Drew Houston: Achieving Viral Growth and Battling Big Tech

Oh, it's great to be here.

Yep, about five years ago, a little under five years ago.

It was Facebook at the time, and then they changed the name a couple years in.

Well, it's been great to watch. And I won't speak for Meta itself, but certainly like seeing the rise, or seeing the power of open source and the impact it's had on democratizing AI has been really powerful. And, you know, whenever there's a new era of computing, there's a question of, is it going to be open, is it going to be closed? And so open source has been super powerful, clearly, in terms of making AI more accessible, and in terms of both being able to use it, build products with it, research. I think, on balance, it helps with things like safety because you have a lot more people working on that problem versus a situation where you know one or a couple companies are the gatekeepers to everything. I think there was definitely an alternate future where we'd all be kind of paying rent to a couple big companies, but instead we've seen the price performance of these AI models plummet, you know, like drop 10 or 100 times better price performance every year. And then for a product developer like me or like Dropbox, that means what we can build is a lot more exciting and more affordable, and things that would have been kind of priced out as being too expensive to do that much inference at our scale suddenly become viable. So it's been awesome to have that vantage point too.

For me personally, it started to heat up. I mean, I started playing with more classical machine learning several, you know, mid-2010s, because I didn't... I wanted to go to, or I at least thought about going to grad school. Did my undergrad in computer science but never did the grad level courses, so I tried to self-teach classical machine learning to see what that was about. And then certainly as deep learning, large language models came on the scene, then there was that thunderclap, iPhone launch moment of ChatGPT. And when they had these instruction-tuned models in GPT-3, then I was a little embarrassed. I was on my honeymoon coding and playing with these things, where I was like, oh my God, it's happening. All the stuff I wanted to build, you know, back when I first started studying this stuff, I hit all these roadblocks because computers basically couldn't understand text. But then the large language model was like, not only does it understand text, it can write text, it can write you JavaScript and then write a sonnet about the JavaScript. I'm like, wow, a lot just happened in like one minute. What seemed like one minute. Very unusual person that could do all of those things, which is powerful to the AI side. So it was a big wake-up call.

Yeah, I mean, the thing I'm most excited about is we sort of have our human intelligence, or our human brain, and then there's this new silicon intelligence, or silicon brain. And just as you saw in computing where adding a GPU to the CPU suddenly made all these new things possible, like AI in our working lives, we're going to have this really kind of wild partnership with our... as we plug in the silicon piece of our brain. And so I think a lot of work will be reimagined, where we'll be able to offload a lot of our busy work and be freed up for more creative or relational tasks. You know, I don't think anybody can really predict exactly what the precise route will be through the fog here, but I think we'll be able to have... or I think we're opening the door to where anybody can be a 10x person, and people that really figure out how to master these tools can be 100x people. So I think it'll be certainly the most transformative change in our lives.

Not a lot. I mean, I'm very excited about voice, or the ability to have sort of an infinite memory in all these different contexts as you're using your computer or your phone, for sure. It needs to become more socially acceptable. I think a lot of those things are very cool, but there's some social element of adoption that exists in the real world.

We do.

So we established pretty basic principles around transparency, around how we use AI, things like privacy, safety, kind of all the... setting forth all the things that Dropbox will and won't do. Because, I mean, for a number of reasons, but one is, you know, customers, and I'd say all of us, are really excited about the good parts of AI, pretty concerned about the things that can go wrong. And then when you intersect that with your stuff, your most important information, your personal information, your company's information, you care a lot about making sure that you can trust your counterparties or the services that you're using. And we had some experience around this with Dropbox 1.0, because people for the first time, or a lot of people for the first time, were putting their most important information into the cloud. And this happens a lot, you know, whether putting your money in a bank instead of under your mattress, or putting your credit card over the internet, or putting your files in the cloud. People start out with a lot of the same apprehension, which is very justifiable. And so this was a way for us to get out in front of that, say here's how we'll use your data, here's how we won't, and reinforce one of Dropbox's big advantages, which is that our incentives are much more fundamentally aligned with our customers from a trust perspective. Because I think a lot of people worry about, you just think about how these models are trained, it's like, oh man, is this service going to take all my stuff and sort of grind it into little pellets to either sell me ads or to train their next foundation model, or have some other kind of undisclosed use of my information? We wanted to draw a bright line and say no, we're not. You know, I think everybody's looking for providers they can trust, and so setting these principles affirmatively was a step we took pretty early on, and a lot of companies have also published similar principles.

Yeah, the most notable one is we have a new product called Dropbox Dash, which is really centered around universal search. So the basic problem we're solving is, you know, I know that thing exists, I can't find it. And this question of, why do we live in a world where it's easier to search all of human knowledge with a Google search, and then when I go to search for my company stuff, or search my own information, I've got 10 search boxes and it's a much worse experience? And then weirdly, despite all this technological progress, this problem is much worse today than it was 20 years ago. 20 years ago, if you wanted to find your stuff, you just searched your hard drive, right, or your email, so maybe one or two search boxes. But now we've had this kind of gone wild. So Dash is really about fixing that, giving you one search box that can search everything. So it'll search your Google Docs, your Slack, your email, your files, but really anything. You don't even need files in Dropbox, so this can be a completely standalone thing or integrate with your Dropbox. But we see that as a very fundamental challenge that a lot of knowledge workers have, especially as we've moved into these distributed ways of working. Like, where's the information I need to do my job, and all the paper cuts that come along with that.

Enterprise search is a market that's been around for a long time, and it sounds like this is both consumer and B2B elements of it. We're focused on work use cases. Yeah, so on the one hand, enterprise search has been around for a long time. On the other hand, similar to cloud storage, when we started, there were a lot of other things that sort of claimed to do it, but in our view, none of them had really done it right. And I think there's a similar dynamic that we've seen with how enterprise search has evolved, where yeah, it exists, but no one really... when you ask the average person, do you use any of these enterprise search products? They're like, no. And, you know, have you tried them? Yeah, it didn't really work very well. And so certainly the generative AI allows you to not just do search but really get answers. And so Dash is also... we also view Dash for a lot of the questions that ChatGPT can answer, because it's not connected to your stuff. So if you ask, you know, if you or I ask ChatGPT a question, we'll largely get the same answer. But then if I want to ask, like, when does my lease expire? When does... where's the slide from last year's product launch? Things like that, you need a product like Dash, which is connected to all your stuff and is grounded in all of your content.

Absolutely. So with search, well, there are a couple aspects. I'd say when you sort of double-click on this specific problem, there's... people often have retrieval use cases where it's like, I know there's a thing, or I'm looking for a specific asset, and I actually just want the thing. I actually don't have a question I want answered. Or I want an answer to a question, like, help me... I'm really at the end of the day looking for, okay, what is Dash's roadmap? Or what's the release date? Or what are the key features? Things like that. So it's a question. And the blue links kind of format is actually pretty good for retrieval, you just want to get the thing quickly. But then when you're discovering things, or you don't know the answer, don't know where it might be, then natural language fits in a lot better. And then there are also some of the underpinning technologies of large language models, these embeddings, or a lot of semantic search, neural search, fuzzy search, vector search, a lot of names for it, same thing. But basically, one of the unlocks was you can now search without having to match the exact keyword. So if you say, like, oh, what's the 2025 strategy or the 2025 plan, these vector search can help turn up the right results for both of those things within the proximity that are close to it. Rather, it could be you could say 2025 plan, and if it was only language-based, it might show nothing if it was actually called whatever company estimates through 2030, but it can get directionally there. And, you know, when you implement these things properly, these can also be learning systems where they're always getting smarter based on how you're... it's becoming personalized to you as you use the service, and then just smarter overall as the world is using the service.

Yeah, lots. I mean, you know, my first love is engineering, and I grew up as a little kid coding. And so when the large language model came along, around the ChatGPT timeframe, I was suddenly like encoding like an 18-year-old again. And I'd say even before that, when I was learning classical machine learning or the pre-deep learning techniques, there's lots of stuff that I wanted to automate. And, you know, even my career, I started out as an engineer. I knew I wanted to be a startup founder, wasn't actually sure I wanted to be a CEO, sort of backed into it. But as I really inhabited that role and the company was really scaling, I was like, man, there's a lot of tedious stuff that managers or executives do. So, yeah, I would have all these little tools to audit my calendar or figure out which of these emails needs a response. Or, actually, the motivation for Dash from my perspective was like, I can't find my stuff. I need a better way of managing this personally. I mean, forget a product. And then I wrote... and I was like, oh, and there's these embedding-based, or you know, these vector search is a new thing. And I built this little prototype of a personal search engine, and I'm like, oh my god, it works, it's super scalable, it's super fast. Everybody's going to be using something like this in a few years. This is maybe 2018, 2019, and I'm like, we should be all over this. So actually, I still write many thousands of lines of code a year,# and that's one of the key ways that I really get more of a tactile feel for, here's what the technology is, here's what it can do, here's what it can't, and this stuff's fully baked, this stuff isn't. And, yeah, I mean, I think it wasn't just my prototyping that led us to Dash, but that's really what caused me to have a lot of conviction pretty early and put a lot of chips on the table.

Yeah, there's a lot of good stuff, and I'd say the best stuff is really timeless. And often the principles are pretty straightforward to describe. And the ones I lean on are some of my favorite books. High Output Management by Andy Grove is one of the best books on management ever written. Another is The Effective Executive by Peter Drucker. I could give a lot more, but just on the sort of productivity and effectiveness angle, I think those two probably get you 95% of what you need to know from a theory perspective. But what's really hard about it is applying it, right? Because it's sort of easy to read the book, and then, you know, it's probably a super disappointing fraction of people actually apply the lessons. And it's just kind of this constant fight. So a lot of my use of AI was really to automate some of these things. For example, one thing that every busy person struggles with is, how am I using my time? And the first chapter after the introduction of The Effective Executive, the title is Know Thy Time. So what you do is you basically do a time audit. And he's like, okay, here's all the stuff in my calendar, and I'm going to categorize it, put it in these buckets, and then you line it up to your priorities. And so, you know, I read the book a long time ago, but 15 years ago I probably did my first time audit with my admin at the time, and I was like, oh, I probably spend most of my time on recruiting and on products, I don't know. And then after the audit came back, we tallied it up and we're like, actually I spend my time on just about everything except for recruiting and product. So it sort of highlights this... it's one of these areas where it's not just that we don't know where our time goes, it's actually worse than not knowing. Where it actually goes and where you think it goes are often wildly disconnected. And that's one of the first findings in the book, is that people are usually shocked by a time audit. And so, you know, it's something you have to do periodically, but it's a really manual process, and so it's something that lends itself to automation. But yeah, all kinds of things like that. How am I using my time? Am I focusing? How am I keeping on top of my communication? Personal search. Yeah, there's a lot of little code repositories in my little bag of tricks.