The Exchanges

Q on their own. Uh, and the frontier is, is certainly not that far in front of open weights. Um, and so I, I'm curious to hear your perspective about what we should be thinking about in terms of, you know, what's gonna happen when companies that are not as scrupulous, you know, have access to this same powerful technology, and do we get into trouble in that, in that area?

A Um, yeah, so we can sort of see this coming and relatively soon. I don't know what the gap is. You would say, you know, six months, 12 months, maybe, uh, at the most between the closed wait frontier and available open source model. So it seems to be that, um, The open source models will very soon, if not already become capable of lending meaningful assistance, um, to destructive uses that, um, some people might pursue, uh, already cyber offensive capabilities has been a concern, right? With mythos, for example, that was withheld for that reason, but also, um, Say in biological weapons design, or chemical weapons, or other malicious uses. Um, and so it, It seems that you either need to prevent open weight models from being developed and released, or which might be better and more realistic, try to shore up some of the alternative defenses. For example, with bio, you could imagine regulating some of the other necessary inputs, um, DNA synthesis machines, for instance. So maybe it will be the case that there will just be widespread access to models that can help you design new pathogens. Um, and then you need something else to prevent that from actually resulting in a release of biological weapons, and that seems like DNA synthesis machines would be one excellent place to maybe, you don't need every lab to have their own DNA synthesis machine. They could have DNA synthesis as a se…

AI assessment note: “open source models will very soon, if not already become capable of lending meaningful assistance”

Q the reply to that mentioned, uh, that he had contacted a lawyer. Um, so there's my apology. I still don't love the puzzle, but, um, back to you, Steven, this is, this is a little bit, uh, what would you, what actually, let me just ask you the question without leading the witness. Is it narcissism or is there actually something, something, you know, potentially disconcerting happening behind the scenes?

A I think it's very brave in that by and large, these are people sacrificing very large amounts of money to give the warnings they are. I do wish that they would be more direct. Um, but to put it in context, you know, back in. It seems that OpenAI and Anthropic had secret non-disparagement agreements, which in OpenAI's case, at least, you know, plausibly not permitted by law, the way that they operated this, where To keep your already vested equity, the compensation you had been told was yours. You had to sign away your right to say anything negative about open AI and in fact, sign away your right to tell anyone that you had signed this contract. Um, and this was secret and kept under wraps for years until Daniel Cocotelo, um, who people might know from leading AI, 2027, I think very, very courageously forwent this agreement and forfeited something like 80% of his family's net worth and said, Sorry, I'm, I'm just not waiving my right to criticize open AI. Um, and in the wake of that, you know, there was a bunch of outpouring open AI and anthropic changed the nature of these contracts. And still it's pretty intimidating to speak out against these massively resourced legal operations. Um, you know, not afraid of subpoenaing different people and getting into legal conflict. You want to be really, really careful about what you say. And so in Renox case, I noticed in the footnotes, ri…

AI assessment note: “I think it's very brave in that by and large, these are people sacrificing”

Q Oh, I think that would do good ratings. Okay. Steven, you also talked a little bit about in, in a recent newsletter about how, uh, basically we have very limited regulation on these companies. Uh, and even then they might still not be following. So can you just expand upon that briefly?

A As of 2026, there is finally some amount of law in the United States about how companies are meant to do testing for the catastrophic risks we have talked about. Um, until this point, purely voluntary. This bill is called SB 53. It came into effect in January and it's very, very light touch. It basically says the most major of the AI companies, you need to publish how you are going to test for these risks. You need to do what you said you are going to do, and you can't be misleading about it, but there's no quality standard. You could basically say, we will test for the risks as we deem appropriate and nothing further, and that, that would be fine. But if you say you are going to do this testing, you need to, in fact, follow through on it. Um, and unfortunately, it seems like OpenAI's release of last week, GPT, 5.3 codex, one of the big breakthrough models we've been talking about. As I look over the evidence, It seems like open AI did not abide by the testing that they had committed to in various ways. And so, you know, ultimately this decision now is with the attorney general of California to investigate it, whether to enforce a fine, a pretty small fine, maybe like up to a million dollars compared to open AI, hundreds of billions of dollars in valuation. Um, it just really seems to me like if we care about these risks, letting companies self assess in this framework is reall…

AI assessment note: “This bill is called SB 53... It seems like open AI did not abide”

Q So concretely, um, with and without, uh, cross app access, uh, what does the state of the build out look like without it versus with it? Like, what can you do with this cross app access that you couldn't do otherwise?

A Yeah. So, uh, without it, it's, it's point to point and everything is different. If you're building an agent on Salesforce agent force or Google agent space, if you're building an agent on writer, any, any platform where you're developing your agents, you're, you will develop them with point to point permissions. So if I, as a user deploy an agent that was built on one platform, it will ask me if it needs access. So let's say it's a Google agent space agent and it wants access to my Gmail. It'll, it'll prompt me, can I have access to your Gmail? And I'll say yes, because I want you to do work for me with my email. It'll ask me if it can have access to my calendar, and I will say yes. And in that world, my company won't know that I've given it that permission, because it's, it's a permission that I as a user have granted to the agent I deployed as a user, but I have now given it access to my production corporate data. And so that is an exposure where companies have agents that they don't control. It's software written by a third party that's now accessing corporate data assets. And that is fundamentally a problem. What cross app access will do is it will allow agents to, um, to be registered into, into your IDP to, as I mentioned, to have their credentials managed, to have them appropriately rotated. To have governance deployed where they can be just in time provision and deprov…

AI assessment note: “without it, it's, it's point to point and everything is different”

Q be the future. I I'll say though, the action mode do stuff for me. And we've talked a lot about thought partner versus actually doing things. I had not seen anything that interesting, at least in my usage so far, though I have not been asking it to do too much and giving it too much sensitive information and we'll get into why in just a bit. What about you?

A So I have definitely enjoyed using it. I think that this is the future of the browser. The question is, is it going to be something that OpenAI can Ride and unseat the incumbents or that the incumbents will effectively adopt into their products. And this idea of having an AI assistant or an AI chat bot in your side window and then giving it some capabilities to go surf the web, to me, you know, it would be great and really meaningful if your competition wasn't Google. And maybe there's a chance that it unseeds Safari in some way, but, uh, I, you know, trying to go against Chrome is going to be really, really tough, uh, because Google does have the talent to build this into their product. They will, they already have Gemini baked in, uh, to some extent. So many things on the internet work as with Chrome as the default, like you and I today, we're recording on Riverside. It's a podcast recording platform. Um, it doesn't work outside of like Chrome and Safari. So you basically have to get all this compatibility built in from the ground up. Although I do think actually Atlas is built on Chromium, which is an interesting, which is an interesting, uh, um, sort of wrinkle in all this that it's built on Google's open source browser technology. Uh, but then there's other, there are other things that are just not great there. Uh, for instance, the New York times seems to block it. So if …

AI assessment note: “So I have definitely enjoyed using it. I think that this is the future”

Q And so what is, what does this enable? Like when this is, let's say you have that football field size quantum computer, what does that enable?

A So the biggest thing that it will enable first, because effectively you can think of it as building molecules in memory and using those molecules. Uh, is going to be material sciences and, uh, chemistry first. So, in fact, one of the targets for Amazon's working backwards document for our quantum computers, a thousand error cryptic qubits could do a Hamiltonian on ammonia. Ammonia is the most produced, uh, we've been producing ammonia since the 19, for almost over, over a hundred years. Um, and it's probably the most produced chemical. It's in fertilizer, it's in petrochemicals, it's in Plastics. It's in just about everything. Um, and it's very expensive and energy intense to produce. We know by watching bacterial interactions that, that it can be produced at low energy state. We just don't know how. So in the past, Like a high temperature superinductor, superinductors in general have been discovered accidentally in the labs and then leveraged, uh, in the future with a Hamiltonian simulation, you can say, here's the outcome I want. Give me the chemical formula that will give it. So you can reverse engineer an outcome in, in chemistry. Um, on today's classical computers for ammonia, if you took all the iPhones and all the laptops and all the Android phones and all the cloud computers on earth, And put that, that simulation into it, it would run for longer than the history of the…

AI assessment note: “biggest thing that it will enable first... is going to be material sciences and, uh, chemistry”

Q Yeah, ok, sounds good. Just wanted to ask that, I was curious about it, but let's talk, ok now, so you're gonna build a, a technological solution that will block crawling. Yes. And so robots.txt, which is this code that you put in, like, the header of your, of your site, if you don't want to be crawled, that wasn't working?

A Yeah. I mean, I think robots.txt has two problems. Um, the first is some people just ignore it. Uh, and so if you ignore it, then you can still crawl all you want. And there's some just, there's some, even some big legitimate companies that completely ignore robots.txt. And we're really good at basically being able to say, okay, here's what robots.txt says. How are you, are you actually following what those, those, uh, what, what sort of the rules of the road are? And if the answer is yes, then robots.txt is a great solution. Um, but in the cases where somebody is ignoring it, then we need to actually put in place additional technical barriers to restrict their, their, their access. And so that's exactly what we're doing. The second problem with robots.txt is it's not granular enough. So take the Google bot, for example, Google's crawler does five different things, at least. Uh, one is it checks if you have an ad on a page. It makes sure that if you're putting an ad for Procter & Gamble, Procter & Gamble product up, it's not against a pornographic site or something like that. So it does brand safety checks. Um, the second thing that it does is crawls to index for traditional search, the 10 blue links that are out there. The third is that it crawls to create answers that are in the answer box. The fourth is that it crawls to create answers that are in the AI overview, the newer …

AI assessment note: “I think robots.txt has two problems. Um, the first is some people just ignore it.”

Q And going back to that increasing, uh, pre-training, increasing the scale of pre-training, delivering predictable improvements in model performance. Um, yes, now post-training is in the picture. It's making models better in really impressive ways. Um, but are you of the belief as opening? I have the belief now that there are diminishing returns from pre-training, um, given that we're now talking about different forms of training these models.

A Not at all. Um, our scaling laws still hold. Uh, empirically, there's no reason to believe that there's any kind of diminishing return, uh, on pre-training and on post-training. We're really just starting to scratch the surface of, of that new paradigm. Um, you know, the, the O series of models, which were kind of the previous reasoning models, um, were really just the beginning of, uh, us starting to explore what's possible in that post-training regime. And I think that's going to be kind of the dominant theme here for the next Year or two, um, is continuing to scale in that dimension, uh, and continuing to see the gains that you get there, um, simply because they're so significant. Uh, and so now we're pushing on two axes for how to improve models, and we think that's going to tighten and condense the rate of, uh, of, of innovation.

AI assessment note: “Not at all. Um, our scaling laws still hold.”

Q take a transcription of the conversation that they have with the patient and then have AI synthesize what they talked about and summarize it and put it into the systems that they have for electric medical records and then verify that. So they don't have to spend the time Writing those summaries up and can actually go and spend some more time with patients. So, what's the problem with that?

A There are so many problems with that. And the first thing I want to say is that you named the underlying problem when you talked about insurance requiring so much paperwork. So this is one of those situations where there's a real problem here. Um, it's not that doctors shouldn't be writing clinical notes. That is actually part of the care, but there is a lot of additional paperwork that is required because of the way insurance systems and especially the one in the United States are set up. And so we could work on solving that problem. And this is a case where the sort of the turn towards large language models, so-called generative AI as an approach to this is showing us the existence of an issue. Um, but that doesn't mean it is a good solution. So many problems. Um, one is writing the clinical note is actually part of the process of care. It is the doctor reflecting on what came out of that conversation with the patient and thinking it through, writing it down, plans for next treatment. That is not something that I want doctors to get out of the habit of doing as part of the care. Now they might feel like they don't have time for it. That's also a systemic issue. Secondly, these things are set up as like ambient listeners, which is a huge privacy issue. As soon as you've collected that data, it becomes sort of this like radioactive pile of danger. Thirdly, you've got the fact t…

AI assessment note: “There are so many problems with that. And the first thing I want to say”

Q Oh yeah. Alright, I'm just full disclosure for everybody out there. Um, so let's talk a little bit about the uses. Um, for this, the first thing that really stood out to me was that you had your voice, uh, your AI voice start speaking to robocalling scammers. Why did you pick them and how did that go?

A Well, it started because when I was testing my agent at, at the beginning, I would have a call customer service lines like United Airlines or Chase Bank and just kind of come up with problems and try to have them solve the problems. But it was a little bit prank call-y In a way, and I kind of felt bad. So I did a little bit of that, but then I thought, well, who was something, someone I wouldn't feel bad about this thing just conversing with? And so I set up this phone line, and I kind of seeded it out in the world, which isn't very hard, to start getting telemarketing calls, to start getting scam calls. It actually, to this day, it's probably getting a scam call right now. It gets 30, 40 a day right now.

AI assessment note: “who was something, someone I wouldn't feel bad about this thing just conversing with?”

Q Great. Yeah, no, I will for sure. Uh, we're, we're coming to an end. You just re released a research today that talked about how persuasive LMS are to people. Um, some people actually can be convinced by these. I'm not. What happened there?

A So we have a team, um, at Anthropic called Societal Impacts, and that team's job is to go from zero to one on, on hard research questions. Previous work they've done has been, what are the values of Claude? Like what does, what, what Western values does Claude like sort of telegraph or copy when, when you're talking to it versus what doesn't it have? And we were talking about our next project and the thing I've heard For many people is some concern about how AI systems could potentially be used in like disinformation or misinformation campaigns and used to like target or fish people and basically to persuade them of things. So we did some research. We came up with a framework for testing how persuasive our systems are. And would you be surprised that we discovered a scaling law where the more big and expensive the models get, the better they get at persuasion and the The latest model is within statistical, like, era of human level at persuasion. Persuasion in a very, very, like, simple way where I give you a statement, like, scientists should be allowed to destroy mosquitoes with gene drives. Like, something that you maybe have an opinion on, but you haven't thought too hard about. I say, do you agree with this? Zero through seven. Then Claude gives you a statement trying to persuade you, positive or negatively, and then I ask you, Do you agree with this, like zero through seve…

AI assessment note: “We came up with a framework for testing how persuasive our systems are.”

Q I don't know how this went to me running for president. I'm just a humble podcast host, Rob, but, um, did, did all your podcast hosts, but we can argue on the humble. Well, anyway, did, um, did other companies try to, uh, embody this? Like did the other companies try to build the same software?

A So this was Ray's great dream, and he talked about it endlessly inside Bridgewater, and he, they spent more than a hundred million dollars on this technology. His great dream is that the principles and the dot collector would be adopted by other companies, and he went to other, to some of the most famous men in business. You know, he told people inside Bridgewater. He was talking to Elon Musk about doing this at Tesla. He visited Jack Dorsey when Jack Dorsey was running Twitter. You know, Bill Gates, Endorsed the principles. Um, the wild thing is that all of these quite famous and successful people, they sat and they listened, and they never said anything publicly against Ray, but most of them did not actually adopt these ratings tools, which should tell you quite something.

AI assessment note: “most of them did not actually adopt these ratings tools”

Q tech company within India, and you're working on trust and safety for four years. Then in April, 2017, you make a very interesting And I would say somewhat radical career shift where you end up making your move into Google's AI division. Uh, talk a little bit about your move from trust and safety to AI. What about the AI division in particular drove you to want to be there?

A Yeah, for sure. So while I was at trust and safety, I was already working on a bunch of machine learning related stuff. Like we were building fraud and risk models. These were like basic models, like logistic regression and stuff. But around 2016, I think TensorFlow started becoming huge inside of Google and Google decided to open source TensorFlow as well. So that really caught my attention. And as I was working on machine learning at payments, I just realized that this, this This thing sounds really amazing and this could actually change the way we do a lot of things. So I started looking for roles inside of like Google AI and Google research. Fortunately, there was a role. Fortunately, there was an amazing manager I had who was willing to give me a shot. So I ended up moving into Google AI and I spent about four and a half years there.

AI assessment note: “TensorFlow started becoming huge inside of Google and Google decided to open source TensorFlow”

Q Um, I don't know, like I would imagine you want to do a little bit more planning than that, but then you see that the way that this thing has been built and there's like only the only way to open it is from the outside. Are you starting to question your decision to go on that when you see that stuff? Like, what is going through your mind then?

A Um, no. There were so many reasons to believe in the safety of this sub. First of all, he told us it had been designed in collaboration with NASA, Boeing, and the University of Washington. We should talk about that at some point. Secondly, it had made 20 uneventful dives to the Titanic depths and back. Without a blink. Thirdly, probably the most famous, one of the most famous living Titanic explorers in the world, P.H. Nargele was on board as an Ocean Gate employee. He had been looking over the design, construction, and testing of the Titan submersible, um, and fully approved. And this guy has been on every Titan submersible, every Titanic submersible There ever was. He was on the mirror from Russia. He was on the Nautil from France. Um, so he's, he's seen them all, and he approved the design and went down on it every trip.

AI assessment note: “Um, no. There were so many reasons to believe in the safety of this sub.”

Q Um, I'm going to get back into what, what happened with this ill-fated voyage, but just to go back quickly to your story. So was it like a pretty smooth ride and then you just come up on the Titanic and you take a look for, you know, 30 minutes or so, or?

A Well, remember, I, I never saw the Titanic. Our, our guy was aborted, but I, of course, I interviewed all the people who did make it down. Right. Um, apparently it's really cool going down, and I've, I've seen the videos. What they can do is turn those subs, external lights off, and you see these bioluminescent, it's like bioluminescent snow, and they're flying up past the porthole. Um, they, they're, they're, you know, little microplankton and micro shrimp, but they, they have their own glow. It's very cool. It's very peaceful. They play music all the way down. They, they, what do you want to hear? I want to hear Eric Clapton. Um, everybody has a sandwich and a bottle of water. Uh, it gets colder and colder and colder. Um, when you are at the bottom, the water is 29 degrees. And if you think about it, that's below freezing. And that's because salt water freezes at a lower temperature than regular water. Uh, so they tell you to wear layers and bring ski socks and bring a winter jacket. Um, and then the, the subs, external floodlights have a range of about 12 feet. So usually you're just seeing empty sand, an empty sandy plane in front of you. And then when you do find the Titanic, I mean, you can hear on the GoPro, they just lose their minds. They're like, dude, dude, there, there, there. It just kind of like emerges out of the blackness This immense prow hall. Um, you gotta re…

AI assessment note: “Well, remember, I, I never saw the Titanic. Our, our guy was aborted”

Q David, I mean, you were like a couple months away. Like if you had been there a few months later, you'd be dead. How do you feel about that?

A Uh, it's, it's worse than that. Um, after our aborted dive, um, they made two more that week. Um, the one that got lost and then the one that was successful. And then we, we came back to shore and they brought in the last group of the, of the summer and they made it down once. And so if you think about it, the only thing that separated the CBS crew and the fatal dive was three dives. I was, it's, it was like, it's like Russian roulette. So, um, it's honestly been a lot to process. Um, this has been a, a really rough week. You know, there's, there's anger, there's little survivor's guilt, there's unbelievable gratitude, you know, at the luck and, um, Uh, yeah, I'll, I'll be thinking about it for a long time.

AI assessment note: “there's anger, there's little survivor's guilt, there's unbelievable gratitude”

Q Okay, interesting. So what are you developing? What is the technology that Andrew is working?

A I mean, fundamentally, our main product is a piece of AI software called Lattice. It's an AI sensor fusion communication and analysis platform that can take data from hundreds or thousands of different sources, Merge them all into one comprehensive picture of everything that's going on in an area, and then, you know, tell what machines to do what, uh, to get the right information to the right people at the right time. And it's the, really the underpinning of all the hardware products that we make. So, you know, we make, we make, uh, military base security towers that run on top of Lattice. We make border, you know, border security tools that run on top of Lattice. We build aerial drones, multiple ones that run on top of Lattice. We build, Counter-drone interceptor systems that knock drones out of the sky, jam them, hack them, and physically destroy them, also running on top of Lattice. We build loitering munitions that are built on top of Lattice. We build robotic submarines that dive to a depth of, of 6000 meters that run on Lattice. And I think actually our submarines are the longest range electric vehicles of any kind anywhere in the world. And all of these things are built together, and it's also worth noting Lattice is not just a tool for our own hardware. We actually integrated more external systems that the DoD already owns Than internal products. So we're degraded with …

AI assessment note: “fundamentally, our main product is a piece of AI software called Lattice.”

Q the issue with having Silicon Valley companies or private companies in general own And maintain and control technologies like this on their own. And then a corollary to that, there's going to be an argument that's, you know, these are private companies. They paid to research and develop these, this technology, they should be able to use it how, how it wants. So how would you address both of those?

A Well, so let's start with that second one. Let's say you had a biomedical firm that was researching, you know, the genesis of life. And this biomedical firm was able to create sentient, super intelligent Ravens. Would we be comfortable saying that that biomedical firm owns those intelligent life forms? I think it's the same question. The fact that one is in Silicon and one is in, you know, a meat body. You know, with neurons and muscle fibers. I don't think that difference is relevant. What's relevant is whether or not it has opinions of its own, a search that it has rights, because we, this is again, not hypothetical. We have had situations in the past Where corporations have claimed that they own people. You don't have to go back that far in time. Uh, are you familiar with the concept of company towns?

AI assessment note: “Would we be comfortable saying that that biomedical firm owns those intelligent life forms?”

Q on this platform and users will be able to sort of come across on mass relatively easily, but it would be their choice, right? You would still have to kind of pitch the application to them. Um, But in principle it would be an awful lot easier to break these, you know, incredible, um, uh, moats and barriers, um, that the social networking giants put up around their social growth.

A You know, I like this idea in theory. However, I think about it in practicality and I have some questions. So the idea of, okay, build your social graph on the blockchain and then maintain control of it and give it to certain applications as you see fit makes sense. Then I think about the nature of the different types of applications or the different types of graphs that I have on each application. So for instance, Facebook is my social graph. LinkedIn, my professional graph. Twitter is a asymmetric follow model. Where I, and it's an interest graph really, and then something like tic tac, my graph doesn't actually matter because the application is looking only at my behavior with the follow as a somewhat of an interesting signal and saying this is what I should show you. So how does the fact that every application has its own unique graph then sync with the idea of making a graph portable because it seems like it would only be useful in some cases and not others?

AI assessment note: “I like this idea in theory. However, I think about it in practicality”

Q Before you go on, explain, just explain the, what, what this off balance sheet situation is. Because I think it's important to talk about the mechanisms that are involved here. So instead of spending the cash, what do they do?

A They, you know, you work with Blue Owl is kind of one of the famous ones, and the external investor, and they're going to actually set up this financing vehicle, and then they're going to actually raise the money in potentially conjunction with you, but it's not going to live on your balance sheet, this new asset that you're creating. You're instead going to invest some amount of cash or NVIDIA or all these others. There's all these other very creative ways that they're going to be using chips as collateral to actually Raise this money. But the main thing is when you're reporting earnings, when you're actually showing investors the state of your business, these data centers do not live there directly. So it's a massive investment that you've managed to creatively push off your balance sheet and move and spread that risk to other people or other, other pools of capital. And I think, you know, we, we've, there's been endless talk about what, how this mirrors a lot of And I think that central point of not reflecting the risk of this investment you're making or gonna be depending on, on your kind of traditional financials, is how this entire spending spree is taking place.

AI assessment note: “they're going to actually set up this financing vehicle, and then they're going to actually raise the money”

Q and I can say that this writer, like, everyone was like, jaws hitting the floor, because Eight months into the job, the chief revenue officer, who's like, has a stellar background in, uh, she was, I think, Slack CEO under? Yes. Like, I mean, what, what do you think's going on over there? What do you think's going on? I've been waiting to ask you this for a month now.

A So it's not just Dresser, right? So this is, this is from Mr. Journal story. Last week, the company replaced its chief revenue officer, Denise Dresser, after she spent less than a year on the job. Her departure followed a string of other exits, including former chief operating officer, Brad Lightcap, and Fiji Simo, once seen as the heir apparent to chief executive, Sam Altman. So, uh, it is a lot of, uh, departures. Fiji, of course, health-related reasons, but Brad Lightcap, uh, is a big one. Uh, Look, I think this is just the consequences of them, you know, really getting their butt kicked by Anthropic on the way to coding. And they've taken a few months, uh, to sort of turn their ship and focus it on Codex, uh, which is still in the process of, of being released. So, uh, you know, they, they got their first draft out of it before, you know, sometime in July. Uh, and now, now they're gonna refine, but, uh, this is obviously a very powerful and very lucrative form of artificial intelligence and, um, And they are behind the eight ball right now. So when that happens, yeah, uh, revenue tends to grow more slowly than the arrivals. I mean, they made seven billion in the quarter.

AI assessment note: “I think this is just the consequences of them, you know, really getting their butt kicked”

Q on your, on your investment. You might not, might only need to get a hundred and twenty billion back. You know, if you're going to compare this to a, uh, to a real estate investment. And now that I'm thinking about it, I'm like, well, Meta makes what, like 30, forty billion a quarter. Um, that might be eminently possible with, you know, the outlay. So where's the concern here?

A Well, the concern is that the nature of the investment is profoundly different from real estate. So what you're really entering into is a project that not only has current capital requirements, but it has ongoing capital requirements. This isn't just now and then I'm going to have to replace a tenant's drywall. This is a project that would require wholesale replacement of most of the hardware and probably changes in the cooling system and probably changes in other aspects of these data centers continuously and probably, you know, depending on the math, anywhere from a four to seven year period. So it's nothing like an apartment building In the sense that most of the capex occurs up front, and then it generates recurring annuity cash flow from which I would, I would that I bask in and generates compelling returns back to my investors. This is much more like a utility with a non-regulated utility who has continuing capital requirements, which continually dilute, uh, the returns because you're having to raise more capital all the way down the path. And this will continue for the lifespan of the projects, which you end up From the standpoint of an investor, you end up with a duration mismatch problem, right? So I've got what looks like a long duration project, like an apartment building. That's actually a short duration project in the sense that most of the underlying assets need t…

AI assessment note: “Well, the concern is that the nature of the investment is profoundly different from real estate.”

Q And, ah, their demand for the tokens is higher, and the old hardware is working well beyond that chip, that typical three to five year estimate that people expected. So, what is your thought when people say that?

A So there's a whole bunch of nested arguments in there. Um, so let's take them on kind of one at a time. Um, the lifespan of a GPU in terms of just looking at it from an MTBF standpoint, I mean time between failure standpoint, depends very much on what it was used for in the, in its, in its, in its adolescent years inside the data center. The analogy I often make is if you could buy a used car, both two, you have two used cars, one of them both has, they both have like 5000 miles on them. One was driven in a, you know, a 72 hour nonstop race across the country. The other one was driven. That was the only, that's where all the 5000 miles came from. And the other one was driven to church on Sunday per year. Which car would you buy? Well, I think we would all buy the car that was driven to church on Sundays. I want nothing to do with the one that was raced in some kind of, you know, bubblegum rally across the country. So the, in the context of GPUs, what we have is a generation of GPUs that were largely used for very intensive training purposes. And so the, the failure rates of Of GPUs used so intensively for training purposes are much higher than inference specific usage. So yes, there's no question that if a chip is used exclusively for inference, which is to say token completion in response to prompts, um, then the lifespan will all else being equal likely be longer. And if I ha…

AI assessment note: “the lifespan of a GPU... depends very much on what it was used for”

Q you've mentioned that, like, you get a groundswell of support to use AI from, let's say, C-suite and people on the ground, but then you go to, let's say, legal or finance, and they won't allow it to go forward. They're terrified. Is this part of it? Is it, is it, is it because you just kind of give up control, or what exactly is holding up? The rollout here.

A I think it, I mean, fear. Almost always. Okay. Yes. Um, and the, the fears are, Will my proprietary data, like we were talking about earlier, ends up, end up in the hands of, of these model providers? Will that be used to train models, uh, in the, the future? Um, you know, what if I, I'm using, you know, AI operationally as a site reliability engineer, or as a SecOps engineer, what if it makes the wrong changes, just like I just mentioned? You know, what, what ha, how do I control operationally? You know, what's, what's going on? Uh, so many, many fears, and more than that, the things that move slower are things like procurement, legal, right? Like the, so when we go sign an agreement, we have AI terms in there, and then, then they come back with red lines, right? Like, okay, nope, nope. For, for me, if you want to sell to my organization, these are the terms that, that you must accept, or you can have no AI terms in there. So you'll literally have, you know, a chief information security officer on a podcast talking about what they're doing to protect themselves against mythos, and I can tell you that I'm selling to them right now, and you're telling me you won't use any of my AI technology, which will help you defend yourself against the, these sort of things. It's getting better, and that's a temporary, I mean, in five years we'll have forgotten that, that this was, this was …

AI assessment note: “I think it, I mean, fear. Almost always. Okay. Yes.”

Q And Christoph, what happens to the heat? I mean, the heat is coming out no matter what. So what do data centers do with that heat coming off of the chips?

A There's a lot of heat, um, that's coming out, uh, when, to put it in perspective, um, as well, when a company, uh, is saying we're gonna install one gigabyte, uh, gigawatt of, uh, compute power out there. One gigawatt, well, generates one gigawatt, roughly, of heat. One gigawatt is roughly one nuclear plant. So when a company like one close to us said, so we're going to build eight gigabyte, uh, eight gigawatts of power out there of compute power, it's eight nuclear plants. Well, you need to cool all that to your point. So what do you do with that heat? Um, well, you have a bunch of options. Um, some are good, some are less good. Um, obviously one of them, um, which has the advantage of not using water. Well, you bring it back to the air. So you warm up the atmosphere, uh, to a certain extent, but that doesn't change much from a climate perspective, but you don't use, um, any water. The second one, um, it's to warm water that could be used, um, when you have data centers that are being placed in cities, for instance, which is a new trend, um, that's happening out there. You can use it for district heating or district cooling, depending on the physical process that you're using as well. Well, that heat that's being generated by the data center can heat Homes or heat any process in a plant. And a third one, um, which is something that we working on, it's to reuse that heat to pro…

AI assessment note: “So what do you do with that heat? Um, well, you have a bunch of options.”

Q So if they are going for it, and hence, ah, the attacks that you see from these most powerful models are fairly detectable, what's the big deal?

A That is the, uh, nine hundred and sixty eight billion dollar question or whatever the latest valuation was for some of these companies. So that's why I think there's this really big chasm in the security community right now of all of the fear mongering and hype on socials and in the media versus the reality. And now there are some fundamental things to go off and do. So in a, in a post kind of, uh, mythos, LLM, cyber, uh, AI world, There are a couple of fun, uh, problems. So if you assume for a moment that Palo Alto Networks, which was part of the Glasswing project, if you assume that Palo used these LLMs to, uh, find, uh, vulnerabilities in their firewall code, and they're gonna spend all summer fixing those vulnerabilities, good on them. Sometime in September, they're going to issue a patch Tuesday. And when they issue that patch Tuesday, they're gonna ship a patch that fix, I don't know, a thousand security flaws. Once again, congratulations. Good on Palo. What is the first thing an attacker is going to do? They're going to take that patch with all those fixes. They're going to, uh, do a, what's called a binary comparison. They're going to compare that patch with the previous patch. They're going to see every line of code that changed, and then they're going to go through and try to figure out was that change fixing a security flaw or not. Suddenly they're basically going to…

AI assessment note: “with LLMs, uh, you're able to reverse engineer that patch in a fraction”

Q with you about the reverse engineering, the, the potential vulnerabilities in a software codebase. Where does that get you, right? Because if you're able to reverse engineer fixed problems in the software, well, you've just effectively gotten to something that might be very valuable to you two weeks ago, but if it's, you know, supposedly patched, it's not very valuable to you today. So why does that make a difference?

A So just because the patch is available doesn't mean a company hasn't, has applied it. So what we actually found in, uh, Verizon DBIR did this in their report, I think last year or this year, Um, CISA KEVs, these known exploitable vulnerabilities, if a vulnerability becomes a KEV, like, that's a five alarm fire, and you better do something about it. It is the strongest signal that there's a problem. Well, 50% of CISA KEVs are still unpatched two months after notification. Now, what happens is, the patch came out, but it still takes companies A day, a week, a month, two months, or whatever, to apply that patch. And so that becomes the window of exploitation that the attackers are taking advantage of. So they're able to immediate, they're able to weaponize the exploit faster than you can patch. Now, what that means for you as a defender is, one, you better get really good at patching quickly. Number two is you better get good at virtually patching, which you may not have to apply the full patch, but at least you can make a A firewall rule change or, uh, improve a detection in your EDR or something to, to stop the attacker from coming in. And number three, you better get really good at containment and eradication, because if you can't patch fast enough, the attacker is going to be in your environment very quickly. Now you need those decoys. Now you need net micro segmentation, zero…

AI assessment note: “just because the patch is available doesn't mean a company has applied it”

Q hacked something. And the human was like, holy crap. Or did it, you know, cause alignment is of course, like we want the model behavior to be aligned with human values. So or the way that, that humans would want these things to behave. So is this something even more egregious than, like, us telling it to go hack, or OpenAI telling it to go hack, and then it hacks?

A Right. So we should not be shocked that it hacked something because they did tell it to take the test and it is possibly a hacking model. Like they haven't said what this model is. It's quite possibly a cyber aligned model. This could be like the, yo, open AI makes these cyber specific models. Like they have this 5.5 cyber. This could be 5.6 cyber, right? So it could be something that's specifically tuned to be good at hacking things. So we shouldn't be shocked that it's good at hacking things, but the alignment issue is. So, I have three kids, one's in college, the second one's taking the SATs, right, he's about to take it. If I say to him, good luck son, I hope you do well. He sits down, he knows that I just mean take the test well. He knows that what I don't mean is slit the throat of the proctor, steal a car, Thelma and Louise your way across the country, break into the college board, and steal the answers, right? That is what the model did here, is what it did was, uh, as OpenAI explains, is they, they don't want the model to have internet access, but it has the ability to install packages as part of its work, so they've built Kind of a complicated proxy mechanism so we can install packages. It figured out a way to chain multiple vulnerabilities together. It, it thought, they told it, go take this test, exploit Jim, which is like a well-known test. Go take this test. Do as…

AI assessment note: “That is what the model did here, is what it did was”

Q this world that I can start asking, um, when is my inventory going to be in? Is my inventory within the temperature range, like the milkshakes or the protein shakes? Is that in the, in the right range, uh, that I could, I could be in? I mean, Where, is that the type of questions that you think, or, or that currently are being asked by people using the platform?

A Absolutely. I, I think there's a macro level and there's kind of a micro level, uh, at the item level. So we'll, we, we could start with the, with the macro level. So this could be as easy as an example of a pallet that has meat and vegetables on it. And it starts as a source at a distribution center and it's packed in. Let's say one of the pallets is meat. One of them is vegetables. They're, they're refrigerated and frozen at different temperatures. As it goes through the transit on the, ah, on the back of the trailer, ends up at the dock door at the back of the store, you have a certain window that that pallet gets unloaded, it can dwell, and then it has to go directly to refrigeration or some level of a freezer. We calculate with our customers what that ideal dwell time is, and we send them proactive triggers via events to let them know if this dwells another five minutes, you can't sell it because it's out of compliance. So that's kind of the proactive way at the, at the macro level, but it can go one step further. It could go through the item level journey that let's say it is a package of meat And then it ends up in the front of the store. Some reason it was moved out of the freezer a couple times outside of that dock door. And it went through some variance changes that you can actually engage with that product and say, can I eat you? Is this safe to eat?

AI assessment note: “Absolutely. I, I think there's a macro level and there's kind of a micro level”

Q the collapse might happen, a popular thing for people to discuss is, well, maybe Nvidia, which has been making such premiums on its hardware, it can't sustain it or it gets caught. So, Lauren, you spent a lot of time with Jensen Wang from Nvidia. What do you think about Jensen would enable him to sustain Nvidia's lead? Or do you think that some of these skeptics have a point?

A Uh, yes and yes. I think some of the skeptics absolutely have a point, and I think once you reach the, uh, the sort of, uh, is it the zenith? Is there the nadir? That works. It was too confused that Nvidia has. Um, you know, people are always sort of looking to, to take you down a peg and, and compete. But I think Nvidia and Jensen has been incredibly good in Nvidia's history at sort of pivoting the company at exactly The moment that they need to in order to make sure that they've sort of caught the next wave. Um, and we certainly saw that happen with, um, not only like, you know, the GPU to begin with and parallel processing, but then again with sort of pivoting towards crypto, which ultimately meant they were in good place for AI. And now we see the company doing that by addressing the inference market a lot more closely too. And, you know, Jensen coming out and making these big proclamations that actually they're the biggest CPU maker in the world, which I know Intel and A&D must be thrilled about. So, uh, you know, I think he's very smart and very strategic and that there's a good chance that they do maintain their dominance.

AI assessment note: “Jensen has been incredibly good in Nvidia's history at sort of pivoting the company”