Loading summary
Transcript109 lines
- [00:00]
A
The frontier model and the open source price wars have arrived in the last two weeks and man, it is incredible when tokens go down 90 or 99% as the case may be.
- [00:10]
B
We used to have FU money, now we have FU tokens.
- [00:13]
C
Price drop is expected, right? As you know, now multiple models are very close to each other, right? Like, you know, I think they're going to start competing in price anthropic.
- [00:20]
A
Put Mythos out. Then rug pulled it. People were like, wait a second, we've been using this thing already. And this back and forth I think woke people up to, hey, we could give you a tool and then take the tool back. This could have cataclysmic ramifications for an enterprise specifically. You got two trust issues and a price issue. It feels like every week speed is increasing. It's causing whiplash. Yeah. Thanks to our friends at PayPal, the exclusive sponsor for this Week in AI. Try the payment and growth platform that's trusted by millions of customers worldwide. PayPal open start growing today@paypalopen.com. all right everybody, welcome back to this Week in AI. This is a dedicated show where I like to ask three founders of leading AI companies to get together with me and talk about what they're working on and what's in the news. You can't get this information anywhere else but this very program. This week in AI search for in your podcast player, YouTube subscribe, get in the comments. We've been doing it for about 20 episodes and we are off to the races. We've got three amazing guests and my trusty partner in crime, my news reader Lonnie Donnie is here. Lon Harris, who do we have on the program today?
- [01:31]
B
It's episode 22, Jason. Actually we made it today. Incredible lineup of guests. We've got Sarah Hooker from Adaption Labs. She's the co founder and CEO. They are helping AI models continue learning and self adjusting based on real world use. Then we've got Manu Sharma from Labelbox. He's the CEO over there. They are a data factory combining human experts to software and AI tools for improved model training. Finally, Spiro Sanos of Resolve AI. He's the co founder and CEO over there. They are an AI site reliability engineer. They've got agents that will automatically investigate and fix bugs on your system without needing human intervention. That's the bit.
- [02:12]
A
Amazing. And the docket has been crazy. The last week in AI has been nuts. I have been vibe coding like a lunatic. I got access to, I got access to the new Grok. I put that into my perplexity computer. I've been playing with GLM 5.2. I got a Bittensor subnet, they gave me a key, so I've had unlimited tokens. Let me tell you something, when you have unlimited tokens, that changes your behavior. But I think this the number one story of the week. I'm sure it's on the docket here, Lon, is the price wars. The frontier model and the open source price wars have arrived in the last two weeks. And man, it is incredible when tokens go down 90 or 99% as the case may be.
- [02:55]
B
Yeah, we used to have FU money, now we have FU tokens. You're correct, Jason. Grok 4.5 from SpaceX AI arrived last week on July 8th. And along with the powerful Chinese model GLM 5.2 from Zai, we're getting dangerously close to frontier model quality without having to pay for frontier model prices. Grok, of course, cost $2 per million input and 6 million per million output token. That's a 60% savings over Opus 4.8 or GPT 5.5. Elon calls it an OPUS class model, but faster, more token efficient and at lower cost. And on benchmarks. Groka, GLM 5.2, they're not quite frontier. They're at like a second step down, which people are sort of calling near Frontier. So on the, the Artificial Analysis intelligence index, opus 4.8 gets a 56 GPT, 55 gets a 55, GRO 4.5 gets a 54, GLM 52 gets a 51. So I mean, we're talking about, they're, they're nipping at the heels here, Jason.
- [03:58]
A
Yeah, and then obviously the cost is the issue. I guess my question, I'll open it up to you, Spyros, is what are your thoughts on the plummeting price of tokens and what does that mean for the frontier models and their massive build out? Now that open source, it feels like it's starting to catch up. And what will that make the next year or two look like for enterprises which obviously don't want people to get their intelligence? We had Alex Karp do a Alex Karp esque CNBC hit where he's talking about AI sovereignty, which we've been talking about a whole bunch here. What's going to happen if open source is this close, you know, three to six months behind the frontier models.
- [04:47]
C
To me, to me, the big news here, Jason, is actually that right, it is the open source getting very close to frontier models. I would say that the price drop is Expected, right. As you know now multiple models are very close to each other, right. Like, you know, I think they're going to start competing in price our own experiments. We see that actually internally, right. Like we see that essentially OpenAI and anthropic models are equally good, almost right. For the frontier. Most hardest tasks we're trying to perform in reasoning. But to me the bigger news, like I said, is, are the open source models because I think what the sentiment is in the enterprise, right, and we're working with some of the largest enterprises in the US right now is that, you know, I pay all this money like Garf said and you know, I get or I don't get the results. There's the whole last mile, I guess work that somebody has to do. But I think the bigger news is that I think people are now worried that their own core business might be disrupted from the closed, let's say from the labs in some sense, right. That are going after everybody. Right. So I think I see a huge, huge change in sentiment and probably like a lot more desire for people to control their intelligence completely. Right. Which means really an open weight model they can run or work with a partner, let's say, that is trusted, that is going to work just on what they want, right. And you know, provide the intelligence they need which without necessarily going after the rest of their business. Right. Which is how they make money.
- [06:00]
A
Sarah, when you look at enterprises and startups alike, they're obviously cost conscious and this is a better deal by a factor. And when the tokens go down, usage goes up and that's a huge win. But the other win is avoiding what happened to figma and what happened to Cursor in both of those cases Anthropic, because they've got a huge valuation, they're going after the application layer. They launched Claude Design. Figma was a bit tweaked about that. They felt they got double crossed. I don't want to speak for them but they, they've got some hurt feelings there. Cursor, same thing Anthropic told them, hey, our internal coding agent is just that, it's an internal coding agent and lo and behold, you know, Claude Code is now trying to eat Cursor's lunch. Cursor obviously then joins XAI and they're doing a great job over there. But what are your thoughts broadly on the frontier models going after the application layer and this massive cost reduction?
- [07:07]
D
Yeah, I'll be cheeky. I think these are like two separate trends which are, you know, the cost reduction. I think is very particular to Grok and what they're trying to do right now. Like, if you think about this really release, it's the first referendum on whether the Cursor team and the data has come into play. It's the first referendum on how that plays out in Groq's future. And also they want more data like that. This is very much an important part. It's some of the most valuable data is like, how are people using this in the real world? So that dynamic of cost decreases. I think we're in a moment for that particular dynamic. What does a set of releases say? One is, I think with a lot of the open source release, that said, hey, China's still going to play fiercely in the open weight space. That's super interesting. The second component is because of that we also have the confrontation of what you're speaking about, which is that companies feel one, that the rip out of very good models of Mythos and everything that happened there has made them much more aware that they have to hedge their risk. Secondly, this temporary cost reduction doesn't really mitigate the overall dynamic that Spiros is talking about. People have massive cost dynamics and so they're saying, we know this is coming. Prices are only going up. We see this unpredictability with closed weight models, we're going to hedge in some way. And with Figma and Cursor, they have learned their lesson, right? And cognition and all these places which are data heavy. They're now investing in either massive buildouts of internal AI researchers as very few, you have to really fight to get them, or they're saying we have to somehow build an internal stack that we purchase elsewhere. And which route you go kind of depends on how much capital you have to invest.
- [09:05]
C
I've seen it happen like in the last week. One was that we had a customer for the first time that told us that, you know, Resolve uses the best models possible for the task. Right. Including our own internal models or whatever models. Generally you don't choose what models. Right. Because that changes things every day. What a customer tell us that they didn't want to use models from one particular lab. Like that was a request. Like, we don't want those models used anywhere. Right. Not like use this particular model because, you know, have a deal, but don't use this one particular lab, let's say, in whatever you do with our data. Basically, that was one. And the second, I think is, you know, what Sarah mentioned also about people investing, having access to table investing a lot and then being Taken away that for. I was talking to an executive at a large company today. That was a huge, huge disruption for them. Right. All that investment went out of the window in some sense for them. Right. When that stopped. And now they're way more careful in where they make their bets. Right. And controlling the best in the other player too.
- [10:01]
A
Spiros, you're referencing Anthropic put Mythos out and then Rug pulled it. People were like, wait a second, we've been using this thing already. And this back and forth, I think woke people up to, hey, we could give you a tool and then take the tool back. And Manu, this could have cataclysmic ramifications for an enterprise specifically, like, oh my God, can we trust these folks? So you got like three or four trust issues. You got two trust issues and a price issue. What are you seeing? You've got a lot of customers. Obviously we won't pick out any specifically here to discuss because we'll protect you from having to cause some channel conflicts there. But broadly, what are you seeing in the space in terms of enterprise? Customers may be using open source, but hey, they need extra data and that's your speciality.
- [10:49]
E
Yeah. So I think it's very clear that majority of the businesses want to have a kind of a portfolio of models for different workloads. So it's very clear that I think there's some, there's some use cases like general intelligence is going to be used for productivity applications in the businesses. Right. So we're using like coding agents to make everyone productive. And I think it wouldn't make sense for businesses to build those applications or tools to go after generalized use cases. However, in the enterprise, at least in America, I would say that majority of valuable use cases are rather core to their businesses. And it's very clear that these enterprises are going towards owning the entire stack, and rightfully so, because for the first time, technology is making it possible that they can actually wield this kind of super AI systems and that can improve over time with their proprietary information, proprietary context, data and confound over time. It wasn't really so clear. Maybe like a year ago it was sort of that kind of a technology that only was in the hands of a handful of labs. But right now we're seeing just insane kind of appetite towards building and owning the entire stack of intelligence. And the nice thing about American businesses is that at the end of the day, they don't have to like anybody, the vendors or people, but they make choices that are rational when it comes to everyday decisions. And when they're using and owning their entire AI stack, we're seeing cost per token with open source roughly on the order of 5 to 10x cheaper than a frontier token. And there are, you know, now that the open source is right near the frontier, there are just incredible amount of use cases, everyday business applications, everyday products and services, things that are powering can be done by open source tech. And that's actually a true trend that is happening that is in the, in the works. And I think in the next couple of years you will see more and more companies building value on top of kind of the compute and inferencing substrate. The thing with open source is that it is all these companies have to make money somehow. And you know I would, I would argue that all of these open source companies are going to go after applications also and offer applications to enterprises to leverage their open source tech. But where I think it's going is that the margins, you would see that in the frontier tokens, maybe I saw in public comment somewhere like it's 80% or so margins that some labs are making and inferencing. I think that is unlikely to happen again in the open source world. So clearly the value is going to be are you driving unit economics for the business? Are you helping a customer increase their gross margins? Are you reducing their costs? Are you reducing their risk? And it goes back to old age world where the value has to be a share of the value you provide to the customer at the end of the day. And there's a lot of innovation that has to happen. A lot of applications are going to be built and the pendulum will likely swing for wide range of applications or technology beyond just inferencing on those side. And that's what we are seeing. Our enterprise customers are applying AI in every part of their business. And you know, you just cannot do that with frontier token pricing today. Right. So it's cost prohibitive in some cases. Yeah, 100%. And in general productivity, employee productivity is one of the, I think like when we are making a model to build a business, you know, we are probably like in a San Francisco company, probably one of the top, up 1,5% when we hire an engineer, we model now $6,000 per month. So let's say it's about $70,000 per year added cost that we are going to plan for as you build kind of scale. Scale.
- [16:13]
A
It's a big number.
- [16:14]
E
Yeah, it's a very big number. You used to have let's say like maybe 5,000 a year. For tools that we would provide to our employees. Now the number has gone quite high. That math doesn't really work in the enterprise with 100,000 people for a business that is selling shampoos, for example, and
- [16:38]
A
the employees all love using these tools. That's the crazy part. They're like, this is awesome. I can get my work done faster. Lon, but you get your work done faster. But does it actually result in roi? And that, I guess, is the core question now on.
- [16:52]
B
Yeah, and it's a lot of companies are using sort of a lot of models together. Like we had Manjil Shah from a Hippocratic AI on here last week. They run 30 models in tandem on some of these medical questions. It's just there's no way to do it if you're paying all of those models these exorbitant prices. I did want to double click on on one more thing before we move on. The SpaceX strategy seems to be going after token efficiency. That is like you don't use as many tokens and that's how you sort of lock in savings. Can the other Frontier labs follow suit or are there business models kind of locked into this? Get you to token max and use as many tokens as possible. Sarah, we'll throw that one to you.
- [17:30]
D
I think to launch a new model which isn't at the frontier, you have to say, hey, the benefit is efficiency. There's a real benefit in that, right? Like there's a real reason why not all tasks require brute force reasoning. Max 20 minutes of deep research. So very valid positioning, especially for agentic tasks which have, you know, if we retrace what's blown up, a lot of the costs, it's people want to do more things integrated in the real world. And when you want to do that, basically you have these long horizon tasks and you have much more tasks that go awry. So you end up spending a ton of tokens and not getting the results. Multiple retries. Will other Frontier labs do this? Efficiency is like one of the most dominant economic drivers on both sides, right? So for users, they're super conscious of how much they're spending. For Frontier companies, inference is a serious cost. So Chinese labs are interesting because most of their computers in training, very few of them are doing inference. And frankly, I don't think they can afford to do inference very well. I think even if you look at the the models that are being served, the latency is not the same at most frontier labs that are very compute heavy in the west, where there's more Access to compute inference is a beast and so anything you can do to decrease that footprint, you're going to do that means the drivers are there. Probably what has happened with SpaceX is they've made a calculated decision. We're going to position this for the user as something that costs less and we're also going to probably drive that price down to drive usage. Because the main lesson I think from all the last releases is that real world data matters and understanding how people use it will help drive that even further in the future.
- [19:18]
A
I had Andrew from Cerebras on the all in interview show the other week when I was here in Paris at the race conference and he was talking about like they cannot build out Sara inference fast enough. And he's got the, you know, best inference chips along with Grok, bought by the other Groq with a Q bought by Nvidia. And they were like, we, we have like 30 years worth of, you know, orders or something. Insane to that extent. But I, I did have one question for you, Spiros, which was I saw Austin Allred, who did Lambda School back in the day, teaching people how to do code and I was an investor in that. I think he recapped it so I don't know how much I have in the new company, but he was saying he's got these interns and he's doing these AI challenges and he said, hey everybody, make a small language model. And he gave them this assignment. Making a small language model, I would think Sparrows is a little bit of work and takes some technical excellence, but apparently people are able to do it. So he had this one person, Nathan, he made a small vertical language model, an SVLM to do one thing really well, read music. And he gave, you know, how much more accurate was. And then another person, Katie, he said, I fine tuned a small Quin 3 model. That's another Chinese open source model. Take a geometry scene with no coordinates and spit it out, tick Z that correctly compiles, yada yada, and you know, again, also huge gains. Do you think we're going to get to the point, Spiros, where people are like, I'm just going to take this model in house. I'm an enterprise. I'm going to build a team around this and we're going to build our own models or we're going to train vertical small language models for the accounting department, for our research department, in plastics, for our distribution, et cetera. Is that the next card to turn over?
- [21:04]
C
I'll approach it from two perspectives. So what we see, and we see this with our own work, right? Like Resolve works in, you know, running and let's say debugging all software production systems where you have a lot of complexity, you also have a lot of data that's specific to your own software system, to your own organization. What we call tribal knowledge, maybe, right. And I think that's true for many, many deep applications. And we see that actually most of the value, let's say as the reasoning of the frontier models becomes great, most of the value actually ends up coming from how you apply this data to your specific business with your specific scenarios. Right. I think that last mile is what actually, if you used well, creates the ROI in the end, right. So most of the value is there and somehow you have to use that. Right. And you know, so far most of the approach is keeping this data outside of the model, but somehow putting it in the context. Right. But I do think now it's probably or it is more effective if you somehow can put that in the ways of the models. Right. Now I don't think though that where we're headed is that every company is going to create custom models for every part of their business. I do think that's going to happen for what is the core business, how they make money and then maybe for the accounting department, like you said, if they're not an accounting company, most likely they're probably going to work with somebody that has done that on their behalf. And that's partially driven by talent. Right. Even if it becomes very easy, I think talent in this domain is always going to be scarce, same way as the chips. So I don't think people are going to be applying a talent in areas where it's not core to their business. But I do think they're going to be working with partners like Resolve, like others in the particular domain which is not core to their business, can provide them true ROI right at the last mile by utilizing the data that they have to them specifically.
- [22:43]
D
I was going to be a bit cheeky and maybe say this. I think the main blocker to people doing customization models again is their hangover from the first time they tried to customize their models. And realizing, I mean, spurs is pointing out the fact that you need up until now AI research staff, you need data, but frankly, tomorrow you might want a different capability. And most people are like, oh well, it's really overfit to now doing. I think your example was like music or geometry. What's more interesting is how do you automate that and how do you make it way more predictable that people can get gains if they invest time. I think this is a major question that most AI researchers are very interested in right now. In most frontier labs, it looks like, how can we do self improvement to automate our own trainings? They're not really incentivized to package it externally. But I think if I think about what we're working on and what others, it's like, how do we actually teach agents to drive customization in a way that's much more predictable and eliminates so much of, frankly, the cycle of regret that people have had previously where they've tried to do something and it's had unpredictable lift. So I think that's super interesting to think about. The same way that code has supercharged. Like, really anyone being able to be an engineer, how do you create automatic harnesses and automatic customization? Why that matters is that most business owners that are not investing in having their own AI research staff, they're always going to ask themselves, prompt engineering is not great, and I hate it, but it feels immediate and like the closer you can get to that and giving predictable gains, I think that's a real testament for how much companies will invest in the long run. And that's really interesting to think about because speed matters here.
- [24:42]
A
It really depends on when you embraced AI fully in your organization. Like, if you embraced it nine months ago and you got a slot machine that put out, you know, sloppy stuff that didn't actually get you gains, you're like, oh, AI is a waste of time. But you haven't seen literally in the last six months how the entire industry has churned over, like two or three revs. Whereas if you're starting today and you did it this week, the same jobs, the same prompts would have actually solved the problem and there would have been roi. I am. I don't know if anybody else is having this inability to keep up and the manicness that I feel that now. When I was building my little. I was building a little podcast player that does, like, deep linking. So it finds, like, the same topic across 25 podcasts and then let you jump between it. So if they were talking about, I don't know, Mythos, it would just jump to the part in the podcast where they were talking about Mythos. It was, like, very helpful for me as a podcaster and as a researcher and whatever. And I just took the same prompt and I put it into three, and I was like, I'm gonna spend 20 bucks on three different ones and see what the output is, I did it on GLM 5.2. I did it on Claude, and I did it on Grok. And I was like, you know what? Grok was the fastest and best. Okay, fine. But I just feel like my head is spinning at this point, Manu. And in terms of my ability to keep up, and then I had fired up Hermes at the same time. I just feel like I'm spinning right now where they're just giving me more and more weapons. I don't know which gun to pick up. I'm like at an armory at a shooting range, and they just put like 20 more guns, grenade launchers, whatever. I'm like, what am I supposed to do? They're like, just pick up the gun and fire. It's going to work. It's chaotic, isn't it, Manu?
- [26:23]
E
It is, absolutely. Yeah. It can be. And I personally am also doing the same thing. I've got just right now, like, 24 agents working in the background.
- [26:40]
A
You are token maxing.
- [26:42]
B
That's a lot.
- [26:43]
E
I don't get time to do this exploration every week, every day. I've picked a handful of kind of portals or tools that allows me to, you know, try different models. And after using all these models, you form a pretty good intuition what model can take on this work and where I need to get the best model to do things. And I suspect that a lot of people are probably going in that direction where they are just picking the models that can do the job and.
- [27:26]
A
Which is crazy when you think of it. We put it on to consumers now, Spiros, to figure out, like you figure out which model works for your query. I do anything that's local, anything that's travel. I go to Gemini immediately because they've connected Google flights, Google Travel Hotel search, and Google Local and Google Maps so perfectly that I'm like, where can I take my daughters for ice cream or chocolate mousse tonight in the eighth and Drosman in the first. Indra, wherever I am, give me a map, tell me if it's open, and it does a perfect job. If I ask Claude to do that, or Grok or whatever, it doesn't have that local data. So I'm a human router now, trying to figure this out as well.
- [28:04]
C
A different version of this, which is like, when you are building on these models, which I think many companies do, whether you're an AI company, like we are, and you're building on frontier models plus your own models, or let's say you're an enterprise and trying to Solve a problem. It is becoming hard in the sense that because things change so fast, like in the past, you would build an application once and you were done. Right. You could use it for years maybe. Right. Maybe you could build a whole company around it. Now you have to actually figure out how to do this monthly or weekly. Right. And to keep up essentially when a new model drops to know what does better. Right. Or when a failure happens, should I go solve this failure now somehow by creating maybe deterministic code? Or should I wait for the next model to drop and it's being solved? Right. It's not worth the investment. So that makes it actually hard when you're building applications on these. Right. So the way you're building our software in some sense is dramatically different than what was in the pre AI era, which also makes it quite hard. And also one of the reasons why you see maybe some of these larger enterprises in SaaS and all of that that are being disrupted. I don't think it's just that the unbundling that happens. Right. It's a very different way of developing software now and they don't know how to do it.
- [29:08]
B
Actually I was reading recently, I don't remember where I saw this, but they're saying a lot of the tests that you would do on a model, it can sometimes take a few weeks to be like, how well can it handle incredibly long term complex test. By the time the test is done running, there's a new model out to test. So we don't even really necessarily know generation to generation how powerful the models are until we get the next model and everybody gets excited about that one.
- [29:33]
A
And the crazy thing, Sarah, is I don't think it feels like it's accelerating. It feels like every week the speed is increasing, like I don't know what, what model next week I'm going to be told I have to use because it's that much better. And it's correct. It is that much better.
- [29:49]
E
It's.
- [29:50]
A
It's causing whiplash. Yeah.
- [29:51]
D
But I honestly think this speaks to the issue of it is that the burden we place on the users to decide this is actually the problem. Like if I would, if I were to characterize this whole conversation, it's exactly why, how do we have self improvement and continual learning is one of the most critical questions. Because that example of it takes enterprises months to figure out are they doing better and then by that time they have a new model. The truth is the complexity of optimization and choice and routing, that should all be automatically designed. It's the same thing with now how every company is building their own harness and everyone's making different choices. I think everyone's ending up with slightly different takes. But the issue is it really should be just an optimization problem. We should know your objective, what you care about and it should just be auto learned how to get there. And I think that's one of the most important. If I think about how conversations have changed year to year, I really hope this time and next year we're having a different conversation which is not about the paralysis of new models, but actually like just about, hey, have I seen improvements on my task? And like not about which model I've been thinking about using because I feel like that's a major kind of prevalent thread I see in all these conversations. And it's super interesting to think about the amount of burden we placed on the end user.
- [31:15]
A
It would be like in the YouTube era, the Netflix era, when Web1Point came out, we would be just expecting people to like, hey, you know, if you had a Seagate hard drive and you ripped it out and used this western digital one, it's going to be much better. So everybody's like, okay, I got to take my Seagate hard drive, like what am I doing here? I'm like changing from a VGA monitor to an ega. Just, it's madness right now. Lon. The next story I think is super important because it relates to, I think how are we going to manage as an industry the perception of these models and the danger and quote unquote, and
- [31:51]
B
the speed at which these powerful new models are arriving. So Google DeepMind chief Demise Hassabis has proposed a standards body for Frontier Class AI. In a long ex post, the DeepMind chief argues that AGI is approaching, he sees it as miraculous, the dawn of an amazing new era of abundance. But he also argues that we're going to need robust safeguards to maintain control of these and I'm quoting here, increasingly agentic, recursively self improving models. So it's basically we, we gotta proceed with caution. There's so much uncertainty, the stakes are so high. He is proposing a standards body modeled on public private partnerships. He cites finra, the Financial Industry Regulation Authority. And he's saying funding would be substantial and would come from the industry itself. So this group would test Frontier Class models for national security threats, cybersecurity threats, that sort of thing. And as well, Frontier Labs would be encouraged to adopt a bunch of best practices, publishing model cards or technical details, maintaining strong internal cybersecurity vetting Key personnel and so forth. He's sort of hoping this starts in the US and serves as sort of a jumping off point sort of internationally trying to sort of codify because as we've talked about a lot, we kind of have this wild west situation now where the government sometimes says, hey, we need 30 days to look at this model. Other models just come out. What I need to know what the panel thinks about this. Is this going to stifle innovation or, or is this just, you know, common sense? We got to do something like this.
- [33:20]
E
I think the devil is in the details and I do think that a better education, a better point of view by the government is generally better about a technology. And however, you know, if we end up with, you know, yet another agency, you know, I think we have about over 300 agencies, it always increases every year. I'm not sure it will actually, you know, if it's not implemented really, really well, that is not necessarily blocking the models, but you know, is providing maybe guidance and, and assessing risk. I could see that to be generally helpful, especially when it comes to national security reasons. But I can't imagine a world where we are training a model and we go to this place to get the model certified or approved to get in the hands of the customers. That would feel like very premature right now. I don't think we are very close to what people are saying that there is a, whatever people call ASI and so forth. Like 10 years ago you showed me these models. I would say this is totally ASI or AGI. We do have an AGI right now and these models are incredibly great. They are going to do incredible things. But at the end of the day these are tools and these tools are being used in everyday applications. And you know, and I think it, you know, I hope that, you know, any government intervention comes with some sort of that understanding. Like it's not, you know, kind of crazy alien technology. It is, yeah, like it is doing some things really well in coding and cybersecurity front and that has international security implications. But imagine some country could launch a really powerful model that is really great, let's say in cybersecurity. And we are stuck in a kind of government process to get a model out to defend ourselves. I mean, that doesn't sound that great either to me.
- [35:53]
A
Spiros, what's your take on this? Should the industry regulate itself?
- [35:58]
C
I think that it's a little self inflicted to some extent what's happening right now, in my opinion. And it's worse than just the government. We see, in the U.S. i think in Europe, for example, everybody now woke up and wants to have completely control of the models and the deployments. And I think all of that came a bit from the doomerism and all the discussions about. I think it was sequence of events that got us to this point. Right. So maybe what David says is potentially reasonable in the sense that at least you want to have predictability on the government suddenly intervening and turning away a model. Maybe if this to happen at least should happen in a way that is predictable and everybody understands it. But I don't think it's necessary and I think it's self inflicted and I think it is actually just. It's worsening than that because now I think maybe everybody realize or maybe think that now they need to control intelligence locally right in their own country. And I think that's possibly hurting in some sense, maybe US based, let's say intelligence, if you wish. So I'm not in favor of this. I think it's too early, probably we don't know enough. And I don't think there is existential risk in any way. And if you grew up in Europe like I did, you know that that's not the solution to our problems.
- [37:09]
A
Well, I mean, Sarah, the government can, is so far behind in understanding it. I mean, if we're using this every day and y' all are building it every day and our heads are spinning at the pace the government is hopelessly behind. So it would, I guess what we're trading off here is a slow government that maybe people would trust more or going fast and the industry regulating itself, like the mpaa, the Motion Picture association regulates itself and consumers seem pretty good with that. But if it came to like nuclear or flying planes around, you know, most citizens are like, hey, I really don't feel comfortable with, you know, Boeing and NetJets and American Airlines regulating themselves. I would rather the government do that. So is this closer to nuclear power and aviation or is it closer to movies, you know, and comic books and music and albums in your mind?
- [38:07]
D
Yeah, I mean, man, I'll give a very nuanced take. Maybe I'll give, maybe I'll give a less nuanced one just because, I mean, I know a lot of the heads of different Frontier Labs. I think some of them are very intent in their desire for like, very authentic in the desire for why this is happening. They truly, they truly believe that the rate of progress is impacting safety. If you truly believe that the idea that an entity would basically control access can be an Attractive one. There's two problems, right? So what models get countered under this? As more and more companies have frontier AI, which I think most of the conversation today has been how, hey, with this technology so important, more and more organizations are going to own it. How is the government going to choose what counts and what doesn't? By the way, this was the same dilemma when they first introduced compute thresholds a few years ago. They said, any model above this amount of flops we're going to audit, we're going to put in this safety pen quickly became a disaster because models became smaller and just as powerful. This is what happens when you try and do these hard takes on fast moving technology. I think you have to ask, what are the incentives of people who really believe in this? And candidly, whether it's authentic or not, I do think it has an impact on who's allowed to build and shape frontier technology. I think that's wrong, given this is one of the most important technological changes of our time. Binary stories always require closer inspection. And I think the idea that everything is completely unsafe and we're hurdling towards this unsafe future is a binary story where there's a clear villain and a clear hero. And that always, I think, makes the average person say, hey, what's happening here? And like, why are. Why are certain people advocating for this?
- [40:09]
E
I would say one. I would add one more thing. Silicon Valley does not realize right now that how close we are towards nationalization of large language models. If Silicon Valley continues to basically say that we have invented this technology, we own it, where it's so powerful, that it might increase all this job concerns, which is partly, I think, the reason why majority of Americans are scared of AI right now. And that is actually a concern, in my view, because, you know, it shouldn't be the thing that most Americans are actually concerned about their jobs right now and that AI might just take away. Take out the things and, you know, kind of believing that, you know, these two things can coexist for too long time and government doesn't come in and intervene is absolute lunacy. Right? Nobody's going to like nationalization. Silicon Valley will be the first people say, like, you know, what the f is this? We don't want this at all.
- [41:21]
A
It would be a pretty extreme. It would be a pretty extreme moment for, you know, Pete. What's the guy's name? Pete Hegseth. Yeah. Like, he's like, you know what? We're just taking. We're taking Mythos. It's ours now. Dario, show up at The Department of War. Here's your desk.
- [41:37]
E
And we just saw, and we just saw a preview of that right now and see how nervous every, you know, people got here. Like, it's like, oh shit.
- [41:45]
A
Well, it's self inflicted, as Spiro said. Dario went out and he said, listen, all jobs are gone. We've created the new God in the sky and we're scared to death about what we built. We're going to give it to you. No, we're taking it back. No, we're giving it to 50 people. Like, if you're panicking, what do you think the government and the populace are going to do? They will panic as well.
- [42:05]
B
They still do it all the time. There were three anthropic people who signed that A is going to take everybody's jobs away letter yesterday that the economists in it.
- [42:12]
A
Yes, it's, I mean, if you run around like the house is on fire, people are going to believe you. Like if you scream fire in a theater, it's like the classic kind of thing. Like, you know, it's your, it's your movie theater and you're saying it's on fire. Okay, we're going to believe you.
- [42:26]
E
Let's say, like these models are going to be that like, like if they're marketed, like they're just so powerful. Just believing that while we're not going to end up in a world where we're clearly government and our military is going to have a unnerved version of these models and so they're going to have to go get those models built like that are military class and for national security. Right. And so just like in aviation, we've got civilian airplanes, we've got fighter airplanes. I think that, you know, we are very, very close. We are testing the boundaries right now with these kind of narratives where, you know, we might end up with, you know, there's, there are civilian models and then there are military class models that, you know, the government just funds to make so that our country can do the things it does with pretty incredible technologies in the military.
- [43:19]
B
Sarah, I did want to throw this to you before we move on at adaption. You're working on ways to get models self learning, self improving without needing us to constantly retrain them and teach them new things. Is this all kind of moot in a few years? Are the models just going to be submitting themselves to the board of authority to make sure they're allowed to keep going? I mean, is, is self recursive improvement going to sort of make this all a secondary concern?
- [43:45]
D
I mean, I think the advocates of that letter would say this is a concern, right? So I tend to take the view that the most powerful tools are powerful in both directions. They can be used for good, bad. So it's, it is important we say, well, the self recursion, how do we bound it? Particularly what we're interested in is how does self recursion help companies own frontier AI? So candidly obvious, how do we all unpredictability with training, with designing harnesses, how do we basically teach models how to do that for a task and do it as fast as possible? I think that's an incredibly important application of AI because it accelerates innovation and it allows people to focus on their question. And I think that will models submit themselves? I think this goes back to the question, what objectives do humans set and what's considered important? Is this body going to be the important referendum on safety? I suspect, given past efforts to do this and to set binary thresholds, no, unfortunately the mechanisms there tend to be incentivized by limiting capabilities to a few, but not necessarily in terms of actual work on safety or guardrails. I'm more interested in how we actually funding research and sensitive domains. How do we actually incentivize that within industry when we're shipping frontier models for medicine or science? There's actually so many super important safety questions there that also have grounded in the reality of making AI work. And I just find that to be a much more refreshing conversation than speculation about like, oh, where will self improvement go in five years? And like, oh, this is all because it's just not precise. So what I like when it comes to safety is, hey, the reality is now billions of people are using this around the world. And like that's incredible. That's, that's the type of technology that's like a mobile phone leap. What are the actual implications of what they're doing? Like, what are we seeing with things like misinformation? What are we seeing with things like the reality of actually doing acceleration of science? Those are conversations that I really enjoy and often contribute to because this is just a bit too unanchored for me. It ends up being a church. It ends up being like, whose God do you believe? Do you believe that doomsday is tomorrow? Or do you believe that this has real impact, that's positive for innovation? And I just don't really enjoy conversations like that to the same degree because it just, there's often no convincing and it's not really anchored to like what we see in practice.
- [46:34]
C
I Completely agree with Sarah, by the way. I think sometimes it will benefit us if all of us who are working in this area, especially those who are more visible, instead of going out and saying like, resolve it. So AI is going to take all jobs or software engineering is done, which is not, by the way, people are hiring more software engineers. Talked about all the benefits that we've seen from technology becoming prevalent and all the problems we couldn't solve and the things that maybe were too expensive are becoming, let's say, accessible to a lot of people now. It will help everybody, actually, including our own future and destiny and what the government really does. Awesome.
- [47:06]
B
Well, everybody on this panel, of course, an accelerationist, but Jason, you got dragged into a bit of a decelerationist face off this week. We're of course talking about Uber and Waymo. They're on opposite sides of a new proposed legislation in Washington D.C. specifically they are thinking about altering their Autonomous vehicle Act of 2012 to allow autonomous vehicles to begin operating even commercially within the city. This would allow for driverless testing and for commercial robo taxi operations. So currently Waymo and Zoox can test their vehicles in dc, but they have to have a human operator behind the wheel at all times. Uber opposes this new bill, but Google owned Waymo supports it. Uber argues that this is going to displace human drivers. It will give Waymo a monopoly in the D.C. area. They also have all kinds of arguments about how it's not good to have a fleet of purely autonomous robo taxis. They stall out on the street, they can't help elderly and disabled drivers get in and out of cars and so forth. So they're arguing for more of a hybrid model, which means you can operate a robo taxi in a city, but it has to be a small part of a larger fleet that also includes human drivers in it and that there should be sort of a balanced model. Waymo says that the bill would allow them to safely deploy autonomous vehicles. And you could still use public transit, you could still use ride sharing. Companies like Uber, they're debating this bill in DC's legislation now. I just want to. Yeah, we're throwing it to the panel. What, what does everybody think? I mean, Jason, what was your, what was your take on this?
- [48:43]
A
Well, I got, I got dragged into this because of course the Uber investment. But I full disclosure, I'm invested in like no less than 10 of these companies and self driving like across the board, whether they're public or private. I've got tons of investments in this because I believe in the category. But I believe this is the first place where we're going to have this battle over jobs. Because in markets where Waymo has hit some level of scale, you're seeing drivers lose 10%, 20% of their revenue, according to some studies. They're getting less jobs and the pricing is going down. They're way ahead of us. There's like almost two dozen players in China with cars on the road now. The ccp, the Chinese Communist Party, they are very sensitive to protests and they are very sensitive to civilization unrest. They've started to have flare ups with the 10 million, 20 million drivers, taxi drivers in these major cities that are losing jobs. So they have put a moratorium on any more self driving cars. It was the wild west for a bit. They are now licensing self driving cars and they're doing it at a pace that will give them time to not lose or not have a lot of people essentially lose their job and have civil unrest. That same thing is about to happen here in the United States. And so what you're looking at is a company like DoorDash or Uber, they are both investing at the same time in building their human driver networks. And they have for 15 years. DoorDash has their own robot. We'll pull it up here. I forgot the name of it, but they have dots. It's dot.it's essentially like a little motorcycle. It's a brilliant form factor because it can ride in the bike lanes, it can go up onto the sidewalk. So at the same time DoorDash has this and door dashers. So they are going to a limit every time they put more of these on the road. What is this thing called Dot? Every time they put more dots on the road, they, their drivers lose their jobs, the unions. And then same thing happening with Uber. Uber has I think 12 partners for AV and they've invested $10 billion in it. The unions now are saying, hey, wait a second, why are you guys eliminating the jobs of the partners? So imagine you're Dara or your, you know, the folks over at DoorDash, you're eliminating jobs at the same time you're trying to keep people employed. And then you put that against cities. And a city like Boston, a city like, you know, or state like Jersey or dc These are the places with the strongest unions representations. And this will be everybody's first. You know, we'll intellectually talk about like white collar jobs and are designers going to lose their jobs or developers? That is nothing compared to what's about to happen to rank and file Entry level employees. Imagine you were 20 years as a driver and then you happen to be in one of the waymo breakout cities and you're going to lose your job in those cities. Uber and Lyft and DoorDash stopped advertising to get more drivers. So I have a prediction. I believe what's going to happen is we're going to move to a licensing regime in each of these, or I think the majority of states with strong unions and they're just going to say you can have 100 cars, it's $30,000 a year for a license, kind of like a medallion, and they're going to slow roll this. Now there's all kinds of accusations that Uber and DoorDash are trying to slow roll this so they can catch up. That all may be true. But this is the first battlefield. This is the one where the public, everybody is going to have somebody in their family who's a doordasher or an Uber driver, and this is going to become the battlefield for AI. And how we handle this is super important, I think. And I think you can have empathy for those drivers and think, how do we give them a soft landing? And I think actually the medallions, $30,000 a year for these and roll them out slowly so you can kind of taper those jobs for, I don't know, five years. Because we'll be sitting here in seven years. I think half the rides will be automated, minimum. And so, yeah, I'm watching this closely and I think it's going to be a huge mess and it's going to be the presidential election in 2028. This will be like a cornerstone discussion. All these trillionaires, billionaires, big tech companies, eliminating this category of work because nobody gives a shit. Excuse my French, but I'm in Paris. Nobody gives a shit about, like a developer losing their job. Boohoo. You know, or they didn't make as much money or, you know, a lawyer or whatever, but you start taking away DoorDash and drivers, people are going to be really sympathetic to that.
- [53:07]
C
Well, for what's worth, I agree with that. Right. I think the whole discussion about developers losing their jobs was completely wrong, in my opinion. It hasn't happened. And this is not like where the focus should be in some sense. Right. Like, we're all great here, actually. And I agree with you. Right. And I'm empathetic too. Right. Like, I think many, many people just get by by being able to drive for DoorDash or Uber. Right. And I think when this actually hit, I agree with you. I don't know the answer, by the way, at all. But I think it's much more serious issue than the white collar. Maybe like maybe potential displacement we see, we discussed, which is not. I don't think it's happening, by the way, but if this happens, agree with you.
- [53:39]
E
Right.
- [53:39]
C
Like probably millions of people who actually depend on it. Right. And I think that probably requires more serious discussion about how it happens. Right. And maybe more government related kind of intervention. And I think it's going to be forced, like you say, because of the election problem.
- [53:54]
A
Sara. This is, I think, one that's grounded in a little bit more reality and less religion. What's your thoughts on this transition here? This one's anchored in like the reality of I have to feed my family, they're dependent on me.
- [54:07]
D
Yeah, I actually really enjoyed your take because I think you're, you're closer to it. You've been thinking about these companies for a while. You know, my dad told me, he said no matter what, no matter where in the world, people care about three things. They care about health, family and income. And I think that this is very entangled with the general sentiment. You mentioned it. But the elections will probably break this too, the surface of how people feel. People want to feel that they will do better than the previous generation. And if they are uncertain about that and regardless of the causal factor, they will feel like anxious about change. And I think that probably what you are saying is true. This is like a petri dish for all those factors coming to bear. The only thing I would say is it's super important to remember the medallion conversation was part of the last massive technological change with Uber.
- [55:10]
A
Yeah.
- [55:11]
D
And you know what's interesting is that there was the same people involved. There were taxi drivers that, I mean if you think about black cabs in London, you spend a decade studying for that. And the, you know, the consumer everyday voter consensus was we like Ubers. And I don't think there's a right or wrong. I actually think this is super important and it will be the topic. But the ultimate answer will probably depend upon how consumers feel, which is interesting. And like we can't forget it in this conversation is that this conversation has been had before and the decision last time was no medallions. Right. And so I think the question is, is the temperature different this time? Are we standing up for that group of people who are feeling impacted and can it be articulated as like, yes, we will preserve medallions or something like what existed before? Because that would be kind of interesting because we're almost returning to the previous social commitment we had for that group of people.
- [56:15]
C
Yeah. I think the difference maybe is that in the previous chain, previous disruption, it was like we have a few drivers and you're going to more democratization. Right. When it came at least to who could participate in this. Now I think the sentiment is going to be like you're going from all these people to technology companies getting all the value in some sense. Right. I don't think that was exactly the sentiment with Uber. It was mostly about okay, I have a medallion or anybody can drive. I think this is going to be worse in that sense. Regardless. I agree with you, the consumer probably sentiment. I know what it's going to be. Right. But I think the popular sentiment might be like the value is being transferred from people to take now.
- [56:51]
D
It's very fair. And I think what Jason pointed out, which is really the core of this is that many more people this time have an experience of knowing someone who's done a gig job. Like, you know, it's very interesting. It's a very large fraction of urban cities who have done some type of gig job. And most of what it's interesting because, you know, I work on building technology and frontier technology. What matters most to humans is like what they weigh as like their experience and you and for that reason it's what they feel about, you know, do I know someone who's been through this and I'm in a very much index and trust that. So again, you're right. I think it could be different just because one of the characteristics of Uber is that it actually has touched millions of people, provided jobs at different times for different people at different stages of their lives. So they have like very concrete memories of what that is. Even if they're no longer a driver,
- [57:45]
A
you're going to have a lot of moms and dads who are like, I dropped my kids off at school, I couldn't get a full time job, but I drove for four hours and then I picked them up and now you're taking that away from me and that's just going to feel, it's going to hit. It's. I predict it's going to hit people in a very different way. But now of course, if you have taken away Mo and you've had that glorious experience of not having to talk to the driver and having your privacy and making a phone call, it's like chef's kiss.
- [58:15]
B
It's bliss. I love it.
- [58:17]
A
So nice, so nice to not have to like worry like somebody's hearing My conversation because I'm a micro. I'm a micro celebrity. Yeah, Manu, what do you think it's
- [58:28]
E
going to be as big of a deal because. Well, first of all, these jobs are something that people subscribe to or basically like they have a choice to do and they're not necessarily kind of understood as like a long term job for anyone as far as I can tell.
- [58:55]
A
Second, you're wrong on this one.
- [58:59]
E
No, so. So your argument is basically saying that, you know, this is a static world and those jobs are going and there's nothing else happening in the world. And what I'm saying is that while in a freelancer economy, people have choice and they're basically actively seeking new opportunities to make more dollars and the technology is enabling people to have more choice. You could start a small business on Internet with these AI LLMs. And I'm seeing so many stories where people are earning a lot of money, building interesting, solving certain problems that, you know, with just asking an LLM to make a application, that's just a very small example. It's not like necessarily meaningful. But what I'm seeing is that there are lot of new opportunities where maybe people are going to create, maybe the creation economy is going to be much even more powerful, have more earning power. And at the end of the day, the rollouts of these self driving cars are going to be dependent on two factors. One is consumers. If consumers like it, that is obviously a litmus test. The second is the city and state officials are going to look at where is it increasing safety in their cities and state, is it bringing more revenue and is it giving people more choice? And generally speaking, that was a core argument Uber made to roll out their technology to all the cities and states and they basically won for most part with a lot of dollars in lobbying experience.
- [60:45]
A
Will back me up on this one. There are people who this is like their lifestyle choice and that's the fabric of it. And the other piece of it is, you brought up Manu correctly, like is this good for the economy there? And then the people running that city make that decision. The argument they're going to make and it is going to land so well is that in Boston you have these, you know, 250,000 drivers and 250,000 doordashers and they live there and they make this money and they spend it in Boston and they go to, you know, Boston Celtics games and they go to restaurants and they have rent and apartments. Now all that revenue is going to go into the Waymo Box and get sent to Mountain View. And that's going to feel profoundly unfair to tens of thousands of local communities that like, oh, wait, all that local revenue doesn't go to the people in our neighborhood, our friends, our neighbors, our cousins. It just gets sucked up and shipped to some, you know, Borg of a company, Spiros. I think it's, I think this one's going to be the one that breaks me.
- [61:48]
C
I agree with you, Jason. I think that's going to be it. By the way, I agree with Manuel. Right. In the long run, we're better off, obviously, not stopping slowing down technology and everybody's going to be better off and more economic value is going to be created. But I think this is happening quickly and I do think that area is going to be more painful than again, software engineers may be being this painful.
- [62:07]
E
Yeah, but that's true. So no state city is going to like a monopoly self driving car situation in their neighborhood. But that also assumes that, you know, Uber is not offering more choices to their freelancers. Last time I checked, Uber is actively trying to offer more opportunities to their freelancers. They are, I think, you know, you
- [62:40]
A
are correct, TaskRabbit type stuff and then data labeling. They were trying to offer that kind of stuff.
- [62:46]
E
And Uber arguably is the biggest platform for work today. So like, if you look at just raw hours that humans spend in a platform to provide value, it is larger than anything else on planet, as far as I could tell. Like it's bigger than upwork, it's bigger than any of the other freelance gig economy things people do. It's bigger than some of our markets we are in. And they are actively trying to go after other opportunities. And I think if they're successful in that, you know, that is perhaps the kind of soft landing, I suppose, where there's certain, certain part of the network is automated. But you know, they're bringing work that cannot be automated and offer opportunities to the freelancers.
- [63:34]
A
All right, Lon, we want to get to this last story. We'll do lightning rounds so our guests can get back to work building the Future.
- [63:40]
B
Absolutely. So two of the biggest CEOs in the AI space are feuding once again on social media over the weekend. Elon Musk reshared an ex poster back in March stating, and I quote, I'm quoting here for the record, scam. Altman is super good at scamming. To which Sam Altman replied, homeboy, again, I'm quoting here, folks. Boy, that's. This is a Sam Altman original. Homeboy, you're the one selling public Market market investors on short term data centers, space data centers. Then Musk shot back, we start flying them next year. Maybe you can come to see them if your parole officer approves.
- [64:17]
D
A pretty good snap.
- [64:18]
B
Elon gets a lot of. Eli gets a lot of trash talk for not being funny. That's a pretty good snap. So setting aside the big personalities involved, nobody has to comment on that. I'm not going to make you guys do that. But SpaceX AI is planning to launch a fleet of orbital data centers that will perform AI investigation inference tasks. And what that was one of the key factors that sort of got the retail investors so excited about the ipo. So I wanted to throw this out to the panel. We'll go one by one. Are orbital data centers really going to launch into space in the foreseeable future? Make your predictions.
- [64:54]
E
I think so. It is. From SpaceX perspective, it is a much easier problem to solve versus the problems they have already solved in the last decade. They have like to make Starlink work, they have to have like, you know, every satellite have a laser system that is pointing to each other, communicating, calibrating at the micron level and so forth. Like the underlying pieces, the foundational technologies to operate such a thing is so immense that, you know, the orbital, orbital data centers don't have to be like data centers on Earth. They don't have to be that big. And they already have solved bandwidth problem and communication problem and you know, and for most part these orbital centers can be, let's say a few racks maybe of a payload that is in the orbit and you can totally put quite fairly good amount of workloads in there. And Norma mentioned they've also solved the hopping of workloads, right? Like the starlink. When you have a Starlink, you're talking with the phased array systems, you're talking to multiple. You're hopping to different satellites to communicate that this used to be actually a dream. I'm an aerospace engineer. It used to be in science fiction. It is totally possible. And if there's one company that can do it is SpaceX. And I think they will more likely than not prove it. It's coming faster than we think, but it will not look like the data centers that are on Earth. And luckily because the intelligence demand is so big that you will start seeing architectures and workloads where you can run modular data centers that can do asynchronous work and just give you an answer after an hour. There's a lot of those workloads today. We are even using A lot of that workloads with managed agents and things like that, where we don't want the answer right now. Just give me the answer an hour from now and just do the work. And those kind of workloads can be done in that architecture that, you know, Elon is talking about. And yeah, I think it's totally possible.
- [67:23]
B
There you go.
- [67:24]
E
And it's coming sooner than we think.
- [67:25]
B
Ask Sarah, what about you?
- [67:28]
D
That was a very spicy set of tweets. I enjoyed that exchange. You know, I think. Here's the thing. What are the main issues with data centers in space? One, you have 3% of GPUs failing each year, very predictably. Newer generations you have to sub in. All this is extremely difficult for space. Right. Because how are you going to remove the components that are not working? The other issue is, frankly, there's a ton of bugs that happen. So that type of data center only works for training because it's co located. More and more of the workloads are moving towards inference. Can they launch it? Will there be a case study? Go for it. I'm sure they'll launch something. Is it going to be a significant fraction of the computer in terms of where Frontier Labs are allocating? It will be a really nice photo shoot. I think that it will be a photo shoot at first and realistically for the next decade, because there's just too much risk. I'm a Frontier lab. The one thing I need to control because it's the biggest factor in my business and cost is my compute. Would I want to put a sizable share of my compute in a data center that's in space? Likely not. It would have to be for very, very good discounting. And I think that's probably what's going to be the framing at first. But that's how it begins, Right? So should they do it? It's a great mashup of their two bets, right. AI and data centers. Is it interesting? Probably not yet.
- [69:03]
A
Spiros, you got to take.
- [69:04]
B
Yeah.
- [69:05]
C
So I'm not an aerospace engineer.
- [69:10]
B
Me neither, believe it or not.
- [69:12]
C
Yeah. So listen, when the IPO was happening, you know, I had the opportunity to buy some shares before the ipo. And like, you know, the people that helped me ask me, okay, do you want to sell right away? And I said no, because, you know, you don't bet against Elon. So my take is different here. Right. I think, you know, this is not gonna. We're not gonna see the end of this until Elon wins. So my prediction is that it is gonna Happen. Right. Wow.
- [69:38]
A
He's not gonna stop.
- [69:39]
D
Yeah.
- [69:39]
C
If.
- [69:40]
A
If history is any predictor, he is the most dogged founder I've ever met.
- [69:45]
B
Drive him up there himself, it might take longer.
- [69:48]
C
It might take longer than maybe quite a few, but I think we're going to see it happening. And also, don't underestimate the cost of physical infrastructure on Earth. Right. And the combination of that and Starling actually is quite powerful, actually.
- [69:58]
A
In seriousness, Elon always gets it right, but he's rarely on time is what I've told him is my kind of joke. So, yes, they're going to be sending him out next year, but to Sarah's point, like, when do they are actually materialized? I would say three years. Right, Four years. But he's going to make his own chips. He's working on that already with his fab. And energy is the blocker now. So there's a lot of chips coming out, there's a lot of data centers, a lot of concrete being poured, but energy is the blocker. And once we start having that energy blocker in space, energy is not the blocker. That's the opportunity. And at the same time, in the next year or two, he will get starship working perfectly. In other words, it's going to land and take off, you know, within the same week, like he does with the existing rockets. There are existing Falcons now that have done 20, 30, 40 jobs, I believe. I think they've like got one that might have done 40 now. And so as that cost goes down, we'll see exactly what happened with Starlink. Starlink was a joke five years ago. Then people got it and they're like, this is good as a backup. Now people have it and they're like, this is my primary and I'm going to pick my airline. I would literally, I did this joke that went viral on Twitter.
- [71:11]
C
That's what I did.
- [71:13]
A
So I asked this question. Spiros is in the same way, would you rather sit in business or first class with regular Internet speed or would you sit in the middle seat in the row above the lavatory with Starlink? And anybody who works in our industry is like, put me next to the lab in the middle seat between two fat people. It's mind blowing how incredible. It changes flight when you have high speed. And these things are now getting to a speed that people don't understand. You can. There is a. There are like two very large size Starlinks. Maybe we could show one of the new ones. Everybody knows the regular Starlink that's like the size of a pizza box. People know the mini, there's a corporate one that's like the size of, I think three pizza boxes. And it might have multiple ones in there. And then you can bind them together and you can start getting gigabit speed. Now, people don't know this. They haven't really. It hasn't really registered on the plane.
- [72:06]
B
What do you need all this power for?
- [72:09]
A
No, no, but here it is. I guess they have it on their website, so. See that one? Yeah. I know people who are, let's just say rich people who, instead of getting like the regular Starlink, they're like, screw it, money's not an issue. I'll get the 300amonth one that I could put on my roof. I'll spend four or five thousand dollars a year. And you can get like these really large ones. And I think when the data centers come out, my guess is that the data centers will have Starlink built in. And so every time one of those goes up, and he didn't tell me this, I have talked to him about these in detail, but my guess is like, if you're throwing them up there, like, you might as well have Starlink on them. And the Constellation will just be all of those. I think there'll be a precursor to this, which is he's got all those supercharger networks which already have a lot of energy at them and are not used from 11pm till 6am Imagine at those places, he puts data centers there.
- [73:07]
B
You imagine at the Tesla supercharger.
- [73:09]
A
Supercharger, just take two of the spots and just dump. You know, put a Starlink on the top and they already have Starlink, but put a big Starlink there. So actually doing distributed compute at people's houses with powerwall. So imagine you put a powerwall in and he gives it to you for free or half price. But it's got a couple of GPUs in it and then they're all connected by Starlink. So I think distributed compute, whether it's on the Bittensor network and crypto, doing it that way or doing it at the superchargers, we need more compute. And the compute has to find energy. Where is energy? That's what's going to happen. Anywhere there's a hydro dam, that dam's going to have a bunch of GPUs added in space, there's gonna be a bunch of GPUs. If you got in a nuclear power facility in some backwater town in China, they're gonna be putting GPU'S there. It's just follow the energy. Right? Is I think what I was told by people who are building data centers, follow the energy. And the energy of the sun's the big win. So. But yeah, it's a three or four year timeline and it's heartbreaking to me that Sam and Elon fight like this. If we were sitting here 15, 20 years ago, I was having dinner or going to, you know, parties with these two individuals and it was all quite
- [74:21]
B
well, you fight fun and awesome. Most with the people who used to be friends. That makes sense. It's, you know, it sucks.
- [74:26]
A
It sucks. I wish the two of them would get along.
- [74:28]
B
I got your answer, by the way. 36 flights. The Falcon 9 booster B1067 completed its record breaking 36th flight on July 9, 2026. It was carrying 29 Starlink satellites into space.
- [74:42]
A
All right, we got a wrap. Sarah, tell everybody what you're working on, what job openings you have, and how people can get in touch. This is time for the plugs. Everybody gets to do a little plug here at the end. Little plug.
- [74:52]
D
Little plug, yeah. This is so much fun. By the way, I still want to ask if Manu wants to actually go to space. I feel like that's a side tidbit that was introduced. So we work on accelerating intelligence. How do we allow more companies and individuals around the world to build Frontier mod? We do that through self improvement and come work with us. We're attracting the best researchers and engineers across the world. So there's plenty of openings. Take a look and maybe I'll pass to Spiros.
- [75:21]
C
So Resolve AI is building models, frontier models for our domain, I would say, and agents to basically help you run software. The simplest way to think of us is we're the counterpart of, let's say, coding agents that take over when software gets deployed and you have to maintain it and run it for years. And when something goes wrong, our agents are on call. So you don't get paid in the middle of the night to wake up and see what's happening. If you're an engineer, our agents do all that work. And in terms of we're in San Francisco, we have a lab, we're hiring researchers, we're hiring also software engineers who want to work in solving kind of their own problem. Because at the end of the day, our users are software engineers primarily. So that's kind of the exciting part about it. We raised $190 million. We're a CVZ company and we're growing fast.
- [76:11]
E
Yeah. So Labelbox is deep in the reinforcement learning game. So we produce some of the most cutting edge reinforcement learning environments for those who don't know what these are. This is how the best frontier models today learn about something new tasks that are happening in knowledge work. And we also have an enterprise business where we have recently announced recursion platform where enterprises can use agents or build agents on the entire open source stack and can post train them with again, reinforcement learning environments that we can help them with. We are a company based in the Mission District. We're profitable and we're hiring lots of engineers and researchers.
- [77:05]
A
Awesome. I have a couple of podcasts and x com Jason. I invest in 100 startups a year. Yada yada. If you have a startup, email me. That's it. Jasonalicandis.com for life. Great job everybody. We'll see you next time on this Week in AI.
- [77:20]
B
Thanks everybody.