Meta Is Building a Prediction Markets App, xAI Bets on Grok’s Racy Side, OpenAI’s Ad Progress

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Wednesday, June 24th. Today on the show, the information has exclusive reporting that XAI is doubling down on its video and image generation tools, even with looser content guidelines. We'll bring on one of the reporters who wrote that piece. Next up, we'll unpack a report that Meta is looking to build a prediction markets app. We'll also break down the latest smart glasses release. We've then got Salmonova and Vista Equity Partners coming on the show to talk about a new compute initiative that they are working on together.

0:48and we will wrap the show with a check-in on the beaches of Cannes, where OpenAI is pitching its ad business. It's going to be a great show, so let's get right on into it. As XAI tries to differentiate itself with Grok, the information has exclusive reporting that the company is doubling down on AI tools that generate images and videos, although some of the content the company has become known for can be quite racy, according to my colleagues, Grace Kay and Theo Waite, who cover all things Elon Musk, including Tesla and SpaceX and XAI. I want to bring on Theo to walk us through his reporting. Theo, welcome back to the show.

1:25It's great to have you here. Why is Grok focusing on video and image content so much these days?

1:32Theo Wayt:So out of all the products and models that they've developed in-house, image and video are kind of the one area where they have significant traction with users. I mean, the core Grok models haven't been updated in a while. They aren't getting that much attention. Encoding, you know, that's a big emphasis for them, but they've done that mostly through the acquisition of Cursor that they're working to complete right now. So when you think about in-house XAI stuff, it's kind of just imagine, which is their video and image tool. okay and what type of content specifically are they focusing on so they want it to be a video and image model that works for everything you know if the if you look at what they talk about publicly um you can use it for commercials for brand marketing they have like a preview of a movie that was generated by grok um but in reality a lot of the usage is for adult content pornography erotica, et cetera, because they have looser guidelines and content moderation.

2:42Theo Wayt:And when you have that on the internet, that's just how a ton of people are going to wind up using your service. That's not really unique to XAI, but what is unique is, you know, are there loose guidelines? And has this been, have they made an intentional pivot to try to diversify away from that content at all or is this just kind of the way users have ended up using the platform? I mean, yeah, I would say they don't mind that it's happening, but it's not what they want to talk about. If you read the SpaceX S1, it's hundreds of pages. And the only time that the idea of any adult content on Grok comes up is in the risk factors section where they, you know, warn that it could pose regulatory issues.

3:25Theo Wayt:But it's definitely not the, you know what they want to talk about is you know xai enterprise grok for business grok for government grok code um but in reality you know adult content is a gigantic part of their business too okay and then you mentioned the loose content guidelines so if this is the way people are using the product have they at all modified their content guidelines to uh to to shift the the the product at all? So they're actually very vague about it. And, you know, Elon had said in the past that, you know, Grok would generate anything that was appropriate for an R-rated movie. But there are definitely examples that are more extreme than that.

4:11Theo Wayt:I wouldn't say that's been consistently enforced. And if you look at like the actual, you know, rules that XAI gives users, they basically say, we can't tell you what the exact rules are. you should experiment and see what you can get away with and figure it out. Now, you talked a little bit in the story about the training component to this. I mean, where are they getting training data for this type of product? I mean, part of it is that there's a lot of NSFW content on X. Something like a fifth of the traffic on X right now is adult content in some form. And then on top of that, my colleague Grace Kay had some great reporting about how, you know, AI tutors that work for XAI training the models have had to read like some, you know, semi-pornographic scripts for their job to help train the models as well.

5:12Where do other AI companies stand on this front?

5:17Theo Wayt:They've totally shied away from it. I mean, last year, OpenAI had, you know, there was this announcement from Sam Altman where he said, we're going to allow erotica, which was kind of vaguely defined for users that they could, you know, verify were adults. But they wound up postponing that indefinitely. So that product does not exist. Anthropic is very clear that they don't allow sexually explicit content. among other more specialized video and image training for video and image models like MidJourney. They don't allow it either. So if that's what you're looking for, Grok is kind of the only place to go at this point.

6:01So let me ask you this now. SpaceX is a public company, and so people are very much paying attention to it from that light as well. Do you think that this focus changes at all? Does the focus on this type of content, does it pose any kind of a risk to the company now that it's public?

6:23Theo Wayt:I mean, it's a tough question, right? Because for a regular company, I think the answer would be yes. They would want to be relatively conservative about this area because it's not launching rockets. It's not what people think of SpaceX for, but it is a gigantic risk. I mean, I just imagine the, you know, SpaceX employee in the EU that normally works on satellite policy who all of a sudden is having to answer questions about, you know, Grok being used to like edit images of children or something like that's not what they signed up for. And that's a huge, you know, potential drag on the company in all kinds of ways.

7:02Theo Wayt:But on the other hand, like this is one of the most successful products from an AI lab that is not considered a front runner. So are you just going to kill it? Like, probably not. So I don't know. You know, there's a lot for XCI to consider. And right now, it seems like they're doubling down. Right. Well, Theo, I want to thank you for coming on. That is Theo Waite, our Elon Musk reporter, here at The Information. Meta's CEO, Mark Zuckerberg, is directing his staff to build a prediction markets app that could rival Calci and Polymarket, according to a report from The New York Times. The app is reportedly internally known as Arena, and it would allow users to potentially wager on events using either a point system or real money.

7:47Joining us now to discuss what this could mean for the company is Jason Helfstein, Managing Director, Head of Internet Research at Oppenheimer. Jason, welcome to the show. It's great to have you here. Thanks for having me. So is prediction markets a good business for Mark Zuckerberg to be getting into, in your opinion? No. I mean, I think we can make a list of a bunch of ideas that seem like kind of throwing wet paper towel at the wall to see what sticks. And we can kind of talk about why that is. But this does not seem like a good idea. I would say perhaps he wishes there was more, I don't know, intellectual content perhaps on Instagram and Facebook that could then be mined for meta AI like X has that can be mined for graph.

8:34but this is probably not the way to get there. Okay. Although we've reported that Zuckerberg has talked about launching something like 50 new apps, maybe even at some point. So it feels like this could get launched in the near term, right? Yeah. I'm not saying whether he is or isn't going to do you that. So I think it's a good idea, a bad idea. I think the problem is you have to, prediction marks only work with their skin in a game with money. So if the people who are predicting are not putting their own hard-earned cash on the line, you don't get then the, we'll call it the engagement value. And engagement value can be data.

9:17It could be the time span. Again, if you just want to capture time span, maybe, but you've got to have skin in the game, right? That's the whole idea. That's why you look at how much, let's say, dollars are being spent per prediction. or the coal amount. In other words, you're saying that they couldn't go half in on this with kind of a points-based system. They would have to really create essentially a replica of a product to Calci or Polymarket. 100%. Okay. And that would then, not that meta is new to regulatory issues, but that would open itself up to a whole new class of regulatory scrutiny possibly.

9:57Oh, I mean, in addition to that, you then have to like get more into payments i mean look if like if you want money why don't they put a fee on facebook marketplace ton of transactions are happening it's free to both sides like why isn't there a fee there right um yeah i i like i don't know i mean ebay that's e they charge a fee right so um yeah it's really hard to understand i listen i think when you kind of take your point about potentially 50 apps like is the real question that they are concerned with time spent slowing on Instagram. And the way to try to reboost time spent is to find more ways for people to engage.

10:38Now, they give you kind of a smattering of kind of time spent quotes, right? And it's hard to pattern it or model it. Again, if you look at this over a period of time, it does look like time spent is slowing on Instagram, still growing, look like slowing we know last quarter daily active people didn't grow they called out um you know iran and kind of whatsapp in russia okay we'll see what happened this quarter but again it just seems like the reason you would do all of these app projects is because you're trying to find a way to boost engagement do you think that meta would buy their way into the prediction market space I'm not sure they can buy anything.

11:25I mean, this is why you've seen these like acqui-hires where we buy 47 % of it and then all the employees join the company. So we're technically not buying it because we're buying the minority. And then all the employees join. I mean, that's sort of on the AI front, but even a predictions markets acquisition? It's hard to say. you would think under the current administration it would be easier than if you were under a democratic administration potentially but it's hard it's hard to say you don't you don't know until you test that i mean i guess we've seen you know pretty small deals blocked under the prior administration um like i i don't know facebook was yeah yeah to buy one of those bigger companies i mean just take us inside zuckerberg's calculus here though a little bit I mean, let's look at the menu of things he's working on.

12:18He's working on AI. He's got this big compute buildup going on. He is also simultaneously saying he's going to launch 50 apps. He's looking at stuff as Wild West as prediction markets. Do you see the vision that he's trying to put together here? Does it feel a little bit scattered right now? What's your assessment of the overall strategy? I mean, creating, if you think that you have the ability using Instagram and Facebook to push consumers into other apps you own and then get them to spend time there and then monetize them that way, that is like logical, okay?

13:05Again, you could argue the biggest opportunity is monetizing WhatsApp, right? and again whether that's at some point with advertising at some point subscription at some point kind of business messaging and kind of business commerce um perhaps seems like the biggest opportunity there but but they're not doing that i mean they are a little a little bit but like they they they never go for the obvious monetization route it seems well i think it's because they don't want to hurt engagement. So I mean, any testing where like, oh, if we put an ad in WhatsApp, will it hurt engagement? Obviously, the conclusion is like, yes.

13:45And we've seen Snap kind of finally do it with the ads and the messaging because they've kind of had no choice, but they've also been kind of forced to go to the subscription route, right? Which is why when Meta announced a subscription offering, there was a market debate. Is this good or bad? I mean, the market tends to put a lower multiple on subscription and advertising. But are you doing subscription because you're running out of room to sell advertising? So and then we can kind of talk about it. And then, you know, right now, it seems like most consumers don't want to pay for AI, right? That that's going to be a free product, whereas businesses will pay for AI.

14:25So perhaps, again, on the business side of WhatsApp, there's the opportunity to sell that offering. Um, so yeah, I, you know, again, but, but now think about it. If you're, if somebody's in Instagram and you're saying, Hey, leave Instagram to go download this app and do this other thing, you're then potentially taking time away from the feed and the ability to monetize ads on the feed. Well, and it's happening already with threads, right? I mean, I, I see, I, I did finally download threads. and I had downloaded it once when it first came out, but I wasn't using it that much. So I got rid of it.

15:02But I, you know, I redownloaded it because the number of times that I was prompted to leave the app, I was like, okay, I gotta, I need to see the content. But yeah, to your point, I mean, that's the last time on Instagram then. And then I guess the question is, well, do you have the same opportunity to sell ads across all of these platforms or were you better off just keeping the one platform, the central advertising ecosystem? I don't know. Yeah, I mean, look, that's what technically they're doing when they convince you to download MetAI, right? Like how you're going to go hang out in that app.

15:40Now, I mean, but now let's take a step back. So as you said in your opener, I mean, all this CapEx, why are they spending all this? It is to have their own version. Most of it, I would say most of the incremental CapEx is so they can have their own version of ChatGPT. why do they want to have their own version of chat gpt it does seem like um that the company perhaps is worried that over time people will spend meaningful blocks of time on an ai talking to an ai engaging with an ai and maybe that's time they other would have spent with instagram and facebook and so you'd rather cannibalize yourself than have somebody else cannibalize you from a time spent standpoint right what do you think of the glasses that they released this week i thought it was kind of a non-event it seems like it's just an unbranded instead of ray-ban branded where they're getting right a cut of it now it's meta branded so it's almost like a generic version of the same glasses because i think yeah glasses are the same so yeah i think that this is kind of a non-event i think with with these glasses it's when can you get the battery life good enough that you can wear them all day and not have to take them off to charge them and we'll eventually get there right and i think most people do believe like ambient computing will be something and whether it's glasses or a pin or a watch or something like the idea of having you know maybe it's a clip on your ear or something but that you'll have compute with you at all time but again you got to get the battery right and then well you know what you know when will the next generation of the glasses that have the heads-up display like will that be something that you know can be accessible broadly at a somewhat normal price point but i mean you say it's a non-event meanwhile we have a number of other companies going at this smart glasses category uh we obviously had snap and and you know that that was whatever it was people people were not excited about that product but google and warby parker have also said that they are working on something together so is meta still the clear leader here in this category to you?

17:45I mean, maybe right now in a unit, but the total number is pretty small. I sense when Google wants to be a player, they'll be a player. I mean, look, they still put up pretty big numbers with Fitbit, and even though nobody talks about that, right? Even though there's a lot of talk about OR and WHOOP, and you kind of look at the amount of engagement, like still a lot of engagement with Fitbit devices. So look, I think that Google has thought that the software is not ready yet, or just let's just call it the whole product, right? You know, the product, like Google smart glasses are not ready for the market yet.

18:20And when they're ready, they will come. And I'm sure they're going to work pretty good because they'll be powered by Gemini and, you know, it'll bring all of the, your personal information with it. Big picture here, Metastalk is down about 20 % in the past year. What do you think investors are looking for? So I look, I think going to last earnings, there was kind of a view that, hey, revenue grows. and cost reductions are going to be enough so that earnings could actually beat Wall Street expectations over the next two years. Then we came out of earnings and people reassessed the view and said, you know what, maybe revenue growth isn't going to be as strong.

18:59So, you know, they need to then, maybe there isn't upside to earnings. And there's probably a debate on like how much cost they can still cut out of the business, but still kind of maintain a culture and get the products produced that they want. Now, probably the debate is probably around how the cost to run AI is getting more expensive, right? We know this. You see it with spot pricing for third-party compute, NeoCloud compute, as well as just all the inputs that go into the right building data center. So the thought is, is Meta going to have to rein in CapEx and maybe go back to thinking about open source model, not having a full frontier model.

19:48And I think that's probably what investors are playing for right now, which is if you own it into that moment, right, when they kind of say, hey, maybe we got a little over our skis, we got to rein it back in, the stock probably reacts well. I would argue we're not there yet. But it doesn't seem like that they're going to back off the very aggressive pursuit of AGI and having, again, their own version of a chat GPT that consumers want to use. So I think that'll be the debate on the stock probably through the near term. Right. Great. Well, Jason, I want to thank you for coming on. That is Jason Helfstein, Managing Director at Oppenheimer here on TITV.

20:35Private equity company, Vista Equity Partners, and venture capital firm Cambium Capital this month launched a new company together. Vector Core Compute, also known as VC2, is a cloud services provider focused on inference. And the new company has made a$3.5 billion commitment to Samba Nova, an emerging chip company taking on NVIDIA. I want to bring on Monty Soroya, Senior Managing Director and Co-Head of the Flagship Fund at Vista Equity Partners, and Rodrigo Liang, co-founder and CEO of Samba Nova, to break down the news. Welcome to the both of you. It is great to have you here. Monty, I want to start with you.

21:13Why did Vista decide to launch VC2? It's good to see you, Kash. Look, I think our goal is to provide solutions to the enterprise. And so as we started to see the AI deployments in our own companies and with our end customers, our focus has been how do we deploy the most valuable, fastest inference at the best cost for the enterprises. And so for us, the hyper focus was on how do we actually lower the cost and actually speed up the process of using agents in the enterprise. And the best way we could think about solving that problem was with the launch of this VC2 company, as well as using the Samba chip set as a part of that.

21:49But I'm so fascinated here. I mean, Vista is traditionally known to be, you know, one of the most prolific software investors in the world. I mean, why decide to go to the great length of creating a new compute provider on its own? Why not just leave it to those who have established businesses in that arena? Yeah, no, it's a great question. Look, we have our 92 companies that are large software providers to the enterprise today, right? So we have over 35 billion of revenue and we've been servicing the enterprise for the better part of over 25 years. And what we're hearing from our customers is that they want full stack solutions.

22:26So historically, we were able to provide some of the software solutions and use some of the best partners we have in the world. So we have Amazon and Google and Microsoft as phenomenal partners that continue to be partners. And then we also have the large AI providers, Anthropik and OpenAI as large providers as well. I think for us, the question became as the Agentech era came on, the needs for these enterprises has evolved. and you're going to hear a lot about this word tokenomics. I did not invent that word. I'm just using it because everyone else is. Good word. It's a good, come on. The word is fine, right?

22:58It's a fun one. You know, in tech, we make up new words to make ourselves feel better every time there's a new wave of technology. Now we've talked a lot about tokenomics and what we mean by that is the actual value of using tokens by ages. The thing that people are missing is a lot of what we've been doing in AI has been basically facing the human, right? So it only had to go as fast as we were able to read. But the reality is the solutions that we're putting into the enterprise now, which are agentic, is agent-to-agent communication. They need a lot more speed than we were able to get from the traditional AI solutions.

23:32So for us, it's truly just going to the enterprise, asking what they want. And what they've come back with is give us a full AI agentic solution. And that includes everything from the software, the system of record, the system of action, all the way down to compute, all fully integrated. And that was the thinking behind VC2. Right. So Rodrigo, I mean, this is$3.5 billion. It's a big commitment. Is that just straight revenue coming to Sambanova over the next couple of years? What does that mean for your business? Yeah, we're super excited about that. Yeah,$3.5 billion is a big commitment, and we're excited to have that as a growth path for Sambanova, but also really the partnership.

24:11Like Monty said, I think you look at all these businesses, they're looking for much more efficient AI, much higher performance, much lower latency AI. And I think having VC2 with Samanova together gives them a really, really good solution, which allows them to actually compete in a very agentic first type of technology ecosystem. And Rodrigo, we had you on last summer, nearly a year ago. we talked about the progress your business was making tell me a little bit about how your offering has changed in the past year or so yeah we're really focused on premium inference if you think of the world you know that's running nvidia everyone everyone has services around nvidia they've done such a great job you know producing the offering that kind of the the broad market wants and yet there's still a demand for differentiation and what people want ultra or low latency, really high performance where the speed matters, like Monty said, you know, if you have applications where agents are talking to agents and it's taking say 20 hours to compile a piece of software that the agents wrote, you know, if you can actually make that two hours, there's significant productivity enhancement.

25:22And so we're focused on the premium and offering enterprises that need that and then providing them the differentiated service. You know, some things you can use, you know, a very large batch size type of inference and actually run that at the lowest cost. But for some, you need premium. You need premium because your models are bigger, your models are lower latency, your models are much higher throughput. Are you still focused on running open source models? That was the focus last time we had you on. Yeah, we'll continue running open source models and these open source models are really, really good.

25:56They're getting bigger and bigger by the day. Everyone's talking about, I mean, it seems like open source is like it's having a moment here. People are realizing that we can't run everything on the most expensive closed source models. I mean, I wonder what that's meant for growth for your business. Yeah, look, I mean, I think the world's heterogeneous. I think the closed source models have done such a great job. You look at Anthropix numbers, there's no doubt that they've done a really, really good job with kind of building services. And I think that there's also a place where open source models are getting really, really good.

26:27and you're able to actually run them, customize them, change them, fine-tune them with your private data. You get into a variety of things with open source models, including actually delivering at a lower cost for certain types of tasks. And so for our business, where we want to run heterogeneous models, any of these open source models, to some extent, even close models, fine-tune models, it becomes a great option for the customers that have that complexity in their business where they want to run a broad range of types of models in order to power their business. And the same way that, you know, NVIDIA has launched its suite of open source models that customers can build on, do you have your own suite of open source models that customers can use to build?

27:08Do you plan to build that out? I think we really lean hard into the open source models that are available by the community. You got some great models like DeepSeq and Minimax, and then you have some new models that are coming in from Reflection and other folks that are putting models out there that are really good GPT-LSS. Look at the Mistral models that are coming back. There are some really, really good models that are out there, and so we lean into the model that the ecosystem is building. Do you think you would build your own open source model in the future? Maybe. I think when you're looking at needs, I think the market is changing really fast.

27:46And like Monty said, one of the beauties of having this partnership with VC2 and Vista is you get access to what the end customers, the enterprise want. And so if there's a need there, what we're going to see certain types of models that the open source community today has not yet offered, yeah, maybe we'll help build those. Right. So Monty, I want to come back to you. So you have VC2 here. So the game plan is what you're going to rent, compute out to Vista's portfolio companies. Is that the business model? Yeah, the business model is that we're able to go brownfield sites. So we're able to put up compute faster than having to build giant boxes and greenfield builds.