The Power of Quality Human Data with SurgeAI Founder and CEO Edwin Chen

0:08Edwin Chen, founder and CEO of Surge, the bootstrapped human data startup that surpassed a billion in revenue last year and serves top tier clients like Google, open AI and top 10. We talk about what high-quality human data means, the role of humans as models become superhuman, benchmark hacking, why he believes in a diversity of frontier models, the scale meta not M &A deal, and why there's no ceiling on environment quality for RL or the simulated worlds that labs want to train agents in. Edwin, thanks for joining us. Great, great. Seeing you guys today. Surge has been really under the radar until just about now.

0:44Can you give us a little bit of color on the scale of the company and what the original founding thesis was? So we hit over a billion written revenue last year. We are kind of like the biggest human data player in this space. And we're about 100, a little over 100 people. And our original thesis was, we just really believed in the power of human data to advance AI. and we just had this really big focus from the start of making sure that we had the highest quality data possible. Can you give people context for how long you've been around, how you got going, et cetera? I think, again, you all have accomplished an enormous amount in a short period of time.

1:22And I think, you know, you've been very quiet about some of the things you've been doing. So it'd be great to just get a little bit of history and, you know, when you started, how you got started and how long you've been around. Oh, yeah. So we've been around for five years. I think we just hit our five-year anniversary. So we started in 2020. So before that, so I can give some of the context. So before that, I used to work at Google, Facebook, and Twitter. And one of the, like, basically the reason we started Surge was I just used to work on ML at a bunch of these big companies. And just the problem I kept running into over and over again was that it really was impossible getting the data that we needed to train our models.

1:53So it was just this big blocker that we faced over and over again. And there was just like so much more that we wanted to do. Like, even just the basic things that we want to do, we struggled so hard to get the data. It was really just the big blocker. But then simultaneously, there are all these more futuristic things that we wanted to build. Like if we thought of the next generation AI systems, if we could barely get the data that we needed at the time to solve, like just building a simple set analysis classifier, if we could barely do that, then how would we ever advance beyond that? So that really was the biggest problem.

2:25I can go into more of that, but that was essentially what we faced. And you guys are also known for having bootstrapped the company versus raising a lot of external venture money or things like that. Do you want to talk about that choice in terms of going profitable early and then scaling off of that? In terms of why we didn't raise. So I think a big part of it was obviously just that we didn't need the money. I think we were very, very lucky to be profitable from the start. So we didn't need the money. It always felt weird to give up control. And like one of the things I've always hated about Silicon Valley is that you see so many people raising for the sake of raising.

2:54Like I think one of the things that I often see is that a lot of founders that I know, They don't have some big dream of building a product that solves some idea that they really believe in. Like if you talk to a bunch of YC founders or whoever it is, like what is their goal? It really is to tell all their friends that they raised$10 million and show their parents they got a headline on TechCrunch. Like that is their goal. Like I think of like my friends at Google. They often tell me, oh, yeah, I've been at Google or Facebook for 10 years and I want to start a company. I'm like, OK, so what problem do you want to solve?

3:23They don't know. They're like, yeah, I just want to start something new. I'm bored. And it's weird because they can like pay their own salaries for a couple of months. Again, they've been on Google and Facebook for 10 years. They're not just like fresh out of school. They can pay their own salaries. But the first thing they think about is just going out and raising money. And I've always thought it weird because they like might try talking to some users and they might try building an MVP, but they kind of just do it in this throwaway manner where the only reason they do it is to check off a box on a startup accelerator application.

3:48And then they'll just pivot around these random product ideas. And they happen to get a little bit of traction so that the VC DMs them. And so they spend all their time tweeting and they go to these VC dinners and it's all just so that they can show the world that they raise a big amount of money. And so I think raising immediately always felt silly to me. Like everybody's default is to just immediately raise. But if you were to think about it from first principles, like if you didn't know how Silicon Valley worked, if you didn't know that raising was a thing, like why would you do that? Like what is money really going to solve for 90 % of these startups where the founders are lucky to have some savings?

4:16I really think that your first instinct should be to go out and build whatever you're dreaming of. And sure, if you ever run into financial problems, And sure, think about raising money then, but don't waste all this effort and time up front when you don't even know what you do with it. Yeah, it's funny. I feel like I'm one of the few investors that actually tries to talk people out of fundraising often. Like I actually had a conversation today where the founder was talking about doing a raise. I'm like, why? You know, you don't have to. You can maintain control, et cetera. And then the flip side of it is I would actually argue outside of Silicon Valley, too few people raise venture capital when the money can actually help them scale.

4:45And so I feel like in Silicon Valley, there's too much and outside of Silicon Valley, there's too little. So it's this interesting, you know, spread of different models that sort of stick. Edwin, what would you say to founders who feel like there's some external validation necessary to, especially like hire a team or scale their team? This is a very like common complaint or rationale for going and raising more capital. I think about it in a couple of ways. So I guess it depends on what you mean by external validation. like in my mind again like i often think about things from a perspective of are you trying to build a startup that's actually going to change the world like do you have this big thing that you're dreaming of and if you have this big thing that you're dreaming of you like why do you care maybe the way to think about it is in sarah's context like if you haven't say you're a yc founder you haven't been at google you haven't been at meta you haven't been at twitter you don't have this network of engineers you're a complete unknown you haven't worked with very many people, you're straight out of school.

5:43How do you then attract that talent? And to your point, you can tell a story of how you're going to build things or what you're going to do, but it is a harder obstacle to basically convince others to join you or for others to come on board or to have money to pay them if you haven't, if you don't have long work at history. So I think maybe that's the point Sarah's making. Yeah. So I mean, I think I would differentiate between maybe two things. Like one is, do you need the money? So first of all, like there's a difference between people who are, yeah, like literally fresh out of school or maybe, you know, I've never gone to for the first place.

6:11And so maybe they don't have any savings. And so they literally need some money in order to live. And then there's others who, okay, like let's assume that you don't necessarily need money because again, you've been working Google or Facebook for 10 years, like, or five years, whatever it is, you have some savings. So I would say one of the questions is, again, like it kind of, the path kind of differs depending on those two choices or those two scenarios. But I think one of the questions is, well, do you really need to go out and hire all these people? Like one of the things I often see, again like i'm curious what you guys see but one of the things i often see is founders will tell me like uh okay so i'm trying i'm trying to think about the first few hires i'm going to make and they're like yeah i'm going to hire a pm i'm going to hire a data scientist yeah these are one of my first five to ten hires i'm like what like this is just wild to me like i would never hire data scientists when the first few people in a company and i say this because i used to be a data scientist like data scientists are great when you want to optimize your product by two percent or five percent but that's definitely not what you want to be doing when you start a company you're trying to swing for 10x or 100x changes, not worrying and nipping about small percentage points that are just noise anyways.

7:14And to some of the product managers, like product managers are great when your company gets big enough, but at the beginning, you should be thinking about yourself about what product you want to build. And your engineer should be hands-on. You should be having great ideas as well. And so product managers have this weird conception that big companies have when your engineers don't have time to be in the weeds on the details and try things themselves. And it's not a role that you come up with the other one before. So I guess with the initial search team, It sounds like you had sort of a small initial tight engineering team.

7:40You guys started building product. You were bootstrapping off of revenue. You know, at this point, you're at over a billion dollars in revenue, which is amazing. How do you think about the future of how you want to shape the organization, how big you want to get, the different products you're launching and introducing? Like, what do you view as sort of the future of Surge and how that's all going to evolve? Before we do that, can you just explain like what the, at whatever level of detail makes sense here, like what the billion dollars of revenue is? Maybe like how product supports the company, who your data, who your humans are, because I think there's just very little visibility into all of that.

8:14So in terms of what our product is, I mean, at the end of the day, our product is our data. Like we literally deliver data to companies and that is what they use to train and evaluate our models. So imagine, you know, one of your one of these frontier labs and you want to improve your model, your model's coding abilities. What we will do on our end is we will gather a lot of coding data. And so this coding data may come in different forms. and maybe SFT data. We are literally writing out coding solutions or maybe unit tests. Like these are the tests that a good piece of code must pass. Maybe it's preference data where it's okay, like here are two pieces of code or here are two coding explanations.

8:49Which one is better? Or these might be like verifiers. Like, okay, here's a web app that I created. I want to make sure that in the top right hand of the screen, there's like a, there's a login button. Or I want to make sure that when you click this button, something else happens. Like there's a bunch of different forms that this data may take At the end of the day, what we're doing is we're delivering data that will basically help the models improve on these capabilities. Very, very related to that is this notion of evaluating the models. Like you also want to know, yeah, is this a good coding model?

9:16Is it better than this other one? What are the areas in which this model is weak and this model is worse? Like what insights can we get from that? And so in addition to the data, oftentimes we're delivering insights to our customers. We're delivering loss patterns. We're delivering failure modes. So there may be a lot of other things like related to the data, But I didn't think it's like this universe of like applications or just like just universe around the data that we deliver and that that is our product. Yeah. And maybe going back to Aloud's question, maybe like product isn't actually the right word here.

9:46But what's what's like repeatable about the company or what are like core capabilities that you guys have that you would say your competitors, you know, fill to meet the mark? The way we think about a company is that, and the way we differentiate from others, is that a lot of other companies in this space, they are essentially just body shops. What they are delivering is not data. They are literally just delivering warm bodies to companies. And so what that means is, at the end of the day, they don't have any technology. And one of our fundamental beliefs is that, again, quality is the most important thing at the end of the day.

10:18Is this high quality data? Is this a good coding solution? Is this a good unit test? Is this mathematical problem solved correctly? is this a great poem? And basically a lot of companies in this space, like just as a relic of how things have worked out historically, it's that like historically a lot of companies, they've treated quality and data as commodity. Like one of the ways we often think about it is, imagine you're trying to draw a bounding box around a car. Like Sarah, you and I, we're probably going to draw the same bounding box. Like ask Hemingway and ask a second grader. Well, at the end of the day, we're all going to draw the same bounding box.

10:52There's not much difference that we can do. So there's a very, very low ceiling on the bar of quality. But then take something like writing poetry. Well, it's like I'm writing poetry. Hemingway is definitely going to write a much better poem than I am. Or imagine, I don't know, a VC pitch deck. You're going to write a much better, you're going to create a much better pitch deck than I will. And so there's almost an unlimited ceiling in this Gen AI world on the type of quality that you can build. And so the way we think of our product is like we have a platform. We have actual technology that we're using to measure the quality that our workers or annotators are generating.

11:24If you don't have that technology, if you don't have any way of measuring it. Is the measurement through human evaluation? Is it through model-based evaluation? I'm a little bit curious how you create that feedback loop since to some extent it's a little bit of this question of how do you have enough evaluators to evaluate the output relative to the people generating the output? Or do you use models? Or how do you approach it? I think one analogy that we often make is think about something like Google search or think about something like YouTube. Like you have, you know, millions of search results.

11:50You have millions of web pages. You have millions of videos. How do you evaluate the qualities of these videos? Like, is this a high quality, like, is this a high quality web page? Is it informative? Or is it really spammy? Like in the way you do this is like you just need, I mean, you gather so many signals. You gather like page dependent signals. You gather like user dependent signals. You gather activity based signals. And all of these feed into, you know, a giant MOI gram at the end of the day. And so in the same way, we gather all these signals about our annotators, about the work that they're performing about like their activity on the site.

12:18And we just feed it into a lot of these different, like we basically have an ML team internally that builds a lot of these algorithms to measure all of this. What is changing or breaking as you are like scaling, increasing these sophisticated like annotations, right? Like if, you know, model quality baseline is going up every couple of months, then the expectation is that like exceeds, you know, what might've been a random human at some point, as you said, can draw a bounding box into all of these different fields where we have modeled better than the 90th percentile at some point. So this is actually something that we do a lot of internal research on ourselves as well.

12:58So there's basically this field of AI alignment called scalable oversight, which is basically this question of how do you have models and humans working together hand in hand to produce data that is better than either one of them can achieve on their own. And so even like even today, something like writing an SIT story from scratch, even today, like a couple of years ago, we might have written that story completely from scratch ourselves. Today, it's just like not very efficient, right? Like you might start with a story that a model created and then you would edit it. You might edit it in a very substantial way.

13:29Like maybe just the core of it is very vanilla, very generic, but there's just so much like kind of cruft that it's just inefficient for a human to do and doesn't really benefit from like the human creativity and human ingenuity that we're trying to add into the response. And so you can just start with like this bare bones structure that you're basically just layering on top of. And so again, there's more sophisticated ways of thinking about scalable oversight, but just this question of how do you build the right interfaces? How do you build the right tools? How do you just combine people with AI in the right ways to make them more efficient?

13:59It is something that we build a lot of technology for. A lot of the discussion in terms of what human data the labs want has moved to RL environments and reward models in recent months. What is hard about this? Or what are you guys working on here? So we do a lot of work building our environments. And I think one of the things that people really underestimate is how it is how complicated it is that you can't just synthetically generate it. Like, for example, you think you need a lot of tools because these are massive environments that people want. Can you give an example of like, just to make it more real.

14:35Like imagine you are a salesperson. And when you are a salesperson, you need to be interacting with Salesforce, you need to be getting leads through Gmail, you're going to be talking to customers in Slack, you're going to be creating Excel sheets, tracking your leads, you're going to be, I don't know, writing Google Docs and making PowerPoint presentations to present things to customers. And so you want to basically these very rich environments that are literally simulating your entire world as a salesperson. Like it literally is just like imagine like your entire world. So it's everything on your desktop.

15:06And then in the future, everything that is, you know, not on your desktop as well. Like maybe you have a calendar, maybe there's, maybe you need to travel to a meeting to meet a customer and then you want to simulate a car accident happening and you're getting notified of that. So you need to like leave a little bit earlier. Like all these things are things that we actually want to model in these very, very rich RO environments. And so the question is, how do you generate all the data that goes into this? Like, okay, you're going to need to generate like thousands of slack messages hundreds of emails you need to make sure that these are all consistent with each other you need to make sure that like going back to like my core example you need to make sure that time is evolving in these environments and like certain like external events happen like how do you do all this and then um like in a way that is actually kind of like interesting and creative but also realistic and not like incongruent with each other uh like there's just like a lot of thought that needs to go into these environments um to make sure that they're, again, like rich creative environments that the models can learn interesting things from.

16:02And so, yeah, you basically need like a lot of tools and balance of sophistication for creating these. Is there any intuition for like how real or how complex is enough? Or is it just like, you know, there's no ceiling on the realism that is useful here or the complexity of environment that is useful here? I think there's no ceiling. Like at the end of the day, you just want as as much diversity and richness as you can get. Because the more richness that you have, yeah, the more the models can learn from. The longer the time horizons, the more the models can learn on and improve on. So I think there's almost an unlimited ceiling here.

16:38If you were to make a five or 10 year bet on what scales most in terms of demand from people training AI models and types of data, is it RL environments? Or is it traces on types of expert reasoning? Or what other areas do you think there's going to be a really large demand for. I mean, I think it will be all of the above. Like, I don't think our environments alone will suffice just because, I mean, it depends on how you think but there are environments. But oftentimes these are very, very rich trajectories that are very, very long. And so it's almost like inconceivable that a single reward, I mean, I think even today, we often think about things in terms of multiple rewards, not just a single reward.

17:16But like a single reward just may not be like rich enough to capture all the work that goes into like the model solving some very, very complicated goal. So I think it would probably be a combination of all those. If you assume eventually some form of superhuman performance across different model types relative to human experts, how do you think about the role of humans relative to data and data generation versus synthetic data or other approaches? Like at what point does human input sort of run out as a useful point of either feedback or data generation? So I think human feedback will never run out.

17:54And that's for a couple reasons. So I mean, even if I think about the landscape today, I think people often overestimate the role of synthetic data. Like I personally, I think synthetic data actually is very, very useful. Like we use it like a ton ourselves in order to supplement what the humans do. Like again, like I said earlier, there's like a lot of cruft that simply isn't worth a human's time. but what we often find is that like for example a lot of times our customers will come to us and be like yeah for the past six months i've been experimenting with synthetic data i've gathered 10 to 20 million pieces of synthetic data actually yeah we finally realized that 99 percent of it just wasn't useful and so we're trying to find right now we're trying to curate the five percent that is useful but we are literally going to throw out nine million of it and oftentimes they'll find out that yeah like actually a thousand even a thousand pieces of high quality human data hydrated, really, really high quality human data is actually more valuable than those 10 million points.

18:46So that is one thing I'll say. Another thing I'll say is that it's almost like sometimes you need an external signal to the models. Like the models just think so differently from humans that you always need to make sure that they're kind of aligned with the actual objectives that you want. Let me give two examples. So one example is that it's kind of funny. Sometimes if you try, So one of the frontier models, let me just say that one of them. If you go use the frontier model, it's like one of the top models or one of the models everybody thinks is one of the top. If you go use it today, like maybe 10 % of the time when I use it, it will just output random Hindi characters and random Russian characters in the model of my responses.

19:23So I'd be like, tell me about Donald Trump. Tell me about Barack Obama. And just like in the middle of it, it will just output Hindi and Russian. It's like, what is this? And the model just isn't like self-consistent enough to be aware of this. It's almost like you need an external human to tell the model that, yeah, this is wrong. One of the things I think is a giant plague on AI is LMSIS, LMRNA. And I'll skip the details for now. But I think right now, people will often... It's like if you train your model on the wrong objectives. So the mental model that you should have of LMSIS, LMRNA is that people are writing prompts.

19:59They'll get two responses. And they'll spend like five, ten seconds looking at responses. and they'll just pick whichever one looks better to them. So they're not evaluating whether or not the model hallucinated. They're not evaluating the factual accuracy and whether it followed the instructions. They're literally just vibing with the model. I'm like, okay, yeah, this one seemed better because it had a bunch of formatting. It had a bunch of emojis. It just looks more impressive. And people will train on basically an LMS objective and they won't realize all the consequences of it. And again, the model itself doesn't know what its objective is.

20:29It's like you almost need an external quality signal in order to tell it what the right objective should be. And if you don't have that, then the model will just go in all these crazy directions. Again, you may have seen some of the results with LOM before, but we'll just go in all these crazy directions that kind of mean you need these external validators. This also happens actually when you do different forms of protein evolution or things like that, where you select a protein against a catalytic function or something else, and you just kind of randomize it and have a giant library of them. And you end up with the same thing where you have these really weird activities that you didn't anticipate actually happening.

21:04And so I sometimes think of model training as almost this odd evolutionary landscape that you're effectively evolving and selecting against and you're kind of shaping the model into that local maxima or something. And so it's kind of this really interesting output of anything where you're effectively evolving against a feedback signal. And depending on what that feedback signal is, you just end up with these odd results. So it's interesting to see how it kind of transfers across domains. These, you know, coarse, as you said, five-second reaction academic benchmarks or even non-academic industrial benchmarks are easily hacked or like not the right gauge of performance against any given task.

21:43They are very popular. What is the alternative for somebody who's trying to like choose the right model or understand model capability? So the alternative that I think all the frontier labs view as the gold standards is basically human evaluation. So again, proper human evaluation where you're actually taking the time to look at the response. You're going to fact check it. You're going to see whether or not it followed all the instructions. You have good taste. So you know whether or not the model has good writing quality. Like this concept of like doing all that and spending all the time to do that, as opposed to just vibing for five seconds, I think actually is really, really important.

22:15Because if you don't do this, you're basically just training your models on the analog of clickbait. So I think it actually is really, really important for model progress. If it's not LMSYS, like how should people actually evaluate model capability for any given task? What all different here labs find is that human evals really are the gold standard. Like you really need to take a lot of time to fact check these responses, to verify the following instructions. You need people with good taste to evaluate the writing quality and so on and so on. And if you don't do this, you're basically training your models on the analytical clickbait.

22:49And so I think that really, really harms model progress. Is there work that Surge is doing in this domain of like trying to standardize human eval or make it more transparent to end consumers of the API or even users? So internally, we do a lot of work actually today with working with all the frontier labs to help them understand their models. So again, we're constantly evaluating them. We're constantly surfacing loss areas for them to improve on and so on and so on. And so right now, a lot of this is internal. But one of the things we actually want to do is start external forms of this as well, where we're helping educate people on, yeah, like these are the different capabilities of all these models.