OpenAI Technique in ‘Astra’ Model Sparks Security Concerns, SpaceX Shakes Up Data Center Leadership

0:13Stephanie Palazzolo:Welcome everyone to the information's TI TV. My name is Akash Pasricha. It is Wednesday, September 2nd. Today on the show, the information has exclusive reporting on OpenAI's new model, Astra, and how it could open itself up to more security concerns. We're then going to talk about a leadership shakeup within SpaceX's data center division. We'll also unpack why Anthropic could be charging more than it needs to be by a significant margin, at least according to one of its emerging rivals, Glean. And to close out the show, we have a new fund manager who is starting her own VC firm coming on to talk about the current state of venture capital and what the chatter is amongst LPs.

0:56Stephanie Palazzolo:It's going to be a great show, so let's get right on into it. OpenAI's new model release, Astra, is imminent, and my colleagues Amir Afradi, Stephanie Palazzolo, and Rocket Drew have exclusive reporting on some of the technical considerations that the company is making that could potentially heighten security risks. I want to bring on Stephanie to share with us what she knows. Stephanie, welcome back to the show. It's great to have you here. Hey, it's great to be back. So Astra is coming out. I'm not sure if we have the exact timeline here, but OpenAI has been very loud on its publicity tour talking about it.

1:33Stephanie Palazzolo:How good is this model supposed to be? Let's just start there. So obviously things are not public about the model yet, but we've heard very good things. You know, Sam Altman earlier this summer kind of teased the model to policymakers in DC. I've also heard from people who have tested it that it is quite impressive. and I think will be very interesting whenever it comes out. So I don't think, you know, we have had cases in the past where maybe model announcements were a bit of a letdown and fingers crossed I'm kind of getting the sense that it won't be the case here, which is exciting. Okay. So your piece today dived into then some of the technical considerations that OpenAI is making and also the trade-offs between making the model better, changing the backend architecture, and then keeping it tight in terms of security.

2:24Stephanie Palazzolo:Tell us a little bit about what you found and what those tradeoffs are. Yeah, so as my colleagues Amir Afradi and Rocket Drew and I found in the story we published yesterday and dove in a little bit more into this morning's AI agenda, Astra basically uses this very interesting kind of newer reasoning technique, which is related to ideas that researchers have been writing about in recent months known as loop transformers or recurrent depth. Basically, what this boils down to is the way that models work today is they basically reason through tough problems by kind of writing out their lines of thinking.

3:03And that's also known as chains of thought. And the good thing about that is that researchers can kind of, you know, read the chains of thought and understand the model's like thinking process and also use that to make sure that they, you know, that they can make sure that the model isn't trying to do anything evil or bad or trying to hack into other companies.

3:20Stephanie Palazzolo:This is like the, you know, when I put in a prompt and I see Claude or Chachy, but you say, you know, thinking or scheming or, you know, whatever other verbs they come up with. And then if I click the dropdown, I can see all the things, right? Is that what this is? Yeah, so it's very similar to that. So the companies actually don't kind of publicly release like the full chains of thought, but they do provide users with a summary of the thinking process. So it is basically very similar to that. And what's different about Astra is they're using this new technique where instead of writing out their thinking, the models instead loop the question that they get through this thing called a layer, which is part of the model.

4:03It basically loops that question through the layer multiple times. And when it does so, it's another way for the model to think more deeply about the question that it's facing. But when it does that, it doesn't actually, you know, kind of write out its thinking process the way that normal reasoning models would. So this is like a new way for the model to think more deeply in addition to kind of traditional reasoning techniques. But the downside is that models may not, you know, write out as much of its thinking process as previous models did.

4:36Stephanie Palazzolo:And when you say it doesn't write it out, so why doesn't it end up writing it out for developers to see? What's the core reason that this looping thing skips that step? So without getting too technical and in the weeds, it's just kind of like the way that this new model architecture works because it's just, you know, the model is made up of these layers. And, you know, in this case, whenever it's looping... It gets too complex. There's too much there. I mean, you keep going back and forth. I mean, I imagine it's repetitive, too, for the model to output that much stuff as well. Yeah, I mean, I think an analogy you could think of is like, you know, if I'm a student working on a tough question, the teacher can ask me, you know, if I need to think about the question more, they can either say, okay, you need to write out more of your thinking and kind of like really logic through this problem that you're facing versus actually maybe just sit there and don't write anything out for a second, just kind of think about the problem more.

5:32so i feel like that's kind of an analogy to like the way humans might think about it that would

5:36Stephanie Palazzolo:make sense right okay so so that so that's that's kind of interesting there are trade-offs though with this you you wrote about what are those yeah so as as i mentioned the main trade-off here is that uh it's possible that models you know like astra that are using elements of like this looping technique are less likely to kind of to kind of write out their chains of thought so what that means is that it's going to be harder for researchers then to look into those chains of thought and make sure, again, that the model isn't doing anything bad, like trying to hack into hugging face, for instance, which we saw earlier this summer.

6:11Again, like huge caveat here though, because, you know, my co-writers and I noticed that a lot of people were really freaking out about the story last night on X, is that OpenAI is kind of like limiting and making tweaks to the way that they're using this approach in Astra to ensure that, you know, this isn't like the end of the world and models are never going to show their thinking ever again. Like Astra does actually show its chains of thought and the way that OpenAI is using this technique is limited so that it's not going to be as, you know, it's not going to be as concerning as a lot of people are making it out to be like on X, basically.

6:47Stephanie Palazzolo:Has OpenAI said anything publicly about this matter or this technique at So they haven't said much publicly. In the past, they actually have written about this sort of research and kind of laid out some of these safety concerns that we were talking about. uh so actually last night in kind of response to our story uh the open ai's chief scientist did post on x and kind of made the argument that the that the idea that we're just going to monitor chains of thought and that's going to be like the forever solution to how to like keep models safe that's actually not kind of the long-term solution here so he was saying that you know So we can't – this isn't a permanent solution, and we need to come up with new ways to keep models safe beyond just monitoring chains of thought.

7:39Stephanie Palazzolo:Which probably – I mean you could say he has a point there because it's – you know, if you're monitoring something that is going ARRI or going rogue, I mean monitoring it is one way to catch it. But the ideal solution would be to stop it from going rogue in the first place. In other words, it's a bit of a Band-Aid solution. This is true. And also, I think it depends on a lot of things such as, you know, the model not lying whenever it writes its chains of thought, right? This is assuming that the model is going to be truthful, which isn't necessarily always the case. And also that B, that the chains of thought are always going to be understandable to humans.

8:16Like, you can imagine that models, like, they're not humans, right? That may be the best way for them to think about problems. They're sneaky.

8:21Stephanie Palazzolo:They're sneaky. They are sneaky. but that the best way for them to think through a problem isn't necessarily in like legible lines of English the way that humans would right so maybe it might write in weird like mathematical formulas and stuff that are like hard for humans to understand another thing that the open AI chief scientist said last night was that okay so he basically made this comment he said that the kind of complexity or like depth of its leading models like Astra are within a factor or two of GBT4. Without getting into like the kind of technical details of that, he is basically trying to imply like, hey, whatever kind of looping techniques we are using here, we are like limiting them greatly, you know, like we're trying to take this slowly step by step.

9:07So I think trying to throw some water on the fire, a bit going on on Twitter last night of people being like, oh, it's over for us.

9:14Stephanie Palazzolo:Which is, again, interesting from a business perspective because on one hand, there's the model race going on and you want your model to be as good as possible and you don't really want to put caveats on the strength of the model or what you've done to develop it. But on the other hand, if people feel like you've taken a not so favorable approach to gaining that advantage, then people won't buy it anyway. So it's like, it's actually, as much as it's a technical challenge, it's also like a comms challenge too. It's like, what do we say about this so as to not shoot ourselves in the foot? Totally, totally.

9:53Yeah, it's very difficult to, yeah, I think it's this big trade-off as you're describing between like performance and safety and how do we communicate all that. And I think a point that a lot of people might've missed from our story last night is that, you know, the big thing here is like not necessarily what exactly is going on in Astra, but like, what does this mean if this trend continues, right? Like in the future, what if like looping becomes more common and like models loop more, they hide more of their thoughts or maybe other AI developers are more aggressive with using looping and aren't using as many safety guardrails as OpenAI might, for instance.

10:35And so I think the bigger question here is not just what's happening right now in the moment, but what does this mean for future models and the way that models are going to be developed and the way we're going to have to try to monitor them and keep everything safe.

10:49Stephanie Palazzolo:Right, right. And so, Stephanie, just to make sure we understand the full landscape here, This technique, I mean, we're talking about it here with OpenAI. Is this also a technique that we know that other AI labs might be playing with? I mean, is it one that we might see play out a little bit more broadly across the sector? Totally, yeah. I mean, this is definitely something that all the AI labs are thinking about. There have been public research papers published about this idea. So all the labs, I'm sure, are exploring it. And I think we'll definitely see similar attempts by the different labs very soon.

11:29I think it's just a matter of like they're all going to have their specific approach to it or how much they kind of like limit how much they're doing looping and stuff. And so expect it to be very widespread very soon.

11:39Stephanie Palazzolo:Great. Well, Steph, I want to thank you for coming on. That is Stephanie Palazzolo, our AI reporter and author of AI Agenda here at The Information. The Information has exclusive reporting that Elon Musk is making some big changes to SpaceX's data center team. I want to bring on Grace Kay, our Elon Musk reporter, to tell us more about what she found. Grace, welcome back to the show. It's great to have you here. Thanks. So what's going on with the data center business at SpaceX? How is Musk changing it up? Yeah, so basically there was a major reshuffling on the infrastructure team at SpaceX AI.

12:15They brought in a bunch of leaders from rocket production and launch operations to replace the existing XAI leaders who've been leading the team.

12:25Stephanie Palazzolo:And who specifically is he swapping here? Yeah. So Jake Palmer, who had been leading all of physical infrastructure for XAI, is out. Dan Roland, who had been leading it, is out. Zach Wells. Pablo Mendoza, who had been leading. Wow. So it's like this is not just one or two people. This is like half a dozen people. Oh, yeah. It's even more than that. But like the leadership and then also some of the people under them. And they've been bringing in this guy, Wes, who is the VP of production for the Dragon and Falcon. And then, you know, someone who is a director for operations for the launch facility for SpaceX.

13:03Stephanie Palazzolo:And is this because people on the data center side have left SpaceX? Or what's the core reason here for this shakeout? You know, I think it's like a mix of things. I think there might have been some cuts. And then I also think, you know, people have gone on to other places. We saw recently that Brent Mayo, who is leading site builds, left, and he's at OpenAI now, working on their data centers. This, you know, engineer, Liz Belk, who had been leading the design for Colossus. She's now at Anthropic. So there's been a lot of changes recently on the team. why do you think these changes are happening from a business perspective?

13:42Stephanie Palazzolo:I mean, how has the data center team been doing? How are the build-outs going in Memphis and, and I was going to say Memphis and Tennessee, Mississippi and Tennessee? Yeah, I mean, they've been, they've gone incredibly fast. The first data center, they put up a hundred, you know, they brought online a hundred thousand GPUs in just 122 days, which is kind of unprecedented for the industry, but I think there's always trade-offs. So when you move at that speed, one of the trade-offs is reliability. So when they were building these data centers, they didn't add the same redundancies initially, you know, that most companies would add when they're building data centers and they had to add those in later.

14:23So that did cause some reliability issues, but that was kind of a known trade-off. And I think now that SpaceX is here, they maybe do things a little differently and, And they might be looking at how they might want to change that now that they're bringing in all these other companies that are leasing compute from them.

14:39Stephanie Palazzolo:So this is kind of interesting because, I mean, reliability, that's certainly something I think a space team would need to know about, right? I mean, making sure that things are done properly, space seems like the best team to do that, you could argue. Yeah, I mean, I think it's a little unique because the people coming in don't really have data center experience. So I think that's a little surprising from my perspective. And I'm curious to see how that'll work out. So it's a change in tactic for sure. Right. Tell me a little bit about Musk's comments at the, there was a G20 related event this week.

15:18Stephanie Palazzolo:Did he give any more color on how he's thinking about the data center build out there? Yeah, he talked a little bit about what he sees as kind of the next hurdles for building these data centers. and you know we've been talking for a long time about like getting chips getting the equipment and now it seems like you know at least according to musk that the main hurdle is power and how are we going to power all these chips that we have now um you know and the information has already done some reporting on this around how elon musk is looking to you know maybe start building their own turbines so he's looking at that and so tell me you know as you think about this shake-up i I mean, one way to think about it is the reliability issues that he's unimpressed with.

15:58Stephanie Palazzolo:Another way to think about it could be financially the importance of the data centers to the company's top line and maybe a shift in priority. I'm not necessarily saying SpaceX is going to focus less on space, but he's taking what appear to be some of his top commanders from the space team and putting them on data centers. Do you think that this is anything about his shifting focus at all within the business? Yeah, I mean, I think it's definitely a shift for his company that they even started leasing out Compute. I think that was something they hadn't initially planned on. And now, you know, they're making, it's over a billion dollars a month from Anthropik.

16:35So I think they're definitely finding out that it's a lucrative business. How much they're going to reshape the company around that, I'm not sure. I think that's something we have yet to see. Great.

16:44Stephanie Palazzolo:Well, Grace, I want to thank you for coming on. That is Grace K., our Elon Musk reporter here at The Information. Glean, an AI enterprise search darling that has been getting a lot of traction, is coming out swinging against Anthropic. My colleague Kevin McLaughlin, who authors our Applied AI newsletter, spoke to CEO Arvin Jan about the dynamics at play. I want to bring on Kevin to share more about what he found. Kevin, welcome back to the show. It's great to have you here. So Glean had their customer conference, looks like, recently, and you learned some stuff about how they are positioning their product against Anthropic.

17:19Stephanie Palazzolo:What did you find? So Glean announced new products, but really the most interesting thing was they had a benchmarking study and they compared the Glean AI assistant versus Anthropics Cloud Cowork in terms of the amount of tokens it consumed and the cost on average per token across a range of 180 or so tasks. I think the broader point Glean is making, and it's kind of interesting, is that Glean has, because its secret sauce is a collection of tools known as context, which includes things like an enterprise graph, which shows the relationships between different kinds of data. We've written about this quite a lot of the information.

17:58But the importance of context is that it actually allows agents to operate less expensively. And this is an important point. And it was one of the first times I think we've seen a software provider that uses Anthropic kind of make the point that they are able to run things more cheaply than Anthropic. And I think the broader point they're making is the importance of context. And context is also something that Anthropic is working on, too. So it was definitely a flex on Glean's part.

18:30Stephanie Palazzolo:Did they give you a sense of how much cheaper they're able to run the models? Sure. Now, the benchmarking study used a different model for Glean's assistant than it did for Anthropic. So it wasn't like – Anthropic could definitely have a point if they were to say this wasn't like an independent study in any way. Right. And they're rivals. So, you know, take it for what it's worth type of thing. Sure, sure. But Glean did say that its assistant consumed 70 % fewer tokens and also was on average 80 % cheaper in terms of token costs than Anthropic. So a pretty big discrepancy. Interesting. Did they give you a sense for if this has changed over time?

19:14Stephanie Palazzolo:I mean, what was this a year ago? I don't think they did the study a year ago. But again, I think the broader point is Anthropic and Glean are definitely starting to go after the same pie, and that is enterprise agents. And specifically, Glean has recently entered the market for long-running agents that can work for hours and perform complex tasks. And of course, token consumption comes into play when you're working with these kinds of agents. And so, yeah, I think the broader point was that Glean is definitely a competitor of Anthropic, and it's, like you said, coming out swinging. Yeah. Last question for you, Kevin.

19:54Stephanie Palazzolo:What's the current status of Glean's business overall? all. I mean, it looks like they have some competitive tools here. Do we have a sense for the company's financials or how growth is going for the business? Sure. They said a few months ago that annual recurring revenue had passed$300 million, which was up from about$100 million 18 months earlier. So they're definitely starting to see some real growth. And I think the reason that that's happening is that Glean's initial product the AI search, which is sort of ubiquitous in the industry. But as they push more into agents, the importance of context, which they've been working on since the beginning of the company seven years ago, becomes more important.

20:34Again, when you're talking about agent accuracy and also token consumption costs. And so, you know, I think that the message I heard was that Glean kind of has this secret sauce that it's been working on longer than a lot of companies have. And now that's becoming more important. And, you know, I think it'll be interesting to see what their revenue growth is a year from now and whether they're able to take big customers away from other companies.

21:01Stephanie Palazzolo:Great. Well, Kevin, I want to thank you for coming on. That is Kevin McLaughlin, our enterprise software reporter here at The Information. Axiom is a new venture fund on the block. The company raised more than$50 million for its first fund about six months ago. I want to bring on Sandhya Venkatechelam, founder of The Fund, to talk us through her strategy right now. Sandhya, welcome to the show. It's great to have you here. Good morning. Thanks for having me. So tell us a little bit about why you decided to go in on your own, launch your own fund. You're a former Coastal Ventures partner. BC Fundraising, from what I've heard, it's a challenging time.

21:40Stephanie Palazzolo:It has been challenging for a couple of years now. Why did you decide to launch this fund? Yes, Akash Fundraising, I think, is fun for no one. And I worked for a truly amazing venture firm at Kostla. But we had a vision of building a different kind of venture firm, one that I call AI native from the ground up, to invest in pre-consensus opportunities. In other words, not what's obvious today and gaining momentum today, but what hopefully will be big and interesting, especially in AI robotics in the future. Okay. So pre-consensus, I mean, AI, is it robotics? I saw real-world AI is something you focused in on.