Ep 160: Box CEO Aaron Levie on Silicon Valley's AI Problem, Evidence for AI Job Creation & What Doomers and Accelerationists Get Wrong

0:00You're telling me this thing is gonna destroy the planet and you're the ones making it and you're not stopping. The actual creators of this stuff are scared about it. Imagine if like Larry David made AI models. Like that's kind of what we're dealing with right now. So where are we in the AI wave? Companies going from zero to 100 million or 500 million in revenue in one to two years. This has never happened before in software. Anything that was slow in a process because of that, we should be able to go in and have agents go and automate. If you believe in job destruction, like in any kind of very large scale way, you basically don't believe that markets adapt.

0:31It's going to become the number one topic in the next presidential election. I think we're in for some very messy few years on this.

0:45My friend Aaron Levy is the founder and CEO of Box. He's been a leader in the technology world for over 20 years. Over 100 ,000 businesses use his products. He's a champion for AI, a voice of optimism against a lot of the doomers. He's also a really funny guy. I think you'll enjoy hearing from Aaron what's going on in the AI world. Welcome to American Optimist. Really excited to have our friend Aaron Levy from Box here with us. Aaron, thanks for joining. Thank you. Good to be here. It's nice to see you out here in California. Don't tell him I'm visiting for too long. Yeah, exactly. Is that like a tax issue or?

1:13No, I come here a lot. Probably about a quarter of my time, actually. We would love to see you more. Yeah, it is the center of the AI world. Although you probably you did see Cerronek in Texas last week. I saw that. I mean, obviously, I think Austin's having a moment and obviously broadly Texas. So it turns out if you can build like real things, it helps your state. So congratulations on that. Thank you. You guys are still ahead in software. Let's dive right in. We're right in the middle of the AI wave. I think it's the most exciting time to build. You know what I'm seeing? And I saw a bunch of them here in the last week, of course.

1:46It feels like every two to three months is like a year now in terms of build time. It's like something's like accelerated. And so business itself is going faster, right? Are you seeing this too? Yeah. So, I mean, we have good kind of nice sort of comparison checkpoints. So if you go back like 20 years ago, you were like, okay, we have an idea. We want to get it into the market. And you might be heads down for like a year or two before the thing actually goes live or seen by anybody. that's obviously shrunk into a matter of months at this point. And so if you kind of take that type of time compression, I think that's sort of changing everything about feature development, new product launches, getting products to customers, obviously of companies going from zero to 100 million or 500 million in revenue in one to two years.

2:33Like this has never happened before in software or kind of technology broadly. So definitely an amazing moment. And it's all driven effectively. you know, the underpinning is model progress. So the downstream effective model progress is everything else just moves way faster. And we're definitely seeing that. Speaking of how long it takes to get into production, I've seen some examples now where an agent will get feedback from the customer and then like go and suggest what to fix, put in their PR, get it approved and fix it, which is kind of crazy. So it's real time almost. Yeah. I still wonder if that's more like a startup phenomenon, but the idea of like this automated software development life cycle is pretty compelling.

3:12So, you know, the moment from bug to then live production fix of that, you could be, you know, you could compress something that used to take days or weeks into a matter of minutes. You know, the main, the main thing is what is the kind of human in the review step that that's needed in that, but it's a very, it's very compelling. As someone from your generation, I think we should be reviewing in person. Yeah, exactly. Yes, exactly. Yeah. We need to keep people in this. What are we talking about. We have so much more taste. But I think that also applies to many other analogous things. So imagine a security incident.

3:42You obviously would want to respond to it as quickly as possible. So that'll become entirely agentic. Any kind of production issue with your systems, you'd want to be agentic because you could take the downtime incident from maybe 10 minutes or 30 minutes or 60 minutes to a matter of seconds. So where are we in the AI wave and where do models go from here? I want to get into the Kimi China debate stuff. But before we go there, like, you know, when you talk to the guys around Anthropoc, like Dario, at least a few months ago was saying these things are going to double every few months and I can already see the next two years where it's going.

4:13And this is one of those things is very unintuitive, even to me, because it's like, you know, if you double eight times in two years, that's that's 256 times. Like, what does that even mean for like where they are now and where they're going the next six months? What is it? What do you what do you see for your business? Yeah. So maybe we'll start first with model progress. So model progress appears to be effectively not slowing down at all and in some domains even accelerating. And so every prediction, like last year around this time, there were these kind of debates of like, have we hit a wall?

4:44Yeah. And I think we've just completely blown through that. And a mix of that is, is the companies are continuing to improve their pre-training, you know, steps, the kind of data that they get into that, and then the post-training steps as well, and then kind of everything in between. So we just keep unlocking, I think, new breakthroughs of how to improve these model capabilities, which is incredible. I don't know if you guys invested in these, but there's a whole ecosystem of new startups that are doing data for the labs. And so if you go back to Dario's originally scaling laws concept, which is just more compute, more data, you'll just see progress effectively, follow that line.

5:24and we are just getting an exponential amount of really, really high quality data and compute keeps coming online to let you go train these models. So if you just, you could just assume kind of a complete exponential curve on model progress. And we're seeing this across all of the labs, you know, obviously Anthropic, you know, kind of really kind of, you know, touching the, you know, the current frontier with Fable, GPT 5.6, you know, being just kind of neck and neck in a bunch of domains. Obviously Grok having really good showing, Meta coming back, you know, Gemini doing quite well. So you've got a minimum of five leading US labs that at a minimum within a very relatively narrow band are going to all be following each other at the frontier, which is fantastic.

6:06And what's cool about that narrow band is you get kind of a narrow band of intelligence difference, but a pretty wide cost variance. So you have different players saying, we're going to be the low cost provider. Other players saying we're going to be the kind of very purpose built provider for high end workflows. That's effectively how things are shaping out in the US. Then, of course, you've got the sort of open weights models, a couple in the US, NVIDIA, Thinking Machines, and a couple others. And then obviously, the sort of China story. The biggest, I think, update for everybody in the past three to six months was, I think the debate would have been probably if you had just gone back one or two years, and we were to take the top 10 AI researchers and say, you know, how far is how far behind is China?

6:44You know, I think general consensus would have been six months or a year, you know, kind of at a minimum, just be based on either access to chips or, you know, kind of the methods they were using. Now, it's very clear, you know, you have something like a Kimmy K3 moment, which is really kind of if you look at artificial analysis and their overview of AI models, it's considered to be basically the third best model, kind of on a composite, you know, sketch of... It's like a near copy of Fable, basically. It's able to do kind of near Fable-esque. And there's debates. People say it's like far less efficient on tokens.

7:19It's got obviously going to have maybe less judgment and taste somewhere in there. But let's just say it's a very good model, like kind of regardless. And even on some things that are more subjective, like this front end design test that people do in LM Arena, it actually was the top model in the world. So on some dimensions, it's actually spiking as the best. So no matter what, you can look at this and say there's an open source ecosystem in China that is near neck and neck. let's say at least still maybe kind of you know one to three months behind but not a year or two years behind and then a really really strong frontier uh ecosystem in the u.s all basically racing ahead so that's kind of model progress and then um what's going to be interesting and dario i'm stealing this from dario a little bit uh like a year and a half ago he kind of framed this i think quite well on some podcasts where he said we as consumers are probably going to start to experience this model progress kind of almost less and less like we're going to go to chat gbt we're gonna go to claw we're not trying to solve hard math and physics problems yeah exactly so like we're not we're we we are not uh you know giving some you know incredibly unsolved math problem to uh you know to a model we're saying you know hey where's like a good place to to you know that's open for uh for food tonight this must be like for your experience where you feel under appreciated why aren't people asking me really hard questions all the time about strategy uh so so what's what's going to be kind of fascinating is you're going to see this kind of bimodal thing, which is consumers probably are already at the point of like, like, you know, I'm talking to an AI almost, you know, five times a day about some personal thing, you know, you know, a kid thing, a family thing, or, you know, somewhere we're trying to, you know, maybe go on a trip.

8:54And like, it's already super intelligent. Like, I don't, I don't think I need that much more. You don't need Einstein versus the physics professor down the street to talk about your pizza order. Exactly. Like we've, we've actually solved that. In fact, you might actually like, there's probably like a, a little bit of a horseshoe, like Einstein would probably get the wrong pizza. So like, so there are some upper limits of like how much intelligence you want to apply to different problems. And so now, you know, I'd say consumers are just like, we kind of know what the consumer thing is going to look like.

9:19Obviously, like when you add robots and stuff, you're going to need that level of intelligence. But for consumer kind of end user products, like we're in a pretty stable position. The flip now is on the enterprise side. And the enterprise actually right now, we're like nowhere near the level of intelligence needed for the vast majority of enterprise tasks. And that's because if you're doing an insanely hard due diligence problem on an M &A deal you're doing, there's just a lot of compute and a lot of intelligence you want to apply to that. You're bringing a lot of data together from a lot of places.

9:47You have mass amounts of data. You want to have the entire history of every M &A deal in the world that's ever happened. You want to be able to understand every single legal framework on the planet that might have implications for this. So there's just a lot of knowledge that needs to be, you know, you know, packed into these models for those types of use cases. And that's, that's more like basic knowledge work. Now you go into life sciences, you go into, you know, material sciences, you go into healthcare, you go into anything dealing with kind of manufacturing, and we still have massive sort of leaps to go.

10:21We're still trying to cure all these diseases, but you know, it is interesting. I do think it's mostly head of doctors now in a lot of diagnosis areas, which is kind of crazy already, you know? Yeah. Yeah. So, so I think for, again, the, the kind of common things you would go into, you know, a doctor for AI plus doctor and, and people debate this as like, is the AI, you know, doing most of the work is that the doctor's judgment on the AI, but, but you can still eke out kind of, you know, positive gains by having doctor plus AI. But, but now I think the real frontier is, is sort of this diffusion of AI in enterprise workloads that actually are, really driving meaningful change in the real world that would actually affect our lives.

10:59And what would affect our lives? It would be if life sciences companies can have, you know, breakthroughs on, you know, a new cure for an unsolved problem. It would break, it would, you know, we would advance our lives if you could have automation breakthroughs in the manufacturing process. You know, we would have more breakthroughs if we had better, you know, software and AI and all of our kind of common ways of interacting with the world. So that's kind of the upgrade cycle that needs to happen. And the only kind kind of caveat is like that's going to probably take like 10 to 20 years um not just because of model progress but because of the role out of interesting because i'm because i'm seeing a lot of things for example one big thing is we bring manufacturing back to the u.s right and there's all these really hard ai problems to do what the manufacturing engineers in shenzhen do and a lot of my friends think we might be able to do a lot of that in the next couple years i guess that's a question are there going to be things along the way that we're solving that that are helping us or you think a lot of it's further out i actually think this is and this is the where this is the spot where you would you would start to veer between the kind of um either like hyper accelerationists with ai and what's funny is actually the accelerationists and the doomers sort of share roughly the same timeline because they they both think that ai is the like intelligence is the only thing that matters and then i'm more on a prag pragmatist sort of timeline which is like intelligence is is a super critical component but then diffusion actually is like your biggest rate limiter and diffusion is like a human has to get the intelligence from the model plus probably their data and then they have to go interact with the real world and do something in the real world and then bring back whatever the feedback loop is back to to the model and the data and that's actually your ultimate rate limiter which is just like the speed at which humans can actually make sense of of what the models are telling them to go do the ability to go do an actual study on humans with a with a new kind of you know biomedical research the ability to actually get the permits to build the thing that the AI helped you accelerate building.

12:49That's all of the kind of limitations of the real world. Right. So maybe, I mean, this would be an argument for like, become the state that sort of lets you actually do all these things. We're actually thinking of putting the permitting into like AI land and having the rules be cleared by humans, but then letting the AI like to go really fast. Yeah. Because you have to go fast. A hundred percent. So think about just even the paperwork kind of processes that slow down the real world. If you can automate all of that, obviously then you could shave things like you know a permitting process for maybe months or you know years in some cases to a matter of days or weeks um hours hours sure there's always this other bottleneck that something like pops up to because there's an escalation that somebody has to go look at and they have to talk to somebody else so i think that like the the really bold bold case on ai is is if you think about it as like everything that is sort of information constrained in a process you should just be able to automate away like like the reading of something the processing of something, the typing of something, anything that was slow in a process because of that, we should be able to go in and have agents go and automate.

13:50And then the only limiters should be like atoms and like human kind of coordination. So you're on my team that bureaucracy is the last great challenge for our civilization. I do. I am. And, but I, maybe I, I overweight that as like, that will just be like an impenetrably, you know, hard problem that will be around forever. Well, we can't be too mean to them because they fund like one side. So, but no, But stepping back, so you're seeing all this data in all these enterprises. Yeah. And so you're saying this problem with diffusion is just like people things. Like, what does that show up as in most of the companies that you're looking at?

14:19I think what it shows up as is you go into an enterprise and they want to go and automate their client onboarding process. Or they want to go and automate their, you know, clinical research process. And they have to, you know, first it's like, well, do you just like deploy, you know, everybody has their little AI agent. in their own personal workspace that they interact with? And that's an interesting strategic question. Or do you deploy something that's sort of centralized, that is like the new machine in the workflow? When we deployed manufacturing plants, we weren't like, oh, let's just give better tools to every single person on the manufacturing line.

14:58You step back and you said, no, let's actually figure out what is the new whole manufacturing process going to look like. Most companies are still in that kind of former process, which is, OK, we're going to enable everybody with tools and that will accelerate each individual, which is incredibly powerful. Or even each department with its own separate areas, but they're not working together. 100 percent. And so and so now you have this issue, which is, well, every single person now has been accelerated in their own kind of domain. But that probably didn't like reinvent the process for a world of agents.

15:27So then you have somebody, you know, wake up and they say, oh, well, what we probably need to do is make sure that we are reinventing this this whole process. end to end, the client onboarding process, the clinical life sciences process, the M &A deal review process, which means we need more of a central system that can both have access to the right data to work with, which is its own very hard problem, and have the right sort of input and outputs for the humans to interact with that workflow. So my latest bias, which is like super obnoxious to big companies, is that it's so hard to re-engineer how some of these things work based on what's possible today, that they're just going to get crushed by new companies started by younger people who are like maybe not even younger necessarily but people who are just doing things fresh in the new way like freshly designed the processes yeah it's like the whole idea like these ai services or these companies just coming in like building competitors i think i think there's gonna be a lot of new big companies right now yeah so uh i i am 100 bullish on that thesis um i with one with one sort of caveat which is um i i i and i tend to now think that what happens is you end up with more of an ecosystem which is which is you know still like on average the the big incumbents like eventually find a way to kind of you know break through especially the capitally intensive ones like the guys that own or regulatory you know control ones like like is is jp like am i am i that dependent on the speed of like you know getting a wire transfer done that i'm going to change my entire bank because because of of you know one bank has agents and the other is a little bit slower maybe not like there's probably like more like i kind of just want to make sure that that what i am on the board of airborne with palmer okay fine okay well then really fast okay then but jb diamond's probably not gonna lose his job and then i think that's the point like like it like you could imagine i'll i'll grant you that that could be a half a trillion dollar company and that would be great but like does jp morgan go out of business you know not sure no problem so so um and we saw this before like like basically I had, you know, I was very, you know, had a very big thesis on like digital disruption in like the early 2010s.

17:29And we talked about this a lot back then, which is like, okay, you're gonna have like a digital version of every single existing incumbent company. So in the early 2010s, you had like Robinhood, Instacart, DoorDash, etc. All the Airbnb, all these incredible digital disruptors, they all became fantastically large businesses. And yet the incumbents, like they still exist in every one of those categories. Job and hotels are still around. And they're serving some purpose. So I think actually both will be true. But I love the idea of the kind of agent first, law firm, bank, life sciences company, new financial services firm.

18:06So this is going to totally happen. And I think it's going to present a huge opportunity for all new startups. What's cool is in AI right now, for the diffusion of AI, you actually need, like, I don't know, there's probably like five different layers of the stack and they're all kind of working at the moment. So you have like the agent first company in that domain. You have the, you have the company that provides agents for all the other companies in that domain. You have the infrastructure providers for the agents in all of the domains. You have these sort of post-training kind of agent, like model, model, you know, um, uh, training companies for all the domains.

18:38And then you have like data and infrastructure and everything else. So it just turns out there's opportunity kind of across the stack for all of these players. But like, Like if you're a cognition, you're going to both arm the new disruptor of like, let's say system integrator companies that are doing software development. And you're going to arm the existing system integrators. Like your job is to make sure both end up growing, you know, fantastically well. And I think there's going to be opportunity for both of those plays. It's a very positive sum view of the world right now. I'm not seeing a lot of zero sum actually.

19:05And like across the board, even with jobs. Like I think actually, I think actually this is a technology that largely ends up being positive sum. I love it. Well, it's called American Optimist. And I'm, you know, I'm on the same page as you. I do think longer term, especially with robotics, there's some disruptions that are very scary, although I think it's very positive overall long term. Yeah, I think the thing that you end up using robots for are going to be a lot of additive things that we just don't do today. Like you're just going to end up getting an abundance of a whole bunch of areas of work done that just were never possible.

19:34Like the thing, some of the top use cases we see from customers with our AI agent and their data are actually things that they never deployed people at before. So it's, I want to go and have an agent read every contract in our company because I want to pull out insights from these contracts that will tell me like, what customers can I go and upsell better? Or where is there a new sales opportunity? It's an entirely new set of work that's being done. Nobody read those contracts previously. So agents haven't replaced anybody's work. And as a result of what the agents are doing, the company now actually is driving more revenue.

20:08They're driving more opportunity. That's going to happen in so many more places than we realize. 100%. I think it's a really fun thing. if you think about your talk and someone from 100 years ago and you describe these as robot servants or something, and then it would be like absolutely comical to them, like how much these robot servants are doing for you for like basic tasks. You're seeing this across the board. One of the new ones is, by the way, if you're in a marketing department and you're like creating material for a customer, you now need to like fine tune it if you're doing your job for every customer explicitly for their use cases, right?

20:37So it's like you're making the robot servants spend like a week to get ready for this one customer email. 100 so so this is the funny thing so uh so and and bill gates and andy grove had this dynamic um of like you know the cpu would get better and then windows would just kind of use all the cpu again yeah um and so there was just this uh like i think it was like andy grove giveth and and bill gates taketh or something so so um uh that same thing is it's kind of happening for like work which is which is you you sort of uh you would assume if you just like you know didn't have any intuition on this and just like in a, you know, kind of looked at a basic model, you would say, okay, this company currently produces 20 marketing assets per week, or this company reviews 100 contracts per week.

21:22Now agents come in and they do those, both those tasks. You'd be like, well, the people that were doing those tasks before clearly can't, there's no work for them to do because the agent just did it. Well, what happens is the economy starts to say that those two tasks no longer have much value. Like we, we end up raising the bar of then what that job is. So to your point, like you don't just now do like, oh, we're going to just like make five ads in this marketing campaign. You make like 5 ,000 ads because you tune it for every single geography, every single market segment, every single industry you're going into.

21:52And the reason you do that is because as a customer, you no longer expect that, that like, just like the basic ad campaign or the basic targeting is going to work anymore. So like all of our expectations have just risen as a result of AI. And we see this actually, we see this in sales as an example. So one would have thought that AI would help every sales rep prepare for every single meeting and instantly make sure that you're super tuned into the meeting. And that means that that's going to reduce all the jobs. Because what about all the work that we were doing previously? What about the solutions engineer that built a, you know, kind of showed a demo or, you know, ended up making, you know, marketing collateral.

22:36Well, guess what? The new thing is you go to that customer and you actually build a working demo prototype of the product that you're actually pitching. And, and so now like you wouldn't just like as a customer, you wouldn't buy software if you didn't see like a working prototype of your environment with that tool. So what we did was AI agents made the previous task really efficient to the point where the previous task is kind of like not differentiating. So then we go and deploy our time at the new much higher version of that task because the market now expects that I think that's going to happen in like so many more areas than I think what people are anticipating So if you if you believe in job destruction like in any kind of very large-scale way You basically don't believe that markets adapt and you and you basically don't believe that some other company will emerge And say i'm going to just do this way better And the way i'm going to do it way better is by adding people time again into the process with really, really high judgment.

23:27And they're going to go and find a way to use these tools to differentiate my product or my service or my marketing better than the competition, which then raises the bar for everybody else. So I obviously agree with you. A lot of people right now are afraid to talk about this in public. There's a lot of doomers. There's a lot of people maybe who don't have the economic imagination and understanding that you do. The polling is getting a lot worse on AI. Bad polling, bad polling. Both sides right now. New York state just imposed a data center moratorium yeah there's red states where they're weighing bands i haven't done it yet we're fighting hard but it's but it's but i so people are really negative they're really worried like i mean are we just like really unpopular nerdy people they don't like and therefore they're coming for us it doesn't matter what we say or like is there is there a better that's probably true no matter what yes like like in all cases that's obviously true obviously we have different politics but we're both part of the whole evil tech tech bro regime it turns out it's a very Bipartisan problem.

24:18Yeah, it turns out everyone dislikes us. What should we be doing and saying? What's the right messaging? Like, what do you want tech people to be doing about this? Well, you know, and I think we probably also agree on this. This is sort of a self-inflicted problem from the industry. And, you know, it's just super obvious. We have been scaring ourselves about this technology for the past decades, obviously. But even in the past kind of four years since the Chachaputi moment, we've been scared of the next model at every single turn. And we have shared those fears with the rest of the world. I'm saying we very politely, you know, as an industry, I've, you know, I think, you know, both of us, we've taken the other side of this conversation, but the industry has kind of let out a very kind of neurotic and anxious kind of feeling about this technology.

25:05It's very antisemitic. And I mean, there is an interesting correlation between the participants, but it's very like, Imagine if Larry David made AI models. That's kind of what we're dealing with right now. So the issue is we've sort of scared ourselves about the AI. And we've told everybody those fears. And so then a lot of people are sitting around being like, wait a second. The actual creators of this stuff are scared shitless about it. Maybe we shouldn't be that into it either. Why is this all happening? And it's a very kind of like a weird cognitive dissonance, which is you're telling me this thing is like going to destroy the planet and the world and you're the ones making it and you're not stopping.

25:48So now I'm just supposed to be very restless about this whole, you know, kind of set of facts. Um, that, that's, that's kind of the industry. And you know, when you have people like, uh, you know, and this is like, you know, I think everybody sort of attributes this to maybe one company or group, but it's a, it's a pretty broad thing. Like it goes back to like Jeff Hinton, like we'll, we'll tell people that all radiologists are going to lose their jobs or whatever. Um, and it's like, well, that just didn't happen. It actually just turned out that we just threw more compute at the radiology problem and you still need people that are in the kind of human and review, you know, loop on that.

26:18So we have a history as an industry with this particular segment of technology of, of sort of not using our imagination about what happens next. And then, and then, you know, kind of publishing all of our fears about it, thinking that, that maybe if we do that, that will kind of create some kind of broad societal progress. And, and I think there's actually an argument that, that it will like, like I I'm actually very favorable to the idea that these are open discussions, but we're getting exactly what one should expect, which is very low popularity about this technology. So then that's going to have a wide range of implications.

26:52It's going to mean data centers don't get built out. It's going to mean that we sort of lock down AI models. It's going to mean potential major new kind of tax regimes, depending on the administration that's in charge. And I think a lot of it will have have sort of been exactly what you should expect to kind of play out. And I don't know, unfortunately, like, I don't know exactly how to change it, because the way to change it is, is, is just everybody experiencing kind of universally positive, you know, outcomes from this technology, and us, you know, putting a lot more emphasis on those, while still mitigating all of the risks that we're afraid of, like, there are real risks with AI, there's no question.