT6, (00:18:12) right? Uh, and you know, right? And so (00:18:15) there there's this kind of lattering (00:18:16) that's happening where the entire (00:18:17) industry is moving forward. Um, the big (00:18:19) models are getting more capable. The (00:18:20) small models are kind of chasing them. (00:18:22) Um uh and then um and then the small (00:18:24) models provide you know completely (00:18:25) different way to deploy um you know at (00:18:27) at at at very low price points. Um and (00:18:30) so yeah I think and and you know we'll (00:18:32) we'll see what happens. I mean there (00:18:33) there are some very smart people in the (00:18:34) industry who think that ultimately (00:18:35) everything only runs in the big models (00:18:36) because obviously the big models are (00:18:38) always going to be the smartest and so (00:18:39) therefore you're always you know you're (00:18:40) always going to want the most (00:18:41) intelligent thing because why would you (00:18:42) ever want something that's not the most (00:18:43) intelligent thing for any application. (00:18:45) You know the counterargument is just (00:18:46) there's a huge number of tasks that take (00:18:48) place in the economy and in the world (00:18:49) that don't require Einstein. you know, (00:18:51) where, you know, where, you know, 120 IQ (00:18:53) person is great. You don't need a, you (00:18:55) know, 160 IQ, you know, PhD in, you (00:18:57) know, string theory. You just like have (00:18:58) somebody who's competent and capable and (00:19:00) it's great. Um, and so, you know, I I, (00:19:02) you know, and I we've talked about this (00:19:04) before. I tend to think the AI industry (00:19:06) is going to be structured a lot like the (00:19:07) computer industry ended up getting (00:19:08) structured, which is you're going to (00:19:09) have a small handful of basically the (00:19:11) equivalent of supercomputers, which are (00:19:13) these like giant, you know, kind of we (00:19:14) call god models that are, you know, (00:19:16) running in these giant data centers. Um (00:19:18) and then and then you know I I I I I'm (00:19:20) not like convinced on this but my my (00:19:22) kind of working assumption is what (00:19:23) happens is then you have this cascade (00:19:24) down of smaller models all ultimately (00:19:27) all the way the very small models that (00:19:28) run on embedded systems right run on run (00:19:30) on individual chips inside every you (00:19:31) know physical item in the world. Um and (00:19:34) that you know the smartest models will (00:19:35) always be at the top but the volume of (00:19:37) models will actually be the smaller (00:19:38) models that proliferate out and right (00:19:40) that's what happened with microchips. uh (00:19:41) it's what happened with computers which (00:19:43) became microchips and then it's what (00:19:44) happened with operating systems and with (00:19:46) with a lot of everything else that we (00:19:47) built in software. Um so you know I tend (00:19:49) to think that's what will happen. Um (00:19:51) just quickly on the chip side um again (00:19:54) like chips you if you look at the entire (00:19:56) history of the chip industry uh uh (00:19:58) shortages become gluts um and you get (00:20:01) just you know like anytime there's a (00:20:03) giant profit pool in a in a new chip (00:20:05) category um you know somebody has a lead (00:20:07) for a while and kind of gets you know um (00:20:09) let's say the the the profits (00:20:11) appropriate to what we u what we call (00:20:12) robust market share um but in time what (00:20:15) happens right is that that draws (00:20:17) competition and of course you know that (00:20:18) that that's happening right now. So (00:20:20) Nvidia's, you know, Nvidia is an (00:20:21) absolutely fantastic company, fully (00:20:22) deserves the position that they're in, (00:20:24) fully deserves the profits that they're (00:20:25) generating, but they're now so valuable, (00:20:27) generating so many profits that it's the (00:20:28) bat signal of all time to the rest of (00:20:29) the chip industry to figure out how to (00:20:31) advance the state-of-the-art AI chips. (00:20:33) Um, and that's, by the way, and that's (00:20:34) already happening, right? And so you've (00:20:35) got other major companies like AMD (00:20:37) coming at them, and then you've got (00:20:38) really significantly, you've got the (00:20:39) hyperscalers building their own chips. (00:20:41) Um, and so, you know, a bunch of the big (00:20:43) a bunch of those kind of big tech (00:20:44) companies are building their own ships. (00:20:45) Um, and of course then the Chinese are (00:20:47) building their own ships as well. Um and (00:20:48) so it's just it's like pretty likely in (00:20:50) 5 years that that you know AI chips will (00:20:52) be you know cheap and plentiful at least (00:20:54) in comparison to the situation today. Uh (00:20:56) which again I think will you know will (00:20:58) tend to be extremely positive for the (00:20:59) economics of of the kinds of companies (00:21:01) that we invest in. (00:21:02) >> Yep. And that startups are also starting (00:21:04) to go after new chip design as well (00:21:06) which is exciting. (00:21:07) >> Yeah. Well, that's the other thing is (00:21:08) yeah, you have these disruptive startups (00:21:09) and actually that just for a moment on (00:21:11) the chips, they were not really big (00:21:12) investors in chips because it's kind of (00:21:13) a big it's kind of a big company thing, (00:21:14) but um it's a little bit of historical (00:21:17) happen stance that AI is running on (00:21:18) quote unquote GPUs um you know which GPU (00:21:21) stands for graphical processing unit. So (00:21:23) um and basically just for people who (00:21:25) haven't tracked this there were (00:21:26) basically two kinds of chips that made (00:21:28) the personal computer happen. the (00:21:29) so-called CPU central processing unit (00:21:31) which classically was the Intel x86 x86 (00:21:34) chip that's kind of the brain of the (00:21:35) computer and then there was this other (00:21:36) kind of chip called the GPU or graphical (00:21:38) processing unit that was the sort of (00:21:40) second chip in every PC that does all (00:21:42) the graphics um and you know and this is (00:21:44) graphics you know 3D graphics for gaming (00:21:46) or for CAD CAM or for you know anything (00:21:47) else you know Photoshop or for anything (00:21:49) that involves you know lots of visuals (00:21:51) and so the the kind of canonical (00:21:53) architecture for a personal computer was (00:21:55) a CPU and a GPU by the way same thing (00:21:57) for smartphones um but by the way. And (00:21:59) over time, you know, these have kind of (00:22:00) merged and so like a lot of CPUs now (00:22:02) have GPU capability built in. Actually, (00:22:03) a lot of GPUs now have CPU capability (00:22:05) built in. So this, you know, this has (00:22:06) gotten fuzzy over time, but like that (00:22:08) that was like the classic breakdown. But (00:22:09) the fact that that was the classic (00:22:11) breakdown, you know, kind of meant that (00:22:12) while Intel had a you know, monopoly for (00:22:14) a long time on CPUs, um there was this (00:22:16) other market of GPUs which Nvidia um you (00:22:19) know basically fought the GPU wars for (00:22:21) 30 years and and and came out the winner (00:22:23) like what was the best company in the (00:22:24) space. But it was like a hyper (00:22:25) competitive market for graphics (00:22:27) processors. it was actually not that (00:22:28) high margin and it was actually not that (00:22:29) big. And then basically it just it (00:22:32) turned out that there were two other um (00:22:34) forms of computation that were (00:22:36) incredibly valuable that happened to be (00:22:38) massively parallel uh in how they (00:22:40) operate which which happened to be very (00:22:41) good fits for the GPU architecture. And (00:22:44) those two basically highly lucrative (00:22:46) additional applications were (00:22:47) cryptocurrency starting about you know (00:22:48) 15 years ago and then AI starting about (00:22:51) you know whatever four years ago. Um, (00:22:53) and so and and Nvidia like I would say (00:22:56) very cleverly set itself up with an (00:22:58) architecture that works very well for (00:23:00) this, but it's also just a little bit of (00:23:02) a twist of fate that it just turns out (00:23:03) that if AI is the killer app, it just (00:23:05) turns out that the GPU architecture is (00:23:06) the best legacy architecture is devoted (00:23:08) to it. And I go through that to say like (00:23:10) if you were designing AI chips from (00:23:12) scratch today, you wouldn't build a full (00:23:13) GPU. you would build dedicated AI chips (00:23:15) that were much more much more (00:23:17) specifically adapted to AI um and would (00:23:19) have I I think would just be much more (00:23:21) economically efficient and you know John (00:23:23) to your point there there there are (00:23:24) startups that are actually building (00:23:25) entirely new kinds of chips uh oriented (00:23:28) specifically for AI and you know we'll (00:23:30) have to see what happens there you know (00:23:31) it's hard to build a new chip company (00:23:33) from scratch um you know it's possible (00:23:34) that one or more of those startups makes (00:23:36) it on their own um and some of them are (00:23:38) you know doing very well um it's also (00:23:39) possible of course that they get bought (00:23:41) um you know by big companies that that (00:23:43) have the ability to scale them. Um, and (00:23:45) so, you know, you know, we'll see (00:23:47) exactly how that unfolds. Um, and of (00:23:49) course, we'll also, by the way, see, you (00:23:51) know, the Koreans are going to play here (00:23:52) for sure. Um, uh, the Japanese are going (00:23:54) to play. Um, and then, you know, the (00:23:56) Chinese in a major way, uh, as well. (00:23:58) And, you know, they have their own, you (00:23:59) know, native chip ecosystem that they're (00:24:01) that they're building up. And so there (00:24:02) there there there are going to be many (00:24:04) choices of AI chips in the future. Um, (00:24:06) and it's going to be that, you know, (00:24:07) that'll be a giant battle that'll be a (00:24:09) giant battle that we observe very (00:24:10) carefully. um and that we uh make sure (00:24:12) that our our companies basically are (00:24:14) able to take full advantage of. (00:24:16) >> While while on the topic of of (00:24:17) international um we you mentioned Kimmy (00:24:20) earlier. So it seems like some of the (00:24:21) best open source models today are from (00:24:23) China. Should this be worrisome to to (00:24:25) folks? How are you thinking and talking (00:24:28) about this topic with with folks in DC? (00:24:30) I know you were just there last week. (00:24:32) How much of this is a concern for uh US (00:24:35) companies particularly just having seen (00:24:37) the rise of China do unnatural things in (00:24:41) solar markets, car markets? Um are they (00:24:44) kind of flooding the ecosystem so that (00:24:46) they can eventually kind of take share (00:24:47) and and increasingly uh own the the (00:24:50) ecosystem? (00:24:51) >> Yeah. So uh you know a couple things. So (00:24:53) one is you know you know you want to (00:24:55) start these discussions by just kind of (00:24:56) saying like you know look there's (00:24:57) there's vigorous debate in in the US and (00:24:59) around the world of look like you know (00:25:00) how much are we in a new cold war with (00:25:01) China you know and exactly like how (00:25:03) hostile you know should should we view (00:25:05) them and it you know it's very tempting (00:25:06) by the way it's very tempting and I (00:25:08) think it's a very good case made that (00:25:09) we're in like a new cold war that's like (00:25:11) that in a lot of ways is like the US (00:25:12) versus USSR um in the in the 20th (00:25:14) century um you know it is the counter (00:25:17) argument would be it is more complicated (00:25:18) than that because the US and the USSR (00:25:20) were never really intertwined from a (00:25:22) trade standpoint Um and and a big part (00:25:23) of that quite frankly was the USSR never (00:25:26) really made anything that anybody else (00:25:28) needed I guess other than weapons. Um (00:25:30) but like you know the USSR's primary (00:25:32) exports were literally like you know (00:25:33) literally like wheat and and oil. Um (00:25:36) whereas of course China exports just a (00:25:38) tremendous number of physical things (00:25:41) right um including like a huge part of (00:25:43) like the entire supply chain of parts (00:25:45) that basically go into everything that (00:25:46) American manufacturers you know kind of (00:25:48) make right and so by the time a US you (00:25:50) know whatever by the time an American (00:25:52) company brings a toy to market right or (00:25:53) a uh you know or a car um or anything or (00:25:56) a computer or a smartphone or whatever (00:25:58) like it's got a lot of componentry in it (00:25:59) that was made in China so there so there (00:26:01) is a much tighter in interlinkage (00:26:03) between the the American and Chinese (00:26:04) economies than there as the American and (00:26:06) Soviet economies and you know may maybe (00:26:08) you know Adam Smith or whatever might (00:26:10) say you know that's good news for peace (00:26:11) and that you know both countries need (00:26:12) each other by the way the other part of (00:26:14) that argument is that the Chinese (00:26:15) basically the Chinese you know the (00:26:17) Chinese governance model is based on (00:26:18) high employment um you know because you (00:26:21) know if if if you know at least all the (00:26:22) geopolitical people say if China ended (00:26:24) up with like 25 or 50% unemployment that (00:26:26) would cause civil unrest which is the (00:26:27) one thing that the CCP doesn't want and (00:26:29) so the corresponding part of the trade (00:26:31) pressure is China needs the American (00:26:32) export market you know the American (00:26:34) consumer is like a third of the global (00:26:35) economy. Uh a third of global consumer (00:26:37) demand. Um and so you know China needs (00:26:40) the US export market or it has high all (00:26:42) of a sudden a lot of its factories would (00:26:43) go kind of instantly bankrupt and you (00:26:44) know would cause mass unemployment and (00:26:46) unrest in China. So so anyway like you (00:26:48) know we there is this complicated it's a (00:26:49) it's a complicated intertwined um (00:26:52) relationship. Um having said that you (00:26:54) know the the mood in DC basically for (00:26:56) the last 10 years on a bipartisan basis (00:26:58) um has been that we need to take we the (00:27:01) US need to take China more seriously as (00:27:02) a geopolitical foe. And you know under (00:27:05) under under that school of thought (00:27:06) there's sort of the sort of you know (00:27:08) there's there's the military dimension (00:27:09) which is you know the sort of the you (00:27:11) know the the risk of some kind of war in (00:27:13) the South China Sea the risk of some (00:27:14) kind of war around around Taiwan and so (00:27:16) that you know that that has everybody in (00:27:18) Washington on high alert um you know (00:27:20) there's also this this economic question (00:27:21) around the kind of de-industrialization (00:27:23) of the US potential re-industrialization (00:27:25) and what that means about you know (00:27:26) dependence on China and then and then (00:27:28) there's and then there's this this this (00:27:30) AI question um and and the AI question (00:27:32) is an economic question but It's also (00:27:34) like a geopolitical question which is (00:27:36) okay you know basically AI is (00:27:37) essentially only being built in the US (00:27:39) and in China. Um you know the rest of (00:27:41) the world either you know can't build it (00:27:43) or doesn't want to which which we could (00:27:45) talk about. So it's basically US versus (00:27:46) China. Um and then AI is going to (00:27:49) proliferate all over the world and is it (00:27:50) going to be American AI that (00:27:51) proliferates all over the world or is it (00:27:53) going to be Chinese AI that proliferates (00:27:54) all over the world and so and I was (00:27:56) saying just generally across party lines (00:27:58) in DC this you know the the things I (00:28:00) just went through are kind of how they (00:28:02) look at it. Um and and the Chinese are (00:28:05) in the game and so the you know the (00:28:06) Chinese are in the game for sure you (00:28:07) know with software u you know deepseek (00:28:09) you know was kind of the big you know (00:28:10) kind of fire the starting gun in the (00:28:12) software race and now you've got I think (00:28:13) it's I think you've got four it's like (00:28:15) deepsek uh which is a deep so deepseek (00:28:18) is an AI model from actually a hedge (00:28:20) fund um in uh in China um it's a little (00:28:23) bit uh kind of took a lot of people by (00:28:24) surprise um then Quen is the model from (00:28:26) Alibaba. Kimmy is from another startup. (00:28:28) Oh, called Moonshot. The company's (00:28:30) called Moonshot. Um, and then there's, (00:28:32) you know, and then, um, you know, (00:28:33) there's also Tencent and BU. Um, and, (00:28:36) um, by Dance, um, you know, that are all (00:28:38) primary, you know, companies doing a lot (00:28:39) of work in AI. Um, and so, you know, (00:28:41) there's somewhere between three to six, (00:28:42) you know, kind of primary AI companies. (00:28:44) And then there's, you know, tremendous (00:28:45) numbers of of startups. Um, and so, you (00:28:47) know, they're in the race on on, uh, you (00:28:49) know, they're in the race on on on (00:28:50) software. Um, they are, you know, (00:28:53) working to catch up on chips. They're (00:28:54) not there yet, but they're working (00:28:55) incredibly hard to catch up. And just as (00:28:57) an example of that, you know, the at (00:28:58) least the common understanding um you (00:29:00) know, in the US is that the reason you (00:29:02) haven't seen the new version of DeepSeek (00:29:03) yet is that basically the Chinese (00:29:05) government has instructed them to build (00:29:06) it only on Chinese chips um as a as a (00:29:08) motivator to get the Chinese chip (00:29:10) ecosystem up and running. Um and and (00:29:12) then the main chip company there is (00:29:13) Huawei, although there could be more in (00:29:15) the future. Um and then there's um so (00:29:18) you know, so so so there's that and then (00:29:19) and then there's everything to follow (00:29:20) which is basically AI in kind of robotic (00:29:23) form, right? And so there there's this (00:29:25) basically global technological economic (00:29:27) robotics competition that's kicking off. (00:29:29) Um and u you know China kind of starts (00:29:31) out ahead on robotics because they're (00:29:33) just ahead on so many of the so many of (00:29:35) the components that go into robots u (00:29:37) because the you know the sort of like I (00:29:39) said this the kind of entire supply (00:29:40) chain of like electromechanical things (00:29:42) you know basically moved from the US to (00:29:43) China 30 years ago and and has never (00:29:45) come back. So so so that's kind of the (00:29:47) the the the DC lens on it. Um and and I (00:29:50) would say you know DC is watching it uh (00:29:52) you know quite carefully. Um uh the the (00:29:55) the the big kind of supernova moment (00:29:57) this year was the deepseek release. The (00:29:58) deepseek release was surprising on a (00:30:00) number of fronts. Um one was just how (00:30:02) good it was and again along this line of (00:30:04) it took the capability set that were (00:30:06) running in large models in the cloud and (00:30:08) kind of shrunk it um onto a um you know (00:30:11) into into a uh into a a sort of a a (00:30:14) reduced size you know a smaller version (00:30:16) of sort of equivalent capabilities that (00:30:17) you could run on small amounts of local (00:30:18) hardware. Um and so there was that and (00:30:21) then it was also a surprise that it was (00:30:22) released as open source uh and (00:30:24) particularly open source from China (00:30:25) because China China does not have a long (00:30:27) history of open source. Um and then um (00:30:30) it was also a surprise um that it (00:30:32) actually came from a hedge fund. Um so (00:30:34) it didn't come from a big R&D you know (00:30:35) sort of university research lab. It (00:30:37) didn't come from a you know from a big (00:30:38) tech company. it it came from a hedge (00:30:40) fund and it it like as as far as we can (00:30:42) tell it it basically is this somewhat (00:30:44) idiosyncratic situation where you just (00:30:46) have this incredibly successful quant (00:30:47) hedge fund with all these you know super (00:30:49) geniuses um and the the founder of that (00:30:51) hedge fund you know basically decided to (00:30:52) build AI um and you know at least (00:30:55) external indications are this was a (00:30:56) surprise to even even the Chinese (00:30:57) government it's it's impossible to prove (00:30:59) you know what the Chinese government was (00:31:01) surprised by or not but you know there's (00:31:02) at least the atmospherics are that this (00:31:04) was not exactly planned this was not a (00:31:06) national champion tech company at the (00:31:07) time that Deepseek was released it was (00:31:09) it sort of came out of left field which (00:31:10) by the way is very encouraging for the (00:31:12) field that it was possible for somebody (00:31:13) to do that kind of who was unknown right (00:31:15) because it kind of means that maybe you (00:31:16) don't need all these you know super (00:31:17) genius superstar researchers maybe (00:31:19) actually smart kids can just build this (00:31:21) stuff which I think is is the direction (00:31:22) things are headed um and so that kicked (00:31:25) off I would say like this kind of I I (00:31:27) don't know copycat's the wrong word but (00:31:28) that that was sort of it feels like the (00:31:30) success of deepseek and the success of (00:31:32) deepseek from China as open source kind (00:31:33) of kicked off a sort of trend in China (00:31:36) releasing these open source models um (00:31:38) you know Look, the cynics, you know, in (00:31:40) DC would say, you know, yeah, like (00:31:42) they're dumping, right? The the they're (00:31:43) obviously dumping. They're trying to, (00:31:45) you know, they see that the West has (00:31:46) this opportunity to build this China (00:31:47) industry. You know, they're trying to (00:31:48) commoditize it right out of the gate. (00:31:49) You know, there's probably something to (00:31:51) that. Um, you know, the the Chinese (00:31:53) industrial economy does have a history (00:31:55) of, you know, sort of, let's say, (00:31:56) subsidized production that leads to (00:31:58) selling, you know, selling things below (00:31:59) cost in some cases. Um, but I think also (00:32:02) it's it like I think that's almost too (00:32:04) cynical of a view also because it's just (00:32:06) like all right wow like they're really (00:32:07) in the race like open source closed (00:32:08) source whatever like that you know (00:32:10) they're actually really in the race. Um, (00:32:12) you know, we we've talked in the past, I (00:32:13) think, on on on LP calls about, you (00:32:15) know, these policy fights that, you (00:32:17) know, we've been having in DC for the (00:32:18) last two years. And, you know, there was (00:32:19) a big pretty pretty big push within the (00:32:21) US government, you know, two years ago (00:32:22) to basically, you know, restrict, uh, (00:32:24) you know, or outright ban, you know, a (00:32:26) lot of AI. Um, and, you know, it's very (00:32:28) easy for a country that is the only game (00:32:30) in town to have those conversations. (00:32:32) It's quite another thing if you're (00:32:33) actually in a foot race with China. Um, (00:32:35) and so I think actually the the the the (00:32:37) policy landscape in DC has I would say (00:32:40) has improved dramatically as a (00:32:42) consequence of sort of an awareness now (00:32:43) that this is actually a two- horse race, (00:32:45) not a one-horse race. (00:32:46) >> For sure. Yeah. Actually on on the point (00:32:47) I'll I'll jump ahead here to policy and (00:32:49) regulation just because it seems like uh (00:32:51) the current stance on on 50 different (00:32:55) set of AI laws by state seems like a (00:32:57) catastrophic (00:32:59) uh way to to put us effectively with a (00:33:02) uh or one of our our hands tied behind (00:33:04) our our back here in terms of the the AI (00:33:07) race. What's a state of plan on that? (00:33:09) Are folks recognizing that that would be (00:33:11) catastrophic for progress and (00:33:13) development? Where do most people at (00:33:14) least stand on that topic today? (00:33:16) Yeah. So it's a little bit complicated. (00:33:18) So I'll rewind to say like two years ago (00:33:20) I was very worried about like really (00:33:21) ruinous federal federal legislation on (00:33:23) AI and there was there was we you know (00:33:25) we engaged you know kind of very heavily (00:33:26) at that point which we've talked about (00:33:27) in the past and I think the good news on (00:33:29) that is I think the risk of that sitting (00:33:30) here today is very low. Um I there's (00:33:32) very little mood in DC on either side of (00:33:34) the aisle uh to really you know (00:33:37) essentially there's very little there's (00:33:38) very little interest in doing anything (00:33:39) that would prevent us from beating (00:33:41) China. Um so so you know on the federal (00:33:45) side things things are much better now. (00:33:46) There there will there will be issues (00:33:47) and there are tensions in the system but (00:33:49) like things are looking looking pretty (00:33:50) good. Um that has translated Jen to your (00:33:54) point that's translated a lot of the (00:33:55) attention to the states and basically (00:33:56) what's happened is you know under our (00:33:57) system of of federalism uh you know the (00:33:59) states get to pass their own laws on a (00:34:01) lot of things. Um and so uh yeah, (00:34:03) basically you know a lot of you know and (00:34:05) and you know with these things it's (00:34:06) always a combination. A lot of (00:34:07) well-meaning people are trying to figure (00:34:08) out what to do at the state level and (00:34:09) then of course there's a lot of (00:34:10) opportunism where AI is just the hot (00:34:12) topic. And so if you're a you know (00:34:14) aggressive up and cominging state (00:34:15) legislator or whatever in some state and (00:34:16) you want to run for governor and then (00:34:17) president you know you want to kind of (00:34:18) attach yourself to the heat. Um and so (00:34:20) there's like a political motivation to (00:34:22) to do state level stuff. Um yeah and (00:34:24) sitting here today like we're tracking (00:34:26) on the order of,200 bills across the 50 (00:34:27) states. And by the way, um, not just the (00:34:30) blue states, also the red states. Um, (00:34:32) and so, you know, I'm I've, you know, (00:34:33) for the last like 5 years or whatever, I (00:34:35) spent a lot of time complaining about, (00:34:36) uh, you know, kind of what Democratic (00:34:37) politicians are threatening to do to (00:34:38) attack. There's also a lot of (00:34:40) Republicans, like Republicans are not a (00:34:42) block on this. And there are quite a few (00:34:43) like local Republican officials in (00:34:45) different states, um, that that also, I (00:34:47) think, have, you know, let's say, you (00:34:48) know, misinformed or ill-advised, um, (00:34:50) views and are trying to put together, (00:34:51) uh, put out bad bills. um you know it's (00:34:56) a little bit weird that this is (00:34:57) happening and that you know the federal (00:34:58) government does have regulation of (00:34:59) interstate commerce um and you know (00:35:02) technology AI kind of by definition is (00:35:04) interstate like you know there's there's (00:35:06) no AI company that just operates in (00:35:07) California or just operates in you know (00:35:10) Colorado or Texas um you know AI of all (00:35:13) technologies AI is obviously something (00:35:15) this this sort of national in scope um (00:35:17) you know it's sort it's sort of obvious (00:35:18) that the federal government should be (00:35:19) the regulator not not not the states um (00:35:22) but but the federal government need (00:35:23) needs to assert itself needs to step in. (00:35:25) There there was actually an attempt to (00:35:26) do that. There was a um there was an (00:35:29) attempt to add a moratorium of state (00:35:30) level AI regulation that basically would (00:35:33) would reserve the right of the federal (00:35:34) government to regulate AI and sort of (00:35:35) prevent the states from moving forward (00:35:37) with these bills. That was I think part (00:35:38) of the negotiation for the quote one big (00:35:40) beautiful bill and then that that there (00:35:42) was a deal behind that and that deal (00:35:43) kind of blew up at the at the last (00:35:45) minute and that moratorium didn't happen (00:35:47) and and you know in fairness the critics (00:35:48) of that moratorum it probably was a was (00:35:51) it was probably too much of a stretch. (00:35:52) Well, it was I'm sorry. It was (00:35:53) definitely too much of a stretch to get (00:35:54) enough support to pass, but it was also (00:35:56) probably too much of a stretch in terms (00:35:57) of restricting the states from certain (00:35:59) kinds of regulation that they really (00:36:00) should be able to do. So, so it just it (00:36:02) didn't quite come together. Um, there's (00:36:03) a very active we're having very active (00:36:05) discussions in DC right now about kind (00:36:06) of the next, you know, the kind of the (00:36:08) next turn on that. Um, you know, the (00:36:10) administration is I would say the (00:36:11) administration is very supportive of of (00:36:12) the idea of of the federal government (00:36:14) being in charge of this as part of it (00:36:15) being an actual, you know, 50-st state (00:36:18) issue. Um, and and and an issue of (00:36:20) national importance. Um, and then, you (00:36:22) know, I'd say most most Congress people (00:36:24) on both sides of the aisle, you know, (00:36:25) kind of get this. Um, so we just we we (00:36:27) kind of have to figure out a way to, you (00:36:29) know, to land this, but but I think (00:36:30) that'll happen. Um, some of the state (00:36:32) level bills are wild. Um, the the (00:36:35) Colorado passed a very draconian uh (00:36:38) regulation bill uh last year. Um, and (00:36:41) against like furious objections from the (00:36:43) local startup ecosystem in in in around (00:36:45) Denver and Boulder. Um, and actually (00:36:47) they're they're now actually trying to (00:36:49) reverse their way out of that bill. um (00:36:50) you know a year later some of the the (00:36:52) nuance of it like the algorithmic (00:36:54) discrimination and like how to mitigate (00:36:55) like what were some of the the extreme (00:36:57) versions of what they they had proposed. (00:36:59) >> Yeah. So the really draconian one was (00:37:01) the the one that we really fought hard (00:37:02) was the one in California which was (00:37:04) called SB1047 and it wasn't it it it was (00:37:06) basically it was modeled basically after (00:37:08) the was called the EU AI act. So the (00:37:10) European Union's AI act. Okay. And this (00:37:12) is the backdrop to all the US stuff (00:37:13) which is the EU passed this bill called (00:37:15) the AI act I don't know whatever two (00:37:16) years ago and it basically has killed AI (00:37:18) development in well it's actually killed (00:37:20) AI development in Europe to a large (00:37:21) extent. Um and then it even it it's so (00:37:24) draconian that even even big American (00:37:27) companies like Apple and Meta are not (00:37:28) launching leading edge AI capabilities (00:37:30) in their products in Europe. Like that (00:37:32) that's how that's how like draconian (00:37:33) that bill was. And it's it's sort of a (00:37:34) classic it's a classic kind of European (00:37:37) thing where they like you know like they (00:37:39) just thought that you know they they (00:37:40) have this kind of view that it's just (00:37:41) like well you know we if we can't be the (00:37:43) leader they literally say this by the (00:37:44) way if we can't be the leaders in (00:37:45) innovation at least we can be the (00:37:46) leaders in regulation. Um and and and (00:37:49) then they pass this like incredibly you (00:37:50) know kind of ruinous uh selfharm you (00:37:53) know kind of thing and then you know a (00:37:55) few years pass and they're like oh my (00:37:56) god what have we done and so they're you (00:37:57) know they're kind of going through their (00:37:58) own version of that. Um, by the way, you (00:38:01) know, I I you know, when I talk about (00:38:03) Europe, I I tend to be very dark about (00:38:04) the whole thing. I will tell you the (00:38:05) darkest people I know about Europe are (00:38:07) the European entrepreneurs who moved to (00:38:08) the US. Um, are just like absolutely (00:38:11) furious about what's happening in in in (00:38:13) Europe on this stuff. Um, but but even (00:38:15) there, like it it's it's so bad in (00:38:17) Europe, like they they shot themselves (00:38:18) in the foot so badly that there's (00:38:19) actually a process now at the at the EU (00:38:21) to try to unwind that. They're trying to (00:38:22) unwind the GDPR. So u anyway for people (00:38:25) tracking Europe uh Mario Draghi um is (00:38:28) the former I guess prime minister of (00:38:29) Italy did this thing about a year ago (00:38:30) called the Draghy report which is the (00:38:32) report on European competitiveness and (00:38:34) he kind of outlined kind of in great (00:38:35) detail all the ways that Europe was (00:38:36) holding itself back and part of it was (00:38:38) overregulation areas like AI. So so (00:38:39) they're trying to reverse out of that or (00:38:42) making gestures you know we'll we'll see (00:38:43) what happens. Um (00:38:46) it in the middle of all that, California (00:38:48) sort of inexplicably decided to (00:38:50) basically copycat the EU AI act and try (00:38:52) to apply it to California. Um which (00:38:54) might strike you as completely insane. (00:38:55) To which I would say yes, welcome to (00:38:57) California. Um uh and um you know, it (00:39:00) was this basically this like Sacramento (00:39:01) political dynamic that kind of got got (00:39:04) crazy. Um it would have you know (00:39:06) completely killed you know AI (00:39:07) development in California. Um (00:39:09) unfortunately our governor vetoed it at (00:39:11) the last minute. Um it did pass both (00:39:12) houses legislature that he vetoed at the (00:39:14) last minute. Um it to Jen to your point (00:39:16) it would have done for it would have (00:39:18) done a whole bunch of things that were (00:39:19) ruinously uh bad. But one of the things (00:39:21) it would have done is it would have (00:39:22) assigned downstream liability um uh to (00:39:25) open source developers. Um and so you (00:39:28) know we talked about you know this (00:39:29) Chinese open source thing. Okay so you (00:39:30) got Chinese out there with open source. (00:39:31) Now you're gonna have American companies (00:39:32) that have open source AI. And by the way (00:39:34) you're also going to have American (00:39:35) academics and just like independent (00:39:37) people in their nights and weekends (00:39:38) developing open source. um you know (00:39:40) which is a key way that all this (00:39:41) technology proliferates and and so this (00:39:43) this law would have assigned downstream (00:39:45) liability to any misuse of open source (00:39:47) to the original developer of the open (00:39:48) source and so you know you're an (00:39:50) independent developer or you're an (00:39:51) academic or you're a startup you develop (00:39:53) and release an AI model the AI model (00:39:55) works fine the day you release it it's (00:39:57) great but like 5 years later it gets (00:39:58) built into a nuclear power plant and (00:40:00) then there's a meltdown of the nuclear (00:40:01) power plant and then somebody says oh (00:40:03) it's the fault of the AI um the the the (00:40:05) the legal liability for that nuclear (00:40:08) meltdown or for anything any other (00:40:10) practical real world thing that would (00:40:11) follow in the out years would then be (00:40:13) assigned back to that open source (00:40:14) developer. Of course, this is completely (00:40:15) insane. It would completely kill open (00:40:17) source. It would completely kill (00:40:19) startups doing open source. It would (00:40:20) completely kill academic research like (00:40:21) in its entirety. Um, you know, anything (00:40:23) in the field. Um, and so, you know, that (00:40:26) like that's the level of playing with (00:40:27) fire. Um, you know, kind of that these (00:40:29) state level politicians have become (00:40:30) enamored with. Um, like I said, I think (00:40:33) the good news is the feds understand (00:40:34) this. I suspect that this is going to (00:40:35) get resolved, but it but it does need to (00:40:37) get resolved because, you know, just as (00:40:39) a country, it just doesn't make any (00:40:40) sense to let let the states kind of (00:40:41) operate suicidally like this. Um, and so (00:40:45) that's what we're doing. You know, we we (00:40:46) talk about this, we call this our little (00:40:47) tech agenda. Um, we're extremely focused (00:40:49) on on on the freedom and starters (00:40:51) innovate. We are not trying to argue, (00:40:53) you know, many many other issues. We (00:40:55) operate in a completely bipartisan (00:40:57) fashion. We have extensive um support, (00:40:59) you know, on both sides of the aisle and (00:41:00) for both sides of the aisle. Um, so it's (00:41:02) it's a truly bipartisan effort. um very (00:41:04) policy based and you know I think very (00:41:06) much aligned with the interests of the (00:41:07) country uh broadly um and so that is (00:41:11) what we're doing and then and then the (00:41:12) other question we get we we get actually (00:41:14) you know in some cases from LP but in a (00:41:15) lot of cases actually from employees um (00:41:17) is like okay why us right like you know (00:41:20) you know with with any sort of you know (00:41:23) policy question like this there's always (00:41:24) this collective action question which is (00:41:25) just like you know tragedy of the (00:41:27) commons which is in theory like (00:41:28) everybody every venture firm every tech (00:41:30) company whatever should be weighing in (00:41:31) on these things in practice what happens (00:41:32) is mo most them just simply don't. Um, (00:41:35) and so at some point it falls on (00:41:36) somebody's shoulders to fight these (00:41:38) things. And we we we Ben and I just (00:41:39) basically concluded that the stakes here (00:41:41) were just way too high. You know, if if (00:41:43) we're going to be the industry leader, (00:41:44) we just have to take responsibility for (00:41:46) our own destiny. You know, for better or (00:41:48) for worse, I think that's the cost of (00:41:49) doing business uh for being the leader (00:41:51) in the field right now. (00:41:52) >> Before we get off the topic of of AI, I (00:41:54) want to go back to one question that (00:41:55) that was submitted in. So, do you think (00:41:57) usage based or utility is a right way to (00:41:59) price an AI compared to seeds? (00:42:02) Ah that is a fantastic question. So this (00:42:04) is one of these giant this is in my my (00:42:05) list of what I call the trillion dollar (00:42:06) questions u where you know depending on (00:42:08) how this is answered will drive you know (00:42:10) trillions of dollars of market value. So (00:42:11) yeah so usage based pricing it's it's (00:42:14) actually (00:42:15) it's actually fairly amazing if you (00:42:17) think about this from a startup (00:42:18) standpoint from a venture standpoint (00:42:19) it's actually fairly amazing what's (00:42:20) happened and I'm trying I'm not really (00:42:22) talking about this in public because I (00:42:23) don't really I because I don't want it (00:42:24) to stop. I think it's actually quite (00:42:26) amazing. Um, which is you have these (00:42:29) technology companies, you know, these (00:42:30) big tech companies with these like (00:42:31) incredible R&D capabilities that are (00:42:33) building these big models, these big AI (00:42:34) models with this incredible, you know, (00:42:36) new new kind of new new kind of (00:42:37) intelligence. And then it it turns out (00:42:39) that they were already in a war. They (00:42:41) were already in the cloud war, right? (00:42:43) And so they were already in the war for (00:42:44) kind of cloud services. And this is like (00:42:45) AWS versus Azure versus uh Google Cloud. (00:42:49) Um, you know, and then all the all these (00:42:51) other all these other cloud efforts. And (00:42:52) so what what what what actually happened (00:42:53) was they sort of like there's an (00:42:56) alternate universe in which they (00:42:57) basically just kept all of their magic (00:42:59) AI secret and captive and just used it (00:43:01) in their own business um or used it to (00:43:03) just compete with more companies um you (00:43:05) know in more in more categories but (00:43:07) instead what they've done is they've (00:43:08) basically you know if I commod (00:43:10) commoditize is too strong a word but (00:43:12) they they have they have proliferated (00:43:14) their magic new technology through their (00:43:15) cloud business um which is which is this (00:43:18) business that just has these like (00:43:19) incredible scale you know kind of kind (00:43:21) of components to But um you know and (00:43:22) sort of this hyper competition between (00:43:24) the providers and these you know these (00:43:25) these prices that that come down very (00:43:27) fast. Um, and so you've got like the (00:43:29) most magic new technology in the world (00:43:30) and then it's basically being served up (00:43:31) by those companies in in in a as a cloud (00:43:34) business and made made basically (00:43:36) available to everybody on the planet to (00:43:37) just click and use and for like (00:43:39) relatively small amounts of money and (00:43:41) then on on a usage basis which means and (00:43:43) usage is great for startups because you (00:43:44) it means you can start easily right you (00:43:46) the the the you know there's very you (00:43:47) know there's basically no fixed co for a (00:43:49) startup building an AI app they don't (00:43:51) have giant fixed cost because they could (00:43:52) just tap into the open AI or anthropic (00:43:54) or Google or Microsoft or whatever you (00:43:55) know cloud you know tokens by the you (00:43:57) know, intelligence tokens by the drink (00:43:58) offering and just get going. Um, and so (00:44:00) it's it's kind of this this from this (00:44:02) from the startup standpoint, it's like (00:44:04) this marvelous thing where like the most (00:44:05) magical thing in the world is available (00:44:06) by the drink. You know, it's absolutely (00:44:08) amazing. Um, uh, I, you know, and, you (00:44:11) know, that model, you know, by the way, (00:44:13) that model's working and those companies (00:44:14) are happy and they're growing really (00:44:15) fast and they're, you know, happily (00:44:16) reporting massive cloud revenue growth (00:44:17) and, you know, they they're happy with (00:44:19) the margins and so forth and so, you (00:44:20) know, I think generally it's working. (00:44:22) Um, and those businesses are, I think, (00:44:24) likely to get much larger. Um and so I (00:44:25) think you know generally that's going to (00:44:27) work but but to to to the question like (00:44:29) that doesn't mean that the optimal (00:44:30) pricing model for for example all of the (00:44:32) applications should be tokens by the (00:44:34) drink and in fact very much I think not (00:44:36) the case. Um you know we spend a lot of (00:44:38) time working we actually have you know (00:44:39) dedicated you know experts on on pricing (00:44:41) in our firm. We spend a lot of time with (00:44:43) our companies working on pricing because (00:44:45) it's you know it's really this magical (00:44:46) art and science that that a lot of (00:44:47) companies don't take don't take (00:44:48) seriously enough. So we spend a lot of (00:44:50) time with other companies on this. And (00:44:51) of course, you know, a core principle of (00:44:53) pricing is you don't want to price by (00:44:54) cost if you can avoid it. You want to (00:44:56) price by value, right? Like you want to (00:44:58) price you price where you're getting a (00:45:00) percentage of the business value um of, (00:45:02) you know, especially when you're selling (00:45:03) two businesses, you want to price as a (00:45:04) percentage of the business value that (00:45:06) you're getting. And so so you do have (00:45:08) some AI startups that are that are (00:45:09) pricing by the drink for certain things (00:45:11) that they're doing, but you have many (00:45:12) others that are exploring other pricing (00:45:14) models. uh you know some that are just (00:45:16) like replications of SAS pricing models (00:45:17) but you also have other companies are (00:45:18) explor exploring pricing models for (00:45:20) example of well if the AI can actually (00:45:22) do the job of a coder or the AI could do (00:45:25) the job of a doctor or a nurse or a (00:45:27) radiologist or a lawyer or a parallegal (00:45:30) right or whatever or a teacher. Um you (00:45:32) know basically can you can could can you (00:45:34) price by value and can you get a (00:45:36) percentage of the value of what of what (00:45:38) of of of what otherwise would would (00:45:39) would have been you know would have been (00:45:40) literally a person. um you know or or by (00:45:43) the way equivalently can you price by (00:45:44) marginal productivity. So if you can (00:45:46) take a human doctor and make them much (00:45:47) more productive because you give them (00:45:48) AI, you know, can you price as a (00:45:50) percentage of kind of the productivity (00:45:51) uplift, uh, you know, from the from from (00:45:53) the from the augment, you know, the comb (00:45:55) symbiotic relationship between the the (00:45:57) human being and and the AI. Um, and so I (00:45:59) I think what we see in startup land is (00:46:01) like a lot of experimentation happening (00:46:03) on on these pricing models. And I and I (00:46:04) and I think again I I think that's like (00:46:06) super healthy. Um, I I you know, I was (00:46:09) in this little speech on this is like (00:46:10) high prices are really underappreciated. (00:46:12) High prices are often a favorite of the (00:46:13) customer. It's actually really funny. A (00:46:15) lot of like the naive view on pricing is (00:46:17) the lower the price, the better it is (00:46:18) for the customer. The the more (00:46:19) sophisticated looking at it is higher (00:46:20) prices are often good for the customer (00:46:21) because a higher price means that the (00:46:23) vendor can make the product better (00:46:24) faster, right? Like you can actually (00:46:27) companies with higher prices, higher (00:46:28) margins can actually invest more in R&D (00:46:30) and they can actually make the product (00:46:31) better. Um and you know most people who (00:46:33) buy things aren't just looking for the (00:46:35) cheapest price. They want something (00:46:36) that's really that's going to work (00:46:37) really well. Um and so often high (00:46:39) prices, you know, the customer doesn't (00:46:41) ever say this. it'll never show up in a (00:46:43) survey. Um, but but the high price can (00:46:45) actually be a gift for the customer (00:46:46) because it can make the vendor better, (00:46:48) can make the product better, and (00:46:49) ultimately make the customer better off. (00:46:50) And so I I'm I'm very encouraged by the (00:46:52) degree to which the AI entrepreneurs are (00:46:54) willing to run these experiments. And I, (00:46:55) you know, we'll have to see where it (00:46:56) pans out. But at least so far, I feel I (00:46:58) feel good about the the uh, you know, at (00:46:59) least the attitude of the industry about (00:47:01) it. (00:47:01) >> Awesome. I actually uh I was, as you (00:47:03) were gone through, I had probably 10 (00:47:04) more follow-up questions, but I'm (00:47:05) actually going to go back to um a topic (00:47:08) you had uh briefly, the trillion dollar (00:47:10) questions. Will open source or close (00:47:12) source win? Feels like we we've come out (00:47:15) on this this debate or where do you (00:47:16) where do you put that? (00:47:18) >> No, I think this is still open. I I (00:47:19) think this is still very open. Um you (00:47:21) know that like the the the closed source (00:47:23) models keep getting better. Um uh by the (00:47:26) way if you generally if you just like (00:47:27) take the temperature of the people (00:47:29) working at the big labs who work on the (00:47:30) big proprietary models like generally (00:47:32) what they'll tell you is progress is (00:47:33) continuing at a very rapid pace. Um you (00:47:36) know there's there's this you know (00:47:36) there's this periodic concern that kind (00:47:38) of shows up on online which is or in the (00:47:40) in the market which is like you know (00:47:41) maybe the capabilities these models are (00:47:42) topping out um and you know there's (00:47:44) certain there's there's certain areas in (00:47:45) which you know there's there's you know (00:47:46) people are working but like the people (00:47:48) working at the big labs are like oh no (00:47:50) we have like 800 new idea like we have (00:47:52) tons of new ideas we have tons of new (00:47:53) ways of doing things. We we might need (00:47:55) to find new ways to scale but like we we (00:47:56) have a lot of ideas on how to do that. (00:47:58) We know a lot of ways to make these (00:47:59) things better and you know we're (00:48:00) basically making new discoveries all the (00:48:02) time. So like I would say you know (00:48:03) generally the people working in the like (00:48:04) across all the big labs are are pretty (00:48:06) optimistic. Um and so like I I think the (00:48:09) big models are going to continue to get (00:48:10) better you know very quickly here and (00:48:12) then you know overall um and then the (00:48:14) open source models continue to get (00:48:15) better. Um and like I said you know you (00:48:17) know every every every I don't know (00:48:18) every month or something there's like (00:48:19) another big release of like something (00:48:20) like this Kimmy thing. Um where it's (00:48:22) just like wow like you know that's (00:48:24) amazing and you know wow they really (00:48:25) like shrunk that down and got that (00:48:26) capability on a very small form factor. (00:48:28) Um uh and so um yeah that's the case and (00:48:31) then you know I maybe just the third (00:48:33) kind of thing to bring up is um the (00:48:35) other really nice benefit of open source (00:48:37) um is that uh open source is the thing (00:48:39) that's easy to learn from right um and (00:48:41) so if you're a you know computer sc if (00:48:44) you're a computer science professor who (00:48:45) wants to teach a class on on CS on AI or (00:48:48) if you're a computer science student (00:48:49) that's trying to learn about it or if (00:48:50) you're just like a normal engineer in a (00:48:52) normal company trying to learn this new (00:48:54) thing um or just somebody in your you (00:48:56) know by the way somebody in basement at (00:48:58) night with a startup idea. Um the (00:49:00) existence of these of these (00:49:01) state-of-the-art open source models is (00:49:02) amazing because that's the education (00:49:04) that you need. Like they actually these (00:49:06) open source models actually show you how (00:49:07) to do everything. Um right. Um and so (00:49:10) like and and what that's leading to (00:49:12) right is the proliferation of the (00:49:13) knowledge about how to build AI is like (00:49:15) expanding very fast. Um again as (00:49:17) compared to a counterfactual world in (00:49:19) which it was all basically bottled up in (00:49:20) two or three big companies. And so, you (00:49:22) know, the open source thing is also just (00:49:24) proliferating knowledge and then that (00:49:25) knowledge is generating a lot of new (00:49:27) people. Um, and so I I you know, you (00:49:29) know, as you guys have all seen sitting (00:49:31) here today, AI researchers are at an (00:49:32) enormous premium. You know, AI (00:49:34) researchers today are getting paid more (00:49:35) than professional athletes. Um, right? (00:49:37) Like, you know, and that's right, that's (00:49:39) a supply demand imbalance there. There (00:49:41) aren't enough of them to go around. But, (00:49:43) you know, again, shortages create glut. (00:49:45) um the the number of the number of smart (00:49:48) people in the world who are coming up to (00:49:49) speed very quickly on how to build these (00:49:50) things u I mean some of the best AI (00:49:52) people in the world are like 22 23 24 (00:49:55) like they you know kind of by definition (00:49:56) they haven't been in the field that long (00:49:58) you know you know they they can't have (00:49:59) been experts their whole lives right so (00:50:01) you know they they kind of have to have (00:50:02) come up to speed over the course of the (00:50:04) last four or five years and and if if (00:50:05) they if they've been able to do that (00:50:07) then then there's going to be a lot more (00:50:08) in the future that are going to do that (00:50:10) um and so just the the the sort of (00:50:11) spread of the level of expertise on this (00:50:13) technology is happening now very quickly (00:50:15) Um, so I yeah, I mean I think it's still (00:50:17) like I said, I think it's I think it's (00:50:18) still a race. And and by the way, you (00:50:19) know, look, the long-term answer may (00:50:21) well just be both. Um, you know, like I (00:50:23) said, if you if you believe my pyramid (00:50:25) industry structure, then there will then (00:50:27) there will certainly be a large business (00:50:28) of whatever is the smartest thing almost (00:50:30) regardless of how of how much it costs. (00:50:33) Um, and then there but there will also (00:50:34) be this just giant volume market of of (00:50:36) smaller models everywhere, which which (00:50:37) is what we're also seeing. (00:50:39) >> Yep. Yep. The another question you had (00:50:41) posed at at that point in time was will (00:50:43) incumbents versus startups went and at (00:50:44) that point in time I think there was a (00:50:46) mixed bag of where the incumbents were (00:50:48) approaching AI. I think that's radically (00:50:50) changed in the last two years. Um and (00:50:52) then on the counter example the the (00:50:55) blossoming of startups increasingly now (00:50:58) maybe migrating into the incumbent (00:51:00) category just how big they since that (00:51:02) time. You you want to take that uh (00:51:03) question and and give uh your assessment (00:51:05) of where where the state of the world (00:51:07) is? (00:51:08) >> Yeah. Yeah. So, I mean, look, you know, (00:51:09) big companies that are definitely, you (00:51:10) know, playing hard. You know, Google's (00:51:11) playing hard. Meta's playing hard. Um, (00:51:13) Amazon, um, Microsoft, um, you know, (00:51:15) there's a bunch of these companies that (00:51:17) are, you know, that are kind of in in in (00:51:18) there, um, you know, very aggressively. (00:51:20) And then you've got these, you know, (00:51:21) what we call the new incumbents like (00:51:22) Anthropic and and, uh, and Open AI. Um, (00:51:25) but you also have like, you know, even (00:51:26) in the last two years, you've had this (00:51:27) birth of all of a sudden like brand new (00:51:29) companies that are almost instant (00:51:30) incumbents. And you, you could say XAI (00:51:32) is one of those. Uh, ML, by the way, ML (00:51:34) is the great outlier to my Europe thing (00:51:37) from earlier. like Mald is actually (00:51:38) doing very well as sort of the European (00:51:40) kind of uh you know French national (00:51:42) European uh continental you know kind of (00:51:44) AI champion um sort of the you know the (00:51:46) exception that proves the rule um but (00:51:48) you know there's there's a bunch of (00:51:49) these now that are like you know doing (00:51:50) quite well and are kind of becoming new (00:51:52) incumbents um and then of course there's (00:51:53) tons of startups by the way there's and (00:51:55) then there's there's actual foundation (00:51:56) model startups right and so you know we (00:51:58) funded uh you know we funded Ilas out of (00:52:00) open AAI to do a new foundation model (00:52:02) company we funded Miriam Maratti also (00:52:04) out of open AI we funded Faith Ali out (00:52:05) of Stanford to do a world (00:52:07) foundation model company and so you know (00:52:09) there you know there's there are new (00:52:10) swings all all you know all early but (00:52:12) very promising um for to kind of build (00:52:14) you know new incumbents quickly um and (00:52:17) so you know that's all happening and (00:52:18) then and then you know what and then on (00:52:19) top of that there's just this giant (00:52:20) explosion of AI application companies (00:52:22) right and so there there's basically (00:52:23) companies that then usually startups (00:52:25) that basically take the technology and (00:52:27) then you know field it in a specific (00:52:29) domain whether that's law or medicine or (00:52:31) education or you know creativity um or (00:52:34) or or or whatever Um but again here it's (00:52:37) just like it's amazing kind of how how (00:52:39) sophisticated things are getting very (00:52:42) quickly. So (00:52:44) talk about the application companies for (00:52:45) a moment. So like an application company (00:52:47) like classic example is like a cursor is (00:52:48) like an application company. So they (00:52:50) take the core AI capability which they (00:52:52) purchase by the drink from you know (00:52:54) anthropic or open AI or Google um you (00:52:56) know to tokens by the drink and then (00:52:57) they they they build a code basically a (00:53:00) code editor what we used to call an IDE (00:53:02) um integrated development environment or (00:53:04) basically like a a software creation (00:53:05) system um so they build like an AI (00:53:08) coding system um on on top of the (00:53:10) anthropic or open AAI or whatever you (00:53:12) know kind of kind of big models feel (00:53:13) that and that the the critique of those (00:53:15) companies in the industry has been oh (00:53:16) those are what are called called GPT (00:53:18) rappers is kind of the pjorative And the (00:53:20) idea basically being is well they're not (00:53:21) actually like they're not actually doing (00:53:23) anything that's going to preserve value (00:53:24) because the the actual the the whole (00:53:26) point of what they're doing is they're (00:53:27) surfacing AI but it's not their AI. The (00:53:29) the AI that's being surfaced is from (00:53:31) somebody else. And so these are kind of (00:53:32) these pass pass through shell things (00:53:34) that ultimately won't have value. It (00:53:35) actually turns out what's happening is (00:53:37) kind of the opposite of that which is (00:53:38) the the leading uh AI application (00:53:40) companies like Cursor I mean f first of (00:53:42) all what they're discovering is they (00:53:44) they're not just using a single AI (00:53:45) model. they're actually they actually as (00:53:47) these products get more sophisticated (00:53:48) they actually end up using many (00:53:50) different kinds of models that are kind (00:53:51) of customtailored to the specific (00:53:53) aspects of how these products work. Um (00:53:55) and so they may start out using one (00:53:56) model but they end up using a dozen (00:53:58) models and then in the fullness of time (00:53:59) it might be 50 or 100 different models (00:54:00) for different aspects of the product. A (00:54:02) and then B they end up building a lot of (00:54:04) their own models. Um and so they they a (00:54:06) lot of these the leading edge (00:54:08) application companies are actually (00:54:09) backward integrating and actually (00:54:10) building their own AI models because (00:54:12) because they have the deepest (00:54:13) understanding of their domain. and (00:54:14) they're able to build the model that's (00:54:15) best suited to that. Um, and then by the (00:54:17) way, also AI open source, they're also (00:54:20) able to pick up and run an open source (00:54:21) models. Um, and so if they don't like (00:54:24) the economics of of buying intelligence, (00:54:26) you know, by the drink from a from a (00:54:28) from a cloud service provider, you know, (00:54:29) they can pick up one of these open (00:54:30) source models and implement it instead, (00:54:31) which, you know, which these companies (00:54:33) are also doing. Um, and so the the best (00:54:35) of the best of the AI application (00:54:36) companies are they are actually (00:54:38) full-fledged deep technology companies (00:54:40) actually building their own AI. Um and (00:54:42) so that you know that's I think (00:54:43) >> small models though right Mark when you (00:54:45) think about god models versus small (00:54:46) models as you were describing that but (00:54:48) that would be small would you categorize (00:54:49) that as a small (00:54:50) >> well some of them I mean we I will let (00:54:52) them I will let them announce you know (00:54:53) whatever they're doing whenever it's (00:54:55) appropriate but some of them are now (00:54:56) also doing big model development um and (00:54:58) again this this is also part of what (00:55:00) this is also part of the learning just (00:55:01) in the last two years well so like (00:55:03) here's a big learning just from the last (00:55:04) two years which is very interesting (00:55:05) which is two years ago or three years (00:55:07) ago for sure you would have said wow (00:55:08) open AI is like way out ahead um and (00:55:10) like it's probably going to be (00:55:11) impossible for anybody to catch up and (00:55:12) then it's like okay well Anthropic (00:55:14) caught up and so but you know they came (00:55:15) out of open AI and so they had all the (00:55:16) secrets you know whatever and so knew (00:55:18) how to do it and so okay they caught up (00:55:19) but surely nobody can catch up after (00:55:21) them and then very quickly after that (00:55:23) there were a raft of other companies (00:55:24) that caught up very fast and and XAI is (00:55:26) maybe the best example of that which is (00:55:27) like you know XAI you know Elon's (00:55:30) company XAI is the company name gro is (00:55:32) the consumer product version of it um (00:55:34) XAI basically caught up to you know (00:55:36) state-of-the-art openai anthropic level (00:55:38) in in like less than 12 months from a (00:55:40) standing start right and So, and again (00:55:42) that that kind of argues against any (00:55:44) kind of permanent lead, right, by by any (00:55:46) one incumbent that's just going to (00:55:47) basically be able to lock the entire (00:55:48) market down like if you can catch up (00:55:50) like that. And then and then as we as (00:55:51) we've discussed the you know the China (00:55:53) part is all new in the last year, right? (00:55:55) The deepseek uh this the deepseek moment (00:55:57) I think was in January or February of (00:55:59) this year, right? So less than 12 months (00:56:01) ago. Um and so and now you've got like (00:56:03) four Chinese companies that have (00:56:04) effectively caught up. And so, you know, (00:56:06) so it's like, all right, I mean, again, (00:56:07) this is these are these are trillion (00:56:09) dollar questions, not answers. But it's (00:56:11) just like, wow, okay, like it's one of (00:56:13) these things where once somebody proves (00:56:15) that it's capable, it seems to not be (00:56:17) that hard for other people to be able to (00:56:18) catch up, even people with far less (00:56:20) resources. Um, and so, you know, I don't (00:56:22) know what that does. Maybe it makes you (00:56:24) slightly more skeptical in the long run (00:56:25) economics of of the big players. On the (00:56:27) other hand, maybe it makes you like more (00:56:29) bullish about the startup ecosystem. Uh, (00:56:31) it certainly should make you more (00:56:32) bullish about uh startup application (00:56:34) companies, right? being able to do (00:56:35) interesting things, which is why we're (00:56:36) so excited about that. Um, you know, it (00:56:39) should make you probably, you know, a (00:56:40) bit more excited about about certainly (00:56:42) about China. Um, (00:56:45) on the other hand, the Chinese (00:56:46) competition putting pressure on the (00:56:47) American system to not screw itself up (00:56:49) is very positive. So, it should probably (00:56:50) make you a little bit more bullish on (00:56:51) the US. Um, and so, yeah, I think, you (00:56:54) know, the these are, yeah, these are (00:56:55) yeah, these are are live dynamics and I (00:56:57) I think we still need more time to pass (00:56:58) before we know the exact answer. I (00:57:00) should say this, but sometime because (00:57:01) sometimes I don't sometimes I freak (00:57:02) people out when I say these are open (00:57:03) questions. Um, when a company is (00:57:06) confronted with fundamentally open (00:57:08) strategic or economic questions, it's (00:57:10) often a big problem because a company (00:57:12) needs to have a strategy and the (00:57:14) strategy needs to be very specific. Um, (00:57:16) and a company has to make like very (00:57:18) specific concrete choices about where it (00:57:21) like deploys investment dollars and (00:57:22) personnel and like the strategy has to (00:57:23) be like logical and coherent or the (00:57:25) company kind of collapses into chaos. (00:57:27) And so like companies like need to (00:57:28) answer these questions and if they get (00:57:30) the answers wrong, they're really in (00:57:31) trouble. Um, venture We have our issues (00:57:35) and venture but a huge advantage that we (00:57:37) have is we don't have to we we can bet (00:57:38) on multiple strategies at the same time (00:57:40) right um and and we are doing this so we (00:57:42) are betting on big models and small (00:57:44) models and prepared train models and (00:57:46) open source models right and and you (00:57:48) know and foundation models and (00:57:49) applications right uh and consumer and (00:57:51) enterprise and so the portfolio approach (00:57:54) the nature of it is like we we are (00:57:55) aggressively basically uh we we are (00:57:58) aggressively investing behind every (00:58:00) strategy that we've identified that we (00:58:01) think has a plausible chance of (00:58:03) even when that even when that's (00:58:05) contradictory to another strategy that (00:58:06) we're investing in and one is just like (00:58:08) the world's messy and probably a bunch (00:58:09) of things are going to work and so like (00:58:11) there's not going to be clean yes or no (00:58:12) answers to a bunch of this like a lot a (00:58:14) lot of the answers to this I think are (00:58:15) just going to be and answers but the (00:58:16) other is like if one of these strategies (00:58:18) doesn't work like you know we're not (00:58:19) we're not trying to hedge per se but you (00:58:21) know we're going to have representation (00:58:23) in the portfolio of the alternate (00:58:24) strategy and and so we're going to have (00:58:25) mult multiple ways to win. So anyway, (00:58:27) that's that's the goal. That's the (00:58:29) theory of why we are, you know, kind of (00:58:31) taking the approach in the space that (00:58:32) we're taking. Um, and that's why I have (00:58:34) a big smile on my face when I say that (00:58:36) there are these big open questions (00:58:37) because I think that actually works to (00:58:38) our advantage. (00:58:39) >> It's a good seg uh to A16Z questions (00:58:42) because we we've gotten a few in so far (00:58:44) and and uh we had a few that uh were (00:58:46) were sent in ahead as well. So uh I'll (00:58:49) start one with a with a broad topic. (00:58:51) What is something you and Ben disagree (00:58:53) and commit on? (00:58:55) disagree commit. Um, you know, we agree. (00:58:58) I mean, we we Ben I was going to say, (00:59:00) you know, we're an old married couple, (00:59:01) so we argue argue constantly, but we've (00:59:03) been (00:59:04) >> where the romance is dead. (00:59:05) >> The romance is long dead. Yes. Yes. Yes. (00:59:07) Yes. The light the fire the fire has (00:59:09) long since gone out. Um, but um uh yes, (00:59:14) if you Yes. We're in the park squabbling (00:59:16) all the time. Um so, um yeah, I mean, so (00:59:21) look, we debate everything. We we argue (00:59:22) about everything. We that that said like (00:59:23) you know one of the things that's made (00:59:24) our partnership work is like we do we do (00:59:26) tend to come to the same conclusion like (00:59:27) each of us is open to being persuaded by (00:59:29) the other one and so we we end up coming (00:59:30) you know we end up coming to the same (00:59:31) conclusion most of the time. Um so I (00:59:33) would say there there aren't like a (00:59:35) there aren't I said specifically sitting (00:59:36) here today there are like zero issues (00:59:38) where I'm sitting here and I'm like I (00:59:39) can't believe you know I just I can't (00:59:41) believe I'm you know I'm putting up with (00:59:42) this crazy thing on on his on his part (00:59:44) that he's doing um that I really (00:59:46) disagree with but I feel like I have to (00:59:47) commit to or I I don't think vice versa. (00:59:49) Um and so so we don't have any of those. (00:59:51) Um, you know, quite honestly, the (00:59:53) biggest thing I say the biggest thing (00:59:55) that I that he and I the biggest thing (00:59:57) that he and I discuss, this this by the (00:59:59) way, this is not this is not the most (01:00:00) important thing we're doing, but it is a (01:00:02) topic since somebody asked the question. (01:00:03) The biggest thing he and I discuss where (01:00:05) I I don't know, maybe I'm always like (01:00:06) second guessing myself or I I I never (01:00:08) quite know where I should come out on it (01:00:09) that he and I talk about a lot is just (01:00:11) like basically the public footprint of (01:00:13) the company. Um so like our pres our (01:00:16) presence in the our presence in the (01:00:17) world in terms of like public statements (01:00:20) uh controversy um uh you know uh how we (01:00:24) vocalize and express our views on things (01:00:26) um and I would just say there like you (01:00:28) know there's a real there's a tension (01:00:29) there's a real it's you know maybe (01:00:30) obvious but like a very important (01:00:31) tension like generally speaking the more (01:00:34) out there we are and the more outspoken (01:00:36) we are and the more controversial we are (01:00:37) the better for the better for the (01:00:39) business in the sense of the (01:00:40) entrepreneurs love it. Uh the the the (01:00:43) founders want to work with is very clear (01:00:46) at this point. The founders want to work (01:00:47) with uh uh people who basically are (01:00:50) brave and controversial and take (01:00:52) controversial stands uh and articulate (01:00:54) things clearly and and they want that (01:00:55) for a bunch of reasons. One is because (01:00:56) it's a demonstration of courage which (01:00:58) they appreciate. But the other is (01:00:59) because it it it it teaches them who we (01:01:01) are before they even meet us. Um and and (01:01:04) and that has just proven to be just like (01:01:06) this incredible competitive advantage. (01:01:07) you know, long long-term LPs will know (01:01:09) like this is why we started with a very (01:01:10) active marketing strategy from the very (01:01:12) beginning and like it completely worked. (01:01:13) Like the the whole thing was if we're (01:01:15) able to broadcast our message and we're (01:01:17) able to basically be very clear in what (01:01:18) we believe even to the point where it's (01:01:20) controversial, like the best founders in (01:01:22) the world are going to understand us (01:01:23) before they even walk in the door, (01:01:25) right? And they're going to they're (01:01:26) going to know us even before they've met (01:01:27) us as opposed to everybody else in (01:01:28) venture, at least at the time, that was (01:01:30) basically just like keeping everything (01:01:31) quiet. Um where they, you know, the (01:01:34) founder just has no idea who these (01:01:35) people are and what they believe. And so (01:01:36) that that like worked incredibly well. (01:01:37) It continues to work incredibly well. Um (01:01:39) it's by the way it's you know it's (01:01:41) generally true across the industry. It's (01:01:43) it's it's like generally the case. On (01:01:44) the other hand, there are externalities (01:01:46) to being you know publicly visible and (01:01:48) and and and to being controversial um on (01:01:50) many fronts. Um we are I would say this (01:01:53) we are we're very much we're trying very (01:01:55) hard to thread this needle. So like (01:01:56) we're we're not backing off of generally (01:01:57) being a a company that does a lot of (01:01:59) outbound. we, you know, we Eric (01:02:00) Worenberg and the team that he's built, (01:02:02) you know, that we've talked to you guys (01:02:03) about in the past, um, you know, is I is (01:02:05) already off to the races. Um, you know, (01:02:07) we're we're going to, you know, we're (01:02:08) tripling down on the idea of basically (01:02:09) being the leaders and articulating the (01:02:11) tech and business issues that matter. (01:02:12) You know, the, you know, the issues for (01:02:14) sure that people need to be able to (01:02:15) understand. Um, and and that's proven to (01:02:17) be very effective. By the way, a fair (01:02:19) amount of our coms are actually aimed at (01:02:21) Washington. Um because again it's like (01:02:23) if you're a policy maker in Washington (01:02:25) and you're sitting there 3,000 mi away (01:02:28) and your entire information source is (01:02:29) like East Coast newspapers that hate (01:02:31) Silicon Valley. Like that's bad. Um and (01:02:33) so you know our ability to like (01:02:35) broadcast, you know, inform points of (01:02:37) view on technology. We just we meet (01:02:38) people in DC all the time um who say, (01:02:40) "Yeah, I you know, most of what I know (01:02:42) about this topic I learned from you guys (01:02:43) because I listened to the podcast, I (01:02:44) read the articles, I watched the YouTube (01:02:46) channel." Um and so, you know, we're (01:02:47) we're going to continue to do that. And (01:02:48) so we, you know, over over over overall (01:02:50) we have a, you know, we're kind of on (01:02:52) our front foot on that stuff. But yeah, (01:02:53) he he and I do he and I do go back and (01:02:54) forth a bit on exactly how, yeah, how (01:02:56) many third rail topics should we touch? (01:02:58) Um, and uh and how frequently. And I I (01:03:00) would say we're we're we are trying to (01:03:02) we are trying to moderate that. (01:03:03) >> As Elizabeth Taylor said, as long as I (01:03:05) spell our name right, um, it's (01:03:07) oftentimes could be good in most (01:03:10) scenarios, particularly when it comes to (01:03:12) little tech. uh double uh and also I (01:03:15) think embedded in that question is (01:03:16) probably uh some degree of of uh uh the (01:03:19) relationship that you and Ben have which (01:03:20) is now going on 30 plus years at this (01:03:22) point. Uh so much so that that Mark has (01:03:24) become uh one person representing both (01:03:28) uh some people refer to Mark as Andre (01:03:30) and Horowitz no lost the mark have (01:03:32) combined just into one person. Uh (01:03:35) >> yes (01:03:36) >> that's the result of 30 plus years (01:03:38) working together. Okay. Um, so it's been (01:03:40) 2 years since you've reorganized around (01:03:42) AI, launched AD. What do you think you (01:03:44) got most right? Uh, and in hindsight, is (01:03:46) there anything that you underestimated (01:03:48) or or missed in that decisioning (01:03:50) process? (01:03:51) >> No, I mean, look, we made we made plenty (01:03:53) of mistakes. I think those were I think (01:03:54) those were the right calls. I mean, AI (01:03:56) was like I said, like you know, the (01:03:58) whole theor back up the whole theory of (01:04:00) venture the whole theory of venture that (01:04:01) we've had from the beginning is that you (01:04:03) know, many people before us have had as (01:04:04) well. that's very correct I think is the (01:04:07) whole theory is like the money adventure (01:04:08) is made when there's like a a (01:04:10) fundamental architecture shift like when (01:04:11) there's like a fundamental change in the (01:04:13) technology landscape. Um and and that's (01:04:15) been true for you know adventure (01:04:16) basically forever. Um uh and and the (01:04:19) reason is because if you have a (01:04:20) fundamental change in technology then (01:04:22) you have this period of creativity in (01:04:23) which you can have basically aggressive (01:04:25) you know very aggressive kind of people (01:04:26) you know kind of start these new (01:04:27) companies and and they have this kind of (01:04:29) shot to kind of come in and you kind of (01:04:30) win categories before big companies can (01:04:32) respond. um if there's no fundamental (01:04:35) change in technology, it's very hard to (01:04:36) make startups work because the big (01:04:37) companies just end up doing everything. (01:04:39) And so you so venture kind of, you know, (01:04:41) sort of lives or dies on on the basis of (01:04:43) these of these waves of these (01:04:44) transitions. Um and and so there's (01:04:47) always there there's always this (01:04:49) question. It's always this question. I (01:04:50) mean, I would just say the best venture (01:04:52) capital firms in history, I I think are (01:04:54) the ones that were the most aggressive (01:04:55) at being able to navigate from wave to (01:04:57) wave, right? And and and look, I was a (01:04:59) beneficiary of this when I came to (01:05:00) Silicon Valley in ' 904. you know that (01:05:02) there was no venture firm in 1994 that (01:05:04) was like the internet venture capital (01:05:05) firm like that it just didn't exist. Um, (01:05:07) but there were a set of venture capital (01:05:09) firms at the time, you know, at the time (01:05:10) our our firm Kleiner Perkins that said, (01:05:12) "Oh, this is a new architecture. This is (01:05:14) a new technology change. It seems (01:05:16) totally crazy. Everybody says you can't (01:05:17) make money on it. Whatever, whatever. (01:05:19) These kids are nuts." But like, we're (01:05:20) going to make those bets. Um, and so (01:05:22) they were willing to invest. And by the (01:05:24) way, you know, KP in the in the in the (01:05:25) '90s invested not only in us, but also (01:05:27) in Amazon and then Google and like in, (01:05:29) you know, company after company after (01:05:30) company. They invested in at home, which (01:05:32) basically made made home broadband work. (01:05:34) um you know they invested in in a fleet (01:05:36) of companies and they were a venture (01:05:37) capital firm that had started in the (01:05:38) 1970s around really around what was at (01:05:41) the time called Minicomputers which was (01:05:42) like a you know three generations of (01:05:44) techn technology back and they had (01:05:46) navigated from wave to wave um and and (01:05:48) you know the same thing is true for (01:05:49) Sequoia the same thing is true for (01:05:50) basically any successful venture firm (01:05:52) has been in business for you know 30 or (01:05:53) 40 or 50 years and so I I think in this (01:05:56) business like of all businesses like you (01:05:58) you just you need you need to get onto (01:05:59) the new thing um you know it it was I (01:06:02) mean quite honestly it was I pretty (01:06:04) amazing that most of the venture (01:06:06) ecosystem just decided to sit crypto (01:06:08) out. Um and and the number of VCs that (01:06:11) we talked to between call it, you know, (01:06:14) the release of the Bitcoin white paper (01:06:15) in 2009 to the beginning of the crypto (01:06:17) war in 2021 who just basically said, (01:06:19) "Oh, we're not going to do crypto." It (01:06:21) was fairly it's I I like I don't I I (01:06:23) never quite know what to do with the VC (01:06:24) who says, "Oh, there's a new wave of (01:06:25) technology and I'm very deliberately not (01:06:26) going to participate in it." And I'm (01:06:27) always like like, "Is that not the job?" (01:06:31) Right? Like so so so like I was fairly (01:06:33) amazed by the VCs that didn't make the (01:06:35) jump uh to crypto. You know they they (01:06:37) looked briefly smart during the crypto (01:06:39) wars I would say of the last you know (01:06:40) three or four years and I think they (01:06:42) they probably look maybe a little bit (01:06:43) less smart now. Um you know AI is (01:06:45) another one of these where there are (01:06:47) certain firms that are are jumping all (01:06:48) over it and there are certain firms that (01:06:49) are just kind of sitting back and (01:06:50) letting it happen. Um and um and and by (01:06:53) the way there were certain firms that (01:06:54) never made it to the internet. I mean (01:06:55) there were there were firms that were (01:06:56) very well known in the 80s um and very (01:06:58) successful that just like did not make (01:06:59) the jump uh to the internet and (01:07:01) basically just petered out. And so (01:07:02) anyway long-winded way of saying I think (01:07:04) I think in this business of all (01:07:05) businesses you have to jump you have to (01:07:06) jump on the new wave. Um and I and I (01:07:08) think we got the magnitude of it of it (01:07:09) right that this is like a fundamental (01:07:10) fundamental transformation inside the (01:07:12) firm. Um you know AD is you know AD is (01:07:14) doing great. Um AD AD itself I believe (01:07:17) is also a beneficiary of AI. um right (01:07:20) because in in two ways one is a lot of (01:07:22) the kinds of products that AD companies (01:07:24) build themselves benefit from AI and (01:07:25) then also AI is a driver of demand in (01:07:28) other sectors of AD like like energy and (01:07:30) materials. Um and so I you know I think (01:07:33) that that generally is is very (01:07:34) consistent and you know is working well. (01:07:36) Um by the way you know crypto's back (01:07:39) back to being a you know I would say an (01:07:41) exciting industry as a consequence of (01:07:43) all the policy changes. Um and then and (01:07:45) then there's even going to be I think (01:07:46) intersections. I I think there's (01:07:47) actually going to be quite a few (01:07:48) intersections between AI and crypto. Um (01:07:50) and then and then biote you know biotech (01:07:52) also bio and healthcare I think are (01:07:54) obviously going to be transformed by AI (01:07:56) both on the healthcare side and on the (01:07:57) actual drug discovery side and you know (01:07:59) and that's underway. And so any anyway (01:08:01) so like the the the individual efforts (01:08:02) in the firm feel good um and suitable (01:08:04) for the time the inter the interactions (01:08:07) between the teams um and the kind the (01:08:09) the hybrid ideas you know the companies (01:08:11) that are coming at these things from (01:08:12) multiple angles uh you know feels really (01:08:14) good um you know maybe the correlarying (01:08:16) question is like you know what do we (01:08:18) feel like we're missing right now um and (01:08:19) I I think the answer is really not like (01:08:21) I don't I don't think like right now (01:08:23) we're not missing a vertical like I I (01:08:25) don't like as of right now like there (01:08:26) there's not like a specific vertical of (01:08:28) like I don't know whatever that like (01:08:29) where we just like, oh, we just need, (01:08:30) you know, we need the equivalent of a (01:08:32) new of a new unit or the equivalent of a (01:08:33) new um you know, new fund or whatever. I (01:08:34) don't I don't see that at the moment. I (01:08:36) think it's more executing extremely well (01:08:38) in the verticals that we have in front (01:08:39) of us. Um and um and then, you know, (01:08:41) being the best possible partner to the (01:08:42) to the portfolio companies. (01:08:44) >> Yeah. Actually, on the point of of AD, (01:08:46) um because uh AI is creating and there's (01:08:50) a lot of talk around AI taking jobs, (01:08:52) etc. Ironically enough, the jobs in AD (01:08:55) sectors have never been more in demand (01:08:57) in the physical world related to energy, (01:09:00) related obviously to data center build, (01:09:02) etc. So like the the pendulum it seems (01:09:03) like also is uh is swinging from just an (01:09:06) accelerant standpoint from from a (01:09:07) society uh point of view. Um you talked (01:09:10) about the importance of society also (01:09:12) needing to be ready for tech adoption. (01:09:14) Like have you seen that accelerating of (01:09:16) recently? what's your sentiment of of (01:09:18) how to actually um increase that just to (01:09:21) also make sure the convergence of of (01:09:23) adoption also falls in line with with (01:09:25) how quickly tech is is actually being (01:09:27) implemented. (01:09:28) >> Yeah. So, you know, look, we've talked (01:09:30) about this before, but um you know, (01:09:31) look, for a very long time, tech was (01:09:33) just not a very relevant look, if you go (01:09:36) back over like whatever 300 years, like (01:09:38) there's just like recurring waves of (01:09:40) like total panic and freakout caused by (01:09:42) new technology. Or even you go back 500 (01:09:44) years, you go back to the printing (01:09:45) press, you know, which basically was (01:09:46) handin-hand with the the sort of (01:09:47) creation of Protest Pro Protestantism, (01:09:49) which really changed things. Um, and (01:09:51) then um, you know, you you go back to (01:09:53) um, you know, there there were just (01:09:54) always kind of, you know, continuous (01:09:56) panics there. You know, there have been (01:09:57) m there have been multiple ways of (01:09:58) automation panics for the last 200 (01:09:59) years. You know, a lot of the (01:10:01) foundational panic under Marxism was (01:10:02) basically a fear of of of of the (01:10:04) elimination of jobs through the (01:10:06) application of automation. um uh you (01:10:09) know a lot of the same arguments you (01:10:10) hear today about like AI is going to (01:10:11) centralize all the wealth in a handful (01:10:12) of a few people and everybody else is (01:10:13) going to be poor and emiserated like (01:10:15) that that basically is what Markx used (01:10:16) to say um which I think was by the way (01:10:19) wrong then is wrong now we can talk (01:10:21) about but um you know and then even like (01:10:23) in the 1960s there was this whole panic (01:10:25) around around AI um uh replacing all the (01:10:28) jobs there was this there's this great (01:10:29) uh it's long long forgotten but it was a (01:10:31) big deal at the time during the Johnson (01:10:32) administration you read these AI pause (01:10:35) letters today you know that this one (01:10:36) that just came out a few weeks ago that (01:10:37) Prince Harry uh headlined of all people. (01:10:40) Um and um uh uh you know he talks about (01:10:44) AI is going to ruin everything and it's (01:10:45) like and 1964 there was basically a (01:10:48) group of like the leading lights in (01:10:50) academia science and uh you know um kind (01:10:53) of public affairs that there was this (01:10:55) thing called the triple committee or the (01:10:56) committee for the triple revolution. If (01:10:58) you do a Google search on it's like (01:10:59) committee for the triple revolution (01:11:01) Johnson white house or whatever you'll (01:11:02) this thing will pop up. Um and you know (01:11:05) it was a very similar kind of manifesto (01:11:07) of like we need to stop the march of (01:11:08) technology today or we're going to ruin (01:11:09) everything. Um and and then you know (01:11:11) even in the course of the last 20 years (01:11:13) there was like a big panic around um (01:11:16) actually outsourcing in the 2000s was (01:11:17) going to take all the jobs and then it (01:11:18) was actually robots weirdly enough in (01:11:20) the 2010s which is amazing because (01:11:22) robots didn't even work in the 2010s and (01:11:24) they kind of you know still don't. Um (01:11:26) but uh you know there's a panic around (01:11:27) that and now there's kind of whatever (01:11:29) level of AI panic. Um and so like you (01:11:31) know I would just say like look that you (01:11:32) know the way I would describe it is you (01:11:34) know we in Silicon Valley have always (01:11:36) wanted the work that we do to matter. Um (01:11:38) you know we spend most of our time quite (01:11:40) honestly with people telling us that (01:11:42) everything that we're doing is stupid (01:11:43) and won't work. Um like that's the (01:11:45) default position. Um you know and then (01:11:47) basically that flips at some point into (01:11:49) panic about how it's going to ruin (01:11:50) everything. Um you know it's it's easy (01:11:52) sitting out here to be cynical about (01:11:54) that. Um especially when you kind of see (01:11:56) the patterns over time. I you know my (01:11:58) view is we need to be actually very (01:12:00) respectful of that and we need to be (01:12:01) very aware of that and and basically (01:12:02) that we you know I use the metaphor with (01:12:05) the dog that caught the bus like we (01:12:06) always wanted to work on things that (01:12:07) matter we are working on things that (01:12:08) matter uh people in the rest of society (01:12:10) actually really do care about these (01:12:12) things um and you know and it's our (01:12:14) responsibility to think that all through (01:12:15) very carefully and to do a good job um (01:12:17) you know both not just building the (01:12:18) technology but also explaining it you (01:12:20) know look you know I think we have a (01:12:21) real obligation to uh you know to to (01:12:23) really explain ourselves and engage on (01:12:24) these issues um in terms of how to (01:12:26) measure how going you know it's it's (01:12:28) sort of the classic social science (01:12:29) question um uh which is like okay if you (01:12:32) want to understand basically you know (01:12:35) patterns of people there's basically two (01:12:36) ways to understand what people are doing (01:12:38) and thinking um one is to ask them and (01:12:41) and then the other is to watch them um (01:12:43) and like every social every social (01:12:45) scientist like every sociologist will (01:12:46) will will will tell you this which (01:12:47) basically is you can you can ask people (01:12:50) right and and the way you do that right (01:12:52) is like you know surveys focus groups (01:12:53) polls um you know what they think Um but (01:12:57) then but then you can watch them and you (01:12:58) can do what's you know called reveal (01:12:59) preferences. They're just observe (01:13:00) behavior which is you can actually watch (01:13:02) their behavior and and and what you (01:13:03) often see in many areas of human (01:13:05) activity including politics and many (01:13:06) different aspects of society and culture (01:13:08) over time is the answers that you get (01:13:10) when you ask people are very different (01:13:11) than the answers that you get when you (01:13:12) watch them. Um and the reason is because (01:13:15) like I mean you could have a bunch of (01:13:17) theories as to why this is the Marxists (01:13:19) claim that people have false (01:13:20) consciousness. the the the the somewhat (01:13:22) the explanation I believe is just people (01:13:24) have opinions on all kinds of things (01:13:25) particularly when they're in a context (01:13:26) where they get to express themselves um (01:13:28) and they'll have a tendency to kind of (01:13:30) express themselves in very heated ways (01:13:31) and then if you just watch their (01:13:32) behavior they're often a lot calmer um (01:13:34) and a lot more measured and a lot more (01:13:36) rational in in what they do and so the (01:13:38) AI that's playing out in AI right now (01:13:39) which is if you pull if you run a survey (01:13:42) or a poll of what for example American (01:13:44) voters think about AI it's just like (01:13:46) they're all in a total panic it's like (01:13:47) oh my god this is terrible this is awful (01:13:48) it's going to kill all the jobs it's (01:13:49) going to ruin thing. The whole thing, if (01:13:52) you watch the revealed preferences, (01:13:53) they're all using AI. So, they're like, (01:13:57) they're downloading the apps. They're (01:13:59) using chat GPT in their job. They're, (01:14:02) you know, having an argument. You You (01:14:04) see this online all the time now. I'm (01:14:05) having an argument with my boyfriend or (01:14:06) girlfriend. I don't understand what's (01:14:07) happening. I take the text exchange. I (01:14:09) cut and paste it into chat GPT and I (01:14:11) have chat GPT explain to me what my (01:14:13) partner is thinking and tell me how I (01:14:14) should answer so that he's, you know, he (01:14:15) or she is not mad at me anymore, right? (01:14:17) So, or like, you know, I have this (01:14:18) thing, you know, I have a skin, you (01:14:19) know, I have a skin condition and (01:14:21) doctors, you know, da da da, and I take (01:14:22) a photo and I and I'm finally like (01:14:24) learning about my own health or I use it (01:14:26) in my job like I, you know, I had to get (01:14:28) this report ready for Monday morning and (01:14:29) I ran out of time and like it, you know, (01:14:30) chat GPT really saved my bacon. Um, and (01:14:33) so people in their daily lives are I (01:14:36) would, you know, just you just look at (01:14:37) the just look at the data you just like (01:14:39) they are not only using this technology, (01:14:41) they love this technology. Um, and they (01:14:43) love it and they're adopting as fast as (01:14:44) they possibly can. So I I tend to think (01:14:46) we're going to the public discussion of (01:14:48) this is going to ping pong back and (01:14:49) forth for a while because there is this (01:14:50) divergence between what people are (01:14:51) saying what people are doing. Um but but (01:14:53) I do think that the what people are (01:14:54) doing part is is is obviously the part (01:14:56) the part ultimately that wins and and (01:14:58) and I think this by the way I think this (01:14:59) technology is going to be exactly the (01:15:01) same as every other one. Um which is the (01:15:02) thing that's going to happen here is (01:15:03) this is just going to proliferate really (01:15:05) broadly. It's going to freak everybody (01:15:06) out and then you know 20 years from now (01:15:08) everybody's going to be like oh thank (01:15:09) god we've got it. Like wouldn't life be (01:15:11) miserable if we didn't have this? um and (01:15:12) or you know 5 years from now or or one (01:15:15) year from now you know people are going (01:15:16) to reach that conclusion. Um so I'm I'm (01:15:19) very optimistic about where this lands. (01:15:21) It's just that you know there will be (01:15:22) turbulence along the way. (01:15:23) >> I'm I'm smiling because I also witnessed (01:15:24) that in the wild. Literally late last (01:15:26) week I was on the plane. The guy next to (01:15:28) me was talking to his chat. I could see (01:15:30) him and he was like help me draft an (01:15:32) escalation letter to United for the (01:15:34) delay on this flight. I was like sir you (01:15:36) are on the flight right now. Like at (01:15:37) least wait until it's over. (01:15:42) It was very good though. I'm sure he had (01:15:43) a great email crafted as a as a part of (01:15:45) that. Uh so, okay, I'm going to switch (01:15:48) gears to uh a few fun questions that (01:15:50) that were sent in uh that uh is intended (01:15:53) to be a lightning round. So, so uh what (01:15:55) what is something you've changed your (01:15:56) mind on recently? Bonus points if it was (01:15:58) someone younger than you. (01:15:59) >> I mean, it's like every day. Um it's (01:16:01) just like it's just a constant, you (01:16:04) know, it's it's almost all like what's (01:16:05) in the realm of the possible. Um, I I'm (01:16:07) I'm terrible at specific examples, so I (01:16:09) don't I don't have one like ready at (01:16:10) hand, but like like I said, it's just (01:16:11) it's it's always Yeah. No, it's it's (01:16:13) often somebody showing up. It's either (01:16:15) something somebody writes or something (01:16:16) somebody says. Um, and yeah, it's almost (01:16:19) Yeah, it's very frequently somebody (01:16:20) who's very young. Um, and um, yeah, it's (01:16:22) just like I would say it's a it's a (01:16:24) routine experience. (01:16:25) >> Good way to stay young. Um, do you plan, (01:16:28) speaking of young, do you plan to be (01:16:30) cryogenically frozen? (01:16:33) Not with current not with current (01:16:35) cryogenic technology. Um the uh the the (01:16:38) the track record of that is not great. (01:16:40) Um uh and um the stories are somewhat (01:16:43) horrifying, but uh you know, we'll see. (01:16:44) >> We'll see. You got we still got some (01:16:46) time. (01:16:48) >> Um how do you stay grounded when your (01:16:50) influence itself may distort reality (01:16:52) around you? (01:16:53) >> Yeah. So (01:16:56) I was just say the good news, you know, (01:16:57) I would say the good news on several (01:16:58) front. So one is look the concern is (01:17:00) real. Um, and it's hard for me to it's (01:17:01) hard for me to talk about with sort of (01:17:03) my Midwestern, you know, kind of, you (01:17:04) know, Midwesterners, we we either are (01:17:06) very humble or we we're really good at (01:17:08) faking it, but um, uh, you know, it's (01:17:10) hard to talk about, but requires some (01:17:11) introspection. But yeah, I mean, look, (01:17:12) the the reality warping effect is (01:17:14) definitely real. Um, by the way, there (01:17:16) is a very big advantage to the reality (01:17:17) warping effect, um, which is being able (01:17:19) to get people to do what you want them (01:17:20) to do. Um, so that, you know, there is (01:17:23) there is another side to it. Um but it (01:17:26) you know it is a concern in terms of (01:17:28) like having an actual accurate (01:17:29) understanding of what's happening. I (01:17:31) guess I would say two things. I would (01:17:32) say one is um you know I mean one is (01:17:34) just you know my partners I think are (01:17:35) quite you know including Ben are quite (01:17:37) forthright um in telling me when I'm (01:17:38) wrong but you know more generally like (01:17:40) we're just we are very exposed to (01:17:42) reality. Um and so and this and again (01:17:44) you know you mentioned I don't know it's (01:17:46) a way to stay younger, make sure their (01:17:47) hair never grows back or whatever. It's (01:17:48) just like you know we run these (01:17:50) experiments you know cuz we make these (01:17:52) decisions about whether to invest or not (01:17:53) invest and we work with these companies (01:17:54) and all their things and like you know (01:17:56) reality kicks in quickly. You know the (01:17:58) the the delusions don't last very long (01:17:59) in this business. Um because like you (01:18:01) know these these things either work or (01:18:03) they don't. Um and you know you have (01:18:05) these like long elaborate you know (01:18:06) discussions about you know theories on (01:18:08) this and that and the other thing and (01:18:09) then reality just like completely smacks (01:18:10) you square in the face you know like you (01:18:12) idiot right you know like you know what (01:18:14) were you you like you know this is like (01:18:17) the you know the ultimate frustration of (01:18:18) the business which is also very (01:18:19) motivating which is the number of times (01:18:20) that you think that you've applied (01:18:21) superior analysis and then you've either (01:18:23) invested or not invested based on that (01:18:25) analysis and it turns out it was just (01:18:26) you the analysis was just completely (01:18:27) wrong right um and you know you just (01:18:30) like completely overrated your ability (01:18:31) to epistemically you know kind of (01:18:32) analyze these things you just you know (01:18:34) basically inflicted harm like I always (01:18:37) the question is always you know it's (01:18:38) sort of you know any activity that we do (01:18:40) is it value add or is it actually value (01:18:41) subtract right and and and I think in (01:18:44) this business of all businesses is kind (01:18:45) of like that and and that applies to all (01:18:47) of my own contributions as well so so (01:18:49) there is that and then and then I would (01:18:50) say um you know maybe the final thing is (01:18:52) just like I do have the entire internet (01:18:53) ready to tell me that I'm an idiot so (01:18:56) that also (01:18:58) that also doesn't doesn't hurt and it (01:19:00) and it does on a regular basis (01:19:04) on on the point of uh your alluding to (01:19:07) earlier about uh decisions on investing (01:19:08) in companies. My favorite line I think (01:19:10) it was from the uh the Cheeky Point (01:19:12) interview that you did uh was you know (01:19:14) when you invest in a company it doesn't (01:19:16) go well at least it goes bankrupt right (01:19:18) if it does if it does well and it does (01:19:20) fantastically well you hear about it (01:19:22) every single day (01:19:24) >> for the rest of your life. Yeah. For the (01:19:26) next for the next 30 years. (01:19:29) validity smacking you in the face saying (01:19:31) you fool. (01:19:32) >> You had it. It's literally It's (01:19:34) literally you had it in your office. All (01:19:36) you had to do is say yes. (01:19:38) And by the way, and this is the thing (01:19:39) like every great VC like if you this is (01:19:42) this is the stories that you know the (01:19:43) VCs tell each other. Every great VC (01:19:45) basically has this history of like my (01:19:47) god I had it was in my office. The thing (01:19:49) was in my office and I said no and if I (01:19:50) had just said yes. Um and so it's yeah (01:19:53) it's very hard to um yes the constant (01:19:55) reminders in the Wall Street Journal and (01:19:56) on CNBC every day that you made a giant (01:19:58) mistake um is yes very good very good (01:20:00) for the the old humility factor. (01:20:02) >> Yeah very humbling helps you stay (01:20:04) grounded uh all the time. Uh last (01:20:07) question do you plan to go to Mars if (01:20:09) and when that opportunity presents (01:20:10) itself? (01:20:12) >> Probably (01:20:13) not. (01:20:16) >> My subliminal Zoom background wasn't uh (01:20:19) sending the positive vibes. This is what (01:20:21) it (01:20:21) >> Well, I'm not even willing to leave (01:20:22) California. Um, (01:20:26) so I'm barely willing to leave my house. (01:20:28) So, um, uh, yeah, I may maybe by maybe (01:20:31) by VR. (01:20:32) >> Yeah. (01:20:33) >> Um, and then we'll see what happens. I (01:20:35) mean, look, having said that, I think (01:20:36) Elon's going to pull it off. Um, and so (01:20:38) I think, you know, I don't know. I don't (01:20:39) know. I don't want to predict. This is (01:20:41) not a prediction, but I, you know, I (01:20:42) would not be surprised if within a (01:20:43) decade there's routine trips back and (01:20:44) forth. Um, so, uh, yeah, we may, uh, (01:20:48) this this may actually become a a (01:20:49) practical question. And and by the way, (01:20:51) I do know a lot of people who are (01:20:52) probably going to go, (01:20:53) >> myself included. Put me on that. (01:20:55) >> Oh, fantastic. (01:20:57) >> The the flights around the world have (01:20:58) prepared me for the six-month journey to (01:21:00) Mars, so I will be just fine.