0:00It's hard [music] to imagine in 2030 that a human is ever [music] doing the first pass of any work.
0:08I don't think we're going to run into an underemployment problem. I think we're going to run into an overemployment problem.
0:13Feels like every six or eight months we need to think about software in an entirely [music] new way.
0:18And the future is just system architecture.
0:26Uh really excited to have you guys on the podcast.
0:29Uh so maybe to start, can you introduce yourselves and uh tell us why you invested in Artemis?
0:36Sure. I'm Andy Treadway. I'm a partner at Theory Ventures. Um Dan and I have known each other for more than a decade. We studied neuroscience and computer science together at Brown University. Um and for many years we've talked about ideas to apply AI to cybersecurity.
0:53Um and one thing I found really interesting was uh in the post ChatGPT era I felt a lot of urgency. Like don't we need to build something now? Don't you feel like you're missing the moment? And you always had a sense of patience. You said I think that these systems and the technology is going to evolve so quickly over the next couple of years and the startup that you could build today is very different from the startup that you'll be able to build in two or three years. And it may actually be in our
1:22benefit to wait and see what new patterns and capabilities emerge. Um and when I think of Artemis I really think of a second generation AI company in cybersecurity and I think a lot of the ways that you thought about building the platform and the company are only now possible in the agentic era and so that's what got me so excited to invest in the company and partner together at the Series A.
1:43Yeah, my name's Tomas Tunguz, uh general partner at Theory. Uh and I I think one of the things that got us most excited about Artemis is whenever you start with a data system you can build a huge platform. Uh and within the world of security the SIM is core.
2:00Uh and on top of it, you can build detections and all kinds of remediations and an end-to-end system. The data volumes inside of AI are huge, right? We see in our analytics businesses, you see query volumes of 20 to 50 X uh coming off those systems when the agents interact with databases. I'm sure it's quite similar in the world of security. And so as Andy said, the second generation AI agentic systems, they need to exist because the time to response is significantly less. There's a lot more urgency within the world of
2:30the CISO to be able to respond and it's almost impossible to do it with people.
2:36Uh so tell us a little bit about Theory.
2:39Yeah, we're about 3 years old. We're about 15 people. Half of us are AI engineers. It's run by a PhD in the field. His name is Brian, CEO, built a healthcare practice at Palantir, two phenomenal sales leaders that built the go-to-market organizations. And we manage a little bit less than 800 million. We're very thesis and research oriented. So Andy and I and the rest of the team will embark and we've embarked on almost 500 different research projects since the inception of the firm. And we don't often find a company that we get really excited about, but
3:07when we do, uh we support them uh to the very limit of our abilities.
3:13So I I'd love to learn a little bit more about this research process and and kind of what does it look like when you're you're approaching something that's you think, you know, this might be really might be something.
3:23Yeah, so I think we think about it in a few stages. First is really trying to understand the problems. We spend a lot of time investing in building communities of technical leaders in different domains from AI to engineering to cybersecurity. And just trying to really understand what are the problems today and what are the problems that they see coming down the road.
3:44Uh and then we think from the technical platform side of things. Uh most of us have technical backgrounds as Tomas mentioned, about half of our team are full-time engineers.
3:55Um and so we really try to understand the contours of the technology and again, what's possible today and what's possible in the and then try to match those uh problems to those technological capabilities, and then overlay a sort of more traditional venture market analysis, understanding where dollars come from, understanding the incumbent market structure to identify where there are what we call technology enabled new category creation. So places where you can not just do something incrementally better,
4:24but where we expect that the entire market structure or structure of some product category will change because of emerging technology, most of it AI.
4:33So let's talk a little bit about cybersecurity. The we as you you gave a little bit of a good context earlier and the I um really think a lot about how we uh spent a lot of time looking at some of the changes and what's possible from a technical perspective and how that then implicates what kinds of security companies that you can can build, what kinds of platforms you can build.
4:54Uh how do you think about the types of security companies that you invest in?
4:58When you're when you're looking at a company that's um operating in a space like Artemis, either uh something that's looking at more operational workloads, something that's looking more at the data layer, something that 's looking more at the posture configuration layer, what what are the things that you really look for? What are the signals that you see that's uh kind of unify between the research and market elements than the company itself?
5:21Yeah, absolutely. So I think the overarching thesis in security that we have is quite simple is that the work to attack a system is extremely similar to the work required to defend a system. And as coding agents get better and better in cybersecurity, that will affect both the attacker and the defender side.
5:41And so as attackers are adopting more and more autonomous, highly capable systems, that will force the defensive side of the market in the cybersecurity industry broadly to rethink how they secure their systems and organizations. And so that's the overarching thesis and I think if you look to every function within a cybersecurity team uh from uh as things upfront, making sure vulnerabilities closed, making sure the security posture is strong, to the operational side,
6:10um even to things like identity, to training, and then of course all of the underlying data systems. Those will need to transform in an era where a majority of tasks will be done autonomously in the future.
6:22And so we've dove deep into uh you know 15 20 plus different areas to try to understand what does the workflow look like today and what does it need to look look like in the future and then what new capabilities or platforms will be required to serve that.
6:37Yeah, I mean trust is a kind of at the core of this ecosystem, right? In the world of security, I think uh social proof is maybe on the surface uh really difficult to ascertain, but then once you start understanding how different CISOs communicate amongst themselves with different problems and which solutions and vendors they choose, a lot of it comes down to trust.
6:56Uh and so within the founders that we looked to back, it's really important to find those teams like Artemis where uh they have an incredible amount of pedigree within the ecosystem. Uh people give them a lot of credibility and credulity uh to what they're doing and and ultimately that is a snowball effect because if uh you work with great companies like Cursor and others, more people want to join the movement.
7:21Yeah, I definitely think uh trust as the core is really uh critical. You know in cyber security you are you are dealing with very very sensitive data, you're solving really hard problems, people are really counting on you to to be able to deliver. And so earning customers' trust and maintaining it and and proving it out, I think um you know has definitely been uh definitely resonates with me.
7:42Um so the the landscape's changing in cyber security, the landscape's changing and um especially when we think about the distribution of attacks and then also the distribution from the defensive perspective, uh how do you think about the way in which agents are interfacing with security teams, both from the perspective of teams increasing the amount of efficiency that have, the way that companies are utilizing AI agents,
8:12and then that introduces new security vulnerabilities, the way that attackers are leveraging AI.
8:17Where where would you say is the kind of largest areas of opportunity? Where would you say is the largest areas of disruption? How how do you break it down both from a kind of market perspective and from a practitioner perspective?
8:28Sure. So I think if you look at the security category historically, you basically have a lot of sensors, so products you deploy to generate information. You maybe have another set of products or tools that you pay for that just aggregate those and present them on a dashboard, create tickets, do some workflow, etc.
8:47And then you rely on people and you wait for those people to have time to look at a ticket, you wait for them to do an investigation, and then you wait for them to remediate something, to close a vulnerability, etc. And it's clear that in the future we just won't have time for that.
9:04When agents can run an end-to-end attack in 5 10 15 minutes, you can't wait 30 minutes or an hour and then hope that someone has time to come in and investigate and resolve something. And so I think what we see as the opportunity is basically compressing that time and moving from a reactive to a proactive posture. I think the simple low-hanging fruit is saying, okay, well, what are our humans doing today?
9:27That would be something like alert triage, that would be sorting through vulnerabilities, etc. And can we just have agents do the same thing, but do it at a broader scale and do it faster?
9:37I think what gets us particularly excited is thinking one, two, or three levels down the road is saying it's not just taking sort of the last step and replacing the human review with an agent review, but it's using agents to affect this closed loop system to learn from incidents and create new detections to identify them better in the past or to identify some category of risk and then proactively mitigate it in the future.
10:04And so I think that's what gets us most excited is not just automating the stuff humans are doing today, but allowing teams to shift to a an entirely different kind of posture where they are continuously getting more and more safe and secure instead of just trying to catch stuff at the end of the funnel.
10:21I I think that there's a a kind of core element there of thinking about what's the difference between how do you take what's possible today and then making that be something that's dramatically more effective versus how do you think three steps ahead into the future of what what are the kinds of things and the kinds of problems that people are going to be solving. Um you know certainly as investors there's very crucial to be thinking multiple steps in the future as as we think about what are the kinds of tasks and kind of work that security
10:50teams are going to be doing three to four years from now, five to 10 years from now. Um I I guess Thomas maybe share your your perspective on on that.
10:59Yeah, I mean I think I think Andy laid it out really well, which is outside the organization clearly threats are happening. They're more sophisticated.
11:05They're faster. The other major dynamic is the cost to attack or to try to attack is being reduced pretty significantly and so the breadth of these attacks is extraordinary. The second thing is so you have speed on the outside and then the thing I'll add is you have speed on the inside, right?
11:20When we work internally now as a result of AI, we can do many many things much faster and as a result of that we pay less attention candidly to like what's actually happening. How are we modifying these databases? What kind of content are we sending via different communication channels?
11:33Um and so inside the organization as well the speed is so fast that it's no longer possible. You can't staff a team to be able to follow what's going on and this is like I think another area of innovation where how do you create systems that really understand what's happening inside of the organization uh that you know an agent may actually do something by accident that you don't really want to have happen and there's no malice in it, but it's still problematic. And so in a sense like that
12:00that boundary between inside and outside is and maybe it's temporary or maybe it's permanent, but the the the dynamics or the the the way that speed or the compression of time changing security is is similar in both.
12:16So building on that and slightly shifting the the topic, let's talk a little bit about the way in which startups are working and how that's changing. Companies in general, but startups are very much at the the forefront of uh working in a very very different way on top of AI agents. Uh Indie wrote an essay uh about this recently talking about a number of companies that you work with including Artemis uh how we have engineers pushing many many
12:44hundreds of pull requests every week. Uh some engineers over a hundred a day uh at at certain points in time which uh is can be uh very very different than ways that companies have worked in the past. Um what would you say are some of the things common trends that you've seen across the companies that you work with and then some things you think uh should be looking out for in the future.
13:03Yeah, definitely. So probably there's been a huge amount of investment and focus on getting agents to be better at and faster at writing software. And there is an assumption that writing more code, more good code faster means you deliver more good product faster.
13:21Uh and I think what we've observed in working in in talking with a lot of startups is that's not necessarily true because there are a whole variety of other bottlenecks and key aspects of the process to delivering great product.
13:32Um and one theme that we think about a lot is the largest and most sophisticated software companies like Google and Meta have for a long time operated very differently from a typical startup or software company. And that's because when you have 10 or 100,000 software engineers all collaborating together on one project, you need entirely different systems. You need new version control. You need new types of project management to make sure that they can write code effectively moving
14:01the product in the same in the same direction. And so we anticipate that more and more companies even at a very early stage will start to look like that where they have effectively hundreds or thousands or tens of thousands of contributors to their codebase, most of them agents, and then we'll need new tooling and ways of working to align those contributors towards some shared goal.
14:23Um, and while I think many investors see the rise of coding agents as eroding differentiation in software, I think at least for now we see it as increasing differentiation where the difference between a very AI native team like Artemis versus another team that might be using coding agents but is not orienting the company around how do we deliver many thousands of PRs every month reliably, um, we'll see that that advantage for
14:52Artemis will start to compound. And so we think it actually increases the difference between the 99th percentile team and the 75th percentile team in terms of how productive they are for developing software.
15:05Yeah, I think that's right. I mean, I think the baseline startup should be expecting to produce 3x more the number of PRs. I think a lot of heads of engineering are pushing in that direction. The revenue per person in a startup is exploding. So the overall efficiency of organizations and all of that means that these businesses are able to generate a lot more revenue per dollar invested than in the last generation, pretty meaningfully.
15:30Yeah, which I think is a a good uh push that is exactly what we want to to see and uh um, you know, being able to to drive the best outcomes. So let me ask a question that's um, you know, maybe maybe a kind of a bit of like building on this, which is um, you know, when you think about the way in which a company works and how that impacts the products that they build, one of the theories that we have internally is by leaning very very
15:59heavily into AI driven developments, we will develop a better DNA internally for understanding good and bad AI workflows.
16:08You know, at Artemis we're building a product that is AI native. It enables security teams to be dramatically more effective at detecting and responding to threats. And a very core part of this is the natural language interfaces. It is the uh that is inherently agentic at every single step of the product. And this is a very different kind of product feel and product look and product experience than what has previously existed. And there's a lot of net new UI work and UX work and product work that you need to do in order to flush this out. And our
16:37our theory internally has been if we invest very very heavily in every single person in the company working with AI, then the product itself will get better as an AI product because the people who are building it will have a better sense of what would they want as users.
16:52But I'm curious what you see across your portfolio, across the the kinds of other kinds of companies, the relationship between these two different factors and things that companies are doing internally that then makes itself really impactful as part of a product. We've always seen this in developer tools, for instance, but this is expanding dramatically beyond that these days.
17:13Yeah, I mean, I think, you know, there's this trope in startups where you ship your org chart. Well, if your org chart is entirely AI pilled, well, your product will be entirely AI pilled. So let's break that down. What does that really mean? Well, once you start engaging with a really great AI product, you don't want to point and click anymore.
17:30That modality's gone. You want to type and say, I want to change a setting or I really need to accomplish this today or what are the three most important things I need to review?
17:40And you suddenly develop an intense frustration with old software. And so if you feel that every day, well, of course you're going to build a product that will feel very native. I think the other expectation as we were just talking about with PRs is the software itself has to move very quickly.
17:57And you have to learn how to use that software in a way where you can see there's increasing sophistication. Like all of a sudden I can issue a very simple query and then I might be able to within a day understand actually I don't want to ask just one query. I want to ask 100 simultaneous queries. And then how do I manage 100 simultaneous agents in a framework that actually makes sense and I can keep it in my mind and the agent's successful. And unless you've used it and seen the failure modes, it's really hard to build for that use case
18:27Yeah, I think both. We see both uh technical uh focus. So how do I architect my system for scale? How do I run my CICD and developer processes for scale?
18:38As well as the process side. So what does planning look like in an era where we're developing a lot more code? I think we generally see a lot more focus shift to upfront planning and alignment with the assumption that execution will be then fanned out across many agents. I think one shared characteristic that I've observed is that the teams that are best at this are very systems oriented.
19:00Um and that runs a bit counter to the traditional, you know, startup experiment fast break stuff where you're sort of doing one thing at a time. I think the best teams think about not how do I ship this feature or how do I improve this workflow, but how do I change this system so that I can more effectively ship and test and deploy the next 100 features or PRs. And you just see such compounding returns to scale with that approach. And so I think I see the best teams sort of continuously upleveling
19:30and reflecting on both the technical and process side. How do we systematize this to do this 10 X, 50 X, 100 X faster?
19:36What's an example of the product that you've seen like where you were because because the team is so AI native, all of a sudden you saw that manifest itself within Artemis?
19:46I really think that a lot of it comes from the natural language interfaces to do everything. That is one that is one core thing. You know, we internally we use both Codex and Claude Code pretty extensively for for code creation. And there's a general sense of, hey, I want to be able to find that old chat. I want to be able to ask questions about things that I've done in the past. I want to be able to change this configuration. I had an issue in Codex where I uh models weren't updating and I was able to debug it entirely within Codex. And
20:15Codex was able to go to my machine and figure out a lot of these these different flows. And I a a of this then impacts how we build Artemis. So as the, you know, customers within Artemis want to change settings, they want to be able to reflect back on the work that they've done in the past, they want to be able to improve the rules that they have in place for detecting threats. They want to be able to audit and see trends and everything that they would want to do within the platform that's helps them do their day-to-day jobs are able to do entirely within natural language. And so
20:44that's one very very core piece. And another is the way in which the product handles asynchronicity and the ability to have multiple threads at the same time. The, you know, within the flow of writing software, it's it's very common that you'll have many agents running at the same time. And that's because large language models are slow and they need to do things that are slow. They need to deploy, they need to run queries and some of these things take time and you can't just sit there and wait.
21:11You could, but you're not going to get as much done as if you do 10 things at once. And so this is something that we see a lot in Artemis as well. We have of our AI mode flows within Artemis involve going and hunting for threats, building detectors, running analytics, building dashboards, building workflows and applications all all within the platform. And sometimes these things can take a bit of time as you're bridging together tens of different data sources in order to perform these analytics. And it's really important to be able to give an experience where customers are able to organize their different chats or
21:40threads within the platform, be able to take things and put it aside and come back to it later, have native notifications when things are completed. All of these types of asynchronous flows are really really important to make first-class citizens in an AI product.
21:55And that's not something that's necessarily obvious to a product designer unless if they've spent a lot of time using AI tools to really appreciate the importance of asynchronicity as a as a core element of development. And there's a long tail of these these types of product insights that that you really kind of develop as you spend a lot of time working with these tools. I mean, another example as as as we're on it is the importance of visibility because ultimately at the end of the day,
22:22when you use a AI tool, the tool is going to go and do a bunch of things and then it's going to give you an answer. And how correct the thing, the answer is, is almost always extremely highly correlated with the things that it did between when you asked it to do something and when it came back. If if the things that it did was very very shallow analysis and checking a couple of random sources, then probably the answer is going to be not great. And if things that it did was going and querying the exact same data sources that you would have queried in
22:52the same order, then probably it's the right answer. And you don't need much more information than that to know how much to trust it. But what that means is that it should be really really easy as a user of the product to be able to see what the system did between when you told it to do something and when it came back to you with an answer. And so having that kind of traceability, having that kind of editability and steerability as first class principles in the product is core for a really truly great AI native experience.
23:15And that's new, right? I mean, you think about like legacy software products, you don't really have you had audits and whatever, I mean, all kinds of lineages associated with it for compliance reasons or long-term logging and debugging, but now we're really steering the product as it goes. And like the robot's actually like looking at its internals and modifying it as as it's working.
23:34Yeah, absolutely. These kinds of self-improvement loops are crucial. And I think it's also the ability to spot mistakes because it's you when you the more that you builds with AI, the more that the it's important that every intermediate step is something that something else can build on top of. And that's only possible if you compensate for the fact that sometimes agents are wrong and you need to be able to spot that and adjust that really easily.
23:59Whether what's spotting and adjusting that is a human or another agent that's that's performing those procedures. And so there's there's a lot of a lot of this you know, thinking from the perspective of what does it look like when we have many many different agents being able to perform these tasks in as sustainable a way as possible.
24:16Yeah, it goes back to Andy's point about the 10,000 engineers, right?
24:20This huge hive mind of people or things working together.
24:22Yeah, yeah, visibility and modularity are important in in an organization of that size. You you know, whether whether it's a bunch of agents accomplishing a single task or a team of engineers.
24:34Um, so we can switch gears a little bit. Uh, so we've you know, at Artemis we solve uh detection response challenges for organizations. Security operations we think is changing very substantially for a lot of the reasons that we talked about earlier. These, you know, actual experiences of working with products is changing. The experience you get uh from your expectations of the tools that you use and then also the difficulty of the challenge itself as attackers are
25:03utilizing these kinds of tools become dramatically more effective, move faster, can perform multi-stage exploitations much more effectively. And so thinking forward to the future, how how do you think about the role of security operations in 2030?
25:18Yeah, I mean I think the the human role in security operations will continue to be critical, um but will certainly change a lot. I think it's hard to imagine in 2030 that a human is ever doing the first pass of any work. I think just from a latency perspective um we need to have agents that as soon as an alert are coming in comes in, they have an investigation triggered and they at the very least collect some
25:48information and make a recommendation if not just taking remediative action themselves. I think the role of humans will start to look a lot more like security engineering. They will be tuning these systems, they will be looking at historical data, they'll be helping interpret new and emerging threats and helping uh monitoring configure these systems so that they continue to work more more effectively over time.
26:16Um and so I think it should hopefully be a really fun and interesting and fulfilling job for security professionals is that they're no longer writing long SQL queries or trying to scroll through logs and find something that happens. Instead they're thinking about emergent risks and then the systems to help combat them and trying to make them better.
26:35Yeah, the future is just system architecture. It's just how do I how do I have agents that are researching vulnerabilities, understanding what's happening in my system, and then creating the systems around the detection and remediation agents in a way that I can get to an answer. It'll probably be you know, it'll look like a hedge fund almost, right? Where you're trying to respond in hundreds or thousands of a second uh to little things that might ultimately be really big and consequential things.
27:01Uh and I think that's fun. Like that is uh it's incredibly exciting. It means a brand new level of scale uh and sophistication.
27:10Uh and I I don't know. I think I really like that future, right? You you're you're having I mean it's a lot like a hedge fund portfolio manager where you're trying you've all these different uh postures or positions and you're trying to mitigate the the downside risk on all of them and with different levels of priority.
27:26Yeah, one one thing that I've seen that's uh been really interesting as I've seen an increasing amount of AI utilization both on the engineering side and the security operations side is how many challenges start to look like data science challenges because so much of once once you when you think about the process of completing a task, you you have you know, first you're assessing of like okay, how am I going to do this thing and then actually performing that thing, whether this is you know, maybe maybe
27:55we're talking about making changes to to a detection rule. I'm going to first think okay, what what kinds of changes might I actually want to make, then we actually make the changes themselves, and then we evaluate the results. And agents will kind of blast through the first two automatically. No problem. We look at this, look at the uh the playbook that we have for modifying these rules, then we make these changes.
28:17But now we need to make a judgment call based on the evaluation of the results. I still need to run a rescoring and and see okay, you know, what before we were flagging these things, this rule, now we're flagging those things, this rule.
28:27Is that what I actually want as this this new this new shift and new change?
28:31And sometimes agents will look at those and make their own decisions, and sometimes we'll be able to make 10 or 20 different changes and gather this up, and then analyst will have to look at these different results. Okay, you know, before the state of the system looks like this, now the state of the system looks like this. Is this a good change or a bad change? And this really is ultimately data science what it comes down to it because you're looking at statistical or exemplar pictures of what is the difference in how your system works from in the past
29:00versus now. And that's you know, it's a mixture of understanding the system design and understanding the the way that the the data actually looks.
29:11Yeah, I think it it should be exciting for security leaders because most today would love to be doing that kind of data science oriented work as much as they can, and every security team is just so capacity constrained that they just don't have the time to do it because everyone is sort of frontline firefighting. And so I think the idea of agents providing unlimited capacity for security teams to go collect information and analyze it and then present that to
29:39humans for to make some decision. It hopefully should help defense in the long run win the war because they really can be everywhere all the time in a way that historically they've had to really be rationing their time and effort and where they actually go collect data.
29:57So we can switch gears to the the rapid fire section. [laughter] So starting with it's been a few years now of theory.
30:09Curious, what has been the most surprising things for both of you?
30:13I think I've continued to be surprised by the pace of progress in all dimensions of agentic capabilities. I think that you know, we anticipated that things would continue to improve, but it has been notable and exciting to me by how many paradigm shifts we've had just in the last three years.
30:39Um, from really good chat agents to agents that started to reason, to agents that now can write and deploy really effective software. And each one of those has entirely changed the kinds of products and companies you can create. And so it's been just a really exciting time where it feels like every six or eight months we need to think about software in an entirely new way.
31:01Yeah, I think I've been surprised by how quickly the enterprise has adopted AI uh and really taken into account how fundamentally different the ecosystem is. And I mean, you look at like the first wave of software and a lot of those pieces of software were little point solutions.
31:19Uh, you know, Salesforce with Salesforce Automation took a component out of Siebel and said, we're going to we're going to optimize this particular workflow. And then over time, you have these businesses that compounded to be these absolutely gargantuan uh software companies. But in the world of AI, I think a lot of enterprise architects and C-level leaders are just wholesale re-evaluating how they work and as a result willing to reimagine extremely quickly.
31:44Uh, and that I would never have imagined would happen this quickly. I I thought the enterprise would be uh slower to adopt or even with the cloud, like when AWS came out, it was like eight years before you had any like significant enterprise, any Fortune 500 actually deploying workloads on AWS, right? And like to think about that now, it's like what enterprise is actually deploying on prem or pushing a lot of workloads there. It might be changing a little bit with AI, but I think that wholesale leap to
32:11re-examine core systems like the SIM or uh or core databases or the CRM or whatever, that I would never have anticipated.
32:22It really is quite shocking both how fast the capabilities advanced then how fast adoption uh really be able to begin to hit. Um, the I think when there's a when there's a change this great and this accessible that it the physics really change in terms of what people do.
32:39Yeah, and then a lot of the enterprises are comfortable with software now, right?
32:42They've been running it for ten or 15 years and so they know, okay, this is the failure mode here, this is the failure mode there. And so they're more sophisticated when they buy, which is really good.
32:51So another rapid fire question.
32:55When you think about the difference between the career that you had and the career that, you know, people will have in the the future, is there any advice that you would give to people earlier in their career based on your experience, based on what you're seeing across the portfolio and based on all the changes that we've discussed in the markets?
33:15I think in every job the ability to plan and manage large number of workers will be extremely important. I think you see in every field from you know, writing to software engineering the core power of agents will be the ability to give yourself 10 20 100x leverage in the volume of stuff you can do. And so then the burden shifts to deciding
33:44where to point that capacity and orchestrating all of the work and reviewing all the outputs. And so I think that that's one meta skill and you really can't get experience doing that unless you actually are just running five 10 15 agents all the time and experimenting with building systems for how to get them to do what you want them to do.
34:03Yeah, I think I think there are two skills you really need. The first is systems thinking and there's this awesome book by an ex MIT professor called Donella Meadows and her book is thinking in systems.
34:15And it's excellent. It's just like, okay, how do I design a loop, right?
34:18Like we talk in the world about AI about like fast loops and small big loops and what Andy's talking about with managing 100 different agents or a thousand different agents. How do I create these loops where I as a person can understand what's happening across them? And the other one is sales. Like I think you'll either be in the business of designing systems or you'll be in the business of selling. And then there's a great book called The Challenger Sale and Technique in Selling, particularly in this era of AI is how do I sell to you in a way where I teach you by challenging one of your core assumptions. Like, yes, okay, this is the way you used to think about
34:48SIM. Let me reframe it because actually the world has changed.
34:53Do you think that uh agents really foundationally changed sales, or do you think it's largely the same as all trust, right? We We buy from the people we trust. And that's just true for It may not be true for a notebook, but it's probably true for a car still. Uh it's definitely true for security systems. Uh and the way that we develop trust is is just by chatting with each other.
35:15Yeah. I agree. I don't think [laughter] I I don't think it I don't think that's really changed much either. Um okay. Uh one more question. Um controversial opinions or opinions that your peers would disagree with um you know about any of the things that we've discussed.
35:31I can go. Uh I think one one controversial opinion I have is that most valuable problems are no longer intelligence constrained.
35:43Um I think we see a lot of focus in business on training and fine-tuning new models, on collecting new data to do them, etc. But I think when you look at most economically valuable jobs, uh we've basically saturated that benchmark and a variety of models are more than good enough to do the core cognitive work. And instead of being intelligence constrained, those problems are either scale and cost or knowledge and context constrained.
36:12And so the implication is there are a lot of really valuable and important products and companies to be built that are systems or harnesses around those that help make models better on a per customer, per category basis. And you don't need to own the model um or you don't need to be a model company to ultimately capture value in some application.
36:38Yeah, I think we're going to run into I don't think we're going to run into an underemployment problem. I think we're going to run into an overemployment problem. I think as we design these systems to be massively productive, there will not be enough people to generate that level of productivity.
36:53And so if we really are, you know, I mean today it's somewhere between 20 to 60% increases in productivity. The very best companies are kind of around 3x and we can debate whether or not the effectiveness is what it ultimately is. But if I can make all of my employees 3x more productive and my competitors at like 40% more productive, well I'm going to hire as many people as possible. That's what the game theory says.
37:13And so I think the fears of a broad scale like underemployment in the fullness of time, it will be the other problem. We need to make many more children to [laughter] satisfy employment.
37:24That that's that's my controversial opinion as well actually.
37:27[laughter] I I strongly believe I strongly agree. Okay, guys, thank you for coming. Really enjoyed the conversation.
37:36Pleasure. Thanks so much.