AI Round Up: Ari Morcos from Datalogy AI and Rob Toews from Radical VC on AI Talent Wars, xAI’s $200B Valuation, & Google’s Comeback

This episode features a deep dive into the current state of AI model progress with Ari Morcos (CEO of Datalogy AI and former DeepMind/Meta researcher) and Rob Toews (partner at Radical Ventures). The conversation tackles whether model progress is genuinely slowing down or simply shifting into new paradigms, exploring the role of reinforcement learning in scaling capabilities beyond traditional pre-training. They examine the talent wars reshaping AI labs, Google's resurgence with Gemini, the sustainability of massive valuations for companies like OpenAI and Anthropic, and the infrastructure ecosystem supporting this rapid evolution. The discussion weaves together technical insights on data quality, synthetic data generation, and RL environments with strategic perspectives on acquisitions, regulatory challenges, and the future intersection of AI with physical robotics and brain-computer interfaces.

• 0:00Speaker 00

  Never bet against Elon is a sound mantra.

• 0:02Speaker 01

  I would also say don't bet against Zuck.

• 0:04Speaker 02

  These are all environments seen all the rage these days on the venture side.

• 0:06Speaker 00

  There's like a tidal wave of cash that is about to come down on anyone branding the barn.

• 0:11Speaker 02

  Is model progress slowing down?

• 0:13Speaker 00

  In the spirit of making this a provocative spicy discussion, I'll disagree with Ari.

• 0:17Speaker 02

  Do you think this MSL experiment is going to work?

• 0:21Speaker 02

  Today, we're introducing a slightly different format to the show that I'm super excited to try. We'll still be doing our one-on-one interviews, but we'll also be adding a news roundup with some great minds in AI today. No better way to start with them with the two co-hosts we have right now. Ari Marcos is the CEO of Datalogy, a really interesting state-of-the-art data curation company. And he previously spent two years at Google DeepMind and five years at Meta as an AI engineer. We also have Rob Taves, partner at Radical Ventures, where he's led investments in companies like Writer, Hebea, and Datalogy. And he also writes a really interesting column at Forbes. We hit on all the big AI topics today. RL environments, valuation of model companies, the AI talent wars, and even hit things like brain-computer interfaces. This was a ton of fun to do. We'll still be doing our one-on-one interviews, but I hope folks really enjoy these kind of episodes as well. Would love feedback on what we've done here.

• 1:04Speaker 02

  Please DM me on any platform and subscribe to stay tuned for all future episodes. Without further ado, here's Ari and Rob.

• 1:13Speaker 02

  Rob, Ari, thanks so much for doing this.

• 1:15Speaker 00

  Thanks for having us, Jacob.

• 1:16Speaker 01

  Yeah, really excited to be here.

• 1:17Speaker 02

  I've been looking forward to this one for a while. I feel like we often have lots of conversations off mic about a bunch of this stuff. And so I figured it would be fun to bring it on to the podcast. Before we dive into, you know, every topic on everyone's Twitter feeds, I felt like it would be good to just let you guys quickly introduce yourselves. So the audience has a bit of context. Maybe Ari, if you want to kick it off.

• 1:36Speaker 01

  Yeah, hi, I'm Ari Marcos. I'm the CEO and co-founder of Datalogy AI. Prior to founding Datalogy, a couple of years ago, I was an ML researcher for about a decade. Spent some time at DeepMind and five years at Fair at Meta AI. And at Datalogy, we work on all things optimizing data for building the best models possible, which can make massive impacts in the costs and performance of the models that you train.

• 2:00Speaker 02

  Yeah, it's tough having an actual researcher on the show when it comes to takes because you really will correct if the VCs go too off what they know. Fun to have around and Rob yourself?

• 2:12Speaker 00

  Yeah, there's going to be like actual substance on this podcast for once, which will be nice.

• 2:16Speaker 02

  At least when Ari talks.

• 2:17Speaker 00

  Yeah, yeah, when Ari talks. It's great to be here, looking forward to the conversation. My name is Rob Taves. I am a partner at Radical Ventures and lead the firm's Bay Area office. Radical is a VC firm, totally focused on AI with close ties to the AI research community, close relationships with a lot of AI pioneers like Jeff Hinton and Fei-Fei Li and so forth. I've won a number of Radical's investments over the years, including in Datalogy and others that I'm sure may end up coming up as we chat. And prior to becoming a VC, I worked in autonomous vehicles for a long time. So yeah, excited to be here.

• 2:54Speaker 02

  Amazing. Well, you know, two great folks to have to hit the topics of the day. I figure no better place to start than the discourse that's been happening around model progress. And I feel like there's some very vocal camps on Twitter and other places around whether model progress has slowed down. There's been people that have said, look, you know, GPT-5 was really just GPT-4.5 with a router. Clearly, maybe you can hill climb on individual areas like coding, but the era of large general improvements in generations of models is slowing. And other people say, what are you talking about? Look at these awards that these model companies seem to be winning in math and coding and all these other spaces. Maybe Ari will start with you. Which side of this debate do you come down on?

• 3:39Speaker 01

  I don't think it's slowing down significantly. I think it might be a little bit harder to measure it and to see it. You know, I think there's a reality that for most of the tasks that consumers want, these models are already quite good. And therefore, the difference that an average user is going to experience might not be as strong. I think we're starting to see progress on a lot of other axes, like Reasoning is obviously a big one. Cost efficient effectiveness of models is a big one. We're seeing smaller and smaller models now match the capabilities of massive models from a year ago. That's a really massive thing for actually deploying models and making them more useful more broadly. I think the one thing that we have seen that has been maybe disappointing to folks, although I would say is actually very expected,

• 4:21Speaker 01

  is that we've seen that just scaling itself has started to hit diminishing returns. And that's something that was always supposed to happen. If you actually go and look at the scaling laws work from Jared Kaplan and others, it says every time you 10x the data or the compute, you get a diminishing return. So we've seen that actually happen with, you know, large models like GPT 4.5 and and Llama 4 and the Behemoth and things like that. But we've also unlocked all these other axes of scaling, which do work quite well. So overall, no, I don't think models are really slowing down. I think it's just changing the exact timbre of where we see the progress.

• 4:55Speaker 02

  Yeah, I mean, I think you're you're how you articulated it exactly. Maybe the the crux of this disagreement where some, you know, some of these areas where models are improving feel like they're for, you know, folks who code or for scientists. Or that's clearly where the focus of the model companies is right now. And I think there's this question of does the general user of chat GPT or Claude, you know, benefit from these areas of progress in the models over the next six, 12 months? Rob, what do you think?

• 5:18Speaker 00

  Yeah, I guess to to in the spirit of making this a provocative spicy discussion.

• 5:23Speaker 02

  Question one, and we're already going for it. Let's go.

• 5:25Speaker 00

  Yeah, diving right into it. I'll I'll I'll disagree with Ari. I take my other side. I do think progress is slowing down.

• 5:33Speaker 00

  Just in terms of the the sheer raw intelligence and capabilities of the underlying models, I think that like the reality is that no exponential lasts forever. And every every exponential in nature ends up looking like an S-here. And I think that's what we've seen with this era of language models and foundation models. I think that the leap from GPT 2 to GPT 3 and then from GPT 3 to GPT 4 was just astounding. And I think we're certainly starting to see a plateauing of that. And a big part of it, as Ari said, is is the diminishing returns from from from sheer pre-training scaling. And that's that's obviously happening and has been pretty well documented.

• 6:15Speaker 00

  And then I think there was a like, there's kind of this narrative right around the time last year where it became clear to people that pre-training really was slowing down and showing good mission returns. There's this narrative of scaling post-training computes is kind of like the next wave that will let us continue this ongoing march of progress and, you know, test on compute and based on reinforcement learning and so forth. And so I think that's like, that was a narrative that got a lot of people excited. And it certainly is the case that reasoning models based on scaling test on compute can demonstrate novel capabilities that are exciting and that have really valuable use cases, in particular in areas like coding. But to me, it's not clear that, like, I don't think

• 7:00Speaker 00

  that that paradigm is enough to continue the pace of progress that we saw in like the 2020 to 2024 era. In particular, on the RL side, and I'm sure we'll end up talking about this more in the podcast, but I think it's, RL is clearly very powerful in areas where there are very clear, obvious, verifiable answers and kind of objective abilities to determine right or wrong, like, you know, like math and like coding, which is why the progress in those areas in particular have been so astounding. But it's less clear in hard to verify domains whether RL really will work as well. And so I do think that,

• 7:43Speaker 00

  I think that we're at a place where to really turbocharge and advance to the next level of intelligent capabilities, like more fundamental breakthroughs are needed that are not immediately obvious to the general public. But it doesn't mean that they're not being worked on in the big labs and so forth. But I do think that, like, the current paradigm, I do think, is slowing.

• 8:02Speaker 01

  I'll just add one thing to that. I think the place where we're really seeing pre-training plateau, right, has been just going to larger and larger massive models, that that doesn't really seem to work. The scale parameter counts up, scale two, you know, trainings that cost billions of dollars, things like that. But all of the benefits we've seen in RL are happening because the base models got good enough to enable that. Right? So I I think that it's also this S curve thing that Rob mentioned, I think is absolutely correct. Right. Is that ultimately any technological progress is a series of S curves. And I think that from the consumer kind of broad, general user, the way your mom might use chat GPT. I think we've definitely seen that start to plateau in the sense of just like, what is it missing now that they really need, right? It's not fully reliable, so we can bring hallucination down a bit more, things like that.

• 8:47Speaker 01

  But it's, it's pretty good for those use cases. And I think that contributes a lot to this perception of, of slowing progress. Cause if you're using these models, is, is GPT five that much better than four O? It's probably a lot cheaper to deploy than four O. But as a user, do I care?

• 9:03Speaker 02

  I mean, I think to Rob's point, the, the question is like, what is the, where are we? And what is the shape of this RLS curve, right? And I think the big questions behind that are, you know, the extent to which this works outside of these easily verifiable domains. And then also the extent to which it generalizes, right? It's, it's kind of unclear to me whether the future is just like, you know, hill climbing on like a data set specific worth of, of domain knowledge. And then moving on to the next one with nothing to show for having done the first one. Or whether there is some generalization across, I'll post to the AI researcher here. What's your like intuition around that, Ari, and how it plays out in the next year or two?

• 9:35Speaker 01

  So, okay, so I spent a lot of my career as a researcher working on, on generalization. And in particular, out of distribution, generalization.

• 9:41Speaker 02

  That a good, good person to ask the question to.

• 9:43Speaker 01

  Yeah, exactly right. And like the problem there always, right, is like, can I teach something in one domain that then will generalize to others? You'll hear me say this a lot. Ultimately, it's all about the data that you show the model. Models are what they need are my favorite catchphrase. And, you know, that's why I started Datalogy. But I think one of the, the tricks of kind of self-supervised learning in the modern era of AI is make it so that nothing is out of distribution. The trick to solving out of distribution, generalization is make everything in distribution. And from training on the entire web, that actually can work quite well. You also are going to look at a lot of data that's not going to teach the model anything. And, and that can make a pretty massive difference in the efficiency of this. But I think it still really remains to be seen with RL.

• 10:26Speaker 01

  Cause I think the challenge is when you build an environment. It is going to be fairly specific. And there are lots of ways to reward hack these environments. And these models are extremely good at finding any difference between the reality and the simulation that you have of the RL environment. This is actually one story that is for monkeys, but is actually very related because like I originally came from neuroscience, actually. So I did my PhD in neuroscience and I worked in a monkey lab. And you would teach monkeys to do these tasks where they would look at a screen and then based off of various things, they'd have to move their eyes to give answers and get rewards. And one of the things we consistently found was that if there was a way to cheat, the monkey would find a way to cheat.

• 11:09Speaker 01

  And one example of this that was really shocking to me was then the monkey was solving a task in a way that just made no sense. If the monkey was actually learning the task, it would never have solved it this way. It would the way based on the pattern of correct and incorrect answers. Turned out that in the correct case, the luminance of the screen, the brightness of the screen, if you perceive it and take a picture, was about half a percent higher. This is entirely imperceptible to human.

• 11:33Speaker 02

  That's what you framed your whole outline with them on.

• 11:34Speaker 01

  And the monkey, that's what the monkey learned. The reason I use this example is that this is exactly what we will see with AI models and RL environments. If the environment is not perfect, it will find a way to overfit to it and then it will never generalize. So there's this real kind of adversarial challenge of how do you build environments that are accurate and representative enough that the only way to solve the task is to learn the general purpose principle. And I think it very much remains to be seen whether that will work or not. There are some domains where that makes sense, where it's easy to verify, like code, as Rob just mentioned, most of the real world is not like that.

• 12:12Speaker 02

  I think these are all environments seem all the rage these days on the on the venture side, Rob, I'm sure you've met plenty as well. I'd say broadly, maybe fall into two categories, or at least how they think about end customers, right? Some are, you know, say, look, all this work is going to look like data labeling where the vast majority of the work happens, you know, within the labs. You want to sell basically to the to the six folks that are that are doing the stuff at scale. Other folks saying, look, actually, given that these, you know, RL tasks don't generalize, there actually will be. A broad swath of people that are going to build and are all their own models for their own use cases and, you know, go serve those folks. Any early instincts from either of you on which one of those approaches is more likely to play out?

• 12:51Speaker 00

  Jacob, as you said, it is, it feels like this past summer, dozens and dozens and dozens of RL environment startups have popped up. It's definitely the topic du jour. I do think, you know, when I think about that first category you described, the sort of companies developing environments to sell to the few big labs that have these massive budgets for it. I do think it's no secret now that the biggest labs, OpenAI, Anthropic, XAI, have set aside massive budgets to spend on RL environments. And they're doing a ton internally, but they're also willing to spend a ton externally. They just want to scale their access to RL environments as fast as they can.

• 13:30Speaker 02

  There is no limit to those budgets, basically. If there's something good to buy, they will, they will go buy.

• 13:34Speaker 00

  There is no limit to those budgets. And so there's like a tidal wave of cash that is about to come down on anyone writing these environments. And so as a result, there's no surprise that all these startups have popped up. And I do think that many of them will scale very quickly in terms of top line. I think the question is like, is there a durable business here? And what does the longer term trajectory look like? Because I do think that RL environment generation will become increasingly automated, or the needs of these labs will shift as time goes by. And so, you know, five years from now, do I think that there's going to be massive budgets spent on this particular product? Probably not. So I think that, I think it really comes down, like thinking about it through the lens of a startup or a VC, I think it really comes down to like,

• 14:17Speaker 00

  what will the act two and the act three be for these RL environment companies? And frankly, like, how effective will the founding team be at bobbing and weaving and finding the next opportunity and finding the next opportunity after that?

• 14:31Speaker 00

  It's common in this category, it's common to analogize this category to scale AI. And you often hear companies describe themselves as like, you know, we're scale for the RL era.

• 14:41Speaker 02

  They do it because VCs like those really simple analogies, right? Oh, great. Huge outcome. Sounds good.

• 14:47Speaker 00

  Yeah, dumb it down for me. Oh, yeah. Great exit. Okay, I'm sold. But I do think at a high level, like, it's a fair analogy. And I think it's also helpful in the sense that scale shape shifted several times on its path to being a successful business because Alex Wang was such a special CEO and was able to see around corners. And, you know, for a while, they were making almost all their money selling autonomous vehicle companies and labeling, you know, drawing bounding boxes around autonomous vehicle data. And then they saw this big opportunity in US government and defense, and they really scaled massively selling to USG. And then obviously, most recently, they really rode the RLHF wave and, you know, labeling preference data for RL.

• 15:30Speaker 00

  And so I think like, I think if you're at Alex Wang, that's an exciting place. Like, right now, this is an exciting place to be, and you'll kind of bob and weave. But so I think it really, the founders that you're betting on, I think, matter so much. And then maybe just briefly on the second category, which I kind of like to think of as like RL as a service, almost, like selling RL capabilities to enterprises, to basically build customized personalized models on their own data that improve and learn their own enterprise environment. I think that category is more interesting, honestly. And it's not easy to do. And I think that like, all of those businesses will involve some degree of professional services, basically.

• 16:15Speaker 02

  We call it forward deployed engineering around here. It's no longer, you know, not services at all.

• 16:20Speaker 00

  Yeah, yeah, yeah. As long as Palantir keeps trading at 100x revenues, the forward deployed engineer model will be trendy. But yeah, so I do, I like that model better. And I do buy the narrative that. RL is like environments are endlessly diverse and therefore there's value in experts that can help enterprises apply RL in their in their setting.

• 16:43Speaker 01

  I generally agree with Rob on the first point. I think durability is the main question for me. I think data collection and RL environments are not entirely analogous. There are a number of key differences between them. And I think it remains to be seen how those differences manifest in the different incentives. One of them that I think is quite big is that in many areas, if you build a lot of RL environments, you will get better at building other RL environments. I don't think that's generally true if you're going for expert data or things like that. So I think it makes more sense to have this in-house. I think this is also pretty straightforward as to how you would go about trying to automate this. And many are trying quite hard to automate this. I think another big thing is that if you look at the data annotation platforms like

• 17:24Speaker 01

  Scale and Surge and stuff, they almost always built bespoke data sets for their customers. They did not build a single data set and then sell it to the same six customers. The RL environment companies are much more generally in that path of I'm going to build this an environment and I'm going to sell it to everybody. So then you don't get a differentiator between them. If you truly believe that RL environments are going to be the differentiating factor between the frontier model labs, then the frontier model labs would have a very strong incentive to buy or and build RL environments that they only have access to. So I think that's a difference in kind of the the economics of things that that would change potentially where this goes. I do expect that this probably won't last for a very long time.

• 18:09Speaker 01

  So whether or not you can build a really great business in that time and whether you can pivot when things change makes sense, but I would expect that that a lot of this would be brought in-house. To the second point, I'm quite bullish in this direction. In general, I think the consumers face and the enterprise face, I think, are very, very different. And I think we often think of them as one, but they're totally different with respect to what does a consumer need versus what does an enterprise need? A consumer needs a model that's a mile wide and an inch deep. It needs to be able to answer a question about anything to a reasonable degree of accuracy, but it doesn't need to be perfect. An enterprise has the opposite need, right? They need in general a model that's an inch wide and a mile deep. It does one thing or one small set of narrow tasks, but it does it to five nines reliability. It does it for as low a cost as possible, et cetera.

• 18:54Speaker 01

  If you want to build models like that, the barrier to entry is it expensive? It's expensive and there's expertise, right? And all of those barriers are getting eroded dramatically. That's actually a whole part, a huge part of what we do at Datalogy. And our mission is that if you want to train your own model, you want to build your own model, you need to have good training infrastructure, you need to have good data. The training infrastructure side has gotten a lot easier over the last couple of years, Mosaic first, but now things like SageMaker and Together and lots of other platforms. The data side is still really hard. But I think there's this myth that in order to, for an enterprise to build or customize their own model, they're going to have to spend hundreds and hundreds of millions of dollars. And obviously, there are many folks who are very strongly financially incentivized to propagate that. And I think that's just not true.

• 19:35Speaker 01

  We have tons of proof now that you can build an amazing custom model for order single-digit millions of dollars, and it's going down. I think that's going to make doing custom RL on these models very desirable and attractive. And I would expect we see a lot of demand for that more broadly.

• 19:50Speaker 02

  I feel like for a while, you know, building your own model was, was around like, figure out what the, the state-of-the-art best OpenAI Anthropic models can do, and then build a lower-cost version yourself. And now, it actually seems that there's some real promise to improve capabilities through, through a bunch of these techniques, and actually push state-of-the-art forward.

• 20:07Speaker 01

  Especially with domain-specific and relevant data, right? If you want to build, and you have a bunch of proprietary data that exactly matches the task that you want to do, that's really valuable. If you want to use that data to build a general-purpose web model, that's not going to be nearly as valuable.

• 20:22Speaker 02

  Well, I want to, I want to hit the kind of broader foundation model space and a bunch of things to talk about there. I figure maybe to start, you know, this past week, Gemini, I think for the, for the first time, got to the top of the App Store, I think largely driven by Banana Banana. But I wanted, you know, both of your reactions to this like, long journey from, from publishing attention is all you need, to being called way behind on everything, to now top of the charts. All right, maybe I'll start with you, because I know you were at DeepMind from, from 2016 to 18.

• 20:49Speaker 01

  Yeah, I wonder if Robert remembers, but I've been very bullish on, on Google and Gemini for a long time. There's too much talent there. They, they have amazing folks across the board. I think the problem that Google has fundamentally generally had, right, is that they haven't Been good at launching new products for a long time and they they haven't had that kind of very focused mindset of we need to win, we need to to really get there. I think that's clearly changed in the Gemini team over the last year. Sergey apparently is spending a lot of time directly with the team. I think that probably leads to some pretty different, you know, and a different energy level. And when you get Google firing on all cylinders, I think you see the sort of rapid ascent that that we see here.

• 21:34Speaker 01

  And then I think you pair that with Google's distribution advantage and they're going to be quite a behemoth.

• 21:39Speaker 02

  Yeah, Rob, do you expect them to have like the the best model in two, three years? Is this gonna gonna gonna continue?

• 21:44Speaker 00

  Yeah, I do. I am very, very much on the same page as Ari. I'm I'm super bullish on Google. In the wake of ChatGPT, there was this narrative that Google lost a step and they got complacent and OpenAI out executed them. And there is a lot of truth to that. But I do think to Ari's point, like, Google just has so many structural built in advantages. Like, they have the resources, they have the computes, they have they have such a deep batch of talent. That, yeah, I'm I'm bullish on them. I think that I'm expecting, there's rumors that Gemini three is dropping at some point in the next couple of months. Like, my expectation is that it's gonna be very good. I think it's gonna be state of the art. I think it's gonna be better than GPT five.

• 22:24Speaker 00

  And, yeah, I I would expect Google to, if I had to pick one of the major labs, I would expect Google to be advancing the frontier as much as anyone going forward.

• 22:34Speaker 01

  I think there's an interesting question about how this will affect Google from a financial perspective, like, because they're gonna, I think they're gonna win this. But they're gonna now end up going to a platform that's much harder to monetize and that has a much higher cost. So they're gonna get squeezed on both sides. So I think there's an interesting note. I think there's gonna be like a challenging time from a revenue and profit perspective, probably as a result of this transition. But then I think probably Google will come out of that. They'll fix, they'll figure out how to monetize it appropriately. They'll figure out how to bring the cost down. They have all the in-house chips to do that effectively. And then you'll start to see all of that kind of start rising again pretty rapidly is my guess.

• 23:06Speaker 02

  Yeah, I mean, obviously, a huge threat to both OpenAI and Anthropic, right? I mean, if you think about these most recent funding rounds that that each of those companies Raised, I think, OpenAI is predominantly based on ChatGPT, which obviously Google has a ton of resources to compete with on the consumer side. And then Anthropic is predominantly around enterprise API revenue, which is really coding. And Gemini 3 comes out and is really good at coding, you know, could easily see folks switching.

• 23:32Speaker 01

  Defaults are really strong, right? Like, so I think that clearly ChatGPT has brand awareness now.

• 23:39Speaker 02

  For consumers, certainly.

• 23:40Speaker 01

  Yeah, for consumers, I think it's huge, but it's also there, right? So like, I actually wonder, like, for ChatGPT, like, I think everyone knows about ChatGPT. They have escape velocity, I think, pretty clearly for the brand. But like, when Apple does deploy a good model on the phones, which they will do eventually, right?

• 23:54Speaker 02

  Hold on, I'll run.

• 23:56Speaker 01

  Like, when they do that and you have this default model on all Google Android phones, it's going to be Gemini and all Apple phones is going to be whatever that Apple model is going to be. Maybe that would end up being Gemini if they follow what happened with search. But there's a question whether regulators would allow that. But like, there was a pretty similar analog that happened. When you have those defaults, that will make a big difference in the consumer space, I imagine, especially if there's not a major difference between the quality of the models.

• 24:23Speaker 00

  I know you're a fearless host, Jacob. Maybe to ask you a question, if that's allowed.

• 24:27Speaker 02

  Only I get to ask the hard questions. You can't ask anything too hard.

• 24:30Speaker 00

  I mean, you mentioned OpenAI and Anthropic in the recent funding rounds. Obviously, incredible growth, incredible talent, incredible products they've built. How do you think about the rounds that those companies are raising?

• 24:40Speaker 02

  Yeah, I mean, I think the real question is, obviously, like, if you expand out these growth rates that these companies are experiencing, just a few years, you easily get into very comfortable territory from like a multiple perspective and what you're underwriting. And so I think there's just a huge question of, you know, what do the growth rates look like in the next few years? And they're different for each business, right? For OpenAI, you know, okay, what, there are 700 million users now? It's like, is there room to continue growing that? Probably a little bit, but it's really all around monetization, right? And what is, you know, what's the ability to really monetize a product like ChatGPT? You know, obviously, there'll be something around ads. There'll be something around commerce. But I think the extent to which they're able to do that is the big question. I think under a pretty reasonable level of assumptions, there's many reasons.

• 25:24Speaker 02

  Reasons to monetize far better than search because you're actually completing like end to end actions. And so the ability to get in the commerce flow is certainly there. But I think that's how you think about that. Anthropic, you know, I think, again, it's how much room is there to run on this coding use case? And then on this whole enterprise API market, how much does it end up growing these next few years? Obviously, if the the next 24 months look anything like the last six, you'll be feeling super smart about that. You know, it goes back to either you'll see some sort of step change improvement in the models, or we're just so slow at diffusion into society that like we could press pause now with these capabilities and for a decade, keep running at these growth rates because it just takes time to get the stuff out.

• 26:04Speaker 01

  I don't know if we can keep the same growth rates, but I certainly would believe that if we stopped all model progress, there would be decades worth of work to actually disseminate it through the entire economy.

• 26:13Speaker 02

  Are you a buyer of OpenAI at 500 and Anthropic at 180?

• 26:18Speaker 01

  I think probably of OpenAI, I think it starts, I mean, the upside starts to get smaller. Once you get to these numbers, I think you have to assume a lot of things. But like, it's not crazy based on the revenue multiples or anything like that. I think the question is more like, can they get to a sustainable unit economics standpoint? And will consumers actually pay for it? Like, a lot of the developers and people in AI pay for it. Like, everyone I know who's not an AI developer who uses ChatGPT does not pay for it. So I don't know whether that actually works. I agree with you on your monetization point. I think the challenge with monetizing answer engines is that if you're showing a sponsored link on top of 10 blue links, that's kind of okay. It doesn't undermine the trust of the 10 blue links. But if I'm going to an answer engine that gives me a single answer

• 27:03Speaker 01

  and it biases the answer it gives me based off of a commercial purpose, that undermines trust in the answer itself. So I'm sure there's a way to do this. And I agree with you, I have the same intuition that there probably is a way to actually monetize it more deeply even than you do search. But, you know, search has had 20 years of optimization for this and we're starting at ground zero. So I think it's unclear kind of what will happen going forward there. And with Anthropic, I think you're right, they're more susceptible to fickle developers. So that could still be very, it could have continued that growth rate going forward. At the same time, like we've seen a lot of our employees, every time there's a new model, they switch all of their coding use cases to the different models and they have zero loyalty to a given model.

• 27:48Speaker 02

  So Datalogy is part of the problem, the staff.

• 27:52Speaker 02

  I mean, the other one in this space that I'd be curious for both your thoughts on, I think literally today, XAI announced this raise at 200 billion. So keeping in the same kind of valuation ballpark, obviously, you know, a little bit less buzz around that one in the journals like guys, but it's Elon. What did you two make of that?

• 28:09Speaker 00

  Honestly, I think like XAI's incredible ability to fundraise at the levels that it's raised at and the valuations like really does just reflect, as you said, Jacob, like how powerful the thesis of never bet against Elon is in the minds of so many investors, which to be clear, like is a thesis that would have served you very well for many years.

• 28:32Speaker 02

  Like probably better than most investing theses, almost any of the last, last 15, 20 years.

• 28:37Speaker 00

  Absolutely. Yeah. He's the, he's the greatest entrepreneur of our era. He's built so many incredible companies. There's no denying that. But I think if you set Elon's involvement aside, like I, I would, I think the 200 billion, $200 billion valuation for XAI, to me, makes a lot less sense than either Anthropic's or OpenAI's. I think they have much, much less of a business. And they're also just playing catch up. And to their credit, they, they have caught up very, very quickly. They've caught up to the frontier. But at least to date, they aren't doing anything fundamentally different than what OpenAI is doing, what Anthropic is doing in terms of their technology approach, in terms of the product. I heard an interesting analysis recently, which was on the topic of Elon

• 29:21Speaker 00

  Musk, that XAI is his first company that's really just like a me too company where he's kind of building another company that, you know, that's in the mold of others as opposed to, you know, starting an electric vehicle startup in 2003, or a space company, or Neuralink, so forth. It doesn't feel like there's anything really distinct about XAI yet in terms of their approach. And they all, they also have been losing a lot of talent. Like it's been publicized, like a lot of their top researchers have left them. In recent weeks and months, and you know, as with a lot of Elon Musk companies, there are concerns about culture there and so forth. So I guess just circling back to where I started,

• 30:02Speaker 00

  like I'd never bet against Elon is a sound mantra. So you know, I will say a thing definitive, but I, XAI does not seem as strongly positioned to me as some of the other labs. And I'm very interested to see if that changes.

• 30:16Speaker 01

  Rob said almost exactly what I would say. I think like, I don't think the numbers justify it. Like they're valued at more than Anthropic. At the same time, it's Elon. So you know, betting against Elon has proved to be a pretty bad decision for one's pocketbook in the past.

• 30:31Speaker 02

  Well, I mean, Rob, you alluded to it, obviously, in some of these researchers moving around from X to, you know, other places. And I feel like this has been one of the, you know, dominant topics of the last like three, six months, the kind of race, talent war for top researchers, the salary packages that folks are getting in an effort to build some of these super teams. I mean, I'm curious for your reaction to this. And also, is this going to just keep continuing?

• 30:55Speaker 01

  Well, okay, I'll start with the latter question and then talk about why. I actually think this can make sense. It has some downsides, but I think it absolutely can make sense to spend on talent like this. So I first entered the field of AI in 2015, NERIP's 2015. At the time it's called NIPS 2015.

• 31:09Speaker 02

  Were people giving you $100 million packages at the time?

• 31:12Speaker 01

  They were not. It was a much-

• 31:14Speaker 00

  There was probably not a single VC there.

• 31:16Speaker 01

  There may have been one or two, you know, but there were not very many. Certainly a very different conference than what we see today. Every year since 2015, I have had other researchers say to me in various capacities, surely this must be the top. This can't get any more exciting. I remember in 2017, Intel hosted a party, I think it was NERIP's in Long Beach, and Intel hosted a party that had Florida at it, which was notable because, you know, there was a Silicon Valley episode where Florida appears. And like that, everybody at the conference was like, surely this is it. That's the top. We're done. Like this is it. That was 2017. You know, at the time, I remember, you know, seeing some friends who got half a million dollar offers out of their

• 31:58Speaker 01

  PhD, and that was completely insane that that would happen. And I know a ton of people who like left academia at that time because they were like, nope, this isn't going to keep on going.

• 32:09Speaker 01

  AI obviously is going to have massive economic value. Like, I think, and all the talk of AI being a bubble, like, it's clearly different than the dot com bubble, it's already having economic value, it's going to have dramatic economic value, there can be questions about, like, is it too hyped or whatever, but like, clearly there are going to be winners that are going to make massive impacts on society here. I don't expect that's going to change, and I expect that the amount of money that AI companies make will only go up, and the importance of AI to the entire ecosystem will only go up, such that it will continue to make sense for the best of talent. Now, why does it make sense to put a lot of money into talent, I think? Fundamentally, what you care about when you build any model, whether you're one of the

• 32:50Speaker 01

  frontier labs, or whether you're an enterprise, is I put a given amount of dollars into the model, and I get an amount, a certain amount of performance out of that model. And when I think about that, that's performance along different axes, that can be task performance, that can be cost of inference, that can be whatever, some combination of all of those. But that's fundamentally the equation that I care about, right? And anything that can make that ratio much higher, anything that can serve as a compute multiplier, is super duper valuable. One of our core bets at Datalogy is that, and what we've seen time and time again, is that high-quality data is a massive compute multiplier. You can get a 10x on the performance per dollar with high-quality data, which can enable you to do dramatically less, sorry, dramatically more with a much smaller compute budget. And that can be a massive impact.

• 33:34Speaker 01

  Well, talent is also a compute multiplier. If you know how to make the GPUs sing, you can easily eke out substantial improvements in performance per dollar that can easily be more than the cost of that talent itself. So, anything that makes your compute more valuable, is, and especially in a multiplicative way, is an immensely valuable thing to do. I think the challenge with these very large salaries is, does it attract the right people with the right sets of motivations that are gonna be aligned with the mission and are gonna really put their all into building this? Or are you gonna get people who are gonna primarily be concerned about not getting fired, or just spend a couple months and then leave because they've made, you know, $10 million in a couple months?

• 34:19Speaker 01

  Or things like that. So I think there's questions around whether culturally that sets up the right incentives and whatnot. But I think from the, from a simple perspective of like, can you justify this investment? Is there a return on it? I think there clearly can be.

• 34:31Speaker 02

  I mean, it seems like obviously if you're, if you're going to allocate a certain amount of budget to compute, which we all can see the capex spend is absolutely massive. And you allocate a certain amount of budget to data labeling and, and RL environments and all this stuff. You know, it kind of makes total sense. Again, if all these forces are coming together and multiplying to, you know, in some ways this was underpriced previously. But I wonder like, does this just continue? Like, uh, obviously compute budgets are going to go up, spend on these environments and, and data will probably continue to go up. Do you, do you see this like just going more and more?

• 34:59Speaker 01

  I think demand will continue to grow faster than supply. So yes, probably. I think that's fundamentally the question, right? Like, will the supply of AI talent exceed the demand for AI talent? And like, I don't think that's going to change anytime soon.

• 35:12Speaker 01

  I think the average will continue to grow up, whether or not you'll continue to see, you know, a hundred million dollar pay packages. I think it's possible that is a moment in time, but I think, but that's not what the average AI talent is, is receiving. The average AI talent is still receiving a lot of money, but it's not a hundred million dollars, right? I expect that will continue to go up. And then I think we'll, it'll honestly, it'll probably depend what happens with MSL. If Meta puts out a really great model, you will probably see more of these crazy salaries. If Meta does not put out a really great model and there are cultural challenges that prevent that, you probably will see fewer of them.

• 35:45Speaker 02

  I remember obviously, you know, a few years ago, when the character acqui hire happened and everyone was like, what is, what is going on? And then in retrospect, you're like, I don't know, for, for Noam and, and some of those folks, like potentially a really smart deal. And I imagine it just, it, it is create this really weird environment in the investing ecosystem. And Rob, I'd love you to speak to this too, where when some of these teams spin out on day one, the floor of the business is so high. Because it's like, well, this group of people, I mean, look at thinking machines, what is that group of talent worth from an acqui? Like, those are some of the best minds in AI, right? And, uh, it, it then, it's such an interesting impact on the startup ecosystem. When you have these teams that clearly are immensely acquihireable. Yeah, I think some investors look at it and they're like, this is just a free call option. Like I'm, there's no way I lose my investment based on just what this team's worth.

• 36:30Speaker 02

  And, and maybe they will figure something out.

• 36:32Speaker 00

  Yeah, totally. I, I think that's, that's exactly right. And that's exactly how you get to such crazy valuations on day one, like thinking machines of $10 billion or SSI initially raising a five and then raising it at over 30. Is exactly as you said, the The downside protection is real because there are big tech companies that are willing to pay many billions of dollars to buy them and that have already tried to pay many billions of dollars to buy those companies and been rebuffed. I do think one other interesting dynamic on this topic around talent acquisitions, you mentioned Character AI as an example, is this trend that we've seen of effective acquisitions happening for the sake of adding talent to big tech platforms, but funky

• 37:15Speaker 00

  structures where it's not technically M&A and there's kind of this like, these contortions of transaction details and if, first and foremost, in order to avoid M&A scrutiny and antitrust scrutiny from FTC and the Department of Justice and so forth. We've seen it with Character, we saw it with Adept, we saw it with Inflection. Most recently, we saw a version of it with Scale AI and Meta. And I think the, there are many interesting things about this transaction, this type of transaction. I think one interesting thing from the perspective of VCs and employees and founders is it's

• 37:56Speaker 00

  very unpredictable how the dollars flow in these transactions, which is like a new development that is different than anything that's existed in the past in startup finance. Previously, an acquisition meant that like, people got paid out according to their cap table. If on the cap table you own X percent, then you get, you know, that amount of money, obviously relative to the preference stack and so forth. Because these transactions aren't actually literal M&A, the parties involved can kind of redefine who gets paid and who doesn't get paid. And in some deals, the investors just get their money back, in some deals they get less than their money back, in other deals, they get a better multiple.

• 38:39Speaker 00

  In a lot of, in almost every deal, the founders get paid very lavishly, but in most deals, the employees don't get paid well, if at all. And so I think. That element of these transactions, I think, is really harmful and a bad thing. And just given the unpredictability of them. And also frankly, like, very transparent attempts to find a loophole to evade M and A scrutiny, which it seems like it's just a matter of time until the antitrust authorities, you know, adjust accordingly. So anyway, my hope is that this particular flavor of transaction is like a blip in time that goes away before long, but it is an important trend that's worth thinking about.

• 39:18Speaker 01

  Yeah, I'll just add to that because I feel quite strongly about this. There's a social contract that, you know, founders make with early employees that if you bet your most valuable commodity, your time and your energy on the vision of this company, you will share in a portion of the upside. And these aqua hire, not aqua hire sorts of situations break that social contract in a way which is fundamentally bad for startups long term, right? It will make people less likely to join startups, more likely to take less risky options, and it will really starve startups in the long run. So I also hope that there's regulatory scrutiny to kind of prevent this. But I also think it's really up to founders to do the right thing by their employees here.

• 40:02Speaker 01

  Like fundamentally, when you enter an acquisition negotiation as a founder, you can choose what you care about and what you want to make sure you preserve. You can make sure that all of your employees get paid out proportional to their ownership of the company and to that social contract. And, you know, I think sometimes you see people acting in great support of their employees. Sometimes you don't. You know, I would hope that we see, if this trend continues, we see more and more founders really going to bat for their employees to make sure that they do well in these situations and that you don't end up with this kind of shell of a company that's basically dead, that now the employees want a bunch of equity in, but it's worthless.

• 40:40Speaker 02

  I'm so curious to see where this ends up from a regulatory perspective because the one thing they will have to solve is just the, especially in AI world where everything is happening so fast, the time just matters so much. Like imagine you buy an RL environment company today and you have to wait a year in review. It might not even be that relevant in a year, or who knows? Like you want that today. I think a big part of the structure has been, hey, this person can walk through our door on Monday and help us with that. And so I don't claim to be a regulatory expert, but I'm curious what they come up with to at least be responsive to that. Because in this AI age, it seems all the more important.

• 41:12Speaker 01

  I mean, to this question, I'm curious what both of your perspectives here. Like, do you think regulators would allow Finkey or SSI to get acquired? Or Mistral, for example? Like, I think there's a real question with these companies that are at these kind of, you know, 10-plus billion valuations that are mostly talent-dense. Will regulators actually allow that?

• 41:33Speaker 00

  It's a very interesting thought experiment. And to Jacob's point, none of us are antitrust experts, but I do think like-

• 41:41Speaker 02

  Can you go to law school, Rob? I feel like this is like a great, you know, put those skills to use.

• 41:46Speaker 00

  I try to keep that as a secret, yes. But I do think by like, conventional antitrust analysis, an acquisition of thinking machines or an acquisition of SSI should be fine because those companies don't have revenue and therefore they don't have real market share. And them getting acquired shouldn't change pricing to consumers one way or the other. I think if you have like a creative slash flexible slash proactive antitrust regulator like Lena Kahn, I'm sure if Lena Kahn were still in, were still running the FTC, she would come down hard on like Google trying to buy SSI. But I think if it's a more staid or laissez-faire

• 42:30Speaker 00

  antitrust regulator, like I think by the books, it should, there shouldn't really be a problem with those companies getting acquired.

• 42:35Speaker 02

  Switching gears, one thing, I do occasionally listen to other podcasts, not just my own. And one thing I heard-

• 42:41Speaker 00

  This is the only podcast worth listening to, for the record.

• 42:44Speaker 02

  That's very, you know, paid endorsement from Rob, much appreciated. You know, I was listening to BG2 the other day and Sherwin Wu from, you know, who runs a lot of engineering for the OpenAI platform was talking, I think he was asked about AI infrastructure companies, kind of around RL environments, but also just generally about AI infra. And he said something that I've heard often, which is, you know, he was short all things AI infra because, God, the scaffolding just changes so fast as these models get better, as the way people use them get better. And I was curious for both of your reaction to that.

• 43:15Speaker 01

  I think it depends on what kind of infra you're talking about, right? Like, is a company like Fireworks gonna experience a massive change in what people care about with respect to fast inference, such that their business is no longer gonna make sense? Like, no, like inference is gonna be required no matter what, that's obviously.

• 43:31Speaker 02

  Inference and data labeling seem like the safe ones in this infrastructure conversation.

• 43:35Speaker 01

  Yeah, I think there's a number of these categories. Actually, hot take I might say is that data labeling, I think, is maybe a little less safe because you don't need infinite data. And something we maybe come back to a little later, but synthetic data is pretty darn powerful.

• 43:47Speaker 02

  To being like, you know, to helping create judges for these RL environments, right?

• 43:50Speaker 01

  Right, they're shifting to this direction. But like, at a certain point, you don't need more human data. You have, or the human data you need gets more and more narrowly specified, such that, you know, it becomes a smaller and smaller market. Like, I think that will happen at some point. Question whether that happens sooner in five years, but like, at some point, you will see that. But like, I think it really depends on like, what kind of infra. If you're talking about building something that takes, that is a model wrapper, where you take, you know, current GPT-5 and you build a bunch of scaffolding around it, so it can do a task that it wouldn't be able to do otherwise. That is certainly something that is in danger of getting overwritten by an updated model, especially if it's for a pretty general purpose sort of task, I think.

• 44:34Speaker 01

  But the core parts of infrastructure that you'll need to have fast inference, that you'll need to make training easy, that you'll need to be able to scale effectively. You know, I think we'll see things like, you know, we'll see a lot of folks that are gonna start coming in to help companies constantly update their models. You know, like, for example, the models that you see at big tech that go into ranking are generally updated every couple hours. And if they don't do that, they lose a lot of revenue. That's obviously where everything's gonna go. It's gonna take some time to get there and you're gonna have to build some infra to do that. But I think we'll see successful companies on that. Things where you're just kind of putting a scaffolding around an existing foundation model. I think that's where I would agree with Sherwin's point here. Yeah. What about you, Rob?

• 45:16Speaker 00

  Yeah, I would largely echo Ari. I mean, I agree with Sherwin that because things are moving so quickly, it's very hard to build a durable infra business in AI and investing in AI infra is challenging right now. And there is a lot of infrastructures being built that will be obsolete. But I disagree that that's the, that's the right conclusion from that is to be bearish on the entire category of infra. Because I do think to Ari's point, like in any major technology shift, there's a need for a whole new suite of tools and infrastructure to support the, the applications and use cases on top of them. And the challenging thing is just identifying to Ari's point, like what are the types of infrastructure that will stand the test of time and be enduring? And, you know, as Ari mentioned, inference seems like a good bet.

• 46:02Speaker 00

  You know, up until recently, you could, for many years, you could say data labeling is a good bet. And obviously, scale rode that wave and, and now maybe there's a debate around with synthetic data and so forth is that, is, has the era of human labeled data peaked perhaps? I think there are certain types of infrastructure companies or even as the technology changes in a million different ways, the need for that, the, the service the company is providing stays the same. And so I think the, yeah, the challenge is just figuring out which companies those are.

• 46:29Speaker 01

  I mean, if you think about, like, thinking about this from Datalogy, because we're an AI infra company, right? Where I've thought a lot about this. Fundamentally, our view is that no matter what you have to do with AI, the quality of the data you show to those models is going to be critical. You know, when, when people ask me, what's your core companies? What's your unfair advantage at Datalogy? What I say is, you know, that we want to be the best in the world. And I think we already are at valuing data with respect to a given downstream use case. So if you tell me I want a model to do X, we can say, here's the data that's going to lead to that, either by filtering existing corpora or by building synthetic data or whatever. That's a very generic framing that is going to always be useful in some way or another. Now, the exact form that that might take might change over time, right? Maybe in two years, we're doing that more for RL than we are today.

• 47:15Speaker 01

  You know, so that might change. And then I think when you're building an AI infra company like this, then one of the things Things we have to think about and I think about a lot is how do we make sure that the core thing that we do better than everybody else maintains, but also enables us to be flexible as the ecosystem changes. And as what is prominent, what is needed at any given time changes. But fundamentally, data quality is always going to be necessary for AI models. And if you can find a place, I think, where there is something that will always be true and then focus on the core aspect of that. I obviously believe that will be a successful strategy.

• 47:53Speaker 02

  Because it feels like otherwise, without that stake in the ground, you're kind of playing whack a mole with like whatever the latest and greatest way people are doing things is. And that feels really challenging at the infra layer.

• 48:03Speaker 01

  Especially in AI right now, like Ray, like what, like the only thing that's like a truism is that like whatever we're doing today is not the same thing as what we're going to do in six months. Like, there's going to be a lot of commonality, but things are going to change. You have to build your business so it's adaptable to that.

• 48:18Speaker 02

  Well, I want to end with kind of a broader set of questions that I'll stuff into the, to the back, the back of the, the podcast here. But we were talking about it earlier and I'm too intrigued not to ask. Ari, you worked at Meta for a while. Do you think this MSL experiment is going to work?

• 48:35Speaker 01

  I think it might. I mean, look, we still decided about don't bet against Elon. I would also say don't bet against Zuck. You know, one of my controversial takes, I think, is that I think that metaverse spending actually is going to look pretty good in the long run. If you look at the AR glasses that they just put out, I think that's where that vision really is. And actually, FAIR and a lot of Meta's AI research was under Reality Labs. So counts as metaverse spending. So about two of the years that I was at FAIR, we were part of the Reality Labs org that counted as metaverse spending. So a lot of, I think actually it's very plausible, I'm not positive, but I think Llama, actually, if you go and you trace, like, where did the budget from Llama come from? It came from metaverse spending, based on how that worked, because we were part of Reality Labs at that time.

• 49:17Speaker 01

  Look, I mean, Zuck has consistently made strategic moves that were widely lambasted at the time and seem like genius moves in hindsight. I think it's possible, like I said, there's a reason why you can think paying these numbers for talent could make sense. My worry is more about what cultural incentives that might create. I have certainly seen that Meta has lost a lot of really incredible talent, folks that I knew for my time at Fair, who I would say were equally worth the amounts of money that were being given around, but didn't get those amounts of money. And even though they were well-paid, that made them feel very bitter. So I think there are a lot of these second-order cultural effects that come from doing this.

• 49:59Speaker 01

  And my main question, and I think if it doesn't work, likely it will be because of those cultural challenges. So I think it remains to be seen, but I think those are the concerns that I would have around this strategy.

• 50:10Speaker 02

  Yeah. Well, you alluded to kind of like the augmented reality stuff starting to look pretty interesting, and obviously the demos this week are pretty compelling and where this is all headed. I'm curious for both of you, actually, what gets you excited around the intersection of hardware and like these models? I think a lot of companies seem to have coalesced around glasses as being the form factor and putting models and speakers in glasses. But would love both of you, actually, to just riff on how you think this ends up entering the physical world more and kind of coming off our phones and computers.

• 50:43Speaker 00

  Yeah, so I'm happy to start. I think I've talked both of your ears off at this point, one time or another, about BCI, brain-computer interfaces, and my bullishness around that category. But it's a field I've been doing a deep dive in over the past several months, and I really do believe that we're nearing an inflection point in terms of real-world functionality and applicability for BCI technology, both invasive and non-invasive BCI. And the two always have very different use cases and kind of customer profiles. Invasive BCI, just as a quick aside, basically entails needing to do surgery, most often cutting open the skull in order to implant a computer chip in the brain.

• 51:28Speaker 00

  So that's obviously a very high barrier to entry, but it enables you to get really granular Neural data and obviously the most well-known invasive DCI company is Neuralink, Elon Musk's company, but there are other very exciting invasive DCI companies. And then non-invasive companies are much easier to to build consumer products around because they're just, you know, you can embed sensors into headphones or into a cap or into, you know, even potentially things like on your wrist and so forth. And I do think that due to improvements in sensors, but really more than anything due to improvement in the underlying AI, we're getting better and better at extracting

• 52:11Speaker 00

  enough signal from the noise of brain data that, you know, that comes off your head thanks to the electrical activity of your brain and magnetic related activity that it's possible to build pretty compelling non-invasive DCI products. I think we're going to start seeing these hit the market over the next 12 to 24 months. And I do think it will be kind of a whole new paradigm in terms of how humans interface with computers.

• 52:36Speaker 01

  The Metarisk band is a version of this.

• 52:38Speaker 02

  Right. From the Control Labs acquisition, right? Which was a really DCI company.

• 52:43Speaker 01

  It's not actually measuring neural activity, it's measuring, you know, muscle signals, but it's, but it's a pretty good example of that. I think the glasses as a form factor make a lot of sense personally. I mean, I think I remember when Google Glass coming out, you know, being super excited about that and like it just was ahead of where it actually is. You think about the other form factors, like what you see in her, where he holds it in his pocket and there's a little camera, the humane kind of had a similar idea. I think glasses make more sense as a form factor than that. I think there's some interesting questions, like do a lot more people end up wearing glasses than they did in the past? Do you see differences where in like sunny places, people adopt these technologies more because wearing sunglasses is more common than in cooler places? I don't know, there may be some interesting questions, but I do think glasses are a pretty darn good form factor.

• 53:28Speaker 01

  And I think then you start integrating DCI stuff with that over time, that makes a lot of sense.

• 53:33Speaker 02

  What's one thing that each of you has changed your mind on in the AI world in the last year?

• 53:38Speaker 01

  For me, I would say that

• 53:43Speaker 01

  The success of RL, really starting to see RL work on language models for the first time in the last year, I think, has been somewhat surprising in the sense that this is something that just like didn't work for a long time. People tried it over and over again. RL, I think, has been a story of a lot of initial promise and then it not really generalizing and working in the real world. We've seen this like multiple times. I think we're seeing kind of the best evidence yet that RL can really work in the real world. Although in hindsight, the reason that this happened was because the base models just got good enough that RL started working. Like, if you take the exact same techniques that you apply, like the DeepSeek R1 techniques, and you apply it to, you know, models from three years ago, they just will never work, or even from a year ago. Even looking at

• 54:27Speaker 01

  Llama versus Quen, like Quen RL is really well because it's trained on a lot of synthetic reasoning traces, whereas Llama wasn't, and therefore Llama doesn't do a good job at this. So I think that's probably one thing that has been surprising to me. And, you know, we're clearly seeing that this can now work quite well in a lot of ways. What about you, Rob?

• 54:45Speaker 00

  One topic that, I don't know if I would exactly say I've changed my mind on it, but it's just become murkier to me over the past year. But I also think is like an increasingly important question is, I guess the way you could frame it is like, when will the GPT-3 happen? When will the GPT-3 moment happen in robotics?

• 55:04Speaker 01

  Yeah.

• 55:05Speaker 00

  And we've, we've talked about AI applied to the physical world a little bit, but I think like, it's helpful to take a step back and remember that like, all of the AI capabilities that we're talking about, and AI agents, and automating work and so forth, pretty much all applies to knowledge work in the digital world. And there's an entire other world of labor and activity, which is the physical world, which today remains largely unautomated. Or if it is automated, it's not with AI-first techniques. It's kind of like simple previous generation robotics. And there's obviously, you know, been a huge wave of research and capital and startups working to build general-purpose foundation models for robotics, both on the sort of hardware agnostic side, companies like Physical

• 55:49Speaker 00

  Intelligence and Skill that have amazing, amazing research teams, and then, you know, general-purpose hardware platforms as well, humanoid robots from companies like Tesla, as well as OneX and Figure and so forth, with, with the vision of building a general-purpose robot that can do anything a human can do. And it does, progress is certainly being made very quickly along that axis. And so it does seem like we are on a trajectory to make that happen eventually. But it's very unclear to me, and I think it's a really important, open question. Like, is that one or two years away? Is that five years away?

• 56:29Speaker 02

  Yeah.

• 56:29Speaker 00

  Is it more than five years away? I think it's, when you get into applying AI in the world of atoms and not just bits, I think it gets exponentially harder. And so I think it's, it's hard for me to tell based on the program. And companies like Pi, you know, publish and open source research, which is amazing. And I think a great service to the community. But as a result, there is, there is information out there that you can look at. And it's, I think it's still, it's unclear to me how, how close that is. But I think, I think that will determine a lot in terms of how quickly AI transforms all of our worlds.

• 57:02Speaker 02

  Yeah. Well, you promoted my podcast, you promoted my portfolio company. You know, you definitely get an invite back. It's interesting, similar to what Ari was saying. Obviously, a huge part of the advancement robotics has been having general reasoning models. And now, you know, really strong video and world models. And it's amazing how, you know, some of these spaces that were traditionally more challenging, it all feeds off each other, right? And the improvements in those areas have really driven a lot of the progress. You know, I was going to say, actually, maybe close to Ari's world. One thing I've like changed my mind on is, I feel like at the application layer, you know, a year ago, I would have said, really don't spend much time thinking about models. Like, build workflow. And you're not going to be able to differentiate much on the model side. Like, just go really deep in, in the domains and kind of, as fast as you can to build a breadth of use cases.

• 57:46Speaker 02

  And it finally feels like actually we're, we're starting to turn a corner around doing, you know, RL and some of these things in your specific domain to improve, you know, both reliability and quality of these models is really starting to matter more and more.

• 57:59Speaker 01

  It all just comes down to cost, right? If you think this costs a ton of money, then it doesn't make sense. If it can be done for, you know, a million dollars, this makes sense for every enterprise. And I think just we're going to see costs continue to go down, and that's what drives all of this.

• 58:11Speaker 02

  Anything that you think's like under discussed today, that either we didn't talk on the pod today, or something that, you know, you wish people in the discourse were, were around more?

• 58:19Speaker 00

  I have one, which I think is a really important theme that not enough people are talking about yet, but I think will become increasingly important. We talked previously about

• 58:33Speaker 00

  the progress in fundamental model capabilities, slowing and plateauing and S-curve and so forth. My view as an observer of the frontier technology space is if there's one kind of development or approach that is most likely to lead to that next leap and kind of continue the exponential upward growth and, you know, put us on a path to fast takeoff for superintelligence and so forth. It's this notion of being able to build AI that can itself autonomously develop better AI and stronger AI, this kind of recursive AI concept. And most kind of like big picture visions of how superhuman AI comes about, going back to like the 60s

• 59:13Speaker 00

  and, you know, the very early AI researchers, is conceptually around this notion of like building an AI that can then build a better AI and it sort of compounds from there. I think in the past year or so, this has become much more of a reality. And there are more than one high-profile groups of researchers and founders that are building companies along these lines that are in stealth now, but I'm sure will be de-stealth before too long. And, you know, it's something that the big labs, no doubt, are taking seriously. But I think it's a concept that like hasn't necessarily broke through to the mainstream or it's a concept that like people are familiar with, but it kind of sounds sci-fi still

• 59:55Speaker 00

  and doesn't sound like it's like on the main path of research that, you know, that the main labs are following. But I do think this is like, if we can, if someone can get this to work and how well they can get it to work, I think will actually be pretty determinative in terms of like, how steep does the slope look from here over the next few years?

• 1:00:14Speaker 02

  That's a good, I like that one. Yeah, I mean, I guess the question on that real fast, I mean, isn't this why we had like, you know, Noam on the pod a while ago and, you know, my sense from talking to him and from a lot of these top researchers is like this question of, hey, does this stuff generalize without, you know, outside of coding? It's almost like unimportant. It's like if you can just hill climb on coding and then, you know, make really good AI researchers or, you know, really augment the quality of AI researchers is like, everything flows from there. And so I almost wonder like if that is the only important, you know, if we get that, then it feels like we're on quite a path of improvement.

• 1:00:49Speaker 01

  Data continues to be the most under invested area of AI research relative to its impact. You know, the second thing I think would be evals, which are data in many ways. You know, we hear all of this discussion of the data wall. We're going to run into too much data wall. I think it's starting to get quite a bit of hype and becoming more prevalent in the zeitgeist. But synthetic data is a huge part of the solution here. There's a lot of nuance to how you do this correctly. There are, there's been a lot of concerns around kind of synthetic data leading to model collapse, where you kind of get this like cycle where the snake's eating its own tail and you kind of get it going towards babbling or things like that. But there are also lots of ways to get around this. And this was kind of called out now as a pretty major part of what made GPT-5.

• 1:01:32Speaker 01

  All the Chinese models have called this out as a major part of what drove their gains. And this is how we actually get more juice out of the existing data sets we have. A combination of curing the data better and then figuring out really the right ways to do synthetic data. It's also going to be really critical as we go to more domain specific use cases. I think people are sleeping in general on how big of an impact this is going to make if you can do it correctly. There are a lot of risks with it, but you can do it correctly. And if so, that can make a huge impact.

• 1:02:05Speaker 02

  Amazing. Well, all right, Rob, this has been a ton of fun. Really appreciate you both coming on and doing this.

• 1:02:10Speaker 01

  Thank you so much for having us.

• 1:02:11Speaker 02

  Hey, guys, this is Jacob. Just one more thing before you take off. If you enjoyed that conversation, please consider leaving a five star rating on the show. Doing so helps the podcast reach more listeners and helps us bring on the best guests. This has been an episode of Unsupervised Learning, an AI podcast by Redpoint Ventures, where we probe the sharpest minds in AI about what's real today, what's going to be real in the future, and what it means for businesses in the world. With the fast moving pace of AI, we aim to help you deconstruct and understand the most important breakthroughs and see a clearer picture of reality. Thank you for listening and see you next episode.