Grateful to Dropbox CEO Drew Houston for joining Thursday Nights in AI to discuss the power of LLMs, AI in the workplace, and seeing past the AI hype. Our top takeaways…

On today's AI boom:

“LLMs are like 10 breakthroughs in one thunderclap. For most of the history of computing, computers couldn’t really parse a sentence. An LLM can parse text, read and understand text, generate text, and code. Any one of those things would have been considered a breakthrough that would have kept us busy for five years. But they all happened at the same time. It’s like the Amazon guy dropped off an alien artifact that is LLMs and we’re all kicking and poking it, seeing what noises it makes.”

On distinguishing real vs hype:

“People have been predicting what AI can do for a long time, more or less correctly. But they often didn’t come true until 10-20 years later… That’s the prism I look at AI through: a lot of cool demos, but more discernment needed to figure out what can really scale.”

On how AI will affect knowledge work:

“I’m interested in how we can re-segment knowledge work so that our human processor can be doing human things and the silicon brain can be doing the other things. We're sort of in an environment where we just invented this other kind of processor and now we need to figure out what we can offload to it.”

On the opportunity for Dropbox:

“ChatGPT is amazing, but there are whole categories of questions that ChatGPT and generic AI can't answer. and it is limited when it comes to answering personal questions. Dropbox is well-placed to build personalized AI connected to your information in a secure and trustworthy way, that can help folks make sense of their working lives more broadly and build more of a self-organizing Dropbox."

 On how to build while the ground is moving quickly beneath you:

“You need a portfolio approach. For us, we think of it in terms of baskets: singles, doubles, home runs, more speculative ideas, more sustaining ideas. This way, we can make small investments and double down on what works. We have a talented team, so we also built our own internal Y Combinator called AMP to explore different ideas. We take a small team and let them rip. Sometimes they explore my ideas, sometimes it’s what they want to do. We want to explore a little bit of everything.”

Watch the full interview below:

Drew Houston: @drewhouston

Ali Rohde: @RohdeAli

Kanjun QIu: @kanjun


This transcript was edited for brevity.

Q: Can you tell us more about Dropbox Ventures and Dropbox’s new products, Dropbox AI and Dropbox Dash?

Drew Houston: Dropbox Dash is an AI-powered, universal search for all your cloud stuff. I was thinking, ‘What's the version of Dropbox I would build in 2023?’. You see, I started the company back in 2007. My stuff was everywhere. I couldn't find it. Back then I had all these files and all these different devices and operating systems that didn't talk to each other, and the solution looked like syncing all your files to the cloud. That’s how Dropbox started.

Now, when you look at the overall system, with all the fragmentation of tools, basic experiences like search are worse today than 20 years ago. If you think about it, like 20 years ago, you want to find something you just search your hard drive. It's there on your email or it's not, but then, we've ended up in this world where it's like, I have 10 search boxes that each search 10% of my stuff, I have to run search ten times. There were a bunch of other issues we could talk about, like how often do we have that experience where we got 100 tabs across the top and we go, "I know I was looking at this document yesterday or an hour ago, and I think I'm typing in an exact keyword match into the Chrome search button. It's not finding it." So like all these little paper cuts like that. Our solution is Dropbox Dash, which is basically building the missing organizing layer for all your cloud stuff. So it lets you search, not just your files, but all your cloud content, your Google Docs, your Notion Docs, all your communication channels, Slack, your email, everything. We’ve launched Dropbox Dash into close beta, reception's been good so far. It's very early. We're just continuing to iterate on the experience and getting ready to tune our growth and monetization and all that.

We also launched Dropbox AI. So our internal code name for it was FileGPT. So kind of as it sounds, basically if you have a big PDF or you have a big slide deck or if you have a big video, we will transcribe all videos and you can basically search, ask questions, and/or summarize your content, pretty much all of the usual things you'd expect.

And then we also launched a venture fund Dropbox Ventures to get more involved in the community of startups and kind of the Cambrian explosion of innovation that's happening.

Q: What’s surprising in the world of AI now?

Drew Houston: I think certainly the biggest surprise is ChatGPT being like the iPhone launch moments of the AI era in a way that it captured everyone’s imagination in the last seven or eight months.

From a Dropbox perspective, the magnitude of this new opportunity is a great surprise. ChatGPT can do pretty amazing things, but you realize pretty quickly, that there are whole categories of questions that ChatGPT or any ‘generic AI’ can't answer because it's not personalized. So if you ask ChatGPT, “What's my passport number again?” Or like, “Where was the launch deck from last year for the spring launch then?”. For things like that, ChatGPT can't answer because it's not connected to your information. So there’s this opportunity in front of us to build personalized AI, where you can be connected to your data in a secure and trustworthy way and help people make sense of their working life more broadly and build more of a self-organizing Dropbox.

Q: What are you prototyping right now, and why are you still hacking despite running a giant company with lots of revenue?

Drew Houston: Yeah, I guess a little background. I started out as that little kid with a PC junior in my living room, just like a glowing screen and all these buttons, and fell in love immediately for life. I started playing computer games, then I learned how to code, then I wrote my own computer games. At first, I wanted to be a career game developer. That was actually my first programming job. I've always been an engineer. That has always been my first calling. And then I knew I wanted to be a startup founder of some kind. I wasn't sure about being a CEO, I sort of backed into that.

Kanjun Qiu: I remember we have a quote that's like: "The journey from founder to CEO from engineer to therapist."

Drew Houston: Yeah, engineer to therapist, engineer to politician, engineer to…. It's, you know, there's a lot of different hats. But when you inhabit the job of a CEO or executive or even a manager, there are a lot of things you have to navigate. There's a lot of repetitive shit, and I'm like, man, I really don't enjoy the process of doing all these manual things over and over again.

So in the last several years, I've been trying to find new ways to automate my job. Around 2016, I wanted to better understand machine learning, since it seems like the next step after my computer science degree. Now, I have been doing things like using machine learning to write little scripts to triage my email, like which of these emails has a request, or which ones are important, in some definition, which ones have some kind of invitation, what do I need to respond to not respond to, etc. I also use ML to classify my calendar and figure out where my time goes, kind of like Mint for your time, and then the list goes on and on.

I love engineering and coding. I write thousands of lines of code a year still, not in production, but like when prototyping. I think it really helps me be grounded and build conviction around like what's real, and what's hype.

Q: Speaking of real vs. hype, how do you distinguish what is real and what is hype?

Drew Houston: I just turned 40. Well, there's not a lot of good things about getting older, but you do get some perspective from having seen multiple cycles, right? For example, when the dot com thing was happening. I was just finishing high school, and entering college. The iPhone launched right after I graduated and the cloud AWS launched in 2006. Those things made Dropbox possible.

When the internet came along, people had these Star Trek ideas like you’ll be able to communicate, and do all these things that sound like science fiction, and often, it did come true like 10 to 20 years later, and we are using these technologies now in our daily lives. I think scaling it to that point is what makes it real. Like, you have these amazing demos, but then certainly anybody building with AI is like thinking it's like easy to make the demo. But then, how do you scale these things up to production? How do you make them reliable? How do you make them secure? Things like that. And then the markets like, who's going pay for what? What's the whole value chain going to look like? Are there any traps? There are a lot of cool demos, but then more discernment is needed to figure out what can really scale.

What's amazing about large language models is that there are 10 breakthroughs, in one thunderclap and I think it helps me to kind of like try to slow down the tape and be like, well, there were a lot of things that just happened with that thing. For most of the history of computing, computers couldn't really parse a sentence. Now, a large language model can parse text, then after that, it can also read and understand text and or understand the meaning, it can then generate text, can code, knows about the world, and has like theory of mind type things. It can do creative tasks that would have been reserved for arts. Any one of those would have been considered a breakthrough that would have kept us busy for like five years, but they all happened at the same time and there's all the multimodal stuff so something I'm very excited about, is the combinations of all these things. This is the world we now live in.

It's sort of like the Amazon guy dropped off this alien artifact of a large language model, and we're all kind of kicking it and poking it. seeing what noises it will make. It's a really cool time.

Q: What are your ideas on automating some aspects of your job with AI?

Drew Houston: Yeah, I use AI or large language models a lot for writing. I use Whisper for audio transcription every day, this morning I was babbling into my computer to make a very raw transcript. I have one pass that cleans up the transcript and makes it a little more coherent and then back and forth…pretty much a peculiar writing process that's very much assisted by large language models. Any kind of repetitive task, right?

Way better to your point, usually, I'm not giving super remedial feedback in a meeting, but sometimes, we'll be like, all right, if you're writing a business case, there are like 50 things we care about. It's like pretty cognitively difficult to keep all 53 things in your head but like, that's no problem for a computer. So they can kind of check through using large language models to check anything where you need a lot of conscientiousness and discipline and detail orientation. LLMs are going to be really good at that.

I'm really interested in like, it’s 2030. How do we re-segment knowledge work so that our human processor can be doing human things and the Silicon brain can be doing the other half of the things? Right now, we're sort of in an environment where we just invented this other kind of processor, like a Silicon brain. And now we need to figure out what we can offload to it cause there are a lot of human tasks that really should be done by computers such that computing went from being a person thing to a machine thing, from a human verb to a machine verb.

Q: Where do you think the value will accrue in AI? Will it be toward the incumbents? or to new startups?

Drew Houston: Yeah. I remember this book by Clay Christensen called The Innovator's Dilemma. It's a classic. I highly recommend reading for everyone here. He talks about some theses like, there's sustaining innovation, and there's disruptive innovation. Sustaining innovation is where you're taking an existing thing and you're just making it better, faster, cheaper, whatever and a lot of what AI does is that…same products, same customers, same business model, pretty easy to adopt. But then disruptive innovation is different in that it's like something that makes life difficult for the incumbent, right? And there are a lot of classic examples of this. I mean the iPhone was disruptive, right? In all kinds of ways, but like, compared to the old, but using like a computer operating system in a phone and all these other things. But then disruptive innovation can also happen in tech and maybe not obvious ways.

I think AI is going to accrue a lot of value for the incumbents since sustaining innovation is easier. I'm excited about making things AI-native. For example, during the mobile-native era, Facebook was able to take and put it into a mobile app pretty well.

I am also interested to see how disruptive ChatGPT will be. Certainly, people are making the argument that search is going to be disrupted because it's a lot more convenient for certain kinds of tasks. Like, I don't need to go through Google or Wikipedia and, you know, rabbit hole to figure out what I want. It does not only disrupt the product but also the business model right? So the 10 blue links, auction, like that stuff kind of goes away or at least becomes different in a messaging context. I think it'd be super interesting.

Going back, I think folks who own the customer relationship will also have an advantage. That's what we believe in the application layer. Because if you're the thing they're looking at, then you control to some extent, like all the steps that happen after that or all the layers in the stack below. I think you could see that aggregators in general, have a lot of power. For example, Kayak or FlightSearch. None of them operated an airline, but they're able to capture some of the value and create value by getting in front of the customer and providing a simpler way to search flights instead of logging into 10 different airlines. So I think you'll definitely have the top of the stack, whoever owns the customer, and then the bottom of the stack who controls the hardware. NVIDIA has a sort of a monopoly right now on the GPUs that we need to train these big models, but we can debate how long that's going to last.

Foundation models have commodity characteristics because they're easy to swap out to some extent. Not at all foundation models are created equal. So you have different sizes of costs, latency, like IQ, like it wrote, you quality context length, specialized models, like what Bloomberg was doing or medical or, and so on. So you can differentiate on a bunch in a bunch of ways, but then LLMs have a particularly tricky thing where you can use their outputs to distill or bootstrap, your in-house models. So it's sort of like you give the blueprints of the model to some degree, just by putting the results out there. So I think that makes it easier. It makes it hard to have switching costs and durability. I think it is already pretty kind of bloody in the foundation model space but if you can be the number one on those domains, if you can durably hold that land, you can generate a lot of value, but otherwise, yeah, I think like my engineer training was helpful for some things, but then sort of my business training on the job ended at the end is another important lens you need to add.

This event is brought to you by @OutsetCap and @imbue_ai

Watch video on YouTube