The Exchanges

Q He's citing Yahshua Bengio. He's, he did his homework. You told him, this was right before GPT-V came out, that GPT-V is smarter than us in almost every way. Uh, I, I thought that that was the definition of AGI. Does, is that, isn't that AGI? And, and if not, has the term become somewhat meaningless?

A These models are clearly extremely smart on a sort of raw horsepower basis. You know, there's all this stuff on the last couple of days about GPT-FW as an IQ of 147 or 100 and 44 or 151 or whatever it is. It's like, you know, depending on whose test it's like, It's some high number, and you have, like, a lot of experts in their field saying it can do these amazing things, and it's, like, contributing, it's making it more effective, you have the GDP value, things we talked about. One thing you don't have is The ability for the model to not be able to do something today, realize it can't, go off and figure out how to learn to get good at that thing, learn to understand it, and when you come back the next day, it gets it right. And that kind of continuous learning, like, Toddlers can do it. It does seem to me like an important part of what we need to build. Now, can you have something that most people would consider an AGI without that? I would say clear. I mean, there's a lot of people that would say we're at AGI with our current models. Um, I think almost everyone would agree that if we were at the current level of intelligence and had that other thing, it would clearly be very AGI like, um, but maybe most of the world will say, Okay, fine, even without that, like, it's doing most knowledge tasks that matter, um, smarter than us, and most, most of us in most ways, we're at AGI. …

AI assessment note: “What I think this means is that the term... is very underdefined.”

Q AI lab leaders, um, like Dario from Anthropic, something that he said is like, you know, we are starting to, um, mirror open source in a way where that customization is possible and the bringing the data in is possible with our, our off-the-shelf models, meaning clawed at Anthropic. So why go open source and customize? Is it just a greater degree of control or What do you get there?

A Well, I mean, when we say greater degree of control, we're actually saying a greater ability to improve performance. It's, it's not control for control sake in that sense. It's really about the performance. Let's go back to chat concierge for, for a moment, right? Like terms of it, what are the different, uh, steps that like something like chat concierge, uh, uh, goes through, right? Um, even simple interactions, you want to confirm, uh, the needs with the user. You then want to simulate The plan that you're going to work against. You want to validate that plan, make sure that it's, it's right, etc. Each one of these steps, ah, is a combination of both the, the data that is required to improve the execution of the, of the step, but also, ah, kind of the reasoning capability and the interaction with the user. How does the business want to execute it? You may have certain things that you say, if these happen, I want With this to be, this to go to my, to my human in the loop, ah, in that thing. In other, some other people may have different things. So there's a fair bit of customization both of the UX and of the, ah, performance itself, right? In this context, ah, what are things we care, one of the things we care about most is speed or latency, right? You know this from your favorite home assistant that you use. If it takes three seconds to respond to you, it's way less satisfyin…

AI assessment note: “when we say greater degree of control, we're actually saying a greater ability to improve performance.”

Q So it's important to digest what you just said, which is that there are similarities between The human brain and artificial intelligence, the way the models work today. Do you think one day we'll be able to merge with AI?

A I, I think you have to kind of be specific about what that means, but yeah, I mean, one of the things that I'm preoccupied with is, is, um, The, the binding problem, which the, so all of your experience, I mean, it's important to realize that kind of all of this, like everything you see and hear and feel and think, like, all of this is brain activity. Like that's, at the end of the day, that's all it really is. Like the brain is the thing that is you, it's the source of, of, ah, your entire experience of the universe. But within that, it's composed of billions of neurons that are distributed over a large area and spread out in time. But you don't experience the activity of these neurons independently. You experience a unified moment. And even within your brain, you've got two hemispheres, each of which is processing one half of your experience. So the left hemisphere is processing the right hemifield and vice versa. But you don't experience two hemifields. You experience a single visual field. And so how does that happen? There's some physics that we don't understand yet. And if you could understand that, then you can imagine adding, you can imagine building, say, a conscious machine, You could add a third hemisphere. You could connect this over the network. This would, in some very fundamental sense, allow you to redraw the border around a brain and change the, like, where you…

AI assessment note: “You could add a third hemisphere. You could connect this over the network.”

Q I imagine for someone like yourself, like the short term numbers, you know, might look better. Like, oh, look at our division. We're highly profitable. But in, in the way that you're talking about the fact that we have this fast moving economy, it's just not a winning strategy, right? To just say, all right, let's just automate and be more profitable. As opposed to keep up with what's next.

A I mean, what I tell my team is if I look at, I mean, our portfolio is pretty wide. Um, but if you sort of go to one solution area or within one solution area, a single product, you can argue we could get to teams of a thousand or 2000 even that are focused on one product area. And I talk to these teams and I say, Hey, listen, in today's world with the tech that's out there, a 200% startup can actually operate at the throughput of what we operate with 2000 or a thousand. Where we sort of maintain and sustain our advantage is if we take our 2000 and now operate it at the pace of 8000, now that's going to be hard for somebody to go do with 50 or a hundred people. So for us, it's as much about, hey, keeping the staff, maybe even growing it so the pace of innovation is just going to be hard to match with anybody that comes up now with some new tools, generative coding, generative development, and say, hey, now we've got this app and that app, but we need to be able to keep that same distance. Um, from an innovation throughput perspective, if that makes sense.

AI assessment note: “So for us, it's as much about, hey, keeping the staff, maybe even growing it”

Q and they'll figure it all out and you don't really need, uh, much structure underneath. And what you're saying is if I'm hearing you, right, basically like, yeah, you can get an ROI from a large language model. Uh, but there needs to be a structure built in. Otherwise it's going to, I guess, hallucinate and get things wrong and not do what you want. Is that the right read?

A Yeah, I think that's the right read. That's one, right? And I think we need to be pragmatic because That end state of everything is going to be run by agents, billions of agents. I mean, I think it's just, it's an unrealistic, hyped up, uh, end state right now that we're not ready to get to, and I think there's a lot of other people that are now saying the same things as well, that, hey, it's not going to be a year of agents, it's going to be a decade of agents, because it's going to take a little bit for us to sort of go figure out the accuracy, the trust, and the reliability on it, and we believe in the same thing, but we have to sort of go through those steps, and one of the things I like to sort of, Make fun of ourselves in some ways is, you know, we're not going out there and announcing like thousands of agents. And I think somebody came to me and said, well, you guys just only announced this many agents. You could have sent an email. You didn't need a conference for it. Um, but this is where our, our point of view is, listen, you want to do the stuff that's pragmatic, that's grounded, that's creating value and not just the hype that sits on a shelf somewhere that nobody really gets any value from.

AI assessment note: “Yeah, I think that's the right read. That's one, right?”

Q able to do this? Like, is that what, is that what's really needed is, I mean, of course some work on, on, you know, the company end, but, um, little better models, like sort of, I don't know, maybe they can solve the captcha so they can log into the system and then handle these requests. I mean, where do you think that that leap is going to be taken?

A I think that, um, better models will certainly be able to do a lot with that. I actually think what's really going to make a difference more likely than computer use models, computer using models is just the improvements in coding models. And that's a little bit basing the future on the immediate past. Coding models have gotten gotten so good recently, and they can make developers so productive that I think that that is going to unlock For a lot of internal builders and a lot of company builders, it's just going to unlock this potential to create, um, so many connections. Along with sort of tools that have agentic AI built in that can interrogate, you know, the API space of systems and create experiences. So I think we're going to see, um, what could be semi-daunting projects for people to take on just to automate another three percent of customer service tickets. Suddenly become things where the ROI is a lot clearer, and they'll be like, yeah, I could do that. I could, I could knock that up in cursor. I could get, you know, opening our codex to do that with me, and it wouldn't take me too long.

AI assessment note: “what's really going to make a difference... is just the improvements in coding models.”

Q speaking confidently about like, well, this isn't what we actually see in the real training run. I mean, how do you know? Right, how do you know that the models aren't faking out anthropic researchers or playing this long run, long con against you, uh, and eventually will, you know, when they get powerful enough and have enough compute, will do the real bad thing that they want to do?

A Yeah, it's a good question. And of course we can't be a hundred percent sure. I think the, the evidence we have today is, uh, you know, fortunately even these really bad models that we've been discussing here that try to fake alignment are pretty bad at it, right? So it's, it's pretty easy to discover the ways in which, you know, this kind of reward hacking has made models very misaligned and it wouldn't, you know, it would be, it would be very, um, It would be very unlikely to miss, miss any of these signals, but you're right. There could be much subtler, you know, more nuanced changes in behavior that, that happened as a result of a real hacking and real production runs that weren't sabotaging our research or any of these things that are really big headline results that are very sort of striking, but maybe more, yeah, more subtle impacts on behavior that, that we, we weren't, you know, we haven't disentangled yet.

AI assessment note: “we can't be a hundred percent sure. I think the, the evidence we have today”

Q that is to tell the model that the bad things it's doing aren't actually that bad, but you're still left with the model doing bad things in the first place. So, uh, it just leaves me in a place of, I don't want to say fear, but like deep concern about like, whether or not AI models can learn to behave in a proper way. What do you think, Monty?

A Yeah, I think when you put it like that, it's definitely sounds concerning. I think, I think what I would say is there's a, you know, I like to think of it as multiple lines of defense against misalignment, right? And so the first line of defense is obviously we need to do our utmost to make sure we never accidentally train models to do bad things in the first place. And so that looks like making sure the tasks can't be cheated on in this way, making sure that we have really good Systems for monitoring what the model is doing and detecting when they do cheat. And as we said in, in this, the research we did here, those mitigations work extremely well, right? They, they remove all of the problems. There's no hacking. There's no misalignment because it's pretty easy to tell when the model's doing, doing these hacks. But like Evan said, we may not always be able to rely on that, or at least it would be nice to have something that would, we could kind of give us some confidence that even if we didn't do that perfectly, even if there were some situations where We accidentally gave the model some reward for, you know, doing not, not exactly what we wanted it to do. It would be nice if we could kind of ring fence that bad thing, right? And sort of say, okay, maybe it learns to cheat a little bit in this situation. It would be okay if that's all it learned. What we're really worried abo…

AI assessment note: “I like to think of it as multiple lines of defense against misalignment”

Q the models and in some cases amongst the hardware because of the TPU thing. And, and when you see commoditization happens, the economics of AI shift. So I'm curious what you think about it. And I I'm curious what you think, you know, if you agree with my premise that we are starting to see this commoditization, um, what do you think the economic impacts of the commoditization really are?

A Yeah, I would frame it's like the taking what you're saying and framing it slightly differently is that, um, If you believe all of that to be true, that you're, that you're laying out there, maybe investors are just looking at it. Like, look at the end of the day, then if that's the case, who has the sort of best position and best business model for all of this, you know, future going forward. And I think it would be hard to argue against Google, just given that they own not only Google cloud, you know, on which a lot of models and everything are running, but they, the TPUs as you're noting, obviously they have the search business and Waymo and all of these other things. Whereas the other companies, if, if AI is starting to, um, commoditize in some way. You know, Nvidia, their business potentially is impacted in a good way in some cases, but for the most part, it's probably going to be looked at as like, well, the arms race is sort of slowing down a bit, and that's going to be bad news for Nvidia at the highest level, because you won't need to necessarily buy, you know, the absolute top of the line. Um, and, or on the flip side, as we're talking about, if Google's able

AI assessment note: “who has the sort of best position and best business model for all of this”

Q robot like I saw, uh, today walking around a factory floor looking for spills, uh, and probably very efficiently through computer vision, making a notice that, hey, there's a spill and maybe saving, you know, a lot of time or a potential, like, Uh, batch of product. Uh, but previously maybe that was something that a person did. So how do you think that this will impact labor moving forward?

A So I'm an optimist. Uh, and I read with great interest and agreement that all we're going to experience is growth. Uh, economic growth for sure. The predictions are one percent percentage point increase to global GDP growth in the coming years. Uh, and I look at You know, research like the World Economic Forum, who did the Future Work Study at the beginning of this year, and they concluded that a hundred and seventy million new jobs will be created by 2030, ok? Uh, ninety-two million of existing jobs will be displaced. My mathematics tell me that's seven to eight million of net new jobs, incremental jobs, and employment by 2030, ok? And what we're experiencing is, in all senses, Economic growth. So you think about growth, you think about some of that research and you think about some of the previous general purpose technology shifts that we've been through, albeit this one is different. All of them have resulted in growth and more labor and more employment and more business models. And I think the same is absolutely true here. When I put that into the real world, so put that to one side for a second, every customer that we deal with, whether it's in North America, whether it's in Europe, Whether it's in Asia and the industries that we serve to a greater or lesser extent are dealing with labor shortage today. And that's driven by aging workforces. It's also driven by re-industri…

AI assessment note: “they concluded that a hundred and seventy million new jobs will be created by 2030”

Q Is that a preview for what's going to happen with us?

A I haven't studied it in detail, but directionally, from what I understand, I think so. And I think, again, we come back to an environment where I genuinely believe there's going to be employment growth over the long term. The nature of the roles will fundamentally change, no question. And I think You know, that's a responsibility that falls onto, you know, governments, uh, higher education to be thinking about what's the shape Of the labor force to come, and how do, what does one plan for it? The nature of the jobs will be different, and there's no question, World Economic Forum said ninety-two million jobs will be displaced, and those individuals in those roles will need to be thinking about how they enhance their skills and how they develop to take on new roles.

AI assessment note: “directionally, from what I understand, I think so.”

Q So as someone who works very close to industry, Uh, you obviously are keenly aware of the need for power and how much these large language models are planning to take out of the ecosystem. What do you think the future is going to be on that front?

A Well, I think the future is going to be on two fronts. Uh, obviously increased power generation is one way to solve the energy crunch. The other way to solve the energy crunch is by getting more at what already exists. And I don't know if you were listening to Sabine from Siemens, the CEO of Grid Software at Siemens, but the stat that caught my eye was that there's a hundred and fifty billion output a year that's lost in the US as a result of outages, right? So if we deal with those outages, and I don't know what the power consumption issues are there, but we deal with those outages, with those types of numbers, something's telling me that we've got more energy available. So how do we look at existing infrastructure and squeeze more out of existing infrastructure? How do we avoid downtime? How can we get the throughput increased on existing infrastructure and then overlay that with new power generation? And there's all sorts of ways clearly you can do that. Um, you know, introduction of new nuclear technology is obviously one that tends to be more carbon efficient. There's clearly carbon intense hydrocarbons that can be used. Uh, there's a proliferation of different things. Uh, I also feel the renewables needs more attention. Some of the technological developments with solar in particular, predominantly coming out of China, Make the cost of generating wattage for that technolog…

AI assessment note: “Well, I think the future is going to be on two fronts.”

Q Okay, so you are a believer then in this vision, and it sort of leads me to a natural follow-up, which is do the chatbots sort of become the storefronts of tomorrow? Are they going to be the place where we do most of our transacting? How does a chatbot fit into the picture of, you know, transacting online?

A Uh, it's, it's, the, the way we think about it, Alex, is, you know, before online commerce existed, um, you know, you walked into an AMC theater to buy a movie ticket. Then you decided to buy sometimes online. Then you sometimes decided to buy it because you clicked on an ad on a social media website through their app. And that doesn't mean that physical commerce went away. It just, they were like, The share shifted toward more channels in the way human beings actually interacted with it. And we think with the Gentic, you're just going to see another channel that's going to be very helpful in certain types of commerce, in certain segments of commerce, um, which doesn't mean that, you know, the existing channels are going to go away. I just think there's going to be a good amount of migration from how things work today to a portion of that is going to move to a Gentic. So like, that's, that's, that's, that's, that's the hypothesis we are working with.

AI assessment note: “you're just going to see another channel that's going to be very helpful”

Q the discovery process could look like, because this idea that, you know, we might provide the LLMs with lots of information about us. I think we're already there. I mean, we probably share more with LLMs than anything else. And we're being honest, some of us are in love with these things, not me, but some people. So how could this discovery process, uh, look as, as Visa gets involved?

A Um, I think the simplest way to think about it is You know, you use your cord to do a number of things in life, and a lot of those are frankly not used at LLMs, right? You walk down the street to like Pete's or Starbucks coffee, and you like to stay at the Ritz-Carlton versus like a Motel Six, and you typically like to fly United versus Spirit Airlines. Now, those are just things that while you're Agentec platform surface area, LLM, may have a lot of information about you. Those are just, like, personal preferences tied to your long-term spending preferences. And what we believe is, if the data is private, if it has full consumer consent, if it is totally tokenized and no raw data has actually moved anywhere, but can there be a bunch of signals that, you know, you, compared to like other spenders, That a agent or a company can sort of like consume based after you've like fully consented to it. And after consuming, say you say, I want to go to Tokyo. I don't want to live in the, I want to stay in the Ginza district in Tokyo for like a week. Um, you know, and give me hotel recommendations. Now, knowing that you like to stay at the Ritz-Carlton or knowing that you generally have a propensity to stay in, like a signal that tells that you have a propensity to actually stay in a Luxury hotel would be very helpful for a LLM to say, oh, you know what, here are the options which we thin…

AI assessment note: “a signal that tells that you have a propensity to actually stay in a Luxury hotel”

Q How long do you think it's going to be until you have that data? Um, I guess that's a way of asking like, how, when is this going to be real?

A Look, I think there's, I think you're seeing, starting to see versions of it become real, but it's again, very early days and these tend to be like, you know, very, very specific transactions. You're starting to see, you know, some of this go live in, we're already seeing some form of agenda transactions in different ways, shapes and forms with like various AI platforms. But again, it's very early. It's like at limited merchants. It's like, you know, in beta is largely in the United States. I I'd say in about six months or so, we'll start to see like some early statistically relevant enough data sets. Where we can start to like draw conclusions. That's like my best guess, but you know, it all remains to be seen how quickly some of these things get adopted.

AI assessment note: “I'd say in about six months or so, we'll start to see”

Q For them to serve. So what's your perspective on this?

A I, I, my guess is it's much simpler than that. These are very smart people who've been in rapidly scaling businesses. And you know, if you're, if you're Sal Maltman or Darryl Amodei or any of that, you're going, wow, I'm sitting on this, I'm sitting on this rocket, you know, this rocket ship and it's kind of taking off. I need people around me. Who understands scale, who can, who can, um, ship products, um, quickly, uh, scale them very quickly and understand how to operate in complex and very fast moving, uh, um, environments. And if you, if you basically take that as your, one of your list of requirements, um, or expectations, then of course people from companies like Meta, you know, feature high up on the, high up on the, on the list. So no, I, I, I wouldn't have thought it's, it's, I wouldn't have thought it's through the, that's my assumption at least, I wouldn't have thought it's through the prism that you've just described, which is one is engagement with commercial upside, the other one is engagement with sort of expensive engagement, because I would have thought at the moment what they're just racing to do, and they're clearly, they're clearly burning a lot of money in order, in pursuit of this objective, is just to, is just to expand and get people using these products. Um, and that in a sense is a bit of a playbook from From Mark Zuckerberg. I mean, you know, Mark, it…

AI assessment note: “my guess is it's much simpler than that. These are very smart people”

Q really good point here, which is that there is insatiable demand for the products, and you do have public company CEOs like Lisa Sue, uh, who, like, have to have some rigor behind the things they say, talking about these major numbers, uh, and even companies like Anthropic, which are losing a lot of money Planning to get profitable in a few years. So what's your read on this Gil?

A There's a lot to unpack there. Uh, and the framework that I do it with is to say both things are true. So AI is the most revolutionary technology that we've had in a really long time, whether it's back to the internet or back to the industrial revolution, we'll only know in retrospect, but clearly the tools are very powerful and are getting better. All you need to know to do to know in order to realize that is just to use them. May ask CHI GPT to do things for you that are hard, that you would ask other people to do, whether it's summarizing, writing, giving you advice, and you see that not only is it incredibly capable, but it's much better than it was a year ago, and it's much better than the year before that. So yes, there is insatiable demand for this product. That is true. That's, there's a lot of healthy behavior around that capability, and the healthy behavior are reasonable, thoughtful business leaders like the ones at Microsoft, Amazon, and Google that are making sound investments in growing the capacity to deliver AI. And the reason they can make sound investments is that they have all the customers. They have all the business customers, And by extension of the relationships with OpenAI Anthropic, they have all the consumer relationships as well. And so when they make investments, they're using cash on their balance sheet. They have tremendous cash flow to back it up.…

AI assessment note: “and the framework that I do it with is to say both things are true.”

Q Is it possible that something can be super intelligent, but not generally intelligent? Like, is it possible that maybe super intelligence happens without AGI because AGI is all about generality and what you're talking about is not?

A It's not possible, I don't think. I think they need to be general. Um, they need to transfer knowledge from one domain to another. They need to, um, you know, um, have generalist reasoning capabilities. But when you apply it, and you put it into production, and you let it have more autonomy to make decisions, or you let it generate, uh, arbitrary code to solve a particular problem, Or you let it write its own evals so that it can modify its own code and generate new prompts to generate new training data, to write new evals, to then iterate on its own performance. These capabilities, autonomy, goal setting, um, writing code, uh, modifying itself, you know, if you add to that then also a perfectly generalized model or, or, or sort of general purpose model, That's a very, very, very powerful system, which today I don't think anybody really knows how we would contain or align something like that. And so it's not to say that we should not do any one of those dimensions. It's just to outline a roadmap of capabilities which we're all working on, which add more risk, especially when they compound with one another and you combine them all together. And so, you know, my claim is that we should just approach this with caution, remembering That we don't want to bundle together all these capabilities so that there's a higher risk of a, you know, recursively self-improving exponential takeof…

AI assessment note: “It's not possible, I don't think. I think they need to be general.”

Q that's, that's really interesting. And you can imagine how many other ways that becomes applied. But then also, at that point, are they taking on the risk as well? And then, like, They're giving away free compute in order to get that longer term rev share, which just adds actually an amazing whole additional layer of risk to the business. Like, I wonder what, like, what could that look like?

A So I actually thought this was much cooler when I saw it for the first time. Um, and then when I think about it, it's like, all right, so let's say, let's just go crazy here. Let's say OpenAI develops medical super intelligence that enables pharma companies to do things they never could before, uh, at human scale. Somebody else is just gonna develop super intelligence as well, and they're gonna get into a price war. So how long can you say, I want a party of the profit of this drug that you're gonna develop with our technology? Where like somebody else can be like, well, here's the platform. Just pay us a licensing fee, right? That only, that pricing model only works if you're the only one that can offer that. I just don't think they're gonna be.

AI assessment note: “That pricing model only works if you're the only one that can offer that.”

Q So can you give me an example of what happens when all this goes right?

A Oh, um, I actually have a really interesting example that one of our customers gave, gave us a couple of days ago. Um, large global financial services institution, non-US based, um, and they have to, and they have transactions that happen around the world with their customers that are flagged as being suspicious. As a regulated bank and with global know your customer anti-fraud and anti-money laundering rules, they have to evaluate each one of those. So originally, and, and they've got, and it's a bank that's come to get through acquisitions. They have a business they bought in Geography A, and they're on different systems with different repositories, and then the systems they have could be securities trading systems, cash machine systems, and, and all these are not tightly integrated because they're all kind of have been separate over time. So you basically have this issue that's flagged, and you have to have a human go investigate it. You know, what were their credit card transactions? Did they happen to buy a plane ticket and let's go to this place? And it would take a thousand people a full-time job to basically on a daily basis, go through these types of issues. And now they've created an agent where they were basically when an incident comes up, they score it based on this agent going, looking, oh my gosh, there's been, I'm, I'm making this up, but there was a cash deposi…

AI assessment note: “I actually have a really interesting example that one of our customers gave”

Q have a hard time trying to understand, again, that the nonprofit has a stake in the for-profit worth a hundred thirty billion and will get more ownership as the for-profit becomes more valuable. It still felt very open AIE, corporate structure-y to me. Like, it didn't help necessarily clear up A lot of the, what's been vague about everything so far. Do you, do you understand what is going on?

A So I'll be the first to admit not fully. Um, why I think this is important. This is why I think it's the most important story is because Microsoft's ownership in open AI is finally cleared up. And yes, there's going to always be some weirdness with open AI when it comes to the nonprofit, uh, and the for-profit, the public benefit corporation. Uh, but up until this point, you had Microsoft in this position where it get, it was entitled to Percentage of the company. Um, and it also had access to all of its IP rights and it could potentially lose access to all that. If opening, I just said, Hey, we've hit AGI. So I think that to me is the biggest thing that needed sorted out. And until that became sorted out until this sort of conversion was complete and Microsoft was baked in as a traditional investor, the company couldn't move forward and now it's been settled and now it can move forward. So to me, that That is the biggest point, and of course, we know Microsoft has been holding it, holding it up, because it was trying to negotiate and strong arm OpenAI for a large percentage, uh, uh, even larger, I think, than it got of this public benefit corporation. That is the big deal to me.

AI assessment note: “So I'll be the first to admit not fully.”

Q Are you seeing a real increase in this type of content showing up on medium because of the scale?

A We saw like a 10 X increase in what people were trying to post on medium. But also at some level, it was kind of a non-issue because the tools we used to battle it are the same tools we were already using to, um, to battle spam and to filter out the filter, the good stuff into the network and leave the bad stuff out. So we got a huge increase, uh, in volume, but not in what the readers see. And I still, I'm like somewhat perplexed that that article made it to print because it was really, A nothing burger for us. It just, it was a little bit of extra work, but the same type of work we were already doing.

AI assessment note: “We saw like a 10 X increase in what people were trying to post”

Q know what these agents are gonna do. Like, uh, you spoke about that fast food chain, um, that, you know, this, again, idea of going to take action, um, is, I guess it's both thrilling, but also somewhat scary, um, Especially if you don't have the right governance in place. So can you just talk about like the state of governance around these, around these agents and why that's important?

A Well, I, I think All of us are working to catch up with the capabilities of the technology, and we believe that we have a responsibility to help elevate the industry's readiness to deploy agents in a way that they can be safe, secure, and productive. And, and the industry's race to innovate has focused more on production than safe and secure, and so we, we believe we have a very important role At helping companies get the balance right. So I spend many of my days every week talking to customers, and I mentioned CISOs, Chief Information Security Officers, and CIOs, um, about these challenges and what they see in their businesses. And one of the most common questions that I hear them addressing within their companies is the issue of data governance. And specifically, what data is an agent able to get access to? And this, this is similar to the problem of authorization I mentioned previously. When you authorize an agent, what is it going to have access to read? But here the concern is really twofold. One is, what data should it be able to see? The second is, what data should it be able to use? And one of the latent concerns people have as they're deploying, deploying third-party agents, is they want to know if that data is going to be used anywhere else. So is it going to be used just for my query to help me do my work, which is usually fine? Or is that data potentially going to b…

AI assessment note: “one of the most common questions that I hear them addressing... is the issue of data governance”

Q So how important then is the long-term research that is detached from the need to innovate right now?

A First, no research is detached. Research, again, as I mentioned, the best research is research that is motivated by either a need that you already know, or by exploring the art of the possible. And when you think about exploring the art of the possible, it's motivated by saying, well, no, if I manage to solve it, that is going to unlock things that are actually going to be meaningful for my business, for my products, for capabilities. So it's always connected. To your question, the importance of long-term research is greater than, is, is more than ever. And here's why we are actually, when you think about our job is really to drive breakthrough research that is going to be transformative, that could enable actually products and capabilities and experience and science and, uh, all societal challenges to actually be solved in a way that is materially better than we can do today. Now, Some of it is something you can actually innovate and find the kind of, ah, the shorter term research. A lot of it is really to find entirely new paradigms. To think about, I mean, think about the transformers that, you know, were developed by Google research back in 2017. It was a new paradigm that once done, it actually created a lot of the industry. Or thinking about some of the work we're doing on genomics or quantum. Quantum, of course, is a very long term as, as we know. So in many areas, actua…

AI assessment note: “the importance of long-term research is greater than, is, is more than ever.”

Q that the actual, the model went through all these different, uh, potential treatments that hadn't been tried yet, and actually found one that would work better than the ones that humans had uncovered. Um, obviously this technology, generative AI technology, is going to be applied in research all across the board. Do you anticipate that it's going to lessen the need for researchers, or are we going to have more?

A Well, we're going to need many more researchers in all disciplines. I mean, think about what's the role of a researcher. It's really to build on what we can and ask the right questions and, and, and build for the next one. Now, the only situation where you need less researchers is if you assume that we practically almost answered all the questions that we need to have. I don't think anybody here in the audience would think that, uh, we're only understanding tiny bit of what we need to understand. In fact, The opportunity that we have with the AI to empower researchers is going to give opportunity, not only for more researchers, but for each of them to ask bigger, bigger question, move faster on the research agenda, have better results. I mean, think about AlphaFold, which, you know, uh, my colleagues, uh, were recognized with Nobel Prize, uh, Demis and John. Um, I mean, we don't have less researchers working on proteins. We have actually have many more, right? Uh, but now they don't need to work on the, Protein folding problem. They're actually using it for bigger questions. With AI co-scientists, again, think about the fact that every grad student, every postdoc, have now their own research lab, which can help them with literature search, looking at hypotheses, so now they are going to ask bigger questions. They are going to ask the kind of questions that previously we expecte…

AI assessment note: “Well, we're going to need many more researchers in all disciplines.”

Q Um, and today, you can't, it's hard for me to even conceptualize that that would get funding, because a VC might just say, why wouldn't I just do this and chat to my team?

A We've seen a lot. We've seen nutritionists, we've seen a bunch of different things that have come out, you know, so I would say there's buckets of, do you need a discreet application or you don't need a discreet application? Certain things, for a bunch of different reasons, including regulatory and compliance in, in areas like health tech, you need a discreet application. But certain things, including general things like, you know, I'm eating this piece of salmon. How many calories does it have? Could you count it in your calories? Is something ChatGPT is great for. So we've found, um, sadly, that, you know, we haven't seen this wave of startups that we believe are sustainable. So there's actually been a handful of startups that are Wrappers on ChatGPT that are maybe a little bit better at travel. They might be a little bit better at being your math tutor, but they're not that step function different. Um, and even if you go back to, you know, the areas of search. If you remember, there was search, and then people said, oh, there could be vertical search where we get really good at something. So, obviously, Indeed is a very large company that was vertical search for jobs. Kayak was a very big multi-billion dollar outcome. That was vertical search for travel. And, you know, you're able to break down that landscape and then think about where, where that goes, because with a more n…

AI assessment note: “we haven't seen this wave of startups that we believe are sustainable”

Q or how to improve the relationship. Uh, if they came to you and said, we have a natural language fitness coach. If they came to you and said, we have, uh, you upload, uh, photos or videos of your, of your soccer practice, and we'll talk to you about positioning and, and form. Each one of those ideas, to me, sounds like they would be, like, billion dollar ideas, right?

A Yeah, very financeable. Very, if, if, maybe not billion dollar ideas, we'll see where that goes, but very financeable. If you think about life coaches, fitness coaches, Sports coaches. Anything where you have a tremendous amount of knowledge, and you could take that knowledge and make it very specific to somebody, which, you know, again, going back to Harvey's, is not that different, right? Law is a huge, huge, ah, pool of knowledge that you have put certain rules around it. You know, historically, they just thought that was a thought exercise in rules. Today, we could call it LLM. And then that produces Better, faster, cheaper results of how to, how to be more efficient in your life or job.

AI assessment note: “maybe not billion dollar ideas, we'll see where that goes, but very financeable.”

Q In order to justify this valuation. So let's just start broad, uh, as we begin the second half here. Um, are we marching towards technology companies like OpenAI, uh, like Anthropic, uh, basically trying to automate all work, all white collar work, and if they're successful, what happens?

A Well, I mean, I would say automating all work, right? Because if you think about some of the robotics things that are happening now, Some factory automation things that are happening now. It's both blue collar and white collar. I think maybe differently than any kind of automation going back to, you know, farming where, you know, there's bulldozers and there's, um, steam engines that are automating blue collar work. You know, this has been very different. Uh, and so I, I do think in talking to the bankers and the lawyers who usually hire a whole lot of folks or, you know, entry level consulting firms, BPO firms, they are Pausing or taking a slower approach or a more thoughtful and cautious approach to how they fill in the bottom of the pyramid. Uh, and that makes them, you know, rethink their, their business. Um, you know, I do believe that they're going to rethink their business. I think you're going to lose some people, but those people are going to be repurposed, right? So if you go back to, uh, the beginning of the 20th century, so beginning of the 20th century, about 93 people, 93% of Americans were in the agrarian economy. Farmers, basically. Uh, at the end of the 20th century, it was about three percent of the American workforce as farmers. And if you looked at just those two stats, you'd be like, oh my god, something horrible must have happened. All these people must ha…

AI assessment note: “I think you're going to lose some people, but those people are going to be repurposed”

Q though that tells maybe not a different story, but an interesting story about where AI usage is going. So Anthropic recently put out this economic index that said that the, uh, percentage of use of its tools in the workplace has actually flipped. It began with augmentation, maybe when the tools were a little bit less reliable and now it's gone to automation. So what do you make of that?

A I think that's right. I saw the pod you do with, uh, Dario Amadei. I thought it was, It was great. And I'm a big fan of what they're doing. Um, I'll tell you when we started, so we had an experiment in a way we, uh, we licensed, uh, AI tools, cloud code from Anthropic and, uh, chat GPT from open AI. We put them out there just to see what kind of uptake there would be. And there was, you know, huge uptake. Everybody wanted to use it. Um, but I recently wrote a script that analyzes like the sentiment, what people are doing with AI and the company. And it was about 35, 40% of the usage was just rewriting things. Right. So it's helping people get more efficient, which was great. I didn't mind that necessarily, but it wasn't like innovative leaps and bounds ahead, but then we're doing some more advanced stuff. That was sort of the control group. If you will, we have people who we've given space and time to like really get trained and really set out like, what do I want to accomplish? How much more efficient do I want to get like a more structured use of AI? And so we're doing, I'll give you an example. Um, a few really cool ones. One is, uh, every time an inquiry comes into our customer service desk, you know, customer services sort of In a way it's like ground zero for AI, right? There's a lot of, a lot of sort of human decisioning that can be replicated there. So every time a mess…

AI assessment note: “I think that's right. I saw the pod you do with, uh, Dario Amadei.”

Q But on the other, you've written scripts on your own. So where do you stand here?

A Yeah. So, uh, I think there's a few ways you can do it. We could start like, Hey, we're going to start with the most foundational, uh, we'll build our own models. We'll train our own models. And that's too expensive for so many companies like ourselves. Uh, or we can say, hey, we're going to tap directly into those APIs. Um, or we can say we're going to use cloud code to build faster. I, I view those as like levels of maturity. What I've said we won't do at Netrix is look, I think AI is going to be so like instrumental critical to the future of the company, like how we grow. And if I believe that it's so critical, I can't outsource all of that. Right. I mean, who are we if in five years we have this amazing automation or two years, incredible automation with AI, we're moving faster. And what I mean by that is it can't just be, we're gonna go turn on the Salesforce Einstein thing, or we're gonna go to, you know, every one of our vendors and check the box and pay them an extra 50% to get their AI module, right?

AI assessment note: “if I believe that it's so critical, I can't outsource all of that.”