The Exchanges

Q And be about Trump's election. Okay. But, but what about the money though? The, the, um, the, the deals for his companies. Did he, did he make out okay from, from those?

A Yeah. I mean, he did great. And I think some of it, it's, of course, it's always hard to, it's always, it's always hard to untangle. Okay. Like why did Palantir get those contracts? Did Palantir, Palantir got a, uh, an eight hundred million dollar army contract, another four hundred million dollar contract. Um, and, uh, uh, he took over this, uh, project maven from Google. That's another forty million bucks a year. You know, it goes on and on. Um, And of course it's, it's very hard to untangle, you know, why does a company get these deals? Um, because of course they are, you know, they're, they're, they're explanations. Well, we were getting the deals cause we have the best product and we've spent a long time, um, selling it. But of course, when you're talking about defense contracts, I think inevitably like politics plays a role. And, um, and I think it's, it's no accident that, that Palantir took off during the Trump administration, that, that, that really had these, a series of breakthroughs. You know, you know, under an administration where Teal had influence. Um, I also, um, I also think, you know, you brought up his, his, his tax status and, and this, um, we could probably have a really long conversation about this, but, um, Teal has been pursuing his whole career, this very, very, very aggressive strategy to pay. It's legal, but, but just exceedingly aggressive strategy …

AI assessment note: “Yeah. I mean, he did great.”

Q Now, yeah. And I know you say it's not a duel, but what goes through your head before you say, hey, okay, we're gonna, we're gonna go do this. Uh, you know, obviously people are gonna look at it as competitive to your old employer. Did you worry about relationships there or, um, you know, how it might be received and what has the feedback been from your former colleagues?

A You know, I obviously do worry about it. I have a lot of close relationships with a lot of people at Google. Um, and I would roughly say that feedback falls into two buckets. One set of people that go like, yeah, we understand why you're doing this and why you didn't think you could do this within Google and a different set that kind of goes, um, you know, we wish you really, you had done this within Google because if anyone could have changed what Google was, it should have been you. Um, you know, both are reasonable points of view. And there are some people that, you know, kind of don't just want to deal with it. This is all too much, um, for them. Um, and I respect those points of view, but at some level, one has to be driven by what one sees is, uh, you know, is the right long-term outcome. I personally do not think of ad supported free products as being good for consumers. Um, good for our country in the long term, um, because it is very hard for them to stay true to what you and I want as users and as customers of these, um, you know, of, of these products. That conflict of interest is just, is just really, really unavoidable. And the fact of the matter, uh, Alex, is that, um, while at one level the products are free, All the benefits of scale for products like this, they go to the creator of the product. They don't come to you and me. You know, when it comes to Neva, for…

AI assessment note: “I would roughly say that feedback falls into two buckets.”

Q working for Google as a head of marketing in the middle East in And then, um, the Egyptian regime killed a man, Khaled Said, and you created a Facebook page. We are all Khaled Said. So can you take, which eventually ended up leaving, leading to the overthrow of the regime. So can you take us back to that moment? What were you feeling then? And what were you thinking?

A Well, it's, um, uh, it was basically, uh, a moment of frustration because Uh, I'm someone who was born in 1980, and since I was, uh, one year old, uh, since I was, like, coming to life, there was only one, um, president that runs the country, and the ruling party was, um, kind of getting old in their positions of power. It's been, like, 30 years at the time, uh, and unfortunately, uh, the way, uh, the country runs, um, uh, has not been ideal in, in the perspective of a lot of people, especially young people, Uh, who, uh, were kind of exposed to the global phenomena, who got together on the internet, who, so, uh, when, when the event of Khaled Saeed death happened, and the response that came from the government about his case, Um, uh, was, was just basically, um, a denial and, uh, and saying that, no, he did not die from being, uh, beaten up by police officers. He just swallowed, um, uh, some drugs and, uh, and that's what basically caused him to die. Uh, when that happened, I started the page and my idea at the time was not to start any kind of, um, massive unrest or any kind of massive revolutions or anything. It was just as simple as, Uh, I am, I know how to communicate. I learned how to communicate, have a background in engineering and as well as a master in business administration, a lot of experience in the street. And, uh, it must be that people like me should be doing so…

AI assessment note: “it was basically, uh, a moment of frustration because”

Q a little bit of optimism and realism versus a sense of resignation, which we see too often. So It's a nice note to, to bring our conversation, um, to a close on. I do want, before we go, I want to hear more about, um, just quickly, what are you up to now and what's next for you? You live in Sanford, you live in the Bay area now too.

A So I live in California and I haven't been working for three years and a half right now. Um, uh, mostly working on myself and trying, I, I went through a very hard depression, um, because of, uh, you know, all the events that I've, uh, experienced in Egypt and then leaving the country. I haven't visited back home for seven years now. Um, I recently got divorced, so I'm kind of like, uh, trying to, uh, to Stabilize my, my life, if you, uh, if, uh, if you would say, uh, and bring it into a better place. And as I'm doing that, uh, I, uh, I'm enjoying it, and I learned to be very honest with everyone. You know, I kind of like, if someone is hiring me for a job, I, I, I don't mind telling them about my depression, and I actually would appreciate if they don't hire me.

AI assessment note: “mostly working on myself and trying, I, I went through a very hard depression”

Q smarter with the, with the actual model itself, um, you know, doesn't seem right to me. That, to me, felt like the weakest part, and also the part that got most people most alarmed, uh, of the entire essay. So Stephen, to you, what do you think about this recursive self-improvement, uh, uh, argument? And then briefly, just on the, on the, the entirety of the essay itself, your thoughts.

A I think Matt's essay is directionally correct, but a bit early, and there are a few steps that we maybe haven't gotten to yet. Um, I think he is largely correct on the automation of engineering within the AI companies. It's like a little overstated relative to my experience, the experience of people I talk to, but broadly there, there has been a huge shift. The job of an engineer at one of these companies now is much more supervising these agents as opposed to writing the code yourself. In AI, 2027, one of the big accounts of how explosive AI growth might happen. That's one step, but then you need to take that engineering and use it to actually automate the AI research. You need to go from being able to implement the ideas more quickly to using that to fuel faster and faster growth in the breakthrough ideas themselves before you can turn that around and say, now make the AI better and better, at least in a really concerning way. You certainly go faster with just engineering. OpenAI talked about that with some of their launches from this past week, how the model played a role in this. Um, but it's not a full runaway train. There are also questions about what happens from there. Are there enough GPUs to go around? What bottlenecks might we encounter?

AI assessment note: “I think Matt's essay is directionally correct, but a bit early”

Q How many of your clients, like, do you think will actually go ahead with, um, the blocking of these AI crawlers? I'm talking about not, not your whole client base, but just the news publishers.

A Just the, just the folks who are ad supported. So again, we've done this that like, if you're, if you're not ad supported, none of this applies to you. Like you can, you can set. But if you're ad supported by default, starting mid September, we're going to, you know, switch this over and you can, and we can opt out, but, but we expect that the vast, vast majority of people that have ad ad or subscription supported. Uh, businesses will, will do this. So there will be millions of sites that will drop off of Google's radar. And at the same time that we will, um, very much, uh, tighten up all of the other controls to on, on AI crawlers. So right now it's basically put sort of a no trespassing sign in the top line and say, don't, don't come and get this content. And most people are actually pretty good at respecting the, the no trespassing sign. Starting in mid September, we're gonna actually make it not just as no trespassing sign, but we're actually gonna have, basically put a bouncer there saying, Unless you're compensating creators, you don't get this content.

AI assessment note: “we expect that the vast, vast majority of people that have ad ad or subscription supported”

Q Um, yeah, no, but do you, do you think they're going after TikTok and YouTube here, especially YouTube to me?

A Yeah, I mean, they know everybody, it's, it's amazing how, uh, how YouTube has become this, this force that's eating into everything, eating into cable TV, eating into Netflix, of course, eating into the social networks, and I would even go further and say there's no such thing as social networks anymore in this world that you're hoping for, where there's no reverse chronological order, or there's no, there's no algorithmic feed. Uh, it, it, it's long past, uh, the point in time where that was even a possibility, and I think that You know, social media was a moment, but it's now just all entertainment. Meta knows that, TikTok knows it, YouTube knows it, and they're all competing on the same product, which is effectively shorts.

AI assessment note: “they're all competing on the same product, which is effectively shorts.”

Q of new problems, which we're going to tackle in a second. But I, I just want to go back to one thing you said before we move to move there. Um, which is, you said that you tried to sort of broker, um, a understanding and a way forward between the publishers and the platforms, right, the news publishers and Facebook, and that it didn't work. Why didn't it work?

A I thought about this quite a bit. Um, I, I really think at the end of the day, and this is actually why I'm very optimistic about news and AI going forward, you're never going to solve, um, I mean, what we were trying to do was get more high quality news on the platform, right? But if you're optimizing for engagement, which is what social media does, you can't also optimize for accuracy and quality, because that tends to not be what people engage with. What do they engage with? They engage with the most hyperbolic, you know, crazy content out there. Just, that's just the human nature. So, if social media is always going to optimize for engagement, it's very hard unless you make a decision, which, you know, is not something that big social media platforms are inclined to do about how, you know, how you're going to rank news content, then it's just not going

AI assessment note: “if you're optimizing for engagement... you can't also optimize for accuracy and quality”

Q this type of build out, what is the return that's going to be necessary to justify these investments? So let's just take like, seven hundred billion. What is, what for, for even investors to like, I don't know, not go under, or I guess a lot of this is big tech, but like, what are the numbers that we need to be looking for, for those numbers to be rational?

A So you have to turn it around and look at it from the standpoint of the providers of capital. So alternative uses of capital and what return I could get on the same capital in another context. So the way that I try to analogize this loosely, and this is very loose, is that data centers from the context of many capital providers are real estate. They're really just multi-tenant apartment buildings. It just so happens there's no humans in the apartment building. There's just GPUs. And so from the standpoint of providers of capital who look at these, As project finance in the, and then by that measure, try to compare the returns they're getting on this to the returns they're getting from doing project finance. So think about it in the context of commercial real estate, uh, a strip mall, uh, a multi-tenant apartment building or whatever else. So increasingly the providers of capital for these things look at it in that context and say, well, what's the yield in terms of I'm contributing on, you know, a hundred billion dollars to some massive meta project. What's my reasonable cashflow expectation very much analogous To what I might expect from the cap rate on a, on a multi-tenant apartment building, and is this competitive on that basis? So that's, that's the, the, the short answer to your question is it's very much a market-based return that's required. The scale of the money is, i…

AI assessment note: “we're looking at cap rates around 6.8%, six percent, is that reasonable?”

Q A million X. Okay. Let's just say that. Um, a million X. Why are they still writing the checks?

A Because, yeah, no, it's, it's a bit crazy, but yeah. So why are they still writing checks? But you might say the same thing. It's back to the Harold Prince line. Back at the, during the financial crisis, as long as the music is playing, I keep dancing. This is that, right? This, as long as the music is playing, they're all going to keep dancing because there is absolutely no incentive as any of the largest capital providers on earth from sovereigns down to private equity and private credit to walk away because you get pressure from ELPs saying, why aren't you participating in this? And then even worse as a sovereign, as a sovereign wealth fund. And I've been inside these folks is that Once you're managing hundreds of billions of dollars, you start looking at opportunities, not in terms of their economic value, but in terms of check size. And you say, I need to write a check for fill in the blank, a hundred billion dollars, because I do not want to write a hundred one billion dollar checks. So this weird filter starts happening where you now, these projects are like, look at my friend, sorry, Qatar. I have this project. That's perfect for you. You want to write 50, a hundred billion dollar checks. Nowhere else on earth can you write it other than these giant data center campuses like the, the, the meta project in Louisiana or take your pick. And so once you become, uh, develop a…

AI assessment note: “there is absolutely no incentive as any of the largest capital providers on earth”

Q for his great newsletter at paulkodroski.com. All right, Paul, uh, you know, we talked a little bit on the other, on the, on the first side of this, this break or before the break about, um, the money will keep coming until the music stops. Um, I mean, I imagine it would take something dramatic for the music to stop playing. What do you think could be the compelling event?

A So the argument I make is, you know, people fall into this trap of saying it's going to be this or it's going to be that. I think it's actually overdetermined in a statistical sense, meaning that there are so many different ways it can stop that. The only thing you can say is that it's going to stop because it could stop because of a macro event that changes the hurdle rate that external capital providers are looking for. If I'm suddenly looking for, you know, high single digits and not six and a half anymore. Well, then all of a sudden data center projects with their deflating underlying token comprising, uh, looks much less competitive. So that changes things dramatically given that more than half Half of data center projects now or half of the capital for data center projects now are external financing. So that changes things dramatically. Um, the, the, so, so the, the, the providers of capital pulling back is a, is a, an obvious source. And then obviously the, the, the post IPO phenomenon of having these, these companies having to generate competitive returns on the back of a deflating commodity and then moving up market and discovering the returns aren't there as they move up market and they continue to spend aggressively on CapEx. Investors become dis, uh, you know, unhappy about it very, very quickly as we know from hanging around this stuff for a long time. So it wouldn…

AI assessment note: “it could stop because of a macro event that changes the hurdle rate”

Q That is scary. So how are you playing it? I mean, you are an investor. Are you like shorting certain things or what is your plan here?

A So I very much, so my day job in large part is in venture capital. And so for the most part, we just don't invest in it. We just, it's, it's not obvious. It's not obvious how to invest around AI because one of the worst things you can do as a venture capitalist is get into a marathon where there's a thousand participants. They're all at the start line. They're all well-funded and well-trained. And it's like, oh my God, I'm going to have to outlast all these people to get to the finish. And so you, you really have to pick your spots and try to stay away from these sectors where people are concentrating capital and doing it in a way that leads to much poorer returns. So for the most part, you know, we've been, we were very active in a host of different areas, but not, not AI, which is perverse because it's not because we don't believe in AI. It's because we believe it's structurally a terrible place to be as an investor. And then on a, on a more personal level in terms of, you know, assets, I just. I'm very loathe to invent, to commit any, haven't committed new capital to any sort of broad index class passive categories in over two years for that reason. Because you just, you, whether I like it or not, prior commitments now amount to a much larger commitment to this asset than I would like already. So I'm already over invested in this stuff just by the fact of having a pulse and …

AI assessment note: “in venture capital. And so for the most part, we just don't invest in it.”

Q of AI model leaves exhaust, uh, and then the AI model companies can come build competing products. Um, I just want to ask you, like, how true is that? Uh, how, like, is this data actually valuable enough for an anthropic or an open AI to see people using their models in certain ways, and then being able to go out and basically go up market and build another product?

A So my co-founder likes to say the value of this data is essentially zero until suddenly it is not. Um, and so, yeah, it's a lot of, like, individually looking at one user's trace is essentially worthless, unless I have some security problem I need to go urgently look at it, but normally it's worth nothing. But in aggregate, it's incredibly valuable, because the being able, and it's also valuable It slices different ways, right? So there is value in saying, okay, hey, watching every developer, what questions they ask, what code gets generated, whether that code is good or bad, will allow me to reinforcement learning, ah, on the, on these models to make them better for everyone. Then there's also, okay, well, there's also organizational specific value, which is the, the model, you know, working inside of an organization. Uh, we believe we will be doing reinforcement learning for specific organizations to tune the models to their specific, uh, environments. Um, and so the, uh, and there's also, uh, there's also value from a security perspective as well. Both individually and in aggregate, understanding what is good and what is bad. So it's valuable, but it's also incredibly expensive. This, this data is emitted at very, very high velocity and very large volumes, and so, you know, we're having to build new technologies for how to process this cost effectively so that we can actuall…

AI assessment note: “individually looking at one user's trace is essentially worthless... But in aggregate, it's incredibly valuable”

Q couple months ago that, like, only five to ten million people use agents, um, but I'm curious to hear your perspective on, like, are you seeing it now in terms of the increase in the amount of data that's being generated, and where do you expect this to go as, you know, these agents aren't used by, like, five or 10 or fifteen million people, but by a billion people?

A Yeah, so I believe it's felt, but I don't think we can measure it yet. So the, the general growth of data as measured by IDC is about a 30% compound growth rate. And that, that's been, that's gone up from, let's call it 25, 27% over the last couple of years. But the, to your point, agents are only in the hands of a handful of people today. Um, It's not only in the data that's output by, by using the agents themselves, it is also the other exhaust as the agents move at machine speed as opposed to human speed. So, if I now have agents that can work 10, 20, 50 times faster than humans, all the things that they are replicating that humans used to do, Now also amplifies. So if that agent is browsing a website and I used to have, you know, let's say I'm an enterprise and I used to have 10,000 internal users of this application, but now I have agents which are, you know, exploring multiple hypotheses and now start to look like 30,000 or 50,000 users. Well, now I just had an increase of a factor of three or five on all of this other exhaust and all my other systems that, that are, uh, emitting this, this data. That being said, it's still too early to be able to point to like an industry example of like, well, we know it's going to be blah. So logically, we're feeling it. We know it's happening. Uh, you can kind of get some samples and some observations of AI native organizations genera…

AI assessment note: “I believe it's felt, but I don't think we can measure it yet.”

Q Okay, so speaking of not caring about the math and going all in, one thing that you wrote recently, I think it was in twenty-twenty-five also, you wrote, one thing has become clear, nothing short of AGI will be enough to justify the investments now being proposed for the coming decade. Do you still believe that?

A Yeah, I do. Um, I think that's when you and I first started exchanging emails, uh, on this topic. Uh, it's funny actually, I, I remember writing that post, I had just read Hyperion, which is the name of Meta's new data center. Maybe there's something, ah, to that. But, um, you, at that time, again, to remind the audience, at that time, everyone was talking about 10 gigawatts, then 30 gigawatts, then a hundred gigawatts of CapEx. And so, and those are just astronomically large numbers, um, far beyond where we are today, to be clear. Um, and so if you kind of think about those dollars of CapEx, the only possible way to pay those dollars back is going to be AGI. And so what I think the market kind of gets wrong, and it's just inherent to Wall Street, but Wall Street sort of always has this view of, like, is the stock market going to go up two percent, or is it going to go down two percent, right? It's like this just constant volatility, and everyone, you know, if the market's down 10% in a month, it's like, oh my god, if the market's up 10% in a month, oh my god. In reality, I think we are reaching this, like, sort of bifurcated path, where path one is, like, we get AGI, we pay back all these numbers and more. It's the greatest technology in human history, all of the things that we hear in the media all the time. And then path two is, We got the timing wrong. There's no next appli…

AI assessment note: “Yeah, I do.”

Q Why do the data centers take so much water and power to operate?

A It's different, um, reasons. So first, um, if we look at, um, the, the, the chips themselves, um, the way the technology works, well, you need power. You need electrons in order to get the system, the logic system work, um, in a chip. And the more of those chips you have, the faster they are, the more data they can ingest, the more power you need, um, to, um, Make them compute, uh, ultimately. So the higher the compute, um, performance of the chip, the more power you need. At the same time, and that's the tricky, uh, component, well, the more heat you generate, you have almost a direct relationship between how much power you use and how much heat you create. So that heat, you need to remove it, because if you don't remove it, well, the chip, It's gonna melt. It's gonna die. Not good things, uh, are gonna happen. And when you cool the chip, well, so far we've been using water to cool those chips through a cooling tower like you have in an air conditioning in a normal building that uses a lot of water.

AI assessment note: “we've been using water to cool those chips through a cooling tower”

Q Talk about that, because the chips perform better when they're cooler?

A It's a big debate, uh, right now, so we've all, um, heard as well, so NVIDIA and others, um, on social medias or in technology publications are saying, okay, we're gonna use chips with much warmer water. It's this, uh, 30 C's to 40 C's to 50 C's, so you need to cool less because you can afford warmer water. I personally I'm not sure that that's gonna stick, because at the same time, we see that when a chip gets cooled further, its performance increases, uh, as well at the same time. So, I'm just thinking, how do we react as humans? Uh, which trade-off are we gonna make? Um, the one that is protecting people in nature, or the one that's giving us the max compute performance, and we'll figure out, uh, The people and planet component later. I've been 30 years in that business, and one thing I've learned is that humans have a tendency to go for performance first and to figure out the problems later. So I would think that in that case, that's probably where we're gonna go, and I see it for a company like Ecolab to find ways to get much cooler environments, so for the chips, while we do it in a way that's using zero. Water because of the closed circuit system that we talked about. So it's a big debate, this one. I think that the cooler chips are going to prevail.

AI assessment note: “we see that when a chip gets cooled further, its performance increases”

Q we consider in their blog post, we consider this incident to be an unprecedented cyber incident involving state-of-the-art capabilities and are responding accordingly. Um, you know, it's sort of like, oh, look at this terrible thing that happened, but, uh, a moment to share, uh, how good our, our cyber capabilities are. That's the argument. What is your response to the notion that this might be some marketing from OpenAI?

A I know lots of people at OpenAI. Every single one of them absolutely hated anthropics marketing around mythos and thought it put the entire industry at risk. Um, this incident has put OpenAI at risk of regulation from the White House, regulation from the EU. It is also an admission of the violation of the Computer Fraud and Abuse Act, as well as multiple European laws. It would be absolutely insane for them to use this as a marketing moment. What you're seeing is them being very, very careful and defensive in their language. They're also very lucky that Hugging Face is being super cool and chill about this. So that is why they are saying these things because, um, you know, Hugging Face initially comes out saying we've been attacked. We don't know who it is, but it does not look like the model was being subtle. I don't know where it was running. It's quite possible. It's like Azure or something. It was probably not covering its tracks, and so I expect Hugging Face got their American lawyers involved, was working with the FBI, was probably issuing subpoenas, and was very, very close to finding out it was just open AI. So, like, or did find out. I do not know the timeline here, but like, The legal issues here are very fascinating and interesting. And because they're all working together, I expect nobody goes to jail. Nobody gets sued. Everybody's going to hold hands and hug. And i…

AI assessment note: “there's absolutely positively no way this was a intentional marketing move”

Q You can't really make mistakes here. But then again, something that we encounter often is that when you're dealing with AI systems, they are probabilistic. They are, you know, they're, they're not, you know, if A then B systems. They are systems that tend to freelance a little bit, for lack of a better term. So how can you then trust an AI system to act accurately in these environments?

A It's a great, great question. You have to start with the source of the data. You know, I've, I've been a victim of ChatGPT where I've asked it questions that I knew maybe 60% of the answer, and then I get an answer that's completely off base. And then you look at some of the sources that the data is being pulled from, you say, okay, this is not credible, or this was a opinion article and it pulled the information from there. So tying it back to the real world of, um, asset tracking and what that means from a data source is that you have to have a source of truth behind the data. And validity and trust that that data coming in is factual data, and then you can rely on the AI to do the, the legwork behind it. And, and for us, from a platform perspective, we really pride ourselves that the assets that we're tagging and the data that comes from it, we really synthesize that data in a right format so that when it is sitting in a platform and AI is now layered on top, Providing the insights that you're working off of really clean, decrypted, trusted data at the end of the day.

AI assessment note: “you have to have a source of truth behind the data”

Q the P and the T and the GPT as your, uh, Twitter bio says, the guardians called you, um, father of AI. Uh, it seems like the media tends to, um, you know, put these labels on researchers. You know, there's, there's a handful of fathers and godfathers of AI. Uh, how do you respond to that and how would you contextualize your role and where the technology is today?

A No single person can create an AI by himself or herself. You need an entire civilization to build an AI. You need not only the guys who are trying to invent algorithms, learning algorithms for artificial neural networks, that is what current AI is about. You also need people who, um, you know, build better computers. You need all the video gamers who are, um, Creating a market, uh, for, um, uh, acquiring more of these fast and faster computers, and you need, and providing an incentive to the computer makers to, um, speed up, uh, the computation per dollar by a factor of 10 every five years. You need all the farmers who are feeding the video gamers, and so on. So, it is impossible for a single person to create an AI. You need an entire civilization.

AI assessment note: “No single person can create an AI by himself or herself.”

Q the trenches working on AI researcher research for a long time, and we're definitely in a period of fast progress today. So there's been all these questions about how much better AI can get and whether it's going to hit a wall and where current techniques will lead. Uh, where do you, what do you think? Where do you land on that question about where AI can go from here?

A So, um, since the 19 seventies, I have been an optimist and I have been claiming That within my lifetime, uh, I want to build an AI that learns to become smarter than myself, such that I can retire, and we are obviously not there yet. And at the moment, the only AI, the only AI that is working well is the AI behind your screen, you know. Yes, behind the screen, there's an AI that can pass the Turing test. Um, what is that? It means if you type, uh, questions to it, it answers back, and, um, and And now the question is, can you distinguish whether the other guy is a human or a machine? And today, this Turing test is passed by many AIs. However, it just means that the Turing test is a bad way of measuring intelligence, because there is no AI in the physical world, outside of the screen, that can do all the things that a little boy can do, that can do the things that a A plumber can do. For decades I have used the plumber as an example. Or an electrician. Um, so all the, the, the things that humans can do with their hands, their physical hands, they don't work well. Only bits and zeros and ones behind the screen, that's the only thing that is working well. However, it's not going to stay like that forever, and, um, there is progress in the physical world, AI for the physical world, And I think the culmination point, the culmination point of that, which I've been talking about agai…

AI assessment note: “there is progress in the physical world, AI for the physical world”

Q Yeah, and so that sort of brings me to this question, which is, as the machines get closer to the human brain, does it change the way we're gonna think about what it means to be human?

A I, I think it will change what many people think about humans. I guess it won't change much what I think about humans, um, because it's more or less what I said, you know, many decades ago, but yes, uh, many people who, you know, who claim that AIs can't have emotions and stuff like that. They, they do that decades after the fact, and they are going to change their minds, I'm pretty sure, and often it's just a, a matter of direct experience, so maybe you have heard of these little sweet cute robot seals that you have in certain healthcare centers, and the people who are interacting with these furry Uh, artificial beings. They get really emotionally attached to them. They really like them, and they, they play around with them, although they are not smart at all. You know, they don't learn much, and, uh, and if even such a simple robot can invoke, um, feelings of, you know, almost love or something, then, um, you can imagine what will happen once you have really convincing, very sweet little robots that, um, are more like little animals, Except that they can do maybe a couple of things that these traditional biological little animals cannot do.

AI assessment note: “I, I think it will change what many people think about humans.”

Q whether, whether the margin of the business can be maintained at the current prices. And I thought, okay, well, forget about it because these labs have such an economically valuable tool for us that they'll raise prices. But we might be in the moment where they're going to get into a price war because OpenAI is rumored to be potentially dropping prices. And so what do you think about that?

A I think we're absolutely right on the precipice of that. I think you actually saw in the audience here by comparison, we saw every hand go up when you said, would you be willing to pay double when we were together three months ago in April? And when, when Alex asked if people were willing to pay four and five times as much, there was still a quarter of the hands in the room that were up. And we're talking a room of about 200 people roughly, right? So it was a lot of people was a good, you know, good, good tea sample, so to say. I contrast that to what we just saw right now, and I think it's a fairly, you know, even distribution, similar subset of people. And the reality is there's there's more skepticism of value that they're getting from it, especially when you start layering on the access and the capabilities associated with Some of the models that are still per seat, as well as some of the open models, which you can get access to and, you know, for free, you can do a lot of really cool things.

AI assessment note: “I think we're absolutely right on the precipice of that.”

Q to see this continue, agents can't be limited. They have to be able to operate autonomously and spend all those tokens and be effective. So what are the limits that you're seeing with agents today? And do you think that the, like, if we could extrapolate a little bit, it means that we're going to see some more speed bumps as the labs try to roll this technology out further?

A Yeah, I mean, you use the term limit. I think it's, um, I think it's, uh, a function of both risk tolerance, um, as well as cost tolerances. And then finally, like, what is it you have an expectations of these things doing on their own? And so the limitations are actually in all in all three of those areas that are coming through. Um, you know, from a risk tolerance point of view, I think people are saying, wait a second, I'm worried that the agent without some level of, you know, Call it control or governance around it. It could go just about anywhere. And what does that mean within my organization, depending on what access I give to it from a data perspective? Um, you know, I would say from client data or whether it's, you know, it's code itself and what can it do to change code? If you ask to do one thing in one area, will it simply think that it needs to do that everywhere else? And there's a, you know, we'll say the ability to extrapolate on a single point and like what control exists there. So that's the first limit. The second limit is on, like I said, on the cost experiments, variance. We went through there a second ago, which is, Hey, look, like there are just things you're not going to want it to do because back to the MIT study, there might be things that humans can do not only better, but more cheaply, especially now, depending on, you know, if you're, if you have a…

AI assessment note: “a function of both risk tolerance, um, as well as cost tolerances.”

Q able to function there. But, ah, and you've built these specialized models to do it. But let me put the question to you. If the entire thing, the entire generative AI moment is built on a model innovation that was meant to translate language, ah, then, then why would we need something specialized to do that as opposed to the bigger models with that foundational, you know, innovation baked in?

A Yeah, I mean, like, you're, you're right. The transformer model and language translation, I think they're kind of very, very, very tightly coupled. But I think also very early at the, at the beginning, it was clear that the transformer models can, can do more. But when they do more, when they're made for, for, for different purposes, they also lose a little bit of the capability that they had maybe initially when they've been made for translation only. This The set of parameters that is available there, this, which, which kind of determines quite often the capacity of the model, um, it needs to be divided into very many different things. And, and therefore, if you're keeping the model very much strictly to do one particular task, however, it's defined in this case, language translation, it can perform better and it can perform on that also more consistently. I think something that you see with generalized models is that depending on, on which kind of input you give to them, they're gonna tend to be better or worse, and, uh, specialized models have a kind of better layer of consistency. They do, um, they are, they're quite often much better, as in, as in ours, that the quality assurance, uh, is, is, is really built so that we can make sure that Whether it's an email that you're translating, or whether it's marketing material, or a technical patent application, or in all of those…

AI assessment note: “keeping the model very much strictly to do one particular task... it can perform better”

Q Interesting. And how much has the, uh, growth and of capabilities of LLMs Enabled you to do this job? Like, and talk, talk a little bit about how we've seen better LLMs, uh, and what they've, uh, enabled DeepL to do in terms of like going from a point A to where you are today.

A Oh, totally. I think there's, there's this, like, this big stack of, uh, of use cases and their complexity, uh, and how, how hard they really are to, to solve with, uh, with AI. And I think it maybe starts somewhere at the bottom of, like, sending out spam emails. Like, you really don't have, you really don't need to have, like, the best translation for, for that. Uh, anything just kind of basically works. Like, we know how those emails look like. Um, and at the, at the top range of that is probably regulated documentation that needs to be compliant, where like there's legal liabilities behind, behind all of that. Maybe think like a, um, I think like a leaflet that is, that is being distributed with a, with, with medicine, um, things like that. And As the quality of models has been rising over the last years, we've been able to unlock more and more and more of those, of those use cases. And honestly, this is always something that is really and truly complicated for our customers to find out is the model quality good enough for doing this in a particular job. And this is also then our responsibility as DBL to come in and help our customers find out Um, what is the quality level that they require? What is the error rate that they're seeing on those? What is, what is the reasons for this error rate? Maybe optimizing the whole setup there. And at the end, bringing them a solution w…

AI assessment note: “As the quality of models has been rising over the last years, we've been able to unlock”

Q is one thing, right? Let's say I'm a U S company. I want to operate in Brazil. I can probably get my company to function my website to function, maybe some of my customer service. In Brazilian Portuguese, but then there's also laws, regulations, customs, so does language get you half of the way, or how far, you know, can this take you when you're trying to operate somewhere else?

A I think it gets you pretty far, and I think it gets you already also pretty far because you can start leveraging local partners at this point in time already. Like you can, you can engage that Brazilian law firm that is going to help you in some aspects that are local, and you can engage them in a good way. And maybe they speak English, so then kind of this is easy, but maybe they're not. And in those situations, you can, you can, uh, already start far quicker. Um, there's going to be definitely things that you're gonna have to set up in your new market. It's not only language. There's also other aspects, uh, but you can, you can get there, I think. And, and hey, even like we're doing, um, in our, in our local markets, uh, for example, in Asia, uh, we are, we are talking to journalists. We are talking to our customers. And all of those interactions are being translated by our, um, by our technology and our AI. So like, if I would be doing this podcast with a, with a, with a, in a, in a Japanese market, we would be totally doing that with deep, uh, running the language layer in the background.

AI assessment note: “I think it gets you pretty far, and I think it gets you already”

Q Yeah. Uh, where do you stand on the AI device then? Do you think that that will be successful? Like an AI wearable or whatever open AI is brewing?

A I think it makes a lot of sense. I think kind of getting devices as small as possible and kind of as, as near to us as, as possible, specifically also in the case of, of language translation, uh, that, that makes a lot of sense. I'm, I'm a big advocate of the fact that for like real time translation, we actually have all of the devices that we need, like the airports that we have, like the phone that we have that actually suffices for, for that. Um, but having more data and having, uh, devices embedded with us all of the time and in a situation, um, like that can also gather a lot of data about us and therefore be more context aware. I think that that is pretty cool.

AI assessment note: “I think it makes a lot of sense.”

Q I know, but that's the flip side of it. Like if you don't have your memory with you, you'll be frustrated, right? That you, you taught it all this stuff to do. And then you go to work and it's like stupid. It doesn't remember anything of how you operate.

A I know, but it's going to be, I, I, I agree. It'll be, there will be a case to be able to bring your personal AI into your work setting. But again, like, let's say I'm at company A and my competitor company B is breathing down my neck. And now my star employee brings their AI to, to me, you know, they're in my systems and then, you know, company B now company B is really trying to figure out what's the best way to, to compete. Um, What company B could then do is make the godfather offer to that employee who not only will they get them literal institutional knowledge in there. Absolutely. They're going to get them and their AI with the memory of how to do that.

AI assessment note: “I agree. It'll be, there will be a case to be able to bring”

Q you might want to do next, but going to do it for you, that's going to happen. And so it makes you effectively an operating system, don't you think? But not the operating system like an iOS, where you would like go open up the your phone and then tap different apps. It's almost as if all interaction with all apps will happen. Through this interface. Is that the ambition?

A I think that you could describe it that way, but I think of it a little differently. Like the way that I think about this is that what is the ideal interface to an AGI where we call it kind of a personal AGI. And I think that it's, again, the same interface that you and I are using right now. You just want to talk to an assistant, right? You want to talk to something that can go and work and operate on your behalf. And so That yes, like that agent, that AGI, that AI will have its own computer, right? It'll have its own access to things that maybe can, you know, like ideal coworker would be they can come over and type things on your computer, too. So some access and delegated access to your own own system. And, you know, maybe you delegate access to your inbox sometimes. Maybe it has its own inbox with some sort of some sort of, you know, window into into The things that it needs, you forward emails to it. These are not actually if you think about it, this is not unprecedented, right? It's like the way that you work with an assistant who's a person that we've we've actually or any coworker really we've spent a lot of time really thinking about how do you build these trust boundaries and make sure that you're able to operate together. And so I think of it as just a different thing. It's not it. You could think of it as an operating system, but an operating system is almost Someth…

AI assessment note: “I think that you could describe it that way, but I think of it a little differently.”