Moonhub AI: The On-Demand AI Recruiter with Founder & CEO Nancy Xu

Full episode

Matt Turck [1:25] Hi, Nancy. Welcome to The MAD Podcast. I've been very much looking forward to the conversation. There's so much to talk about. You are the co-founder and CEO of Moonhub AI, which is an AI copilot for the world of recruiting. But you're also a very active investor in the world of AI. You are the editor-in-chief of The Gradient, which is an AI newsletter that I would absolutely recommend to anyone listening to this, that I consume pretty much daily. And last but certainly not least, you were recently named to Time's list of the 100 most influential people in AI.

Matt Turck [1:51] So amazing. Congratulations on all of this. I'd love to start those conversations typically with your background. You've done some really interesting things leading you to all of this. So, yeah, I would love to hear your story.

Nancy Xu [2:16] Sure, absolutely. And thanks for having me here, Matt. Really appreciate it, and very excited to chat with everyone here. I will say I was the editor-in-chief. I helped found The Gradient, but I'm less involved nowadays on a day-to-day basis there. So, hey everyone, I'm Nancy. I am the founder of Moonhub, where I'm also the CEO today. At Moonhub, we are building AI-powered recruiters. So think about every company in the world. Our goal is to help enable those companies to scale their best ideas through the power of people.

Nancy Xu [2:36] And my personal story prior to Moonhub, I was over at Stanford. I was a computer science PhD there, working on some of the early generations of foundation models and LLMs that you see out there in the AI world today. And when I was at Stanford at the time, I had this really interesting insight, which was I heard a lot of people talking about AI through the lens of fear, like it's going to automate everyone's jobs.

Nancy Xu [3:12] People were concerned about what that would look like for the future of work. And I fundamentally believe there's a world where we can actually use AI to help people in that future actually find better opportunities, as opposed to people seeing AI as a source of taking away their jobs. And so that's fundamentally what we really work on at Moonhub. And we do this in the form of creating an AI recruiter that helps these companies scale.

Matt Turck [3:29] Great. So you have just launched. This is a timely conversation. So tell us more about how that works, what the product actually does.

Nancy Xu [3:52] Sure, absolutely. You can go to Moonhub AI and try out the product yourself. There, we do have a get-started version that anyone can try out without paying or being a paying customer. And what we've done at Moonhub is, if I look at the recruiting process today, I break it down roughly into what I call five different stages. So there's the initial stage of calibrating on what you're looking for. There's a stage of actually finding those candidates.

Nancy Xu [4:17] So, what is traditionally called sourcing. And then there's the outreach phase. This is where you're reaching out to everyone and trying to see if they're interested in your opportunity. Then you're screening them, and then you're scheduling, managing the interview process and everything downstream of there. And at Moonhub, our goal is to build an AI that acts as your copilot through that entire process. And today, if you log into our website, we just launched this recently, you can actually go and talk to our sourcing AI yourself, chat with it like you would a normal recruiter, say, "Hey, I'm Matt, and I'm looking for an amazing software engineer."

Nancy Xu [4:55] I need someone in the Bay Area, and they need to have some experience with Python, and I want someone who worked at a Series A company before, whatever it is you're interested in. And Moonhub will go out there. We index about one billion people's data across the public web, bring back those profiles that are relevant for you, and then also work through the calibrating and actually answering your—helping you strategize around your search. So you might ask, "What are some companies that use Python that I should be looking at for my recruiting purposes?"

Nancy Xu [5:12] And Moonhub will help you work through that. And then, if you become one of our paying customers, we will help you through that entire end-to-end process past the sourcing component as well.

Matt Turck [5:29] Is part of the idea that you can, one, accelerate the process of what a manual recruiter would do, but also perhaps find people that would be less obvious? Is that part of what you're trying to do?

Nancy Xu [5:51] That's a great point. Yeah. So today we work with about 100 companies. Actually, before we launched, we worked with about 100 companies, helping them hire and scale their teams—public companies out there, as well as some early-stage startups. And these companies work with us because we help them hire at half the cost, three times faster. And this is enabled by giving them a lot of the—the AI gives them control of running the process.

Nancy Xu [6:22] So if you've ever used an external recruiter yourself before, Matt, you'd know that normally you wait a week, you finally get on that call with the recruiter, they ask you some questions, two weeks later they send you a couple candidates, turns out they're not a good fit, and then a month rolls by and you're finally starting to find maybe a couple of people that you think are good. And I think the real power of AI in the next one or two years is enabling these non-experts to be able to participate in the workflows that traditionally experts will do.

Nancy Xu [6:57] So at Moonhub, what we do is we make it possible for you as a hiring manager or recruiting leader to talk to the AI without knowing how to use Boolean searches or the traditional lingo and complexities of recruiting, to be able to work and collaborate on that step of the process so that the entire recruiting cycle is shorter and faster for you. And that, of course, leads to cost savings for you as well. And so that's the first part: you can get started faster with Moonhub, you can hire faster, and you can get ultimately better-quality candidates.

Nancy Xu [7:24] And then, to your point, Matt, on the second half, a big reason why companies work with us is because we help them find candidates they wouldn't otherwise find. And this goes to the part of AI being a knowledge aggregator. We crawl information on people across the public web, so whether that's your Google Scholar page, your GitHub page, your LinkedIn page, various other pages that have information that might be relevant to whether or not someone is a good fit for an opportunity.

Nancy Xu [7:50] We run models that aggregate all of this data. We train our LLMs on top of this and ultimately surface that as a conversational interface, which you can try out today, and helps you find the right people at the right time.

Matt Turck [8:11] So ideally, I may not have gotten to Stanford or whichever other top school, but maybe you'll find that I've contributed a lot to open source on GitHub, and therefore you'll surface me.

Nancy Xu [8:37] Yeah, I think one of the big pain points in recruiting today is, if you are not a recruiter, oftentimes what I've observed is you're searching for—you’re really just looking for proxies of what you really want, right? So let's say what you really want is someone who has shown signs of excellence in their last job. You can't really tell a system, "Show me people with signs of excellence today," and it'll show you those people. So you end up with these very, I would say, very rough proxies or approximations that are: they went to an Ivy League school, they got promoted recently, they worked at one of these five companies that you think highly of.

Nancy Xu [9:12] But that itself is a very small subset of all the people that could potentially have shown signs of excellence in their last role. And I think the power of AI is it allows you to better navigate that sort of what I call the gray data, and help people understand and find the right hidden gems within that gray data.

Matt Turck [9:21] Yeah. And I assume, hopefully, a big part of the hope is that, from a diversity perspective, it will be fairer and more open, right?

Nancy Xu [9:44] Yeah, absolutely. And we've seen this already. We work with a lot of companies who, diversity is top of mind for a lot of them. And when they use Moonhub, they tell us, "Hey, you've been able to show me, like, 80% of the candidates that I found through you, I haven't found on any other platform." And they tend to be more diverse. They come from more underrepresented groups as well.

Matt Turck [10:12] Yeah. And on the topic of diversity, the sort of good old question, but obviously that's a fundamentally important one, around how AI models are trained: how does that manifest for something like Moonhub? In terms of, on the one hand, you want to increase diversity, or you hope that better diversity is a positive consequence of what you do. But on the other hand, AI models may have been trained on data that indicates that is not as fair to diversity.

Nancy Xu [10:44] Yeah, absolutely. And we care a lot about diversity at Moonhub and how we think about it, especially in the context of recruiting. My big hope is, in the AI future, where I think a lot of jobs are going to change, I think most people underestimate how much AI will impact their daily lives, and especially what their work will look like in five years from now. And we want to help prepare those people for that future and help make sure that they have the right meaningful opportunities that are meaningful to them.

Nancy Xu [11:19] So in terms of diversity, we do a couple of things. So one is, we do a lot of data cleaning and just data quality, so measuring our data quality, making sure our data is accurate, that it's well deduplicated, that we make sure any models that we run on the data, we only push them into production if they've hit a certain threshold of accuracy. So for example, today, if you wanted to go on Moonhub, because diversity is such an important thing, we actually support our users doing sort of diversity-oriented searches.

Nancy Xu [11:54] So you can say, "Hey, I really care about gender diversity. Please boost my search so I see more diverse candidates." We will do that. And the model that actually powers that behind the scenes, we run it against a lot of validation sets and we make sure it's, like, 95, 99% plus accuracy before we push it into production. And we're constantly evaluating the models that we run. But that said, I think in this space there's a lot more we can be doing, and we're constantly sort of scanning the horizon for what else we could be doing on the engineering and product side to make sure that we responsibly use AI.

Matt Turck [12:25] And to the extent that you can talk about it, can you maybe take us under the hood in terms of how it works? Yeah, under the hood or behind the scenes, whatever the analogy is, for the models you use or the stack or whatever you can talk about.

Nancy Xu [12:54] Sure. Yeah, it's a great question, and a lot of people ask us this question. So at Moonhub, we build a combination of our in-house proprietary LLMs. We do use a couple of third-party LLMs for very specific use cases. And I think the sort of really foundational innovation that we've developed here is we've created a whole system for a central LLM to orchestrate our entire conversational interface and all the actions that it's able to perform. And this is actually something that we have filed in a patent recently.

Nancy Xu [13:29] So underneath the hood, to the extent I can share, you can think about pulling all the public data about people, companies, and relevant enterprise news and articles on the internet, aggregating all of that, putting it into various data stores. So we have, for example, structured data stores, like your more standard things like Elasticsearch. We have vector data stores, and we use various other kinds of data stores as well. And then we actually have an LLM that coordinates how to retrieve data from the right data stores.

Nancy Xu [13:57] So this allows you as the end user to look for things like, "Hey, show me people who currently work at an early-stage startup in the Bay Area who's familiar with Python." Different parts of the sentence, some of this is more structured, right? Like early-stage startup means seed, Series A, Series B. Knows Python, that can be a little more fuzzy, right? So it could be they mentioned PyTorch or they mentioned building some backend system that is traditionally made in Python, whatever that might be.

Nancy Xu [14:39] But our LLM actually learns how to route to the right data stores and how to appropriately map the natural-language queries that you provide into retrieval. And then on top of this, we have built some custom retrieval pipelines on top of our data. And we ingest into our core LLM recruiting-specific knowledge that we've collected from recruiters around the world, including our internal recruiters, so that our LLMs can answer recruiting knowledge questions with more accuracy than your typical sort of world-knowledge LLMs.

Matt Turck [15:04] Great. And just to play it back to make sure I understood that: so that's external data, but that's also internal data a little bit. So I can tell Moonhub, "Hey, find me a Rust developer that looks a little bit like this," and I can show the profile of my best Rust developer.

Nancy Xu [15:22] Yeah. So today, if you use the platform, you'll get immediate access to public data. So you'll get access to some of our public data stores. And then for some of our enterprise customers, we help them do integrations with their own internal data stores. And that allows them to, it makes the AI smarter. So you might be able to say, "Hey, this candidate is more likely to join your company because I observed, based on your ATS, that traditionally people at large companies tend to do better in the interview process and they're more likely to accept an offer."

Matt Turck [15:59] Okay. So the vector database is also to pull internal data from the customer. Okay. Really interesting. And then, just to drive it home, the sort of execution layer of this is, what, you send an email automatically to the candidate, or you connect with them through LinkedIn, or how does that work? Or you make a recommendation and then people should do their own reach-out.

Nancy Xu [16:25] Yeah, it's a great question. So if you log in today, you'll be presented with what we call our sourcing AI. It's a conversational interface: you talk to it, it'll answer strategic questions and then help you find candidates that are relevant. Once you find, let's say, 100 candidates you think are a good fit for your project, what you can do from there is reach out to those in a somewhat automated fashion. So you can write email sequences to them. The AI will help you craft email sequences.

Nancy Xu [16:53] And then our—I should comment by saying our product is not a—there's the core AI, but we also augment it with actual expert recruiters behind the scenes as well to enable you to do end-to-end. So customers that work with us, they're using the AI to source and to help outreach. But if they don't want to do that, our human recruiters can actually help them with those components too, as well as helping you screen the candidates. So our recruiters will actually hop on calls, screen the candidates for you, give you the data back, help you prepare data for compensation and offers as well.

Nancy Xu [17:26] So we do offer that end-to-end. And Matt, traditionally, I think a lot of people—for us, we see our main, sort of the biggest parallel to what we build is probably the external recruiting agencies of the world. So if you've ever worked with a boutique external recruiter or a big one like Robert Half or TrueBlue, those are somewhat of an analog to what we build. And oftentimes, we see companies who have used those are extremely frustrated.

Nancy Xu [17:36] They come to us and then they use our AI-first experience and they get better results there.

Matt Turck [17:46] That's very interesting. And is that, you think, what the business is going to look like in five years from now, or is that a first stage as you keep training the AI?

Nancy Xu [18:11] Yeah, it's a great question. So our ambition is not to build another—I think there are a lot of companies out there trying to build tools, like recruiting tools. Our ambition in the long term is really to build the AI-powered end-to-end recruiter for you. And so that's what we have offered from the beginning, is an end-to-end recruiter. And today, the AI can't do all five stages of the process I mentioned to you. It might do stage one, stage two.

Nancy Xu [18:42] We'll probably get to stage five by mid-next year. But in the interim, you have a human who's actually helping you with all those other stages and making sure you have a seamless end-to-end process. And so when companies work with us, oftentimes they're looking for that end-to-end recruiting experience. And so, quote unquote, our competitors tend to be these 50-year-old legacy organizations that many of them still have fax machines, and it's a total human business, basically.

Matt Turck [18:58] And you think eventually the end-to-end includes interviewing? There's a screen, I think you mentioned, but do you think the AI can ultimately determine, in a few years from now, okay, this is the best candidate for the job?

Nancy Xu [19:22] I think in a few years from now, we'll definitely have AI that is smart enough to do it. But I think the question is more so whether or not candidates will be into it, how much it'll impact the candidate experience, so that people actually want to deploy it, right? And for us, if you look at what today you might engage an external recruiter for, external recruiters actually don't do any of the interviewing for you. They help you screen, they help you source.

Nancy Xu [19:40] And so we think that's the really big opportunity. And we look at what that particular segment of work does today. We think that is something our AI plus really talented experts on the team can—well, we've shown that they can do it together end-to-end very effectively already.

Matt Turck [20:00] And you mentioned you already have 100 customers, which is sort of amazing considering you're a seed-stage company and you just launched. Any sort of tips, strategies, lessons that you could offer to other seed-stage companies about how one gets from zero to 100 customers?

Nancy Xu [20:02] Yeah, it's a great question.

Matt Turck [20:03] Hard.

Nancy Xu [20:26] Not sure if it's repeatable, but happy to share my two cents. So before we launched, we had about 100 customers. After our launch, we've, as you can imagine, gained a lot more. But pre-launch, one of the things that we were very specific about building, unlike I think a lot of SaaS companies today, is we wanted to really understand exactly the end state of what it is we wanted to build. So I recently said this online somewhere, but I think a lot of the opportunities for AI startups is to find something that is traditionally very difficult to scale.

Nancy Xu [20:55] But very easy to sell. And then you try to scale it with AI. And I think recruiting is exactly that. So we work with an industry where there's more than 100 recruiting agencies in the United States that each make more than $100 million in revenue a year. They're all basically human capital businesses. They're services businesses. And for us early on, one of the biggest things that we did was just invest in actually understanding how all of these very human-oriented businesses operate.

Nancy Xu [21:25] So we actually built out our own, basically, end-to-end services business. It takes a lot of work early on. But from there, you get so much of the product learnings. We now have a feedback loop where everyone on our team is essentially a pilot customer of our product and our AI. And that's enabled us very early on to acquire customers even as the AI is just starting to develop. And then over time, these customers use more and more of the AI.

Nancy Xu [21:55] And that's been really helpful for us. And of course, I think you want to make sure your early customers have an incredible experience. So a lot of them, I hop on calls personally with them quite often. And within the 100 customers we had pre-launch, I would say at least 50% of them came through referrals from the existing early-access customers. And we were very targeted in terms of who we would allow early access to as well.

Nancy Xu [21:57] Okay, very cool.

Matt Turck [22:30] And as I'm sure you're actively recruiting and all the things, maybe tell us a little bit about the company. $8 million from Khosla, Google Ventures, AIX Ventures, which is Richard Socher, who actually introduced the two of us. Thanks, Richard. Day One Ventures, Susan Wojcicki, Ram Shriram, and so on and so forth. Marc Volpe, Christopher Wray, Chris Wray. So, awesome group of investors and sort of AI folks. Anything you can share, like how many of you are there in the company?

Matt Turck [22:38] Who are you recruiting for? Like, anything?

Nancy Xu [23:07] Yeah, absolutely. So we just actually announced this with our launch, that we have raised about $10 million so far. We raised a round earlier last year, and then we had a lot of interest over the summer in raising a round pre-launch. So we did something small internally. So as part of that, the team at Khosla has been incredibly helpful. Folks at GV, you mentioned AIX. We also recently brought on Marc Benioff as one of our larger investors as well.

Nancy Xu [23:40] So from Salesforce, and we're looking more and more, as we scale, towards working with larger enterprise customers. And so, of course, Mark has been really helpful there, as well as some other additional investors, Barry Eggers, who is one of the founding partners of Lightspeed, Mike at Index, as well as some other folks there. And in terms of where our team is at today, so we're 15 people. We're trying to keep the team lean, but we are actively hiring. So, of course, internally, we're looking for ML engineers, full-stack engineers, recruiters, designers, PMs.

Nancy Xu [24:10] A lot of the sort of functional areas we're all hiring in. And I will say, as a company that works on AI in the recruiting space, we actually dogfood our own product. So we don't have, like, a specific internal recruiting team. We essentially act as customers to our own recruiting product and our recruiting org. And they treat us just like any other customer. Yes.

Matt Turck [24:18] So should people still apply, or you'll find them? Like, if they're good, you'll find them.

Nancy Xu [24:36] Please apply. We will try to find you, but it's always better when folks apply and they're interested, right? And we do actually review our Lever on a daily basis. So if you do apply, you can expect to hear back. And that's, I feel like, always something people should know.

Matt Turck [24:42] And are you distributed, or are you all in the Bay Area or San Francisco?

Nancy Xu [24:54] We're, I would say, hybrid. So most of the team lives in the Bay Area or New York, and we're remote-first, but we do offsites, like, four or five times a year. So quite frequently.

Matt Turck [25:31] Okay, very good. Okay, well, amazing on Moonhub. Sounds super interesting. What a wonderful application of AI, and I just love the way you position it and phrase it. All really, really exciting. Maybe if you could switch hats for a second and put your investor hat on. I saw you're a founder and GP of your own firm, and with a really interesting list of companies. So maybe walk us through how this came about, why you do it, and how you do it, I guess.

Nancy Xu [26:05] Yeah, sure. So for context for folks who are listening, I run a fund, Chum Ventures. It's my own personal venture fund. And a couple of years ago, I had the opportunity to essentially invest in several of my friends. And the way it came about was, I have a lot of friends who have since built incredible companies. And maybe four or five years ago, before starting Moonhub, I was observing all my friends starting these companies. I'm like, man, it would've been really great to invest in them.

Nancy Xu [26:31] And some of them are people that I knew from Stanford when I was there for either undergrad or PhD or MBA. And some of them I knew from before. Turns out, when I was much younger, I was on the U.S. Math Team, and it turns out there are a lot of people there who have since started their own companies, or Physics Olympiad teams, Computing Olympiads. It's kind of a small community, but it turns out a lot of those people ended up starting companies.

Nancy Xu [26:45] And being very successful. So I feel very lucky to have known them. And so I originally met—you might know, do you know Mike Volpi at Index?

Matt Turck [26:51] Yeah, of course. He actually spoke also on this podcast, Data Driven NYC, which is—

Nancy Xu [27:16] Okay, cool. Yeah, so Mike's amazing. And I met Mike maybe five, six years ago through Alexander Wang, who founded Scale. And Alex and I knew each other from way back. And Alex had introduced me to Mike at the time through a different contact. So a couple of years ago, Mike was like, oh, you should definitely think about—Index has this thing called—they make funds for people, basically, like individual funds. So I think, for example, Dylan at Figma also has one of these funds, and some of my other friends do too.

Nancy Xu [27:45] And so Mike was like, oh, you should come. We'll help set up a mini fund for you. And so I really owe it to Mike for helping me get started there. And because of that, I was able to invest in many of my good friends' companies early on. It's definitely not the rigorous investing process that you probably have, Matt, or any other major venture fund has. It was really just investing in my good friends, and I was already helping them out with their startups, helping them think through early on.

Nancy Xu [28:18] And I love helping them sort of just think through early on how to set up the company and stuff. And then I later had the opportunity to invest, which was awesome. And then after that, the first couple investments went pretty well. And so a couple of my friends just came in and they're like, hey, we'll put in some money. You can just keep investing from your own fund. And that's really how it started, Matt.

Nancy Xu [28:35] And I feel very lucky to have been able to be on the journey for a lot of great companies, many in the AI space, some of them like Replicate.com, some are seeing a lot of growth recently. So that's been great.

Matt Turck [29:02] And maybe, like any proper VC, we will use any opportunity to shamelessly plug their companies and talk about how great they are. So just to give you that moment, any company that you would highlight that you either—they're doing super well, or they're particularly interesting, or you just like them for whatever reason?

Nancy Xu [29:17] I just like them. Yeah, sure, I appreciate you giving me that chance. So definitely a couple of companies that I think really highly of. So I mentioned the folks at Replicate there. I think they've been in the news recently for building—

Matt Turck [29:21] So what does Replicate do, maybe for anybody in the audience that may not have ever heard of them?

Nancy Xu [29:45] Yeah, for sure. So Replicate essentially builds these AI models that they allow you to run open-source ML models. So a lot of folks will go to them and run their own ML models for some of the more recent models that are coming out. And they've been really good at helping run ML models in the cloud basically immediately after they become open source. And then we have some other companies, Perplexity.ai. They're an incredible company building consumer search, and everyone should try them out.

Nancy Xu [30:19] And then some more companies on more of the healthcare side, for example, Vyond. So it's a team that's building essentially AI-powered medical scribing for the telehealth and healthcare industry. And that team, Michael and Nikhil, both incredibly brilliant. And the company, I think, is really doing a lot of good in the world today. And yeah, the list goes on, but I think you can probably find online somewhere the list of companies that I've invested in. I will say nowadays, I'm very much doubling down on the companies we've invested in already.

Nancy Xu [30:31] And so opportunistically investing in some, again, good friends, but less active in sort of investing in new companies.

Matt Turck [31:09] Yes, there's only so much one can do. But maybe taking a step back as both an operator and an investor, there's so much going on in AI. Any sort of interesting area or project or company that you're intrigued by that people listening to this may want to look into, or any trend or anything that you find interesting?

Nancy Xu [31:31] Yeah, I think a lot of my interests stem from building Moonhub AI nowadays. And I think as we build the company, I'm seeing a lot of very interesting, what I call, ML-in-production trends that we're following. So one, for example, is obviously agents are a big thing. I consider what we build to be, in some cases, an instance of an agent. But I think, for example, the type of retrieval framework across semi-structured data, there isn't really good open source to do this today, as far as I'm aware, and we had to build some of this stuff in-house.

Nancy Xu [31:56] But I'm constantly assessing this trade-off of: should we build this in-house, or will someone build a company specifically to do this in six months, right?

Matt Turck [31:57] Yeah.

Nancy Xu [32:18] An example of this is there's the Pinecones, Weaviates, Chromas of the world that build vector DBs, and it's great for what I call more free-form text data, but it's actually not as good for some things that perhaps are more like your traditional Elasticsearch-type problems. And I see a lot of people in the industry building their own hybrid solutions of these two things. But I think if someone could really combine all these different data sources together with an LLM that's smart enough to access them, which I do know some of the bigger LLM companies I've talked to are building internally right now, it's just not released yet, that is something I think is very exciting.

Nancy Xu [33:00] And I think some of the other exciting challenges that I'm always on the lookout for are interesting approaches to UI/UX of agents. Building Moonhub AI has helped me really appreciate the power of a good UX. And inherently, I think a lot of these interfaces, like the conversational interfaces, are not very intuitive for new users, but it is very clear to me that some of this is going to be the future. And so it's like, how do you build those UI/UX interfaces that bridge what is intuitive for users today with what will be intuitive for them tomorrow?

Nancy Xu [33:31] And if anyone knows any great agent interfaces, I would love to see it. But I think the default I've seen is like half the screen is the AI, the other half is the normal application. Or for some of the more AI-native companies, what I consider Moonhub AI to be, it's like the full thing is the conversational interface, and then you actually have widgets within the conversation. But it's not clear to me that's the final ultimate UX that's the best for users.

Matt Turck [33:39] Yeah, really, really interesting. So do you think that, in terms of opportunities for startups and new products to be launched in the market and not being immediately killed by big tech, which obviously is a topic that everybody thinks about, this kind of either vertical or discrete problem, but in a full-stack way, is the way to go, where you do work at the model layer, but ultimately you deliver an application?

Matt Turck [34:14] Is that where you think most of the opportunities are, including some of the healthcare stuff that you've invested in, great companies that you've invested in?

Nancy Xu [34:31] Yeah, short answer is yes. You hit it on the nail, Matt. I've always been a big believer in vertical companies, and I think especially today, where you see the OpenAIs of the world, it's very hard to compete on a platform with a platform-level company unless you have substantial funding. And Chris Ré, I'm not sure if I ended up publishing this, but my PhD advisor Chris Ré and I actually wrote a piece together maybe four or five years ago on building vertical AI companies and some of the common challenges that they face and how you should think about them.

Nancy Xu [35:21] But basically, my entire career has been one big bet on building vertical AI agents, like the company I worked at before my PhD, AKASA, also a great company where we were building essentially AI to automate back-office healthcare, so revenue cycle. Amazing company. During my PhD, Chris and I worked a lot on these foundation models for more agent-like workflows. We have some papers from automating customer support and all these other vertical applications. But I do think for a startup to be competitive today, a lot of people talk about, oh, there's these new AI moats.

Nancy Xu [35:52] And what are moats in the AI era? I think there are definitely some moats that arise that are different from what you traditionally see, but I think a lot of the traditional software moats still apply, right? Like, you have deep integrations with the company, you have longstanding sales contracts with them. Elad Gil has a great piece on what are moats for startups that you can read on this. But I do think that vertical applications give you the best chance at creating a longer-term competitive moat relative to, I almost just assume that large language models will become commoditized, and then it's how do you actually nail the use case so the customer stays with you for many years and decades.

Matt Turck [36:37] Yeah, great. What's the reality of the whole AI agents thing? I mean, it's the, I guess for the last few months, the topic du jour that everybody, when you ask, oh, what are you excited about? I find people typically, my venture capital colleagues, saying, oh, AI agents. But what is the reality of it in production? And what are some examples, like real-world examples of AI agents actually at work?

Nancy Xu [36:55] Yeah, absolutely. And I'd love to hear your thoughts here too, Matt, because I'm sure you see a lot more companies on a day-to-day basis than I do. And I think the short answer is: don't be fooled by demos. I think a lot of people look at demos and they think immediately, like, wow, AGI. It can play chess, and then extrapolate from one intelligent task to: you have to realize the progression or hierarchy of knowledge is different for AI than it is for humans, right?

Nancy Xu [37:35] So for humans, you think if I can play chess, then I must also be able to walk, talk, eat, and do everything else. But it's not the same for an AI. If it can play chess, it doesn't entail it can do all these other things. And I think that is one common assumption which is wrong. Like, if you talk to laypeople, they think, oh, it can do this, now it must know everything else. I think going back to AI agents in production, there's a lot of cool demos out there, but getting something to really be production-ready is extremely difficult.

Nancy Xu [38:04] And this is something I've come to appreciate building Moonhub AI. Like, there's just some of the nuances. For example, it's great if you can get it to work 80% of the time, but there's probably no value until you get it working 99% of the time, right? And that last 19% or 5%, as people say, is extremely hard.

Matt Turck [38:05] Yep.

Nancy Xu [38:29] And it oftentimes is not like you just retrieve more data for the LLM. It's probably like you're going to have to build some hacky fixes on top of it in corner cases and resolve all the corner cases for your specific use case. And that's, I think, nailing the vertical application can give you somewhat of an edge against these more broad LLMs.

Matt Turck [38:54] As someone who's in the trenches, talking to customers, selling to customers every day, what's your sense of the level of ultimate receptivity of people to deploying AI, or is that not a question for Moonhub because you basically abstract away the whole AI and you just deliver great candidates?

Nancy Xu [39:09] I think people are, at least in our space, incredibly excited about the potential of AI, but no one wants to go first. Yeah. So it's like everyone wants to do it, but no one wants to be the first person to do it and see what could go wrong. And so I think for Moonhub, we've sort of benefited from this psychology in that many companies come to us because they see us as a way to help them transform into AI-first organizations.

Nancy Xu [39:45] And oftentimes talent is one of the first places and the easiest places where you can become sort of AI-first. So we help a lot of companies who are going through this transition where they have some mandate from their CEO, and their CEO says, "Hey, we all have to be an AI company by the end of this year," right? And then everyone's like, "Well, how do I become an AI company?" And everyone's just kind of running around not sure what to do.

Nancy Xu [40:11] And so, from Moonhub's perspective, we help a lot of those companies sort of become those AI-first companies. That said, though, we do somewhat, like you said, Matt, give people the option to use our AI to help better collaborate with their recruiter, but if they don't want to use it, they don't have to when it comes to Moonhub, right? So almost everyone chooses to use it, but it sort of gives them that optionality so that, I think, when you start selling to larger and larger customers, compliance becomes an issue.

Nancy Xu [40:32] So having the optionality of being able to abstract the compliance behind something that's more services-like is very helpful for selling into larger organizations.

Matt Turck [40:47] Yeah. Are you sensing that this big wave of hype we've had for the last nine months is starting to slow down a little bit from your perspective, or is that just as vibrant as ever?

Nancy Xu [41:07] Yeah, it's a great question. I think there's the hype from the investment perspective, and then there's the hype from what I call the talent perspective, and they're very correlated but not the same. So we have so much talent data that in the last couple of weeks, we've actually been releasing what we call talent insights. So, for example, our team has mapped out where all the ML engineers in the U.S. are, and it turns out the majority of them are in the Bay Area.

Nancy Xu [41:43] And that's why, if you want to build an AI company, you should probably go to the Bay Area, right? And I think another fun one that we've looked into recently is just this title of prompt engineer. A lot of people want to call themselves prompt engineer because it's like the cool thing on the street to be called, right? And if you look at the data, it's like basically until 2023, no one called themselves a prompt engineer. And then in 2023, all of a sudden, 1,000 people in SF are now a prompt engineer.

Nancy Xu [42:09] And that number is actually relatively stable in the last couple months. It wasn't like everyone's trying to become a prompt engineer. Maybe they're going for something else. I don't know. And maybe now it's like a RAG expert or something. I'm not sure. But it could be something like that. I think on the investor side, we raised a round recently. Most of it was—we had a lot of external interest, but we didn't want to raise that big of a round.

Nancy Xu [42:34] So we decided mostly to go internal where possible. I think the excitement is still there, but I think there is more of a reservation around people wanting to find, like, I often hear people saying, "I want to find applications of AI agents with real economic value," is sort of what I hear. And I think people are starting to realize, like, well, ultimately these companies do have to be real companies, and there needs to be a real customer proposition there.

Nancy Xu [43:08] And I'm really curious to see how these sort of large open-source projects will evolve, because I'm not naming any particular ones, but there are definitely some that have tens of thousands of stars on GitHub. But if you look at where those stars are coming from, it's mostly hobbyists. And I'm sure you can monetize hobbyists and sort of grow from there. But I think it is a very big stepping stone for these AI companies to go from hobbyists to becoming something that enterprises will actually want.

Matt Turck [43:51] Now that makes a lot of sense. Okay. Well, it sounds like in this fast-evolving world, you are incredibly well positioned as a full-stack vertical application, as we discussed. So look, it's been super interesting. Congratulations on the launch. Congratulations on all the progress and everything you've built so far. And I'm very excited to see how you continue to progress. Where can people find you online and learn more about the company and all those things?

Nancy Xu [44:06] Well, you can learn more at Moonhub.ai. Click Get Started. I will personally be your customer support if you need it for the product. And of course, feel free to shoot me an email if you're looking for a new opportunity. We're always hiring.

Matt Turck [44:19] And on Twitter, LinkedIn, wherever, do you—

Nancy Xu [44:29] Twitter, it's—yeah, on Twitter it's @NancyXUS. That's me. On LinkedIn, search up my name, you'll find me. You can always shoot me a DM or send me an email.

Matt Turck [44:44] Okay, wonderful. All right, well, I really enjoyed the conversation. Thank you so much, Nancy, and we'll look forward to seeing how you grow. Very excited for the company and the positioning thing. It's super interesting. Thank you.

Nancy Xu [45:17] Thank you so much, Matt. Thanks, everyone. Appreciate it. Thanks for joining us for The MAD Podcast. We're back here every Wednesday with new conversations with leaders in the machine learning, AI, and data space. And if you like this show, you can also find a video recording of not only this episode, but many, many more over on the Data Driven NYC YouTube channel. Thanks again, and catch you next week.