AI Is Already Replacing ENTIRE Company Functions (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) you know at some big companies that are (00:00:02) very tech forward you know 50% plus of (00:00:05) customer support is already done by AI (00:00:07) and that's a $400 billion industry and (00:00:10) then if you know what AI is great about (00:00:12) is persuasion that's sales and customer (00:00:14) support and so of the functions of a (00:00:17) company if you think about them them (00:00:18) they're to make stuff sell stuff and (00:00:20) then support the customers so right now (00:00:23) maybe you're in late 26 you're going to (00:00:25) be pretty good at two of them (00:00:27) >> so that's Gavin Baker a legendary (00:00:30) investor who spends most of his time (00:00:32) these days studying semiconductors, (00:00:34) compute, and the realworld economics of (00:00:37) AI. And in this podcast episode with (00:00:40) Patrick Oshanosy, he talks about the (00:00:42) coming wave of AI labor disruption, the (00:00:45) companies and countries set to benefit (00:00:48) the most, the industries that are about (00:00:50) to get obliterated, and the surprising (00:00:53) way compute itself is being redefined. (00:00:56) Let's get into it. All right, so for the (00:00:59) first 20 minutes or so of the podcast, (00:01:01) they mostly talk about chips. Gavin says (00:01:04) Nvidia's new Blackwell chips, which (00:01:06) should start coming online next year, (00:01:08) are going to be genuinely game-changing. (00:01:10) According to him, Blackwell is what (00:01:12) finally allows us to scale pre-training (00:01:14) again, which has been stalled up for a (00:01:16) while now given the compute constraints. (00:01:19) So this is when Patrick asks, "Okay, but (00:01:22) so what? What does the scaling of (00:01:24) pre-training even further even mean? And (00:01:27) what might it unlock, especially as all (00:01:30) this new infrastructure comes online (00:01:32) over the next few years? Here's his (00:01:34) response. (00:01:34) >> If I were to posit like an event path, I (00:01:37) think the Blackwell models are going to (00:01:38) be amazing. the dramatic reduction in (00:01:41) per token cost enabled by the GB300 and (00:01:44) probably [clears throat] more the MI450 (00:01:45) than the MI355, you know, will lead to (00:01:48) these models being allowed to think for (00:01:50) much longer, (00:01:52) >> which means they're going to be able to, (00:01:54) you know, do new things. Like I was very (00:01:55) impressed Gemini 3 made me a restaurant (00:01:57) reservation. (00:01:58) >> It's the first time it's done something (00:02:00) for me. And I mean, other than like go (00:02:03) research something and teach me stuff, (00:02:05) >> but you know, if you can make a (00:02:06) restaurant reservation, you're not that (00:02:08) far from being able to make a hotel (00:02:09) reservation and an airplane reservation (00:02:12) and order me an Uber and (00:02:14) >> all of a sudden you got an assistant. (00:02:15) >> Yeah. And you could just imagine, (00:02:17) everybody talks about that, but you can (00:02:18) just imagine it's on your phone. I think (00:02:20) that's that's pretty near- term, but you (00:02:22) know, it's you know, it's some big (00:02:24) companies that are very tech forward, (00:02:26) you know, 50% plus of customer support (00:02:28) is already done by AI. And that's a $400 (00:02:31) billion industry. And then [snorts] if (00:02:33) you know what AI is great about is (00:02:34) persuasion, that's sales and customer (00:02:36) support. And so of the functions of a (00:02:39) company, if you think about them, (00:02:41) they're to make stuff, sell stuff, and (00:02:43) then support the customers. So, right (00:02:45) now, maybe you're in late 26, you're (00:02:47) going to be pretty good at two of them. (00:02:49) Um, I do think it's going to have a big (00:02:51) impact on media. Like, I think robotics, (00:02:53) you know, we talked about the last time (00:02:54) are going to finally start to be real. (00:02:56) You know, there's an explosion and kind (00:02:57) of exciting robot robotic startups. I do (00:03:00) still think that the main battle is (00:03:01) going to be between uh Tesla's Optimus (00:03:03) and the Chinese because, you know, it's (00:03:05) easy to make prototypes. It's hard to (00:03:06) massproduce them. But then it goes back (00:03:08) to that what Andre Karpathy said about (00:03:11) AI can automate anything that can be (00:03:13) verified. So any function where there's (00:03:16) a right or wrong answer, a right or (00:03:17) wrong outcome, you can apply (00:03:19) reinforcement learning and make the AI (00:03:22) really good at. (00:03:23) >> So this is where they start talking (00:03:25) about reinforcement learning which (00:03:27) became a much bigger focus right around (00:03:29) the time pre-training started running (00:03:31) into real constraints. This shift is (00:03:33) what gave rise to reasoning models like (00:03:35) OpenAI's 01 and 03. Although at this (00:03:38) point reasoning is basically just (00:03:40) embedded into every frontier model. (00:03:42) Reinforcement learning is also what (00:03:44) enables systems like Alph Go and Alpha (00:03:46) Fold where the model is effectively (00:03:48) learning on its own through trial and (00:03:51) error guided by a clear objective. And (00:03:53) this is why Carpathy says AI can (00:03:55) automate anything that's verifiable, (00:03:58) anything where there's a clear signal (00:04:00) for success or failure. We've already (00:04:02) seen this with how good these models are (00:04:04) at math and coding, where there's a (00:04:06) right answer and the model can (00:04:08) iteratively improve toward it. But this (00:04:10) logic extends way beyond math. Customer (00:04:13) support is verifiable. Did you solve the (00:04:16) user's problem or not? Sales is (00:04:18) verifiable. Did you close the deal or (00:04:20) not? Accounting is verifiable. Did the (00:04:23) books balance or not? And what makes (00:04:25) this especially important is that we're (00:04:28) still extremely early. Not only will we (00:04:30) see a ton of jobs in these verifiable (00:04:32) domains be automated, but if you think (00:04:35) about it, the models themselves will (00:04:38) also keep getting better the more (00:04:40) they're used in the real world. Did I (00:04:42) actually do what the user wanted or not? (00:04:45) Imagine when you scale that across (00:04:46) companies, across products, and across (00:04:49) consumers worldwide. you know, in in (00:04:52) 2023 and 24, I was fond of quoting Eric (00:04:56) Visria and Eric Fishria's statement, our (00:04:59) friend, um, brilliant man. And Eric (00:05:01) would always say, "Foundation models are (00:05:03) the fastest appreciating assets in (00:05:04) history." (00:05:06) >> And I would say he was 90% right. I (00:05:08) modified the statement. I said, (00:05:09) "Foundation models without unique data (00:05:12) and internet scale distribution are the (00:05:13) fastest appreciating assets in history." (00:05:15) But reasoning fundamentally changed that (00:05:17) in a really profound way. There was a (00:05:19) loop, a flywheel to quote Jeff Bezos (00:05:22) that it was at the heart of every great (00:05:25) internet company and it was you made a (00:05:26) good product, you got users, those users (00:05:29) using the product generated data that (00:05:31) could be fed back into the product to (00:05:32) make it better. And that flywheel has (00:05:34) been spinning at Netflix, at Amazon, at (00:05:37) Meta, at Google, you know, for over a (00:05:40) decade. And that's an incredibly (00:05:42) powerful flywheel. And it's why those (00:05:44) internet businesses were so tough to (00:05:46) compete with. It's why they're (00:05:47) increasing returns to scale. You (00:05:49) everybody talks about network effects (00:05:51) much more and you know network effects (00:05:52) are they were important for social (00:05:54) networks. I I don't know to what extent (00:05:56) meta is a social network anymore. It's (00:05:58) more like a content distribution (00:06:00) >> but they just had increasing returns to (00:06:02) scale because of that (00:06:04) >> flywheel. And that dynamic was not (00:06:07) present in the pre-reasoning world of (00:06:09) AI. You pre-trained a model, you let it (00:06:12) out in the world and it was what it was. (00:06:14) And it was actually pretty hard. They (00:06:17) would do RLHF, reinforcement learning (00:06:18) with human feedback. And you try and (00:06:20) make the bot model better, and maybe (00:06:22) you'd get a sense from Twitter vibes (00:06:23) that people didn't like this, and so (00:06:25) you'd tweak it. And you know, there were (00:06:27) the little up and down arrows, but it (00:06:28) was actually pretty hard to feed that (00:06:30) back into the model. With reasoning, (00:06:33) it's early, but that flywheel started to (00:06:36) spin. And that is really profound for (00:06:40) these frontier labs. So one reasoning (00:06:43) fundamentally changed the industry (00:06:44) dynamics of Frontier Labs. (00:06:45) >> And just explain why specifically that (00:06:47) is like what what is going on? (00:06:49) >> Because if a lot of people are asking a (00:06:52) similar question and (00:06:55) they're consistently either liking or (00:06:57) not liking the answer, then you can kind (00:07:01) of use that like that as a verifiable (00:07:05) reward. That's a good outcome. And then (00:07:06) you can kind of use feed those good (00:07:09) answers back into the model. And we're (00:07:12) very early at this flywheel spinning. (00:07:15) >> Like it's hard to do now, (00:07:17) >> but you can see it beginning to spin. (00:07:20) >> So yeah, this is the part people still (00:07:22) seem to miss. Obviously, the real world (00:07:25) is way more complex and abstract than (00:07:27) something like a game of Go. But when (00:07:29) you zoom in on company functions and (00:07:32) cognitive labor more broadly, a (00:07:34) surprising amount of it is verifiable. (00:07:37) There's usually a clear outcome. Did the (00:07:39) task get done? Did it work or not? (00:07:42) That's why these companies believe a (00:07:43) large share of this work can be (00:07:45) automated over time. They just need a (00:07:47) lot more data and realworld experience. (00:07:50) This flywheel Gavin is describing though (00:07:52) honestly sounds very close to (00:07:54) self-improvement. Not in a sci-fi sense, (00:07:57) but in the very literal sense that the (00:07:59) system gets better the more it's used. (00:08:01) Now, this is where the conversation (00:08:03) really zooms out. If this reinforcement (00:08:06) learning flywheel has only just begun, (00:08:08) then the obvious question is who (00:08:10) actually gets to run it? Because (00:08:12) learning at this scale isn't free. (00:08:15) Reasoning models are expensive. Long (00:08:17) horizon tasks are expensive. And (00:08:20) learning from realworld use at global (00:08:22) scale is extremely compute inensive. (00:08:26) Which means the AI race stops being (00:08:28) about who has the best models and starts (00:08:30) being about who controls the (00:08:32) infrastructure. And this is where Gavin (00:08:34) makes a pretty strong claim about China (00:08:36) and Nvidia. Check this out. (00:08:38) >> And what's even more important, every (00:08:40) one of those labs, XAI, Gemini, OpenAI, (00:08:44) and Enthropic, they have a more advanced (00:08:47) checkpoint (00:08:49) internally of the model. Checkpoint is (00:08:51) just um you're kind of continuously (00:08:53) working on these models and then you (00:08:54) release kind of a checkpoint and then (00:08:56) the reason these models get fast (00:08:58) >> the one they're using internally is for (00:08:59) >> better and they're using that model to (00:09:01) train the next model (00:09:03) >> and if you do not have (00:09:06) >> that latest checkpoint it's (00:09:08) >> you're behind (00:09:08) >> you're it's getting really hard to catch (00:09:10) up. Chinese open source is a gift from (00:09:13) God to meta (00:09:14) >> because you can use Chinese open source (00:09:17) >> to try and that can be your checkpoint (00:09:20) and you can use that (00:09:21) >> as a way to kind of bootstrap this and (00:09:24) that's what I'm sure they're trying to (00:09:25) do and everybody else um the big problem (00:09:28) and the big a giant swing factor I think (00:09:30) China's made a terrible mistake with (00:09:32) this rarest thing you know I think China (00:09:34) because you know they have the Huawei a (00:09:36) sin and it's a decent chip and verse (00:09:38) something you know like you know the the (00:09:40) deprecated hop reserving something. It (00:09:42) looks okay. So, they're trying to force (00:09:44) Chinese open source to use their Chinese (00:09:45) chips. Uh they're domestically designed (00:09:48) chips. The problem is Blackwell is going (00:09:49) to come out now and the gap between (00:09:52) these American frontier labs and Chinese (00:09:54) open source is going to blow out because (00:09:56) of Blackwell. And actually, DeepSeek in (00:09:59) their most recent technical paper v3.2 (00:10:01) said like one of the reasons we struggle (00:10:03) to compete with the American Frontier (00:10:04) Labs is we don't have enough compute. (00:10:06) That was their very politically correct, (00:10:09) still a little bit risky way of saying, (00:10:12) you know, because China said, "We don't (00:10:13) want the black wells, right?" And (00:10:14) they're saying, "Guys, that might be a (00:10:16) big mistake. That might be a big (00:10:18) mistake." And so, if you just kind of (00:10:20) play this out, these four American labs (00:10:23) are going to start to widen their gap (00:10:25) versus Chinese open source, which then (00:10:27) makes it harder for anyone else to catch (00:10:29) up because that gap is growing. So, you (00:10:31) can't use Chinese open source to (00:10:32) bootstrap. And then geopolitically, (00:10:35) China thought they had the leverage. (00:10:37) They're going to realize, oh, whoopsy (00:10:38) daisy. We do need the black wells. And (00:10:40) unfortunately, they'll probably for them (00:10:43) um they'll probably realize that in late (00:10:44) 26. And at that point, there's an (00:10:47) enormous effort underway. DARPA has (00:10:49) there's all sorts of really cool DARPA (00:10:50) and DoD programs to incentivize really (00:10:53) clever technological solutions for rare (00:10:55) earths, you know, like using enzymes to (00:10:58) refine them or there's all sorts of (00:11:00) really cool things happening, you know, (00:11:02) and then, you know, there's a lot of (00:11:03) rare earth deposits in countries that (00:11:05) are very friendly to America that, you (00:11:07) know, don't mind actually refining it in (00:11:09) the, you know, traditional way. So, I (00:11:11) think rare earths are going to be solved (00:11:13) way faster than anyone thinks. You know, (00:11:15) they're obviously not that rare. They're (00:11:16) just misnamed. they're rare because you (00:11:18) know they're really messy to refine and (00:11:20) so geopolitically I actually think (00:11:22) blackwell is pretty significant um and (00:11:25) it's going to give America a lot of (00:11:26) leverage as this gap widens so this is (00:11:30) why infrastructure ends up deciding so (00:11:33) much of the AI race I mean if learning (00:11:35) depends on scale and scale depends on (00:11:38) compute then whoever controls the (00:11:40) infrastructure will just compound faster (00:11:42) that's why a company like OpenAI which (00:11:44) is only generating tens of billions in (00:11:46) revenue today is willing to commit to (00:11:48) infrastructure deals that are in the (00:11:50) trillions. And it's also why Gavin (00:11:53) thinks China risks falling behind unless (00:11:55) it continues to rely on Nvidia's most (00:11:57) advanced chips. So that's the story at (00:12:00) the geopolitical level. But the same (00:12:02) mistakes are being made at the industry (00:12:05) level, too. And according to Gavin, one (00:12:07) group of companies is especially (00:12:09) vulnerable right now. Take a look. (00:12:11) applications ask companies are making (00:12:13) the exact same mistake that (00:12:14) brick-and-mortar retailers did with (00:12:16) e-commerce. So brick and mortar (00:12:18) retailers um you know particularly after (00:12:20) the um you know the the telecom bubble (00:12:23) crashed you know they looked at Amazon (00:12:25) and they said oh it's losing money you (00:12:27) know e-commerce is going to be a low (00:12:28) margin business you know how how can (00:12:31) just you know from first principles how (00:12:33) can it ever be more efficient as a (00:12:36) business right now our customers pay to (00:12:38) transport themselves to the store and (00:12:41) then they pay to transport the goods (00:12:42) home. How could it ever be more (00:12:44) efficient if we're, you know, sending (00:12:46) shipments out, you know, to individual (00:12:48) customers, you know, and Amazon's (00:12:50) vision, of course, well, eventually (00:12:51) we're just going to go down a street and (00:12:52) drop off a package at every house. And (00:12:54) so, they did not invest in e-commerce. (00:12:56) They they clearly saw customer demand (00:12:58) for it, but they did not like the margin (00:13:00) structure of e-commerce. That is the (00:13:03) fundamental reason that essentially (00:13:05) every brick brick-and-mortar retailer (00:13:07) was really slow to invest in e-commerce. (00:13:09) And now here we are and you know Amazon (00:13:11) has higher margins in their North (00:13:13) American retail business than a lot of (00:13:15) retailers that are mass market retailers (00:13:17) you know so margins can change and if (00:13:19) there's a fundamental transformative (00:13:21) kind of um new new technology that (00:13:25) customers are demanding it's always a (00:13:27) mistake not to embrace it (00:13:29) >> and that's exactly what the SAS (00:13:30) companies are doing they have their 70 (00:13:32) 80 90% gross margins and they are (00:13:36) reluctant to accept AI gross margins you (00:13:38) know the very nature of AI is you know (00:13:40) software you write it once and it's (00:13:42) written very efficiently and then you (00:13:44) can distribute it broadly at very low (00:13:46) cost and that's why it was a great (00:13:47) business AI is the exact opposite where (00:13:50) you have to recomputee the answer every (00:13:52) time and so you know a good AI company (00:13:56) might have gross margins of 40%. (00:13:59) Now, the crazy thing is because of those (00:14:01) efficiency gains, they're generating (00:14:03) cash way earlier than other people, you (00:14:06) know, than other than SAS companies did (00:14:08) historically, but they're generating (00:14:09) cash earlier, not because they have high (00:14:11) gross margins, but because they have (00:14:12) very few human employees. And it's just (00:14:15) tragic to watch all of these companies (00:14:17) like you want to have an agent, it's (00:14:20) never going to succeed if you're not (00:14:22) willing to run it at a sub 35% gross (00:14:25) margin (00:14:26) >> because that's what the AI natives are (00:14:27) running it at. Maybe they're running it (00:14:29) at 40. So if you are trying to preserve (00:14:31) an 80% gross margin structure, you are (00:14:35) guaranteeing that you will not succeed (00:14:37) in AI. (00:14:39) >> Absolute guarantee. So yeah, what this (00:14:42) all essentially boils down to is if (00:14:44) you're not first, you're last. If you (00:14:46) wait, you fall behind. And if you fall (00:14:49) far enough behind, you don't get to (00:14:51) catch up, especially when it comes to (00:14:53) AI. That's true for companies and for (00:14:56) countries. Now, toward the end of the (00:14:58) interview, Gavin brings up one last (00:15:00) idea, and it's not really a prediction. (00:15:03) It's more like a consequence. If (00:15:05) intelligence really compounds on (00:15:06) compute, then eventually the constraint (00:15:09) isn't software. It's energy. It's land. (00:15:12) It's cooling. And that's where things (00:15:15) get a little wild. Just take a look at (00:15:17) this. (00:15:18) >> What other I'm I'm always so curious (00:15:19) about the polls of things. Like one poll (00:15:21) would be the other breakthroughs that (00:15:22) you have your your mind on things in the (00:15:24) data center that aren't chips that we've (00:15:26) talked about before. as as one example. (00:15:28) >> I think the most important thing that's (00:15:29) going to happen in the world in this (00:15:31) world in the next 3 to four years is (00:15:34) data centers in space. (00:15:36) >> And this has really profound (00:15:37) implications for everyone building a (00:15:40) power plant or a data center on planet (00:15:43) Earth. Okay? (00:15:45) >> And there is a giant gold rush into (00:15:46) this. (00:15:47) >> I haven't heard anything about this. So (00:15:48) please (00:15:49) >> Yeah. You know, it's like everybody (00:15:49) thinks like, hey, AI is risky, you know, (00:15:52) uh but you know what? I'm going to build (00:15:53) a data center. I'm going to build a (00:15:55) power plant that's going to do a data (00:15:56) center. We will need that. But if you (00:15:57) think about it from first principles, (00:15:59) data centers should be in space. Okay. (00:16:03) What are the fundamental inputs to (00:16:05) running a data center? There are power (00:16:07) and there are cooling (00:16:09) >> and then there are the chips. (00:16:10) >> That's like the total if you think about (00:16:11) it from a total cost perspective. (00:16:13) >> Yeah. And just the the inputs to making (00:16:15) the tokens come out of the magic (00:16:17) machines. (00:16:17) >> Yeah. (00:16:18) So in space you can keep a satellite in (00:16:22) the sun 24 hours a day (00:16:24) >> and the sun is 30% more intense. You (00:16:27) know you can keep it in the sun just (00:16:28) because like if the sun's here's this (00:16:31) you know you can have the satellite you (00:16:33) know always kind of catching (00:16:34) >> catching the light (00:16:35) >> catching the light. The sun is 30% more (00:16:37) intense and this results in six times (00:16:39) more irradiance in outer space than the (00:16:42) high than on planet earth. So you're (00:16:44) getting a lot of solar energy. Point (00:16:45) number one. Point number two, because (00:16:48) you're in the sun 24 hours a day, you (00:16:49) don't need a battery. And this is a (00:16:51) giant percentage of the cost. So the (00:16:54) lowest cost energy um available in our (00:16:58) solar system is solar energy and space. (00:17:01) Okay. (00:17:03) Second, for cooling in one of these (00:17:05) racks, a majority of the mass and the (00:17:07) weight is cooling. (00:17:09) >> And the cooling in these data centers is (00:17:13) incredibly complicated. You know, I (00:17:14) mean, the HVAC, the CDUs, the liquid (00:17:17) cooling. (00:17:19) In space, cooling is free. You just put (00:17:21) a radiator on the dark side of the (00:17:23) satellite (00:17:25) [laughter] (00:17:25) >> and it's as close to absolute zero as (00:17:28) you can get. (00:17:29) >> So, all that goes away and that is a (00:17:31) vast amount of cost. Okay, let's think (00:17:34) about um how this these, you know, maybe (00:17:37) each satellite is kind of a rack. It's (00:17:39) one way to think of it. Maybe some (00:17:40) people make bigger satellites that are (00:17:42) three racks. Well, how are you going to (00:17:44) collect connect those racks? Well, it's (00:17:46) funny. In the data center, the racks are (00:17:49) over a certain distance um connected (00:17:51) with fiber optics. And that just means a (00:17:53) laser going through a cable. The only (00:17:55) thing faster than a laser going through (00:17:57) a fiber optic cable is a laser going (00:17:59) through absolute vacuum. So, if you can (00:18:02) link these satellites in space together (00:18:05) using lasers, you actually have a faster (00:18:08) and more coherent network than in a data (00:18:11) center on Earth. Okay, (00:18:13) >> so yeah, I guess this isn't really a (00:18:15) wild take anymore. I mean, we're already (00:18:17) seeing some of the biggest names (00:18:19) seriously pursuing this, like Google, (00:18:21) Nvidia, and others. But if you bring up (00:18:24) space data centers to the average (00:18:26) person, they'll probably look at you (00:18:27) like you're crazy, even though this is (00:18:29) very much a real thing that's actively (00:18:31) being worked on. And it makes you wonder (00:18:33) how far off are we really from fullon (00:18:36) Dyson spheres? Because this seems like (00:18:39) the path we're clearly headed toward. (00:18:41) Anyways, huge shout out to Patrick (00:18:43) Oshanesy and Gavin Baker for this (00:18:45) incredibly insightful conversation. They (00:18:47) did a great job capturing where AI (00:18:49) actually is right now and where it might (00:18:51) be heading. But I'm curious what you (00:18:53) guys think about all this and what part (00:18:55) stood out to you the most. Was it the (00:18:57) part about China's mistake, the (00:18:59) reinforcement learning flywheel, or of (00:19:01) course the space data centers? Let me (00:19:03) know in the comments. Also, if you (00:19:05) enjoyed this breakdown, please consider (00:19:07) dropping a like, hit that subscribe (00:19:09) button, and as always, I'll be catching (00:19:11) you guys in the next