AIで組織内のデータをつなぐ：Dropbox CTO アリ・ダスダンが語る Dash とコンテキスト腐敗

Connecting the dots between data sources across your organization to enhance productivity sounds promising, but how does it actually work? Find out on today's episode. I am Ali Raspin at Dropbox, and you are listening to Me, Myself, and AI. Welcome to Me, Myself, and AI, a podcast from MIT Sloan Management Review, exploring the future of artificial intelligence. I'm Sam Randbotham, Professor of Analytics at Boston College. I've been researching data, analytics, and AI at MIT SMR since 2014, with research articles, annual industry reports, case studies, and now 13 seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policymakers join us to break down what separates AI hype from AI success. Today, my guest is Ali Dazen, Chief Technology Officer at Dropbox.

Most people know Dropbox as a place to store and share files. Now Dropbox connects organizational knowledge to AI and gives teams a landing spot to review, collaborate, and advance their work. What I find interesting about Ali is the path that he took to get here. He started in chip design, moved to web search, and now technology at a 20-year-old company. And I think he's unusually straight about what AI can and can't do. So let's get into that some. Ali, thanks for coming today. Happy to be here, Sam. Thank you for inviting me. Of course. Let's start with a quick version. What does the CTO of Dropbox actually do every day? Multiple things. Definitely setting the, you don't do it every day, but setting the technology vision of the company. There's a big team that I manage, product development, responsible for platform and security and IT and all aspects of that. Day in, day out, it's always busy. Issues come up, product updates, sometimes incidents happen. There are exec team discussions that are happening weekly.

It's not the same every day, but very busy. I got familiar with Dropbox years ago, starting synchronizing files. I admit that I constantly, it's a fundamental part of how I work and keeping all my machines in sync. But up to now, I haven't seen a lot of artificial intelligence in the history of that project. So tell me about Dash and how that works and what you're thinking there. That's true. Yeah, I mean, Dropbox being a file sharing site, you know, a storage site, it's very fundamental to so many different businesses, especially in the businesses that deal with lots of media files, for example, construction companies or advertising companies or creative space and all. I think with the rise of Gen.AI, there was a focus on bringing AI to solve our customers' problems. And Dash started with that. Very simple example is you talk to one of these customers, they have a huge number of videos that they have stored. They have difficulty organizing them, difficulty to find what they are looking for.

It's not only sometimes finding the video, but finding a certain frame within a video. Now with Gen AI, actually, many of these problems have simplified, and we ended up actually bringing with Dash those kinds of capabilities. And then this is basically for the content on Dropbox, but going beyond that, we realized basically our customers need, when they do search or their work, they need far more than their Dropbox context. They need their whole work context. As a result, Dash also has capability to connect to third-party data sources, bring all that content in a way that it's all accessible through the same interface for searching, for chatting, for question and answering, et cetera. Now we are incorporating it into more and more of our product. Hopefully very soon most of our users or all of our users will have access to it. It brings up an interesting problem that, speaking personally, so much of my content is within that platform. But like you say, there's a lot of stuff that is inherently not file-based and not in that platform.

I feel like I see a lot of people in the ecosystem right now trying to get at that cohesive look of all the data across all kinds of different products. Tell me how that's going to play out. I mean, it seems like it's a little bit of a scramble with everybody trying to make this connection for us. Yeah, first of all, it's a very difficult problem. Anybody in a company, let's say there are users, they access so many third-party platforms. They have let's say sales data and sales force They are using Slack They have email They have I don know Notion or Confluence or things like that for documentation et cetera et cetera Also very ambitious to claim that we have a single place to collect all that information and serve But in a way, with Dash, we get close to that because in the end, our users are actually dealing with that kind of context. So what happens is basically as long as you connect these data sources, users have to agree and they have to initiate that. Then we pull the content in real time

and also make sure that we honor all the governance, you know, security, permissions. We try to index that content as fast as possible and then enable the user base, not only do search, but, you know, they can literally talk to their data using Gen AI. You know, these are now easier. We have a container called a stack. You can group multiple, let's say, content pieces into one unit and you can actually now point AI at that unit and also ask questions and figure out exactly what's going on. So basically it just creates a new way of not only interfacing with that content, but for the users literally to find answers that were not possible before. So Dash basically enables all of these through the connector, through indexing and connecting AI actually to answer questions and provide search results. Of course, there are lots of other pieces to that, but that's the high level. Yeah, I mean, it all sounds simple. You just synchronize this and connect that. How hard could it be? But on the one hand, that seems like something that Dropbox has a lot of experience about, this whole latency and connectivity and syncing problem.

And when I was thinking about the product, that struck me as something that you folks might have a better understanding and experience about the suffering that's going to come with that aggregation problem in those, like you said, real time. And, yeah, just how that kind of falls apart. You had this idea of context rot. So first, actually, I like the phrasing. Explain a little bit about first what you mean by context rot and how you think that's one of the maybe underreported productivity problems that we've got right now. I think the source of that is to realize that, first of all, what's fundamental is customers are trying to solve a real problem, independent of, you know, what are the pieces are standing for. and then let's say we are going to use a JNI capability, a large language model as a result of that, that has been trained, let's say, on data that did not include your private data. So there's that kind of a mismatch. And second thing is now when you are using this in production,

let's say in our case through a dash, first you've got to make sure that for the model, for the AI to be able to answer your questions, it has to understand your work context, right? So that is the first part of it. So the work context is different from the context window that LLMs have. So these are, unfortunately, using the... Unfortunately, same for NTS. But LLMs also have that context window, in a sense, that basically, by design, they have to have enough information that you are providing because of your prompt or additional information. So a model can connect your work context information, what you are providing, as input to the LLM together with what it has in the model itself. and we can actually create a good answer to your question. So the right part comes basically if the context is not relevant or it's too much, then basically what you're getting is not relevant, right? So as a result, it might not be relevant because you did not give the right context, right information, or it may be relevant because you gave too much

and the context window being limited, right? It doesn't know which one to focus on. If everything is equal, right, then I do not know which one to select. There is some research, for example, like in 2023 or so that Stanford did, that they showed, for example, LLMs might be focused on, you know, the first data as well as the last part, middle part might be ignored. Definitely LLMs are now trying to adapt to that as they go, as you know. They try to compress the context in a way that hopefully they can keep the relevant parts of it. But if it's not done well, all aspects of that, then you would call it this is a rock. Basically, it's not actually solving your problem. It's a difficult problem, right, to solve that, but that is the high-level view of the problem. So how does that show up for somebody? I mean, how does a normal person, you know, who's not a computer geek, how do they experience that as a problem? I think the simplest term is basically you might get completely irrelevant answers, right, hallucinated or irrelevant to what you're asking, or the information that you have provided to the model might not be used at all, right?

The model might be just pulling it from its own training data that is embedded into the model. Sometimes I am experiencing this myself He tries to remember something from even previous sessions even though it not relevant to the current conversations So there are elements of that But in the case that you wouldn know because let's say you don't have enough background information about the area, you might not even detect it, which just comes back to being cognizant about, hey, you are using an AI. It has limitations, and you should be aware of those. In our case, basically, we try to make sure that we have a good handle on the work context. Do we have it collected? Is it up to date? Are we honoring the permissions and all that? So first of all, make sure that the information that we're going to provide is going to be relevant. There's no perfect solution to that, but you've got to be basically doing best of those. And as long as you can control it to that manner as much as possible, results are pretty good. And we also have continuous definitely testing just to make sure they're not diverging.

I'm curious, how are you putting all these pieces together? We have so many tool providers out there. You have the ability to create tools on your own. How are you making decisions about where you put your own efforts into making tools, where you're using tools from other people? You don't have to get into giving us all the secret sauce, but I think a lot of people are grappling with these decisions about how much to take from vendors and how much to create themselves and how to keep those in sync. Yeah. Yeah. I mean, this is a classic buy versus build dilemma, as you know, applies to many things that many companies do. It's beyond AI, and we constantly face that problem. We try to apply a very principled approach. Basically, sometimes if you have a problem and if there's a solution, good solution out there, we try to adopt it because there is no need, you know, given limited resources in any company almost, Why would you build something that is, you know, it's already, there's a good solution for?

Even if you build something, how are you going to iterate with that version? But sometimes we are, I think, because of the urgency or because of the scale that we have and these tools cannot work, we end up actually creating something on. And even in that case, we create our own solution and for a while we're going to work on that one. And in the meantime, if there's a solution from the third party that we can use, we try to again adopt it. So we have a very, I think, very good balance between build versus buy. But in certain cases, basically what happened is generally certain tools, even the best tool that was in the market could not work with our scale. We have over 700 million users, you know, tens of exabytes of data. Even our code is millions run millions of code. So some of these tools cannot work at that scale. For example, we had to build our own search engine. This was before AI even, because we had to work across a huge number of users. There's no single basic answer to these.

It's just all of the above. But as I said, if there is a good solution, then we got to do, we adopt that. Now, with respect to AI, it's very similar. I mean, we don't do foundation models. We try to adopt the best models that are out there. But then right away, a need appears, just like you said. These models are changing. if there's no solution or in the beginning of Dash, for example, like it was not there and we're like, we're going to build something ourselves, right? And when we also build it, we are building it with the mindset that a good solution might come out and we can adopt that. So we don't cut corners at the same time. Don't build it in such a way that it's going to live with you for 100 years. For those cases, there are certain parts that we are building for good. So anyway, that balance is always top of the mind. And what matters is how fast we are moving and whether we need to have these tools be ready for our use or deployment. Again, if it is ready to buy, we would do that. If not, we're going to build. You mentioned that you have millions of lines of code.

That's not hard to imagine. I mean, I think cars these days have millions of lines of code. So actually, at the rate at which Cherniv AI is generating code, I think every company is going to have millions of lines of code, whether they want to be inheriting that or not. What's your take on generating code and how well you're able to use these tools for your development processes? I think you went from, say, low double-digit AI used to something like over 98% of your users were using that. Tell us about that transition and what was easy about it, what was hard about it. You're right. So 2025, early, we were around, I think, less than 40% adoption beyond even like, you know, in what capacity we are using that fund. This was basically AI for coding And there was a conscious effort because we ourselves I mean one thing about Gen AI especially in the coding context as you know Sam basically it like early web Sometimes I use the same analogy. It is useful for N equals 1. Like, literally, you can use it yourself and you can see the benefit.

You don't need a committee or, you know, the whole market to convince you. You can actually judge yourself. And this was the same thing on our side. And you can see that the benefit is there. So the first thing was like, let's increase the adoption. And as we shared in some of the blogs that you are referring to, there's definitely leadership support, right, just to make sure that we have access to these tools, education, right, training of the developers, sharing experience. It's not just like top-down. There are people who are excited about using these, and they were sharing, you know, this is how I use it. And that actually created other people to copy them. And we were able to reach almost 100% of the people, not only using actually these AI coding tools, but even going beyond. And maybe you can ask later, but like how we go even beyond coding for software development lifecycle because coding is not the only part of it. With respect to generating code, we still have the same, you know, software development lifecycle, basically guardrails in place. You know, we check for security and code reviews are still done by a human,

even though, you know, we try to provide as much context as possible using AI. We have quality, you know, gates and scans and things like that. So all of these are in place, you know, in whether a human generates code or code is generated by AI. And the nice thing about AI is, like, you can generate actually tests on top of it and additional documentation. Actually, it actually helps improve. You can test many things very easily, and you can use AI actually to check other AI code, AI-generated code. So it's a whole solution in a way with respect to that. And we are open to that, and the benefit is there. But as long as you have the right guardrails with respect to security, safety, quality, and all that, it's under control. I think we're very early in this, trying to figure out how these things play out. No human species has had this technology before, and so we're trying to figure it out as we go. It's really interesting to be part of this, as you mentioned. What are you most excited about? They say the most important function is the exponential function, this compounding effect. I think that is, you know, this tool is such a tool that, you know, it can generate, it can improve itself, right?

So that feedback cycle, that virtue cycle and that compounding effect, I think is something to be first scaled up as well as to be super excited about. I think that's the most exciting part. Whatever area that we're going to apply, first of all, you know, things split things up. Even in some of these foundation labs, they're using their own tools to actually write these new models. And I think that was never before. You know, yes, there were code generation tools in the past, even decades ago. But it is not to a level that creating this compounding effect. I think that is one underlying thing that I am super excited about. And the other thing is basically applications of these, right, tools, science and other areas. You know, imagine many of these problems that we are facing right now can be solved maybe faster, even if AI does not solve completely. And some of these, you know, human will be in the loop, let's see. But at least it can help you explore the space far faster, get you up to speed far faster, you know, create these hints here and there, right, so you can actually attack it far faster.

Hopefully, because it is representing a version of the entire, let's say, world knowledge, it hopefully can go beyond humans with respect to finding connections between different knowledge bases, like different areas, different science or other areas. And this was never possible before, right? I think the speed of discovery is going to be drastically changing and improving and accelerating. So those are all exciting things that I'm so happy to be alive, to be witnessing. Well, that feels like a nice positive note to close things on. Excitement to be alive and it's a good time. And the idea that we can take context and we can make connections between different areas in ways that we've never been able to before is pretty exciting. Thanks for bringing that to the table and bringing your engineering eye to the problem. Thanks. Thank you so much for inviting me. I really enjoyed it. Thank you. Thanks for joining us today. Next time, Sam is joined by Suresh Venkataralayu, CTO at Honeywell. We hope you'll tune in. Thanks for listening to Me, Myself, and AI.

Our show is able to continue in large part due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcasts review or a rating on Spotify, and share a show with others you think might find it interesting and helpful. you