The Exchanges

Q Can I ask you, when we think about that enterprise adoption, I think one of the biggest problems that we have is data structures and data cleanliness. Um, I interviewed a guest the other day and they said, we'll have like data cleaner as one of the most important jobs in the next five years. Um, is data structure and data cleanliness the biggest barrier to enterprise adoption?

A Well, I agree in part. I think that certainly the models need to have access to data to perform their jobs effectively, but the caveat is that they'll be able to clean the data themselves fairly effectively as reasoning capabilities go up. The thing that humans will need to contribute to is all of the tacit knowledge within the organization that isn't written down, because I've found that when I try to get agents to do all of these workflows Um, throughout Mercure, there's just an enormous amount of context that lives in people's heads that the agents need to have access to, to perform effectively. Um, and so much of that is going to be the new job of employees of how do we codify all of this knowledge? How do we train agents so that they're able to perform these tasks effectively across every function in the organization?

AI assessment note: “Well, I agree in part. I think that certainly the models need to have access”

Q I think mothers are the most important things in the world. And Victor told me that I had to start with your ability to sell early and why your mother was nervous about it. Can we just start there?

A Absolutely. So I had a dozen different side hustles when I was growing up selling things in, in one form or another, but one of my favorites is that in eighth grade I loved selling donuts, where I saw that Safeway was selling donuts for five dollars a dozen, and so I would buy Safeway donuts, I would bike to my middle school, and sell them for two dollars each. And I saw it was working, so I wanted to scale it up. So I asked my mom to drive me to Safeway. She said that she didn't want any giveaways, so she would charge me 20 dollars to drive me in her minivan to Safeway, buy 10 dozen donuts, go to my middle school, sell them for two dollars each. I had all sorts of things happen where competition popped up selling Chuck's donuts, which if people aren't familiar has like a one dollar cost basis, and so, but they're higher quality donuts, and so I dropped my prices to one dollar For two weeks to run them out of business, because I knew that middle schoolers would care more about, uh, price as the comparative advantage. I had my, uh, principal called me into their office to try to shut down my donut stand, saying that, uh, you know, I wasn't allowed to sell food on school campus, and so I moved my donut stand 20 feet over off of school campus so that they couldn't police me, so to speak. And tying back to your question, Harry, After my mom saw all of this when I was in eighth grad…

AI assessment note: “she was very nervous that I would start selling drugs, right?”

Q Do you buy sovereignty as a reason why a model provider wins? You know, we've got Mr. Allen Europe. You have cohere in Canada. Is sovereignty a reason why a model provider wins?

A Maybe wins in a scoped part of the market. Like, I could see why, for example, uh, there would be a lot of benefits to having Mistral be an expert in European law that might have nuances from other kinds of law, and they've just invested far more in having the best Model there where it doesn't make sense to use other models, but I don't think that The largest companies per se are going to be those that invest in a specific geography. I think it's going to be a broader set of capabilities and the general purpose models that people use every day to code or to, uh, build products, uh, or do their day to day work.

AI assessment note: “Maybe wins in a scoped part of the market.”

Q Do you buy sovereignty as a reason why a model provider wins? You know, we've got Mr. Allen Europe. You have cohere in Canada. Is sovereignty a reason why a model provider wins?

A Maybe wins in a scoped part of the market. Like, I could see why, for example, uh, there would be a lot of benefits to having Mistral be an expert in European law that might have nuances from other kinds of law, and they've just invested far more in having the best Model there where it doesn't make sense to use other models, but I don't think that The largest companies per se are going to be those that invest in a specific geography. I think it's going to be a broader set of capabilities and the general purpose models that people use every day to code or to, uh, build products, uh, or do their day to day work.

AI assessment note: “Maybe wins in a scoped part of the market.”

Q Do you buy this new sexy category? I mean, venture investors are wonderful people, but like this new sexy category that like AI enabled services is like the future gold mine.

A I think in a large way I do. Uh, I think, I think the key thing is that You need to make sure that they're actually going to leverage AI. Like, I think there are a lot of companies that are just like building services and not gaining a significant competitive advantage from AI and using that. That's the thing you've got to be careful about, but I think it's very rational. Like, I'll, I'll give an example in the context of Mercor, which is that we, within this process of turning human time from the talent network into building these super rich environments that mirror Everything that people could do in their jobs. There is a lot of human coordination of how do we answer people's questions? How do we track the KPIs of the project and manage it effectively? How do we build the bespoke tooling for that project? And we have about a hundred people, uh, or call it a 150 people in our delivery organization that do that for deployed work of helping to go the last mile for the customer. But now we have an AI project manager that just completed its first project managing that entire thing end to end, where it's able to hire the experts. It's able to answer their questions. It's able to build the annotation tool using its coding tools, uh, within our platform and produce the end data type. And the experts all had a really good experience on the project reporting to the AI project manager t…

AI assessment note: “I think in a large way I do.”

Q Final one for you, dude. What's the kindest thing that anyone's ever done for you?

A One that really stuck with me is I remember, and I'll probably attribute this to the entire prod community, uh, namely, especially a couple of people like Rob Walken, Ben Spector, and Richard DeHaan. But prod was this nonprofit that got started at MIT and Harvard. And I was sort of a blow in because I didn't get into those schools, but I went to Georgetown. And for the first year of the business, like they would meet with us every week. Ben became a big customer. Richard would Give us like tons of money just as to float working capital. And, uh, Rob gave incredibly valuable advice and they had nothing in it for them. They took no equity. I tried to give them equity and they, they wouldn't, wouldn't accept it. And Mercor wouldn't exist if it weren't for any of those individuals, I would say. And I think that that is something that I'll always be grateful for, for the rest of my life.

AI assessment note: “they had nothing in it for them. They took no equity.”

Q Okay. Let's say a billion because it's easy for my brain. Um, so we have a billion, like, is that like sales for Airbnb and then they get 20% of that?

A So the revenue is between a 30 and 40% gross margin, but the key distinction and why it's not GMV, but it's revenue, is that the experts are actually only one part of the broader value chain that we deliver to customers. So when a customer comes to us, they're generally buying tasks where they would say, Hey, they'll pay a thousand dollars for this task that delivers model improvement. And then we do the end to end process associated with how do we find the experts? How do we hire the experts? How do we build the platform that the experts work on so the experts can do the work? How do we have our AI project manager manage the experts to automate all of the, uh, coordination of helping to, uh, produce this data? How do we Uh, have automated quality checks, et cetera, to produce the end product of the task that we're delivering for a customer. And so that's the large distinction of how we're powered by a talent network in the same way that Uber is powered by a driver network, but that's not the end product, uh, in the same way as some of those marketplace businesses.

AI assessment note: “why it's not GMV, but it's revenue”

Q Is it harder than ever to run the company?

A I don't think so. Like, to give a frame of reference, we were 40 people and fifty million in revenue run rate last year, at the start of last year. Since then we've seven or eight X head count and we've increased the broader scale of the business by, you know, 25, 30 X. And it's definitely been very stressful to keep up with the growth along the way. But I think that now we have the supporting functions, like we have HR or not really HR, but we have finance and we have legal and we're building out HR. And that brings some sense of stability where I don't have to deal with all of these, um, little escalations, and I'm able to just spend my time focusing on building great products, research, and time with customers, and that I think has made it easier, significantly easier to run the business.

AI assessment note: “I don't think so.”

Q Do you buy this new sexy category? I mean, venture investors are wonderful people, but like this new sexy category that like AI enabled services is like the future gold mine.

A I think in a large way I do. Uh, I think, I think the key thing is that You need to make sure that they're actually going to leverage AI. Like, I think there are a lot of companies that are just like building services and not gaining a significant competitive advantage from AI and using that. That's the thing you've got to be careful about, but I think it's very rational. Like, I'll, I'll give an example in the context of Mercor, which is that we, within this process of turning human time from the talent network into building these super rich environments that mirror Everything that people could do in their jobs. There is a lot of human coordination of how do we answer people's questions? How do we track the KPIs of the project and manage it effectively? How do we build the bespoke tooling for that project? And we have about a hundred people, uh, or call it a 150 people in our delivery organization that do that for deployed work of helping to go the last mile for the customer. But now we have an AI project manager that just completed its first project managing that entire thing end to end, where it's able to hire the experts. It's able to answer their questions. It's able to build the annotation tool using its coding tools, uh, within our platform and produce the end data type. And the experts all had a really good experience on the project reporting to the AI project manager t…

AI assessment note: “I think in a large way I do.”

Q With the greatest of respects, does that not change so quickly? You know, when you spoke, when you saw Andre Capathi talk about how he uses coding agents, it was like, oh, I use it for 20% of the work, and then it's like, oh, it does 80%, and I do the final 20% within a six month period.

A Definitely. Well, even another example on that is on Apex, the Frontier model right now is at about 40%, and 12 months ago the Frontier model was a one, which was scoring one percent. And so that's been the progress of the last 12 months, and obviously we expect it to continue and be fairly significant. Um, but I think that the key thing is that everyone underestimates the elasticity for demand and increased productivity in the economy. Like, ultimately, over the last 250 years, we've increased productivity by 25 X, equivalent to automating about 96% of someone's job. And during every technology revolution, ranging from the agricultural revolution to the industrial revolution to the computer revolution, people feared that there would be this enormous job displacement because of the lump of labor fallacy, where people assume that there was a fixed amount of things that had to be done. And when we made people more productive, that would all of a sudden mean that there were fewer jobs. Yet, 250 years later, there's more jobs than ever before. And, and it's because we have No shortage of problems to solve as a society, right? We still need to solve climate change and cure cancer and do all of these other new things.

AI assessment note: “Definitely. Well, even another example on that is on Apex”

Q the world, it's such a large market that you're seeing the unbundling of it into, like, such verticals. I met the other day a medical, real-world medical data provider to them, where basically they have surgeons that kind of, I don't know, have video cameras on and they record all the real-world data. Do we see the mass unbundling of the data providing market? Is that how it plays out?

A It's interesting. We're, we're doing a ton of data collection in the physical world as well, especially across skilled domains where you have electricians and mechanics and scientists strapping cameras to their head to record things. I think that there's always going to be some degree of, uh, value in some of these like niche vendors that are able to go really deep in a Specific vertical. But what we're finding is that there's enormous value to aggregation and economies of scale, and that when we have this talent network of over five million people that are able to refer their friends, it's just so much easier for us to find the marginal doctor because we have that enormous talent network that can refer us to their friends. And even more importantly, that the kind of data shapes that we would build for a lawyer Are often very similar to the kinds of data shapes that we would build for a doctor. And so all of the tooling that we build is very, very cross applicable. And that's the way that most labs have been scaling out their data quite horizontally. And so for that reason, we are finding that the labs tend to prefer partnering with a very horizontally capable vendor that is able to flex across all of the different Verticals and scale extremely quickly rather than working with a hundred different vendors that they have to train for the same data shape and a hundred different do…

AI assessment note: “labs tend to prefer partnering with a very horizontally capable vendor”

Q What tasks are super high value? Is it like the medical, the financial modeling style?

A It corresponds extremely closely to economic value. So think if you go through the top five domains that we serve, it would be software engineering, it would be finance, medicine, law, consulting, et cetera. And the super long horizon tasks within those. And so I think we're moving away from the paradigm of how do we get a investment banker to prepare a financial model and moving towards the paradigm of how do we get a banker that can talk with five different colleagues and wait to hear back their responses and prepare, uh, an entire slide deck with a deliverable that includes the financial model, the analysis in, A multi-week long project. Those are the kinds of tasks that we need to be building to push the frontier of research and evaluation so that those are the capabilities that people are able to use in the models in six to 12 months.

AI assessment note: “software engineering, it would be finance, medicine, law, consulting, et cetera. And the super long horizon”

Q If we take that extrapolated further, how effectively can we build Slack internally, agent led entirely? That would very much concur with the SaaS is dead, because if you're a large company needing Maybe small customizations, integrations, say you're a real estate company and you need very specific integrations to pricing providers, you'd build your own.

A I generally agree. I think that the caveat is when those companies have network effects, there's probably a significant moat that isn't being priced in fully. Uh, for example, Salesforce has tons of companies that are building integrations on top of their platform that creates this Uh, almost marketplace and network effect around it, or Slack has Slack connect, right? And I think even CART is another great example of, right, this whole network effect of the people that use it and want to use the same platform across all of their companies. Um, I think that the companies that have network effects will be able to in some ways generate more value because they can iterate 10 times faster while leveraging those network effects to create more value for their customers and therefore Um, build more valuable products, charge more money, et cetera, and increase revenue. The companies that don't have network effects are going to struggle very significantly because then there's not really a defensible moat in the pure software associated with the products that they build. And so to me, that is the litmus test that determines whether this company is going to become worthless or whether this company is going to gain dramatic value from their ability to 10 X product velocity.

AI assessment note: “I generally agree. I think that the caveat is when those companies have network effects”

Q kind of say that they knew scale was shit for, for a while. Um, I'm, I'm British and very direct, which is quite anti-British to be honest. But they say that they kind of all knew it for a while, and that actually it, it wasn't a surprise seeing now that other people think that they're shit too. Did everyone kind of know that they weren't a great quality provider?

A I think people broadly knew. Um, I, I think Alex was phenomenal at so many things and, um, distribution and sales. Um, but in some ways, scale lost the focus on product, uh, on scaling quality. Um, and that was one of the largest challenges of the business. Um, but actually if I had to choose the most important thing, it would be the internal link to quality, which is that having phenomenal people that you treat incredibly well is the most important thing in this market and getting those people to refer all of their friends and, and actually help to improve the frontier of models. And so I think that Marcore started out really with this obsession on phenomenally talented people. Uh, like our average marketplace pay rate is 95 dollars an hour, to put that in frame of reference, whereas scale and search generally pay about 30 dollars an hour. Um, and, and so it's just a radically different approach to the way that we think about what kinds of capabilities we want models to achieve and how we want to Treat the people, um, that ultimately help to achieve those capabilities.

AI assessment note: “I think people broadly knew.”

Q dollars, and I was like, great, I'm out of here. Like, this is not, and I, I hated Laura as well. Um, and so, yes. Dude, I wanted to start. So when I had Edwin on the show, he said that everyone in the space was essentially a body shop. Direct quote. Um, is that a fair summarization of the space? And how would you kind of respond to that?

A I don't think it's fair at all. I mean, we operate as close research partners to all of our customers and helping them to mobilize some of the Highest caliber people in the world to push the frontier of model capabilities. But I think so much of our insight on the market, uh, is radically different and understanding how important high caliber people are rather than leaving them out of the narrative. And I'll give the backstory of sort of how we really got involved in the market in the first place, which is that scale AI came to us, uh, and they used our platform to hire thousands of people. And we realized that There was this enormous transition underway, moving away from this crowdsourcing paradigm that scale and surge pioneered of how do you get low and medium skilled people that write barely grammatically correct sentences for early LLMs and very quickly moving towards this sourcing and vetting paradigm of how do you find the Goldman bankers, the McKinsey analysts, the FAANG software engineers, uh, the top doctors and lawyers that can work directly with researchers to help them build The highest complexity data on earth and understand what that data is, because when we were dealing with undergrad level math problems, it meant that the researchers could easily look at the math problem and understand why the model was making a mistake. But when we're dealing with the, you know…

AI assessment note: “I don't think it's fair at all. I mean, we operate as close research partners”

Q Speaking of the world's most valuable company, you're seeing this concentration of value towards, like, the top eight names more than ever before. 84% of the year-to-date rally was driven by the top 10 names. Do you worry about the concentration of value to such a small number of players?

A Maybe to some extent. I, Definitely worry about how do we smooth out the benefits to society? Like, how do we ensure that every enterprise and every individual is able to reap the full benefits of AI rather than just like a handful of people in San Francisco? But ultimately, I also think that there is some natural dynamic associated with capital allocation where it is going to be more valuable to give the compute to A anthropic where they have the marginal demand and can use that right away versus a less successful company that might not be able to create the most value with that. So I think that it's probably good from a capital allocation and efficiency standpoint, so long as we are able to manage the societal implications of increasing inequality.

AI assessment note: “Maybe to some extent. I, Definitely worry about how do we smooth out”

Q Totally get you there. The other element is margin, and the margins are pretty terrible in a lot of cases, especially when you take into account free user giveaways, which there's a lot of. Should we give a shit about margin structures given how early we are in the cycle? Or yes, we should. It's always fundamental.

A I think that the answer is yes. Like, both of those matter, and it's very contextual. On one hand, I am a huge believer in capital efficiency. Like, we have very positive growth in net margins, unlike most AI companies. But on the other hand, like, I also see the case that if you're able to distill models and make them an order of magnitude more efficient in 12 months, then it could make sense to Uh, run really aggressive, uh, margins on, on serving models. It, it really comes down to the stickiness and whether those subsidies today are driving, like, large LTVs that make sense long term. Uh, but, but I think the case where I would be hesitant is when there is very competitive markets with low switching costs so that people are pumping hundreds of millions in subsidies Uh, maybe billions in subsidies, uh, and then all of a sudden they switch over, the customers are switching over to a competitor if those subsidies dry up.

AI assessment note: “I think that the answer is yes. Like, both of those matter”

Q Totally get you there. The other element is margin, and the margins are pretty terrible in a lot of cases, especially when you take into account free user giveaways, which there's a lot of. Should we give a shit about margin structures given how early we are in the cycle? Or yes, we should. It's always fundamental.

A I think that the answer is yes. Like, both of those matter, and it's very contextual. On one hand, I am a huge believer in capital efficiency. Like, we have very positive growth in net margins, unlike most AI companies. But on the other hand, like, I also see the case that if you're able to distill models and make them an order of magnitude more efficient in 12 months, then it could make sense to Uh, run really aggressive, uh, margins on, on serving models. It, it really comes down to the stickiness and whether those subsidies today are driving, like, large LTVs that make sense long term. Uh, but, but I think the case where I would be hesitant is when there is very competitive markets with low switching costs so that people are pumping hundreds of millions in subsidies Uh, maybe billions in subsidies, uh, and then all of a sudden they switch over, the customers are switching over to a competitor if those subsidies dry up.

AI assessment note: “I think that the answer is yes. Like, both of those matter, and it's very contextual.”

Q What new role will we have in five years that does not exist today?

A One of the largest things that people underestimate both in the context of AI labs as well as within the enterprise is how significant of a job category it is going to be to train agents. Like what we're seeing is that all knowledge work is converging on training agents because it is structurally more efficient to do something once. Instead of having a customer support representative that is redundantly responding to hundreds of tickets, they're going to train an agent how to do that once. Instead of having A lawyer that is redundantly doing dozens of similar red lines on commercial contracts, they're going to train an agent how to automate that. And even probably when you're playing around with Claude, you see that there's so many repetitive workflows of how you prepare for a meeting or draft emails or whatever it is where it's just much more efficient for you to train the agent how to do that activity so that you can amortize that over the entire useful life cycle rather than doing it redundantly yourself. And so I think that there's going to be this enormous paradigm shift as agents enter the workforce and everyone begins to manage them.

AI assessment note: “how significant of a job category it is going to be to train agents”

Q With the greatest of respects, does that not change so quickly? You know, when you spoke, when you saw Andre Capathi talk about how he uses coding agents, it was like, oh, I use it for 20% of the work, and then it's like, oh, it does 80%, and I do the final 20% within a six month period.

A Definitely. Well, even another example on that is on Apex, the Frontier model right now is at about 40%, and 12 months ago the Frontier model was a one, which was scoring one percent. And so that's been the progress of the last 12 months, and obviously we expect it to continue and be fairly significant. Um, but I think that the key thing is that everyone underestimates the elasticity for demand and increased productivity in the economy. Like, ultimately, over the last 250 years, we've increased productivity by 25 X, equivalent to automating about 96% of someone's job. And during every technology revolution, ranging from the agricultural revolution to the industrial revolution to the computer revolution, people feared that there would be this enormous job displacement because of the lump of labor fallacy, where people assume that there was a fixed amount of things that had to be done. And when we made people more productive, that would all of a sudden mean that there were fewer jobs. Yet, 250 years later, there's more jobs than ever before. And, and it's because we have No shortage of problems to solve as a society, right? We still need to solve climate change and cure cancer and do all of these other new things.

AI assessment note: “Definitely. Well, even another example on that is on Apex”

Q Can I ask you, when we think about that enterprise adoption, I think one of the biggest problems that we have is data structures and data cleanliness. Um, I interviewed a guest the other day and they said, we'll have like data cleaner as one of the most important jobs in the next five years. Um, is data structure and data cleanliness the biggest barrier to enterprise adoption?

A Well, I agree in part. I think that certainly the models need to have access to data to perform their jobs effectively, but the caveat is that they'll be able to clean the data themselves fairly effectively as reasoning capabilities go up. The thing that humans will need to contribute to is all of the tacit knowledge within the organization that isn't written down, because I've found that when I try to get agents to do all of these workflows Um, throughout Mercure, there's just an enormous amount of context that lives in people's heads that the agents need to have access to, to perform effectively. Um, and so much of that is going to be the new job of employees of how do we codify all of this knowledge? How do we train agents so that they're able to perform these tasks effectively across every function in the organization?

AI assessment note: “Well, I agree in part. I think that certainly the models need to have access”

Q The next 12 months will be dramatically better for infrastructure companies upstream of anthropic and open AI than for application layer companies downstream of them. This was your tweet. Why do you believe that?

A The reason I believe that is that the application layer company's businesses are not far removed from the foundation model company's businesses. Like it is not a far leap for Claude Cowork to add capabilities across medical and legal. Obviously they did it with software engineering and do that, can do that across finance. And so I feel like building defensibility in the software layer On top of the models is going to be incredibly difficult. Whereas on the other side of things and the infrastructure side, it feels like there are meaningful moats that are getting built. Um, like we're compounding enormous network effects in the business and a pretty significant data moat as we build out the inventory for our customers. Um, compute companies obviously are able to, um, build moats through these very long R and D cycles. Um, and so I think that There are going to be high margins that get achieved at the infrastructure layer and, uh, sort of sustainable, profitable businesses in a way that it's less immediately clear at the application layer.

AI assessment note: “The reason I believe that is that the application layer company's businesses are not”

Q You mentioned adding security as a third or seventh pillar there. We've seen so many hacks. It's almost become normalized, as awful as that sounds. Are we about to enter a golden age of cyber, given the new threats awakened by AI?

A I think so. I mean, we're even seeing this on the customer side where our customers obviously are very focused on how do we improve the model cyber defensive capabilities so that we can have the best AI security engineer that is able to defend every enterprise from all of these attacks. Um, because in our, our incident, it was the attacker that used a swarm of coding agents to help get access to the system as is happening in a lot of these. Um, and so I think there's going to be An enormous boom in AI security engineering tools and various forms of defense that are able to help protect companies against all of the increasing waves of cyber incidents that are just getting started.

AI assessment note: “I think so. I mean, we're even seeing this on the customer side”

Q What new role will we have in five years that does not exist today?

A One of the largest things that people underestimate both in the context of AI labs as well as within the enterprise is how significant of a job category it is going to be to train agents. Like what we're seeing is that all knowledge work is converging on training agents because it is structurally more efficient to do something once. Instead of having a customer support representative that is redundantly responding to hundreds of tickets, they're going to train an agent how to do that once. Instead of having A lawyer that is redundantly doing dozens of similar red lines on commercial contracts, they're going to train an agent how to automate that. And even probably when you're playing around with Claude, you see that there's so many repetitive workflows of how you prepare for a meeting or draft emails or whatever it is where it's just much more efficient for you to train the agent how to do that activity so that you can amortize that over the entire useful life cycle rather than doing it redundantly yourself. And so I think that there's going to be this enormous paradigm shift as agents enter the workforce and everyone begins to manage them.

AI assessment note: “how significant of a job category it is going to be to train agents”

Q Culture challenges. Before the show, we said that, um, after the show with Adash, a couple of people thought that, like, nine, nine, six was the way that, like, McCore is run, and it's like clock in, clock out. Um, Why is that not true? And how do you think about that?

A So the reason it's not true is that we've never mandated hours at the company. And obviously I work extremely hard. A lot, Adarsh works extremely hard. We work from when we wake up until we sleep pretty much all the time, aside from like maybe working out, but I'm still working, thinking about work during that time. And most of our leadership team of course does as well. But at the same time, majority of my leadership team has kids and we want them to be able to like Go home and see their families and all of that. And so I think that it's some combination of knowing that building a legendary company requires immense dedication to the mission of the business, while also recognizing that we need to ensure that it's a sustainable environment for the best people in the world to do their life's work.

AI assessment note: “the reason it's not true is that we've never mandated hours at the company.”

Q You mentioned adding security as a third or seventh pillar there. We've seen so many hacks. It's almost become normalized, as awful as that sounds. Are we about to enter a golden age of cyber, given the new threats awakened by AI?

A I think so. I mean, we're even seeing this on the customer side where our customers obviously are very focused on how do we improve the model cyber defensive capabilities so that we can have the best AI security engineer that is able to defend every enterprise from all of these attacks. Um, because in our, our incident, it was the attacker that used a swarm of coding agents to help get access to the system as is happening in a lot of these. Um, and so I think there's going to be An enormous boom in AI security engineering tools and various forms of defense that are able to help protect companies against all of the increasing waves of cyber incidents that are just getting started.

AI assessment note: “I think so. I mean, we're even seeing this on the customer side”

Q One thing that powers obviously the agents that we use is the tokens that power them. And I thought the whole point was that we have increased token efficiency and token costs come down. Um, token costs are rising for everyone. Help me understand how you see token costs changing in the next six to 18 months. And why that is?

A Well, it's a fascinating case study in Javon's paradox, similar to what we were talking about in the context of making humans more efficient, leading to more jobs, right? When we make models improve by 10 X year over year, that has just been causing the total consumption of the models to go up and up and up as the cost per performance go down. Um, I think insofar as how it's going to develop is that This trend is going to continue very, very significantly before we start seeing any leveling off of token consumption within the enterprise. Like right now, we're spending more on tokens for our internal agents than we are on employee headcount. And I think most businesses are going to look like that.

AI assessment note: “Well, it's a fascinating case study in Javon's paradox”

Q the majority of startups today, especially on the West Coast, use frontier models to see where they can go and how far they can push them, and then they use open source, often Chinese models, To get as close to that as possible at a much better cost basis. In which case, OpenAI and Anthropik are inherently challenged by that much more cost efficient open source model. Right or wrong?

A I think both are true. Like there's going to be many, many orders of magnitude more demand in five years than there are today. Maybe four or five orders of magnitude more demand. But there's also going to be increased competition with people just distilling and Having fine-tuned open source models that accomplished their workflows. Ultimately, I think OpenAI and Anthropic are incredible investments, and it seems like they're starting to be consensus around that, uh, in a way that there wasn't just a couple of years ago, but at the same time, I think that majority of inference in five years is going to be using a open source or custom fine-tuned or distilled model, not using a frontier model.

AI assessment note: “I think both are true. Like there's going to be many, many orders of magnitude”

Q Do you buy the sovereignty argument of we need sovereign models because we don't want our data going to US or China or wherever that is?

A Maybe in some cases, like there is value in localization. And I'll give an example, which is that oftentimes labs will come to us and say they need their models, not just to be good at American law, but also to be good at British law or good at French law or whatever the jurisdiction is in the world. And I think that that is going to be an important last mile in making the models useful in whatever jurisdiction that they're operating in. That said, the labs are just going to hire 10,000 people in France to teach the models how to be better at French law. And I don't think that there's so much that others are going to be able to do to stop that because The transfer learning capabilities from all of the other domains that they're focusing on are just so powerful.

AI assessment note: “Maybe in some cases, like there is value in localization.”

Q In 10 years, do the models still need humans to help train them?

A I very much believe so, and the reason is that the question comes down to when we'll have super intelligence. Once we have super intelligence and models are better than humans at everything, then, of course, that means that humans won't be able to contribute to models and measure that frontier that models aren't able to do, but I still think it's a very long road. Like, these models have gold medals in Olympiad math, and they're better than the best PhD at reasoning, but they can't Draft an email for me. They can't, uh, schedule a meeting. They can't do so many of the basic things of just using a handful of tools to do a task that takes me a few hours. And that entire road to automating the entire economy and building agents for everything is paved with humans creating evals for all of those workflows.

AI assessment note: “I very much believe so, and the reason is that the question comes down”