Sarah Guo and Elad Gil: The Future of AI Investing

26:16And it's really interesting, like where they're getting the most traction. There's some uptake amongst smaller and larger customers, but there is definitely a segment of the, let's say like analyst audience that does not like this, right? They're resistant to the idea of automation and like, you know, over a very small training period, we're like, oh, we can do 70 % of this, 80 % of it, but that may not be that compelling to them. It might be very compelling to their boss, right? And so I think a part of it is also where in the organization you are selling to, what the incentives are within that organization.

26:58And SMBs are interesting because like, they often don't want to do a bunch of the other functions, right? They love some part of their business and not necessarily like the marketing or the asset creation or anything else. And so it makes that sale much simpler. But I do think we will, I do think like economics and the democratization of capability wins out over the end. To Elad's point of taking it to the extreme, if you take one step back from that, it is interesting to see like this year, this like 2023 and the end of 2022 was certainly a first time in the last like decade and a half. I've seen people so ambitious to build companies efficiently in terms of headcount.

27:49Like we just backed a repeat entrepreneur who is trying to figure out how few people he needs to get to 100 million of ARR. And that's just a very different from like 2019 mentality. Totally. I think people just forgot, right? Because if you actually look at the history of tech, it's largely been capital efficient businesses, right? Microsoft was profitable and it's bootstrapped and it took in a round of funding right before the IPO because Bill Gates wanted a friend of his on the board, right? From a venture fund. and then you go into the 80s and Dell was bootstrapped and they took around record for the IPO, but they didn't need the cash.

28:26And then you go into the 90s and Yahoo never touched a dollar of venture capital at rates. Same with eBay. Google was incredibly capital efficient early on. And then you go into the 2000s and Instagram was 12 people and it was bought. And YouTube was like 25 people or something, 30 people, whatever it was. And so I just feel like people forgot that capital efficient businesses are often the best businesses because capital efficiency is a reflection of the willingness of people to really pay up relative to sort of human capital leverage inside of a company, right? To ratio of revenue per person is quite high.

28:59Or you have enormous pricing power. So I just think everybody forgot that and raised tons of money. And the metric of progress was valuation versus anything else. And I think capital efficient businesses traditionally have done very well and there's counter examples to that, but I think in both directions, but I think in general, people just forgot about that. And I don't think it's an AI wave thing. I think it's just the AI founders that I've seen and have been working with are just more hardcore in general. They've driven. They're really smart. In some cases, they've been working in this area for years and years and years.

29:31And this is their dream come true. And they work nonstop. And I feel like that's just a return to Silicon Valley as it used to be versus like something that new. But it does feel like AI can contribute, right? And accelerate this and like bring the ratio up. But when people say, oh yeah, you know, Copilot, like yeah, Copilot or ChatGPT makes 10 engineering person team one engineer. You're like, no, it doesn't. No, no. Like, what are you talking about, right? Have you ever actually tried the thing? Like, what are you talking about? And you hear this on podcasts all the time, right? Yeah. You know, Punbit's kind of talking stuff like this.

30:06You're just like, that's just completely false. Unless maybe your engineers are so bad that that's true. I don't know. You know what I mean? Like, you know, I think separate from that, there's the SMB question. And in general, people who run SMBs, if it's a true SMB, like a five person company, they're really busy and they're only going to buy like three or four things that they absolutely need to run their business. They're going to have payroll. They're going to have health care or other benefits. They're going to pay taxes, which is why Intuit exists, right? It's like Rippling, it's Intuit, it's Gusto, it's they need HubSpot because they need a market or they need some form of CRM.

30:39There aren't that many things that SMB buy and therefore SMB tends to be an awful market. for most things, right? And it's possible there may be some AI, SMB-centric things, but I think that the current leverage on human capital is a bit overstated, although I think that's gonna expand as we hit different levels of GPT, five, six, seven, eight, whatever. But also I think there are some verticals where it's really gonna eat away at the vertical faster than we all anticipate. And MidJourney is a great example of that. We're actually thinking it's displacing people faster than you would have guessed for certain types of jobs, as well as expanding the market.

31:13And I think there's going to be a handful of those things that hit with each successive wave of capability to their pre-diffusion models and transformer-based models. But I think today it's kind of overstated for other things like CodeGen. Changing it up a little bit, we've talked about MidJourney a little bit, OpenAI. We both, we all mentioned Pika a little bit. All of these companies are sort of these like full stack AI companies in that they have products and then they also have their own foundation models. Is this kind of the silver bullet for an AI company right now and that you have basically the whole thing?

31:46Is it the most defensible way to build an AI company right now? Or are these companies not defensible? Does it not matter that they have their own model and their own product in the same stack? I think when you talk to most researchers today, there begins to be a blurry line between fine-tuning on a sufficiently large amount of unique data and pre-training, right? Yeah. You know, assuming you're starting from some open source base model that is useful and significant, which is not really available before nine months ago. And so I think there will be companies of that form built and there are increasingly like good technical teams doing this in particular domains.

32:35Um, so I, I, I think like, you know, some of the companies you mentioned, like they're not language models, right? They're video models, they're image models. They work in a different domain. So we're just like, these things are apples and oranges. And so there wasn't a different option. Um, but I, I think that there will be application level companies that start from, um, some base model that are defensible in different, different ways. Right? I think this question is a good one about defensibility in AI applications because it's just like, you know, new market dynamics. But if you asked, like, what makes a SaaS company defensible a few years ago?

33:16Like, you know, there's one version of this, if you're just pure nerd about it, where you're like, well, I run Postgres in the cloud. And then I got like crud, you know, on my objects that look like kind of CRME objects. And like that describes a lot of companies that are pretty valuable. And so all of the things that like were valuable for SaaS companies before understanding your customer workflows and then like giving them some real ROI on something they were trying to do and actually being able to do distribution and go to market to those customers, like they remain really hard and really valuable.

33:51And I think the opportunity to be creative about rethinking those workflows is really interesting and the thing that like threatens incumbents, at least when talking to the incumbent founders. But interesting what you guys think here too. Yeah, what do you think, Alad? You know, I think the first wave of founders were all researchers who wanted to build their own models for everything, including companies that didn't need them. And so I think that first wave, every company raised like 20, 30, 40,$50 million. And half of those companies, I think, will make it in part because they took us on model building instead of customers.

34:27Someone should have been took us on customers and there's others who integrated it well, like Apika or, you know, I think when OpenAI launched ChatTPT, nobody thought it was going to be that big, including the OpenAI team. Low-key research preview. Yeah, it was a research preview that just took off like crazy, right? Because it was such an amazing thing. One of the reasons I backed character was because they had an internal version of that at Google, because I just know I'm being amazing, right? Which was MENA. And, you know, it was just so compelling to anybody that you talked to, interacted with.

34:58I don't know if you remember, there was like engineer Google who thought that the chat was sentient, right? Because it's so interesting and compelling. And that was basically ChatGPT in some sense. So I do think the first wave of people will build their own models. I think there'll be a subset of people that will continue to do so. Diffusion models are dramatically cheaper to train and build. than these very large language models. And so I do think on the image, video and audio side, we may continue to see that as a trend because you may differentiate on that. You may start with stable diffusion and then fine tune it and then decide you're just gonna train your own model over time.

35:31That's kind of the evolution you see a lot of these teams do. And then for the large language model, eventually you'll have a war between generalizability, scale and performance. And people will make different trade-offs along those curves. And in some cases, I don't mean I'm just going to use GPT-6 because it gives me really strong sort of logic capabilities and generalizable knowledge and capabilities to do things. And in some cases, you say, I just need this small performance thing that I'm just going to like, you know, add some rag to or whatever and just go, you know, and show. I think it's going to be a mix.

36:06And I think a lot of the application level companies won't need their own, their own bespoke bottoms up build model. and then again separate for them there's robotics and science and different aspects of science and physics and materials there you'll probably see more vertically integrated companies simply because of the domain that they're in and there may not be good capabilities and it may be a big differentiator so i think it's going to fragment it almost seems like every week you know there's a new model that puts out their benchmarks and says hey we're better than every other model And then a week later, you're seeing the same thing from another company.

36:42To me, that almost feels like everything is sort of converging and we're experiencing sort of a commodification of this technology. And thus, the most important thing, perhaps as it's always been, will be customer adoption, retention, sort of back to basics about what makes for a great product. Is that how you two think about it? I'd start with like maybe a minor part of what you just described, which are the benchmarks are a like weak proxy for how these things do in real world use cases. Right. And so like, you know, yes, Demi and Chen Lin at Pika, like, well, how do you know the models better?

37:26It's a very complicated answer. And it's not like this academic benchmark. Right. The way we do testing at Heijen with enough scale is user testing. right? What do people want when we deliver new versions of the model? I think in code, like instead of the academic benchmarks, which are competitive programming and like human eval is handcrafted Python data set from like one guy, right? Not to trivialize that work, but that doesn't represent, can I write code that executes in general production environments? And And I think the benchmark performance is interesting, but like still super academic research oriented and overstates like sort of the understanding that most people have of how different these models are in behavior.

38:22um and and so i think going back to um part of our conversation earlier i don't think we're gonna end up in a fully monolithic world anytime soon and people still want extremely different capabilities with very different data sets right um like can you get agents to work and take actions in software can you make a robot arm that does kick and pack robustly um just to take two things that are like not fully reliable today. It is, you can shape like a, that to, you can shape part of it to be a next token prediction problem, but it's not at all clear that gets us to the promised land. And so I think there's still lots of room for value at that layer.

39:08But then if you look at like text and structured data as a form of text, as a mostly enterprise investor, that is an area where like the core LLMs are gonna be very productive. And I think we're still really just like in the early exploit phase of it. And so I don't know that because people have access to slightly better models from a benchmark perspective that you won't see application level teams get things to work in real world environments in like dramatically different performance levels. They do need to figure out how to communicate that to the customer. Yeah, totally. That makes a lot of sense.

39:52They have to be able to convey the actual value of the different models. I think all of this assumes that, you know, chatting in a text interface is the way that people want to experience this next wave of applications. And I could see that happening. Open chat GPT has gotten very, very large. On the other hand, I could say that, you know, the history of computing suggests that people don't want to just interact with a chat interface. Right. That's why there are that's why the GUI exists. Right. That's why we tap on buttons with our fingers and our mice. Right. What do you two think about that?

40:25Like, do you think do you think that chatting with an interface can serve as a sufficient enough platform for, you know, a whole new crop of applications to exist? Sarah, what do you think? I think there's existence proof of this already, right? Like, man, there's a lot of aggregate revenue in the AI girlfriend apps, right? So I think, like, can chat be a compelling interface for an application? World says yes. Is that going to lead to an operating system? I can see both sides of this argument, right? I think it is unlikely to be chat only. I think it like if we see a new one, it'll be multimodal.

41:07The bull case argument is like, you know, this is the most natural interface. Everybody understands how to use it. We didn't understand intent. We could not have our computers understand intent before. And there is a there will be a killer general consumer application, not like everything, but something that will be the first app. and maybe it's something productivity related or search related with different UX. And it requires new hardware management and new resources. And then like, you know, we go from that to a platform. That's kind of the bull case. I don't know if I believe it yet, but you, I can see why the model providers or the founders want it to be true.

41:54And part of it is if you own the hardware and you own the operating system, you are not subject to like dealing with the mess that is Android compatibility or the like privacy and security constraints of Apple, including in experience, delivering experiences that consumers might actually want, right? Versus like siloed apps and being able to offer all your data to, all your relevant data to a particular application. I will, I'm very open-minded to believe it when I see it, but I don't think it will be chat only. yeah multimodal what do you think a lot do you think chat can be the interface for the next wave of applications and products and companies i think there'll be a series of interfaces i think chat actually works for a bunch of stuff um i think there'll be forms of multi-modality i think gptv is under discussed as like a really interesting um api slash product where you can take images and ocr them or understand the context of the image and then use them in different ways and So I think that's really powerful for all sorts of like enterprise application, defense applications, architecture, chip design.

42:59It could be used in all sorts of really interesting ways. And so I do think there's other things that matter. But I think fundamentally, you know, there are some things that just work right now. And usually people will modify those and change those and iterate on them and come up with completely different paradigms. But I do think it's something that people do naturally and that they like. And, you know, conversational interactions actually work really well for much of what you do as a person. That's how you interact with friends and family and others. And, you know, there's no reason to assume that you'd want to not do that with the machine, although there's other things that you do with the machine as well.

43:32So, you know, again, in the extreme, you probably assume that a lot of these interactions online collapse into agent-electing on your behalf. And so eventually you're not really that involved in many interactions, right? It's sort of the extreme of a couple of years from now or 10 years from now, whenever it is where you have agents representing corporations and governments and people in different ways. And then you end up with more like programmatic interactions across agents versus you having to interact that much with an interface. So I think these things will evolve and it really comes down to the underlying technology capabilities.

44:05And if you remember in the 90s, people came out with the first like PDAs, the personal whatever devices. It was like the proto smartphones. And they kept trying to do handwriting. Yeah, because they thought people would write everything, right? And of course, now we just type everything, right? But people thought, oh, no, people don't like typing, which everybody was doing. And instead, we need to write everything. And they came up with a special language called graffiti, where you learn to write the L a certain way and the R a certain way. because the machines weren't smart enough to understand how they can do.

44:38And that was really dumb in some sense. It was lauded as this brilliant breakthrough in how you think about human machine interface. But of course that doesn't use it all now. Right. And so I think there's a lot of fake hearings where people kind of overthink this stuff. Typing works really well and speaking works pretty well. I do think that part of the skepticism that's often common here is like, I think people conflate, like, is it a good interface for the user and is it possible, right? And if we talk about like a very specific example, just take like sales reps and CRM, right? Everyone hates updating their CRM workflow.

45:21And if you wanted to rethink that, updating their CRM records, right? Like Friday by 4 p.m. before my pipeline meeting, I must do this and it's a chore. And if you think about that operational toil and making it easier, like if you ask people, do you want to do it in fewer clicks? Like they're going to say yes. And so I think like to sort of picture a future where your software had more intent understanding and guess the next action you were going to take all the time. It's not just the text trigger of Mike saying, please update, you know, my company record. it's where you were and what your calendar says and what that meeting, the actual content of that meeting and what does update mean.

46:10But if it happened magically and you could just tell your computer to do things or an application to do things, like there's a really big question of whether this works, but it's kind of unquestionable that's easier. And so I think when people say like chat's a dumb interface, I do think chat without other data sources and assuming you have no multimodality is unlikely. But I think what they're really saying is like a command line without a man page that doesn't work the way Alexa did in like, you know, like seven years ago. Like, yeah, that was dumb. It didn't, it wasn't a really compelling interface.

46:47Yeah, it's super interesting. Like, you know, I can't really see myself, you know, I know there's been examples of GBTs and plugins and things where, you know, know, you book travel or you buy something. Personally, I can't really see myself booking a trip to Italy through a chat interface. But I think to both of your points, if there was an agent that just knew my preferences and knew what I wanted to do and then could just go do it for me, that I think that's a future I could see happening. And I think that makes a lot of sense. And it feels like that is hopefully a world in which we're headed.

47:19What are you about to say? But how do you communicate with a travel agent? It's a good point. The way you communicate preferences with a travel agent is kind of text. Or maybe if you're like really expressive, a mood board, but seems solvable. Yeah. Almost everything that we do is we call people for certain things. And, you know, it's interesting if you're interacting with an executive, you know, their minions will write these really long emails and then the exec will call and say like three sentences, right? Because they don't want to respond in detail and it's way easier to talk about certain things.

47:55And then similarly, there are some times where like text makes a lot of sense. And so I just think it's kind of like the midwet meme. People make all this stuff super complicated. And like often it just boils down to like, what's a dumb thing people already do for specific use cases? And that's probably what they'll do in the future. And we make humans do many things, including call people, where it's not a particularly intelligent communication, right? And so I think calls and chat, if you can break it down into the use cases, it's actually pretty interesting as a AI enabled capability. We are about to launch applications for our next batch of our accelerator for AI companies in bed.

48:36And, you know, we've gotten people doing early applications. And I think we've seen like three different companies in this area, all of who have traction, like in the last two weeks. And so I think there's demand for automation of some of these types of communication. Yeah. And I think per your points there, like this experience, if it's multimodal can be very, very powerful, right? Like the travel agent example. Yes. I could probably book a trip just by talking to somebody, but I probably want to see what they're about to book for me. Right. Or I probably want to, you know, I don't know, look at some of the nearby, I don't know, sites I can go check out.

49:16And obviously I need to actually look at that stuff. So that's really, really interesting. Speaking of interfaces, we are all of a sudden seeing a lot of hardware products, right? Hardware and AI. How are you both thinking about this? Do you feel like there will be new hardware products that emerge as a result of AI? Or do you think that there are going to be new experiences on existing hardware that could not happen without AI or both of those things? How are you looking at the combination of AI and hardware moving forward? I've been investing in this area for a long time. So like Samsara, as an example of that, they're a fleet management company that's not public that has in-device, in-cab, like hardware for fleets or, you know, I've been involved with Android since its earliest days and that's, you know, defense hardware plus AI or machine vision and machine learning.

50:09Square was basically initially a hardware device, right? So when I invested in Square, they were, they still had those little things that connect to your phone for taking the credit card. So I think in some cases, it's a real enabler and it's like a powerful way to gain capabilities, particularly if you're interacting in the physical world, which is what Sam Sara and Andrew will do. In the absence of that, I think a lot of the, for example, consumer applications will largely collapse into the existing platforms, there undoubtedly will be some counter example, right? And you also see applications where people use, you know, that that warring for sleep.

50:42And it's not necessarily just completely built into your iPhone or Android device versions of it that are on. But I think you really need a brand new capability separate system and do this other thing better. And so in the short run, I'm a little bit I think it's really cool experiments, and I'm excited to see where people take it. But I think a lot of the early iteration will actually end up being on the devices themselves. And standalone devices are going to be less performant unless they serve a very specific physical purpose, like what Andrel or Samsara or Square or some of these other folks did.

51:17I think that makes a lot of sense. And I think it especially makes sense given that the OSs and the existing players, they keep a lot of these capabilities kind of locked down, right? Like it's not easy to do things with Siri. iOS, right? It's not super easy to tap in to basically have the AirPods on 24 seven or something like that. I don't know if you remember this, there's actually an era where people said that there would be AirPod companies. Totally. Yeah, of course. All these businesses being built around AirPods, right? So it's kind of fascinating. It was hard to do. It's hard to do. And also the question is what's the real functionality gain and what are you, you know, generating off of it and all the rest.

51:51And so I just think like some of these things are also sometimes memetic. Yeah. And sometimes memetic startups ended up being very large. Like when Instagram started, there was like 2000 different photo app uploading companies and they really nailed it. But then there's often memetic companies where nothing comes out of it or memetic waves. And so I kind of wonder if a subset of these things are just these memetic waves that unfortunately won't necessarily translate. Although they made plant really interesting ideas for the future. Back to the mobile example we discussed, many of the ideas of what mobile phones would actually do were developed in the nineties.

52:22We just didn't have the capability to actually do it. And this feels to me like in some cases, perhaps overlapping. So I guess, you know, Sarah, if you agree with that, I guess then what do we think the potential new experiences on existing hardware platforms might look like? You know, I think of Uber could not have, you could have apps on an iPhone, but Uber could not have existed without the GPS, right? Or, you know, Instagram's another great example of an app that only could have existed because of the camera. Are we going to see new types of apps on existing platforms that only could exist because of AI?

52:57And what could those be? I think the emergence of these is like both guaranteed and like really difficult to predict in the specific. Right. I think there are going to be different versions of exactly what I said, where the workflow is kind of hard to picture because people don't believe in the technology because they've seen a version of it that doesn't work. Right. In the CRM example. and you can imagine like for example action taking that is cross application on your device but you know those commands don't work today in terms of like a lot's point of view like another framing I have here is should there be new hardware because it's for the user like there's some application or is it because someone wants to train the model with the data from the device, right?

53:50Because the latter is not an answer. Like it's not a reason anybody's going to buy the device. And so like if you're doing something new, like do you actually need new hardware? Maybe, right? And obviously if you give people like new hardware capabilities, then you're going to get new applications if you already have the distribution. But we're just talking chicken and egg on the killer application. Yeah, I can imagine just to like make explicit, I think what you were implying, like battery management, different sensors, different permissions management, contextual intelligence that enables these better experiences is the biggest.

54:29Yeah, it could be glasses. It could be always listening. It could be anything that is like passive image and audio output or just data from your different web services. and like being able to like reach that in some new or existing form factor. But I think that's the biggest argument for it. And the question is just like, can you do that in an application? Where does the compute happen? Like, will the ecosystems permit it? It's complicated today. I'll throw out three other examples to each other examples. I think robotics is one of those areas that's becoming increasingly interesting. There's been a series of research papers that have come out even over the last like four to six weeks that I think are kind of fascinating there in terms of moving more and more into like more standard deep learning world for this stuff.

55:16And then the other area, obviously, self-driving. And, you know, that's just a form of, again, robotics and hardware and everything else. And so I do think there are these big areas that will be increasingly tractable as we apply some aspects of foundation models to some of these domains in the physical world and have specialized hardware for them. But it isn't that, you know, consumer hardware device that seems to be on trend right now. Yeah. Let's talk about everyone's favorite topic, Web3. There have been a number of people coming out recently and companies being started and launched that is basically saying that one of the killer applications of Web3 is actually going to have a lot to do with AI and content and content authenticity.

56:07I think Fred Wilson even wrote something about this. Scott Belsky has talked a little bit about this. If we have companies and models from companies like Pika and Midjourney and, you know, name your product that generates some form of media, and it's all trained on some other form of original media, at some point, we will need a way to track the authenticity or the origin of this stuff to enable an equitable right structure. Web3 could be the way that that is done. How do you both think about that? And what are you thinking about the opportunities for Web3 and AI? I think a couple of the concepts you mentioned are classic application areas, right?

56:53Identity, providence. I think the distance between we are going to have creators or existing enterprises that own IP, use crypto technologies to prove provenance is pretty far from a usability perspective. but I think like from a technical perspective, it seems obvious like as one of the solutions that is possible. I think in the short term, you're just going to get declarations from companies that have rights agreements and give creators certain commitments, use this as a way to differentiate their offerings. What do you think, Alad? Yeah, I've long thought that there'll be some blockchain resident form of identity that'll be used potentially for both the ascertainment of origin of content, but also specifically for credentials for agents.

57:58And so if an agent is representing you, how do you know that it's actually representing a specific individual? And can it partially reveal data or aspects of that person cryptographically in ways that are secure? So, you know, that may be your healthcare data, that may be aspects of who you are or other things. So, yeah, it's an area I've been interested in for a while. We actually had Ilya, the co-founder of New York, who also was the last author on the Transformer Paid for Heart podcast. And I also did a fireside with him in Canada. And you know, I think a lot of these concepts are really interesting areas, but it seemed like beta evolved over time.

58:34It's a big question of timeframe. I think the technology is there for the identity, not necessarily for the agents. From a provenance perspective, the content stuff, I think it's a little bit complicated because say, for example, you have an enormous amount of derivative art on the internet already for Van Gogh. Say that you actually removed the original Van Goghs from your corpus of data that you train on, you'd still have enough signal in the style of Van that it doesn't matter. That's right. And those works are all in the clear from a copyright perspective, right? Or from a, uh, usage perspective.

59:11And so I also think there's things like that, that people aren't really discussing that in terms of provenance and, you know, what actually exists and what is a training set, because it's possible you could remove all original art for one, one artists and it won't matter. Right. Because it's just already out there, right? It's like, it's, it's already been implied enough such that somebody else is inspired by it. Correct. In a legal way, right? Like you are allowed to do fan art and you are allowed to do things in the style of Van Gogh by hand. And so I just think a lot of these issues are going to be, um, perhaps they may play out a little bit differently from what people think.

59:43That's obviously very different from, Hey, you know, what's happening with the New York Times article or whatever, and is it being copied verbatim or not? And there's nuances around that as well, because of the metadata that's being exposed on some of these websites. but i think um at least on the image gen side it's a little bit uh trickier i think than people think and it may also be harder to uh protect certain artists because there's already so much derivative art yeah it does seem yeah i agree with you the new york times thing is is is different in that it's it's very clear when the output of this thing is verbatim uh to the to the content it was trained on um however as you said with images video other forms of media it's going to be a lot more complicated and it does seem like you know there there is starting to be a discussion uh in dc about um about the legality of of training obviously none of us uh work in dc but like how do you two see this thing playing out i mean theoretically it kind of makes sense to to me that that training would be fine because that's how humans learn right we read we get inspired We create.

1:00:50Why should machines be any different? But it does seem like this conversation is starting to happen. How do you think it's going to play out? It's a political question. So it depends on who's in office as well. But I think there's already like a strong case for fair use, assuming that there's a set of safeguards around what model outputs look like. And I think people will go actually fight these battles at the application layer and at the proven output capability layer. And I think that's the right place for it. But I do think going back to what Elad said, there are like one of the other core use cases for blockchain based technologies is enabling microtransactions cheaply.

1:01:36Right. And there's still work to be done here. But I think one of the things that Ilya said that still really resonates with me is a blocker at some point to continued progress in model training is collecting data we don't have yet. and getting people to contribute to that. And having, you know, systems for those transactions for labeling and contribution of data that is resistant to abuse, which sounds like, you know, an identity and reputation system, I think is going to be really valuable if it can be figured out. A lot, Sarah, this has been awesome. If people want to find out more about the podcast, No Priors, where can they find it?

1:02:21Snow-priors.com, or you can search for it. You can perplexity for it. Anything else either of you want to plug? No, it's a great conversation. I really appreciate the time of that. Cool. This has been awesome. Thank you both so much. Thanks for listening to Generative Now. If you liked what you heard, please rate and review the episode. That really does help. And if you want to learn more, follow Lightspeed at LightspeedVP on YouTube, X, LinkedIn, and everywhere else. Generative Now is produced by Lightspeed in partnership with Pod People I am Michael McDonough, and we will be back next week with another awesome conversation.

1:02:58Thanks so much.