We chat with Deedy Das, a Principal at Menlo Ventures, who began his career as a software engineer at Facebook and Google. He then dipped a toe in the startup world, spending time at the company now know as Glean. More recently he started a career as a venture capitalist, investing in AI and Infra out of the Anthology Fund, a partnership between Menlo Ventures and Anthropic.
You can find Deedy on Twitter and LinkedIn.
You can learn more about the Anthology Fund here.
You can learn more about Menlo Ventures here.
Congrats to Stack Overflow users Bobince for earning a Populist badge with their answer to the question: What does sorting mean in non-alphabetic languages?
[intro music plays]
Ben Popper Monday Dev helps R&D teams manage every aspect of their software development life cycle on a single platform– sprints, bugs, product roadmaps, you name it. It integrates with Jira, GitHub, GitLab, and Slack. Speed up your product delivery today. See it for yourself at monday.com/stackoverflow.
BP Hello, everybody. Welcome back to the Stack Overflow Podcast, a place to talk all things software and technology. I am your host, Ben Popper, Director of Content here at Stack Overflow, and today I am joined by Deedy Das, who is an investor at Menlo Ventures focused on early-stage investments in the AI/ML and next generation infrastructure world, as well as some enterprise software, but he has also been an engineer himself and spent five years in various technical roles at Facebook and Google, holds a bachelor’s and master's degree in Computer Science with a focus on AI/ML and information retrieval from Cornell. So lots of stuff for us to chat about. Deedy, welcome to the Stack Overflow Podcast.
Deedy Das Thank you so much for having me. It's a pleasure to be here.
BP So most of our listeners are themselves working software developers, so why don't we start there. Tell us a little bit about how you got into the world of software and development, what led to the focus on AI/ML at the academic level, and then maybe what some of the most interesting or fun parts were of working at a Google or a Facebook.
DD Well, it's kind of a long story. Let me give you the shortened version. I grew up primarily in India, but I grew up a little bit in the Bay Area. Our family moved back to India. I came here for college, and when I first came here I wanted to study physics. That lasted about six months. I switched to CS, and the thing that I wanted to do in college was I was obsessed with computer graphics. I wasn't one of those hardcore video game people, but I love movies, games, I appreciated graphics, but then I quickly realized –I told myself at least– that unless you get a PhD in graphics, I don't think you can really do the greatest work. I sort of switched over to doing more machine learning, information retrieval type stuff towards the end, and then this is a common choice a lot of CS majors have to make, which is, do I want to go to a startup, do I want to go to a big company, or do I want to go into something else, which could look like finance or something else. And at the time I was just so enamored by this idea that Facebook and Google are some of the companies that have shaped our lives and I just want to see how the tech is built. And most of my decisions are made with a high dose of idealism. Obviously there's other good things about that job, I'm not saying that's why I took it, but I just wanted to learn more about the stacks and that's why I joined Facebook and Google.
BP So you were going to school in what era? When would you say you graduated?
DD I graduated in the end of 2014.
BP Okay, gotcha. So by that time kind of the exciting stuff that had been coming out of some of the universities in Canada and on the West Coast, the explosion of interest in neural nets and machine learning had happened, and so your pivot to that at the end of your academic career makes a lot of sense. You spent a year at Facebook full-time as a software engineer. I think something people have said a lot when they come on is just that it takes six months to get to know your way around all the in-house sort of tech and systems they have. How long did it take before you got to write your first line of code and what was it like to work with systems that have been built from the ground up by Facebook for Facebook?
DD Well, I think Google took longer. So at Google the line I was given was, “I don't expect you to be productive for six months,” which I thought was absurd. Facebook is a lot, or at least was at the time, a lot quicker to ramp up. So I think I was writing code by day one, frankly. You write code in the very beginning. You have this pretty intense process called bootcamp. I really enjoyed my time at Facebook. I think in the first six months I was doing pretty well. There was a funny story where my boss called me into a meeting six months in and I didn't think I was doing a particularly good job. I had no calibration on what a job looked like –it was my first job– and he was like, “Hey man, we should do a meeting.” And I'm like, “He's going to fire me. I screwed something up.” And he drags me into a meeting, and this was the process at the time, and he's like, “You've been doing so well that we want to give you a promotion.” I didn't think that was on the cards, but I just did a lot outside of my main job. I've kind of always done that. So I did a lot of building random internal tools that helped other people.
BP Then you became a senior software engineer and a tech lead at Google, and it sounds like you worked on some machine learning and infrastructure projects there. Search, knowledge engine, triggering and quality. Are these things that folks might be familiar with from the consumer side or things you can talk just at least a little bit about in terms of what it was like to work on and to build them?
DD Of course. I can nerd out about search for hours. So the way to look at it is, if you think about what people are searching for, about 10 percent of what people search for every day is something Google has never seen before, so that's something completely new. Then there's a huge torso or middle ground where, I don't know why I use this example, but Doom 3 cheat codes, which is like, okay, it's kind of specific, but I'd imagine a lot of people could use it. And on the very left or sort of the head of that distribution is things people search for all the time, so people, places, movies, sports, finance, stock prices. And for those kinds of structured queries, or for even queries like “Tell me how tall Obama is,” all of those queries need to be understood and served through structured knowledge, not just the internet public web documents. And Google has invested a lot of money in understanding what they call a ‘knowledge graph–’ what things are and how they relate to each other. So in knowledge engine, one of the tasks we were doing is, when people enter a query that looks like a set of characters, how do you really make sense of it, and then how do you figure out what to serve as a result of that? How do you know that Obama means Barack Obama? That's who it means. It doesn't mean his wife in this case. It doesn't mean his daughters. That's who it means. It doesn't mean some movie called ‘Obama.’ That's kind of the challenge of knowledge engine.
BP Since then, things have evolved so much. Now if I go to Google and I search, I'll get an AI answer. Obviously, again, like you said, it's trying to understand me in a natural language way, not just in a lexical keyword way, and it's trying to provide me answers in natural language, not just top 10 blue links. Things have really changed in the last year, probably the most fundamentally that they've changed in terms of Google Search since it was created. There have been obviously more answer boxes and images and trending news arriving, but the AI search we're seeing now is a pretty fundamental shift. Stack Overflow is involved in that. We announced that we're licensing our data to Google so that they can train Gemini, which is at the heart of giving those answers. And for us, it's all about attribution. If you're going to train on this knowledge and that's what you're going to give the answer, then make sure you attribute it back to the people from our community who created the knowledge or whatever. So for you, did you see the seeds of that when you were there? You're mentioning that we want to understand the query not just in terms of the words that are said, but exactly what the person's intent is, and then we want to provide them some kind of knowledge back, not just a set of links. Is that what you felt like you were working on at the time as well?
DD Not really. So the goal was different. Then, the goal that we had in pre-LLM world was really a small set of structured intents– so people's names, specific asks about people, not unstructured queries. And the way LLMs work are sort of more in that torso layer of queries. I think the history of events is as follows. Google spent a lot of time focusing on these head queries and serving results on the search result page. That has been a work in progress since, I don't know, 2010 maybe to 2018 or so.
BP I guess the evolution of providing their own knowledge and their own information at the top, that was kind of an evolution, and AI answer seems like the final sort of result of that. So let's move on a little bit. You were there for a number of years and then you went to a company called Glean. I don't know a ton about it. I did a little bit of looking, but a Gen AI product, an assistant that looks at all the docs and the Slacks and the emails inside of your company, and then you can ask a question: “What are the changes in V5 of the project and what should I know about that in terms of our go-to-market plan?” And it'll give you back an answer because it has the entire corpus of what your company knows. In some ways, I would say that's probably competitive with what Stack Overflow does. We have Stack Overflow for Teams. Everybody goes in and asks and answers questions. If you have a question that's in your Slack or Microsoft Teams, it pulls it in, and then you can go and say, “Hey, tell me about the changes in our latest API and how will that affect enterprise customers,” and it’ll give you the answer back. So both of them have this version of universal enterprise search where the knowledge within a company can become more accessible thanks to an AI agent. What was it like to be there from the beginning? How did you see it evolve? What were some of the things that were sort of bleeding edge when you left?
DD This was 2019. Glean was not called Glean then, it was called Scio– very bad name. And I joined the early team before we really had built anything out. What was very compelling for me at the time is enterprise search, especially the problem that you just described, I feel like it wasn't as hot back then. Right now, everyone's talking about it because of the whole Gen AI wave, but it's a problem that wasn't new. Ever since the nineties people have been trying to sell people boxes in companies to search through the data in their enterprise. But for various reasons, we felt like it wasn't great anymore and all the current solutions were not fitting market needs.
BP I think I've been with Stack Overflow now for five years. When I first showed up we wrote a big piece, “Documentation is dead. Here's the better way to do it.” It's kind of like technical debt. No matter what company you go to work for, at some point you're going to run up into this issue of, “We have so much information but I can't find it quick enough, or I don't know where to access it, or the person who wrote it down has left, or the wiki's gone stale.” Everybody knows these problems. Enterprise search or some kind of version of it, what is the best solution? There's been a million and it does seem like LLMs are uniquely suited to it because what do they do well? They can reason over a large corpus of text that is easy for them to memorize or have in a RAG system or whatever it may be. But you were there in 2019, so pre-LLM. What was the approach to serving it up back then?
DD Fundamentally, there was various problems to solve, but some of the core problems were that most companies use over a hundred SaaS apps. So crawling and indexing all of those in nearly real time–
BP The knowledge ingestion, pulling it all in.
DD And then updating permissions. Because say you have a Google Doc, every time you change the permissions, you don't want that to be searchable anymore, so keeping permissions up to date is hard. And then maybe the hardest or most interesting piece of the entire thing was how the ranking algorithm works. Because one of our learnings from Google was that Google had its internal search engine which was a super popular tool inside of Google called Moma, but Moma didn't really work well. It was kind of the butt of a lot of jokes where they were like, “Google works so well, but Moma kind of sucks.” And if you thought about it, we had a hypothesis on why Moma sucked. Moma sucked because Google works so well because it has a high volume of click data. People are clicking through stuff. Those are the signals that feed the ranking algorithm at the end of the day. And it's a very head-heavy distribution. A lot of people query for a few things. In the enterprise, that's not true. People query for all sorts of things. Everyone's job is very, very different. They're not all querying for benefits every day.
BP They're mostly looking for edge cases. They're like, “Wait, why did this happen? This isn't working. What's the deal with that?”
DD Exactly. And there are not that many clicks, so it's about how you design the algorithm.
BP That's one area where Stack Overflow, just by dint of having come out of this forum idea with the wisdom of the crowds, has an advantage in that, let's say you have a company and it's got Confluence and it's got Jira tickets and it's got wikis and it's got Google Docs and it's got email and it's got Slack and you ingest all of that. How do you know which of that data is accurate, which is the most relevant, which is the most up to date? One useful way to do that is to have a system where everybody votes on it and it’s got all these tags and metadata that says, “Well, this was the answer that was most-liked, or this answer was popular, but this new answer is now more popular.” And so there's essentially reinforcement learning with human feedback happening through all the voting mechanisms and the commenting and the tags and stuff like that. So that can be a really useful way to do that, especially when you're digging into some deep well of a project that was built six years ago and nobody's asked a question about it until it breaks and then everybody wants to know how do we fix this.
DD Absolutely. I think the other signal that we found very valuable is obviously likes and human feedback, but one of the things you get in the enterprise is you also know how many people have seen the Jira ticket, who else has seen the Jira ticket, have other people on your team seen the Jira ticket? It turns out it's a very good signal to know that the person you work the most closely with and you do Slack with all the time, they've been looking at a Jira ticket all day. If you're searching for something, that might be kind of relevant. And so it was really engineering more enterprise signals and ranking that are more applicable to the workplace than not consumer search, which was one of the core challenges.
BP A lot of work goes to the admin of a Stack Overflow for Teams community to organize things, and we're trying to do things, I think to your point of, “Okay, we see this person is always the one who leaves the most upvoted answers about Node.js questions. They are now designated automatically as the subject matter expert, and next time a question comes in on that, can we shoot them an email? Can we automate somehow, to your point, this person was looking at the Jira ticket, they might have an opinion on it. This person is always the one who answers these kinds of questions. If you've got a question about it, maybe we can connect the two of you. All right, so you did a bunch of time at these big tech corps, then you did a bunch of time at a startup and now you've moved into the world of venture capital. What made you decide to leave the operations and business side of things and go into VC?
DD At Glean, I did primarily two or three different roles. I started as building stuff, then I started managing engineers, and then I went and I sort of led product on their second product line, which is Glean Assistant, and then that product scaled to 10 million+ ARR. After those four years, I felt like the company had just become big and the challenges were different. A lot of my job was going in and doing the same thing every single day, which is talk to customers, make them happy, negotiate what you will and will not do for them, and then deliver. And I felt like I wasn't learning much. The growth had sort of plateaued. And the company had become big so it comes with a little bit of– I mean, big is relative. The company at the time had become 400 or so people and I felt like that wasn't necessarily the same Glean that I joined. I got what I wanted out of the experience. I thought about what I could do after. Of course, one of the options is to start a company, but I never considered being a VC in my life. I always considered that sort of the dark side. I didn’t want to do finance. I never wanted to do an MBA or anything like that. But it was so compelling the more conversations I had to work with startups doing zero to one. I felt like I could add value. I found out all these lessons at Glean of how do you hire, how do you scale, how do you do GTM, how do you do all of these different things that startups kind of always need that help with, especially first time founders, and bringing the technical expertise with that startup expertise, I just felt like I could be valuable. That was kind of the reason that I moved over.
BP And so my understanding is that you now are focused in on something called the Anthology Fund that you helped to launch, and that's in partnership with Anthropic which is one of the big players in the space. Obviously, Claude is one of the most well-known Gen AI assistants. Can you just talk a little bit about why create this specific fund? How does the partnership with Anthropic work and what are some of the unique and groundbreaking AI solutions you're seeing from startups that you're invested in or you just want to share with our listeners because you think they're cool?
DD So I think to give you a brief history in how a lot of VC firms work, a lot of the traditional VC firms, Menlo being one of them, they do a few deals a year per partner. So they like to do lead deals and we're multi-stage firms. So when you do a deal, you partner with a high conviction startup. You try to take a high ownership and then you never fund the competition and you bet on the winner. That's sort of the classic model of venture. Now with the way that startups work, when you have a wave today where there's so many new companies being launched, we wanted to capture more than just that. Deals were moving quicker, people were getting funded out of the bat. There's so many interesting ideas going around, and the Anthology Fund was a way to say, “Hey, you know what? Anthropic is developing a developer ecosystem and they're a big company that we back. We want to connect them with startups that are using their API, align them with what the product vision should look like based on how startups are using them.” But also we want to fund more companies at an earlier stage and get them in because we think that this wave could lead to some generational defining companies that everything from healthcare, consumer, a lot of spaces are going to be transformed by the new developments that we see, at least we think that. So that's kind of the premise of the Anthology Fund.
BP Gotcha. So just so I can clarify, maybe that means more like instead of doing some of what maybe Menlo has done in the past, which is high conviction, invest, take a board seat, pro rata, follow every round, try to get them to liquidity, you're saying that we're already an investor you mentioned in Anthropic, we want that platform and ecosystem to grow, so create more of a seed fund where we can find lots and lots of startups, we connect them with this ecosystem, we try to get all of that to grow together, and it's not like you're taking a board seat on every single one.
DD Exactly. Because traditionally, multi-stage firms are not going to do a lot of seeds. So now we can.
BP So tell me about a few of the companies that you've seen that are building on top of a platform like Anthropic, or it doesn't have to be. But I'm familiar more with some of the really well-known players in the space who are creating foundation models, basically. I know about that and I know some of the players in this space who are doing super interesting things in the world of image and video generation or audio generation. But beyond that, I'm not sure I have my finger on the pulse of which startups are really creating cool stuff or what some of the use cases they're discovering are. Why don't you name your top two or three that you've been thinking about and maybe that's a reflection so that people can get a taste of what you feel the zeitgeist is, or when you sit down and you talk with other partners at Menlo, you're saying, “This is the area I think we're seeing a lot of potential in.”
DD I think the way we've structured the fund is that we want to invest in the cutting-edge AI companies of today. We think this era, this small one to two year era, is going to give birth to many long-lasting AI companies. And sort of the large broader categories we see them in is that you have foundational models, so we already kind of have a big bet there, and there's a high capital at bet. We have the infrastructure layer, and then we have the application layer, and then you have a couple of oddballs that kind of straddle the lines between those three layers. In the infrastructure layer, one of our big bets were data processing platforms, so an ETL tool called Unstructured. ETL tools really transform data, make them ready and consumable by LLMs. We also have a lot of bets in the application layer. And in the application layer, you can think of applications of Gen AI as anything from that we have companies that are doing consumer applications. So we've just recently made an investment in a company that was doing the next generation of fashion that's not announced yet. Another company in the infrastructure layer, for example, that people use a lot is Pinecone, which is in our portfolio. And then in the application layer, we have climate companies that are happening. A lot of companies are saying, “You know what? The new models completely change the game for how weather is predicted across the globe.” And weather predictions control so much of the GDP and the economy. Even getting one weather prediction wrong can mean hundreds of thousands, millions of dollars of delays just at airports, for example. We have climate, we have healthcare. There's so many new things happening in terms of drug discovery and digital health in the healthcare space, and just so many different pieces of that pie. The last one I want to call out is that some of the companies that straddle the line that do more than just an application or infra or foundation model are complete new rethinks of certain categories. So there's a lot of bets in the space of robotics. The ideal thing that everyone has wanted to create in robotics is a model that takes any robot, any input, and then just figures out without any specialized training how to operate that robot. So that's a bunch of ideas across the spectrum in the AI space that we're looking at.
BP When you say that, it sparks a lot in me. I think about folks I've talked with like Timescale where they're trying to sort of build an open source Postgres stack for AI applications. So that's sort of similar to a Pinecone where it's like, “Hey, do I need a vector database? Can I use Postgres to build that? Do I have to change out to a different technology?” And we've talked with some other folks on the ETL side where so much of what you're doing when you're trying to apply this technology in-house is saying, “What data do I have? Is it unique from what was in the foundation models? And how do I clean it or annotate it so that my results can be accurate?” Personally, before ChatGPT came out and sort of made Gen AI something in the zeitgeist to where LLMs is a word people know, I was playing around with Midjourney and using that as a consumer and paying 10 bucks a month for it because it feels totally worth it. I can imagine if there's a fashion startup where it's like, “Upload a few shots of yourself and then try these clothes on, try these glasses on. Tell us what color you want these glasses through Gen AI and then we'll make them for you and send them your way.” And then definitely, we've talked about this a bunch on the podcast, but there are these industries that feel tangential like weather prediction, for example, for agriculture could be huge, for transportation could be huge. AlphaFold is always the one I bring up when it's like, “All right, well, LLMs are cool, but are they really intelligent and what do we need better chatbots for anyway?” And it's like, “Well, if this kind of technology is going to accelerate drug discovery by 10x, then someday we're going to look back on it like we do vaccines or something.” It's going to have this enormous impact. So I like the way you're thinking about the portfolio. It makes a lot of sense with what we've been talking about on the podcast.
[music plays]
BP All right, everybody. It is that time of the show. We want to shout out somebody who came on Stack Overflow and shared a little knowledge or curiosity. Awarded to Bobince a Populist Badge. That means they came in and they gave an answer to a question that already had an accepted answer, but their answer was so good that it is now two times more popular. Bobince provided an answer for the question: “What does sorting mean in non-alphabetic languages?” So congrats, Bobince, on your Populist Badge. Over 5,000 people have been helped by your question. As always, I am Ben Popper. I am the Director of Content here at Stack Overflow. Find me on X @BenPopper. Email us, podcast@stackoverflow.com. We have listeners come on as guests. You can suggest topics, you can tell us what you want to hear about or what you're sick of hearing about. And if you enjoyed today's episode, leave us a rating and a review. It really helps.
DD My name is Deedy Das. I invest out of Menlo Ventures in early stage AI, infra, and SaaS, and work at the Anthology Fund with Anthropic. You can find me on Twitter @DeedyDas, and you can reach out to me on superdm.com/deedy to message me. Thanks.
[outro music plays]