The Startup Powering The Data Behind AGI

In this episode of Gradient Dissent, Lukas Biewald talks with the CEO & founder of Surge AI, the billion-dollar company quietly powering the next generation of frontier LLMs. They discuss Surge's origin story, why traditional data labeling is broken, and how their research-focused approach is reshaping how models are trained. You’ll hear why inter-annotator agreement fails in high-complexity tasks like poetry and math, why synthetic data is often overrated, and how Surge builds rich RL environments to stress-test agentic reasoning. They also go deep on what kinds of data will be critical to future progress in AI—from scientific discovery to multimodal reasoning and personalized alignment. It’s a rare, behind-the-scenes look into the world of high-quality data generation at scale—straight from the team most frontier labs trust to get it right. Timestamps: 00:00 – Intro: Who is Edwin Chen? 03:40 – The problem with early data labeling systems 06:20 – Search ranking, clickbait, and product principles 10:05 – Why Surge focused on high-skill, high-quality labeling 13:50 – From Craigslist workers to a billion-dollar business 16:40 – Scaling without funding and avoiding Silicon Valley status games 21:15 – Why most human data platforms lack real tech 25:05 – Detecting cheaters, liars, and low-quality labelers 28:30 – Why inter-annotator agreement is a flawed metric 32:15 – What makes a great poem? Not checkboxes 36:40 – Measuring subjective quality rigorously 40:00 – What types of data are becoming more important 44:15 – Scientific collaboration and frontier research data 47:00 – Multimodal data, Argentinian coding, and hyper-specificity 50:10 – What's wrong with LMSYS and benchmark hacking 53:20 – Personalization and taste in model behavior 56:00 – Synthetic data vs. high-quality human data Follow Weights & Biases: https://twitter.com/weights_biases https://www.linkedin.com/company/wandb

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:02Lukas Biewald?

  You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host, Lucas Biewald. Today, I'm talking with Edwin Chen, who's the CEO of Surge. I was really looking forward to talking to Edwin for a long time for a number of reasons. I mean, one thing is that Edwin has built an incredibly valuable business with no venture backing in a really short amount of time. So I think he started this data collection business in 2020. In 2024, he crossed a billion dollars in revenue, which is just spectacular success. But that's probably not even the top reason I wanted to talk to him. The human data collection business is this really important part of building AI systems that's not talked about enough.

• 0:45Lukas Biewald?

  I got actually into that business in 2026, 2027. I started a company called CrowdFlower that did this data collection in the early days. And around the time I sold CrowdFlower is around when Edwin started Surge. And the data collection business has changed a lot over the years, but one constant has been that it's a huge spend that people building high-quality models do. So Edwin has these front row seats into what most of the foundation labs, what's the foundation model builders, what they're doing and how they're thinking and how they're building their models. So he has a lot of insights. He doesn't do a lot of podcasts, so we were really lucky to get him. I hope you enjoy this conversation.

• 1:27Lukas Biewald?

  So you're the first guest we've had in the human data generation space. And it's a space, obviously, I was in for a long time. So I'm really curious about this. Could you, first of all, start by telling the story of Surge, what you were thinking when you started it, and how it's gone?

• 1:48Speaker 01

  Yeah, I can give the founding story. So basically, I used to be an MO engineer at a bunch of these big companies. And the problem I kept running into was that we just kept facing all of these issues, getting the data that we needed to train our models. So for example, I used to work on our search and ad systems at Twitter. And one of the first things I wanted to do was build a sentiment classifier. So sentiment analysis is super simple. Problem and all we needed was 10,000 tweets to train our models. But our human data system at the time was literally two people we'd hired off of Craigslist, working nine to five. So we had to wait a month to get started. And then we had to wait another month for them to label the tweets in the spreadsheet because our tools are just terrible. And when we finally got the data back, we were just seeing that was complete junk. They didn't understand slang like, you know, she's such a bad bitch. So they were labeling this negative.

• 2:33Speaker 01

  And they didn't understand hashtags and all these other aspects of tweets. And so what ended happening was I just spent a week myself labeling all these tweets because that was actually so much faster and better. And like, I think one of the things that we often said was that this was really simple things. Like, yeah, at the end of the day, sentiment analysis isn't all that complicated. And at the same time, we just had this bigger problem going to solve around how we wanted to optimize our analysis for the right objectives. Like when I first started working on Twitter, this was the old days when it was purely chronological timeline. And so one of the things we wanted to do was make it easier for users to discover tweets they really cared about. And so the question was, how do we, how do we train our recommendation algorithms? And the obvious choice was clicks and retweets at a time. Like you just train your algorithms to produce as many clicks and retweets as possible. But the problem is we tried doing these things and it turned out to be this incredibly negative feedback loop.

• 3:19Speaker 01

  Like, once you optimize for clicks, you get the most clickbaity content in the world rising to the top. You get lots of racy content, you get lots of bikinis, you get lots of listicles about 10 horrifying syndishes and so on. And so we wanted to train our models on all these deeper principles instead, where we'd ask the human, human raiders to label tweets and recommendations according to our product principles. But if we just couldn't even get simple sentiment analysis right, we definitely couldn't get this more complex data at the quality and scale we needed. So, um, yeah, so I guess eventually, uh, there's a problem that happened over and over again at Google and Facebook too. And so eventually I just realized it's just something that I needed to go out and build, build, build myself.

• 3:53Lukas Biewald?

  All right. So what, what year did you start, uh, Surge?

• 3:57Speaker 01

  So we started Surge in 2020, uh, right in the middle of pandemic.

• 4:01Lukas Biewald?

  Oh, nice. And so, um, I feel like 2020, I guess, you know, there were some, you know, established data labeling, um, data generation companies at that time. Did you have a particular, um, you know, take on the space?

• 4:17Speaker 01

  Yeah, so our take on a space was that all of these solutions out there, they were basically focused on this idea of very commodity labeling, a very low skill. Like the example I often give is take the problem of drawing a bounding box around a car. Like, yeah, you and I, we can all draw bounding boxes around cars. Like, almost like a three-year-old can draw bounding boxes around a car. The bounding box that I draw isn't going to be any different from the bounding box that Terence Tausch draws or Einstein would draw. It's like a very, very low ceiling on the complexity of data that is required. In contrast, if you think about all the things that we want to do today, like, yeah, we want models that can write poems. We want models that can solve relativistic equations.

• 4:59Speaker 01

  There's almost like an unlimited amount of intelligence that we want to feed our models. And so all the other solutions at the time, yeah, they were designed for like a very low-scale commodity style of labor. And so it was just, it was not focused on quality at all. Instead, it was focused on scale.

• 5:16Lukas Biewald?

  I see. And so how did you source the people in the beginning?

• 5:21Speaker 01

  So literally, the first few people that were on our platform were like, they'd actually just been, it's kind of funny, but I've kind of been working on this problem for a long time. And so I already had a network of like people who are really, really interested in doing this kind of work. And so when they heard that I started Surge, I mean, a lot of it was also me in the beginning, but it was also, you know, all these laborers that everybody, in some sense, accumulated throughout the years.

• 5:47Lukas Biewald?

  That's cool. And then who were the early customers?

• 5:52Speaker 01

  So the early customers were a lot of like tech companies. So there were a lot of, so we had this idea that we wanted to really, really focus on engineers and research scientists who really, really understood the quality of data. And so these were a lot of, just again, like, I kind of been working in this space for a while. And so there were just like a lot of friends or just contacts I had at all these companies, who, yeah, had been dying for this kind of higher-quality, kind of like next-generation solution. We could just do far more advanced tasks than what were possible at the time. And so, yeah, like there were a lot of tech companies like, I think like Airbnb, Twitter, like a lot of the just like startups in the search and ad run space.

• 6:31Lukas Biewald?

  And at that time, maybe 2020, I mean, that was actually when I, right around when I was leaving the space. I remember like, it seemed like what was really taking off then was sort of like autonomy and, you know, robotics, a lot of vision applications. But it sounds like you were more focused on text. Is that fair?

• 6:50Speaker 01

  Yeah, so we were always focused on language and what are called behavior from the beginning. I wouldn't say it was necessarily text per se, like there are a lot of complex problems in the image space too, especially nowadays. But what we didn't want to focus on, like what we've never done is like very, very simple image labeling tasks or a very simple, like, you know, bounty box style tasks where I just don't really think that there's any intelligence really needed to build such such solutions. And so we always focus on this like higher complexity, higher, kind of higher scale, higher sophistication space.

• 7:27Lukas Biewald?

  Well, I would say, I think the, I think the image labeling, when you actually like dig into it is, is more, more complicated, you know, for the record, but I, I take your point. I guess, do you, it seemed like you really built Surge under the radar, like, was that an intentional decision? I mean, I think you might have said recently that you're over a billion dollars in yearly revenue. Is that, is that right? Can, can we say that?

• 7:53Speaker 01

  Yeah, yeah, yeah. We were over a billion last year and we kind of like hit or building number ARR a while ago.

• 8:01Lukas Biewald?

  I mean, that's like incredible achievement, you know, in just like five years and you did it with no outside funding. Is that right?

• 8:09Speaker 01

  Yeah.

• 8:10Lukas Biewald?

  That must be, that's a historic level of growth of that funding, I think. I mean, did you sort of intentionally avoid the VC path?

• 8:22Speaker 01

  So I think one of the things that was really important for us was we wanted customers to be buying us because they really, really believe in having high-quality data and not because they saw us mentioned in some tech article. Like, we really wanted partners who had the same vision we did and who could go out and show the world how important data actually was. But again, if you contrast with the kinds of data people were looking for five, 10 years ago, people, again, really weren't focused on quality. They treated data as this kind of commodity where they would, like, the researchers themselves, they would barely look at the data, they would just let their vendor management teams kind of handle the entire process because they just wanted to outsource it. And so, yeah, we just had this idea where we really, really wanted to focus on people who believed in quality or were buying us for that reason.

• 9:07Lukas Biewald?

  But so, I mean, to get to a billion dollars plus in revenue, you've obviously had to kind of scale your system. So that goes beyond, you know, like, hiring a few people that you've known for a long time. So can you talk about like how you've scaled your processes as your scope has expanded? Like, do you focus, do you think it's more important to like, you know, kind of scale the technology that's enabling this? Or is it like, you know, internal human processes? Or is it all in like the hiring? Or is it in this kind of like management of what's going on? As much as you could say about how this works, I'd love to hear it.

• 9:44Speaker 01

  Yep. Yeah, so I think the thing that people underestimate in this space is how much technology you actually want to build. Like, people tend to think that humans are smart, and so if you just throw 10,000, 100,000 humans a problem, they'll solve it. Like, it's a little bit crazy to me. Like, sometimes we'll, I'll interview candidates from some of the competitors in our space, and when they describe to me what they're building and how they operate, it's just incredibly, incredibly manual. Like, they are kind of just body shops at the end of the day, and they literally have no technology. Like, if you ask them... Could you tell me the quality of this worker? Could you tell me how good this worker is at this particular task?

• 10:28Speaker 01

  Could you show me a dashboard with how your quality is improving over time? And what A B tests you run and what like algorithms you're building to improve it? And they literally can't because all they're doing a lot of times is like even some of these so-called technology companies, they're literally just dumping data into spreadsheets and their employees, their engineers are literally either creating the data themselves or reviewing it themselves. And so there almost is like no technology that's going on. And so I think what we have to realize of this space is that quality control actually is incredibly difficult. And so if you want to get the highest quality out there, you need to find or you need to create a lot of these kind of really sophisticated algorithms to detect the highest quality,

• 11:13Speaker 01

  highest quality data that you can. And it's incredibly complicated because in this space that we have with L L M's a day, you really want L L M's to be good at every task in the world. So it's not just a single domain. It's making sure that they're really good at poetry and at the end of the day, you really want to make sure that they're really good at physics. And so how do you find the highest quality, highest quality data in order to train the models and then how you also remove the worst at worst? And so, yeah, you kind of can't just throw warm bodies out of your brain. You need to build a lot of technology to manage it.

• 11:46Lukas Biewald?

  Look, I mean, you know, I'm kind of coming from a different era. I think of this, you know, labeling where like a lot of labels were simple, you know, kind of like, you know, yes, no, or, you know, kind of simple tasks where it's really clear, you know, what's right and what's wrong. But I think, you know, you're doing much more complicated tasks. And like you're saying, I think, you know, even getting two people to agree on high quality and low quality as the task gets more complicated, I think it's like, you know, more and more difficult. And so can you talk a little bit about like what technology you actually have to build to, to manage quality for like a generic, you know, complicated expert task involving language?

• 12:29Speaker 01

  So let me start actually by contrasting with a very old school take on data and data quality and how we just think about the problem differently. And so, yeah, so let me give you an example. So let's say you wanted to train a model to write an eight line poem about the moon. And the way most companies think about it is, okay, well, let's just hire a bunch of people from Craigslist or through some recruiting agency. Let's ask them to write poems. And the way to think about quality, again, going back to the image annotation days, it's like, okay, is this a poem? Is it eight lines? Does it contain a word moon? They just check all these boxes in the same way you might check, okay, this is a cat, it's a dog. They're just checking boxes and then they're saying, sure, this is a great poem because it follows all the instructions, it follows all these checkboxes.

• 13:15Speaker 01

  Then what happens is, okay, you get these terrible poems that feel like they're written by kids in high school, right? Like a kid in high school, sure, they can write an eight line poem about the moon. But is it a great poem? Is it like an evocative poem? Is it a type of poem that a Nobel Prize laureate would write? No. And so checking boxes doesn't work. And so some other companies might be like, okay, sure, these people on Craigslist don't have any poetry experience. And so what I'm going to do instead is I'm going to hire a bunch of people with PhDs in English literature. But what they don't realize is this is actually also terrible. Like a lot of PhDs, again, going back to what I was saying about even people with MIT CS degrees, a lot of PhDs, they're not good writers or poets. Like you think of people like Hemingway or Emily Dickinson, they definitely didn't have a PhD. I don't think they even completed college. And so I think we think about quality completely differently.

• 14:01Speaker 01

  Like what we want isn't poetry that checks some boxes and uses complicated language. Again, we want a type of poetry that Nobel Prize laureates would write. And so what I think you need to do is like you need a mindset shift. Like one of the things we all want to think about is like there are certain people who are trained up in this very objective, very, like this domain of computer vision in a sense, that lacks all of these nuances and lacks all this inherent subjectivity. And so what we want to do instead is we want to recognize that poetry is actually really subjective and rich. Like maybe one poem is a haiku about moonlight on water and another poem is something with internal rhyme and meter and another one is somehow focusing on emotions behind the moon rising at night. And so you actually want to capture that there's a thousand ways to write a poem about the moon.

• 14:47Speaker 01

  In a way that you, there aren't a thousand ways to draw a bounding box on a car. Like, there are a thousand ways to label something as a cat or a dog. Like, there isn't a single correct way to, to write this poem. And each different way that you write it, or each different preference that you have for different types of poetry, that should give you different insights into language and the mind and imagery and human expression. And I talk about poetry a lot, but it's not just poetry. Like, if you think about math as well, there are a thousand ways to prove the Pythagorean theorem. And again, each way you prove Pythagorean theorem, it's kind of based off of different insights into the mathematical reality of the universe. And so, like, one thing that happens is that when you think about quality the wrong way,

• 15:31Speaker 01

  you get kind of commodity data that optimizes for things like inter-rater agreement. So, again, going back to the computer vision world, like, sure, if all you're doing is labeling images of cats and dogs, you want high inter-rater agreement. Like, you want people to agree that this is a cat, and you want people to agree that this is a dog. Like, one of the very, very easy ways that you go quite control world in this old school world is just by, by seeing whether there's agreement with a majority. But in this new gen AI world, again, there's no way that you can ask people to write a thousand poems and then take a majority vote and get the best poem. Like, what actually happens when you optimize for a high inter-rater agreement is, like, you get the lowest common denominator of things that people want,

• 16:15Speaker 01

  which actually turns out to be kind of trashy and low quality and unengaging a lot of time. And so, you just need to think about quality in a different way in order for your data to kind of really embrace human intelligence and creativity.

• 16:30Lukas Biewald?

  I mean, one of the things that, that we found, you know, I was like working in this space, is that maybe the biggest issue was actually eliciting

• 16:40Lukas Biewald?

  You know, from a customer, what they really wanted like I think, you know, and and obviously you're doing simple examples, but you know, you think about putting a bounding box, you know, around a car, it seems so simple, you know, to like a ML researcher, especially one who hasn't looked at a lot of, you know, specific data examples, it is really hard. I mean, I think you're one of the rare ML researchers that really wanted to look at data. You know, I was too, and that's why I got into the space, right? But like you think like, okay, putting a bounding box around the car is really simple. But okay, like, what if the car is occluded, right? Then, you know, do you put it around where you think the car is? Like, what if the car is in like, you know, a reflection of a car in a mirror? Do you still want the bounding box, you know, around that car? What if it's a billboard with a picture of a car and it's not a real car? You know, do you want this?

• 17:22Lukas Biewald?

  So like, even even in the simplest cases, as soon as you start to look at real world data, you know, it gets way more complicated. I remember working with customers, there would actually be like a long process of eliciting, you know, from the customer what they wanted. Some customers would write these giant, you know, documents trying to like enumerate every case and exactly what they want. But I think those two kind of would end up, you know, it'd be hard to kind of reason about exactly, you know, what you want. So if you if you followed that, um, you know, there's instructions to the letter, you'd end up in some ridiculous cases, which isn't actually really what, you know, the person wanted in, you know, the first place. I'm kind of curious, like, um, how you how you think about that problem?

• 18:02Speaker 01

  What we always try to do inside is try to understand the goal or the principle that the researcher has. And so when we understand the goal or principle or how they're going to use the data, then we can almost, you know, put ourselves in their shoes and instead of asking them, okay, what happens when this car is occluded? What happens when this car is in a reflection? And we just think about, okay, uh, kind of like deriving from first principles, given given what they told us. What do we think that the user, what do you think that the researcher wanted? Like, so I mean, this is why I don't know if I can envision, but okay, maybe if you know that the data is going to be used for some LiDAR model, then you would just know that this makes sense and this doesn't. And so we can have that higher level goal instead of like a mechanical list of instructions

• 18:46Speaker 01

  that it does kind of always what we what we aim for. And it is hard because, you know, at the end of the day, sometimes researchers don't know either. But this is where I think we as a company, we often also try to have a strong opinion on what their best type of data is as opposed to, um, As opposed to almost like blindly accepting whatever people tell us. And oftentimes, we actually do like to get into like a feisty debate about

• 19:14Speaker 01

  about what type of data would be most useful to bird models.

• 19:17Lukas Biewald?

  So then, do you hire a lot of ML researchers? Like, is that, are those the people that you want working directly with the customer in order to do?

• 19:23Speaker 01

  Yeah, yeah, yeah, we have a lot of ML researchers. Like, I think one of the things that we often think about is, we almost consider ourselves a kind of like a research company. Just that instead of researching like IRLMs and in a way that, you know, other French labs might, we're more about researching the data.

• 19:40Lukas Biewald?

  Totally. And has it been, I mean, I, I really admire, you know, how you've kind of avoided, you know, status games and float under the radar. But I would think that in doing that, it might make it harder to, to hire. Like, do you have a different strategy for hiring? Or do you think that maybe it doesn't matter, you find the people that really connect with your mission?

• 20:01Speaker 01

  Yeah, so I think one of the things that we often think about is, we don't want people who are just joining us in order to, in order to like notch another brand or their resume. Like, sure, there are other people out there who do that. And we're missing out on those candidates. And I think they also tend to want to build, like, they tend to be people who want to build large teams. They tend to want to be people who want to build empires. And

• 20:27Speaker 01

  it almost turns into a hiring for the sake of hiring? When you, okay, so like, yeah, like, why do you want to, why do you, why do you want to hire this person? Well, I needed to hire this person because the other person doing this job is spending all their time engineering. And so why, why is that person spending all their time engineering? Well, it's because somebody told them that they needed to hire a internal tooling team. And so why, why, why did that person need to hire an internal tooling team? Well, it's in order to make the engineers more productive. Well, why are the engineers not productive enough? Because they're spending all their time in meetings. Why do they spend all their time in meetings? It's because we hired all these people and they communicate with them. And so I think there's like a lot of benefits if you start from having people who really, really believe in your mission.

• 21:10Speaker 01

  Then, at least for us, we've been able to stay smaller with like a, with a much smaller team.

• 21:15Lukas Biewald?

  How big is your team?

• 21:17Speaker 01

  So we're a little over 100 people.

• 21:19Lukas Biewald?

  Wow, that's an incredible revenue per employee. Congratulations.

• 21:26Lukas Biewald?

  Have you, I'm curious, like, you know, you've gone through this really fast change, I think, from being like an ML researcher essentially to running a very significant, large company. Where have you felt stretched the most? Like, what's been challenging?

• 21:49Speaker 01

  Certainly, the thing that I found most challenging is sales. I mean, probably any researcher would do that, just the concept of having to go out and hawk your product is a little foreign to me.

• 22:03Speaker 01

  So I think it is, I think we're lucky in a sense that our product actually, at the end of the day, is it is for researchers. And so kind of having that research mindset, like, at the end of the day, like, we, like, what we're trying to build, which is not, it's almost like that Disney quote, you know, we don't, we don't make movies to make money, we make money in order to make movies. So in similar way, what we're doing, it's, we're not trying to generate revenue. We're trying to generate data that will help AGI. Like, one thing that we'll often do is sometimes when customers or like when you, when new companies come to us and they ask us if they can work with us. And if their goal is kind of just like unaligned with AGI, then we'll have

• 22:47Speaker 01

  to just say no to them because we don't want the revenue and we want to focus on the AGI companies in a sense. I think, like, again, going back to not raising, the fact that we don't have a board or like that we don't have an external board, the fact that we don't have VCs who are just dying to, you know, make as much money as possible. I think that gives us sort of freedom that allows us to focus on the most important problems, which has allowed us to kind of maintain that research focus.

• 23:15Lukas Biewald?

  Totally. So you only work with companies focused on building AGI?

• 23:19Speaker 01

  Yeah. So like, for example, if a company came to us and they were saying, like, uh, like, yeah, we just want to train a, what could that was, let's see, I just want to train a, like, I'm a newspaper and I just want to train a, I don't know, like a, like a category algorithm, category classifier. Yeah, we just say no.

• 23:40Lukas Biewald?

  What if I like, what if I want to make like a video, like auto video generator? Would that be in your wills or is that too?

• 23:50Speaker 01

  Uh, yeah, I mean, we do it in that sense just because like building such video generators is like a like a part of building AGI.

• 24:01Speaker 01

  So yeah, we're doing that sense.

• 24:02Lukas Biewald?

  Look, I mean, I imagine, you know, five years ago, the data being collected is pretty different than, you know, the data collected now, like I would think probably some of the tasks, well, tell me if this is true, but I would think some of the tasks that you'd be doing five years ago would be like easily automated by LLMs today. It's just been such an astonishing, you know, pace of improvement. Can you talk about how the, the types of tasks have changed over the last few years?

• 24:28Speaker 01

  Uh, yeah, so, so many of the types of work that we do are very, very different. So when we first started, a lot of our work was in tasks like search evaluation or content moderation. And like, you know, where as they, it's almost purely LLM work. So that's one big difference. And then even within, even within LLMs, there has been this obvious trend towards higher complexity, higher sophistication, higher expertise. So I can give a couple of examples. So like, for example, there's been a big increase in multimodal complexity. So a few years ago, it was all text data, just conversational text assistance. But now we do a lot of work with images and audio and video.

• 25:09Speaker 01

  And I think the interesting thing is like you actually want the models to understand all of these modalities simultaneously. Like one of the things you might want to do is like, okay, I'm taking a video of something on my phone. And now I'm asking the model to create a program based off of the video on my phone that simulates this in real life. So, um, yeah, there's been a lot of multimodal, um, increase in complexity. So there's also been a big expansion in languages, like, you know, first people were naturally focused on English only work. But we actually work in over 50 languages now. And I think what's also interesting, it's like very hyper specialized. We, we support coding in Argentinian Spanish and we support legal and financial expertise in Bolivia. And it's because each, like even today, I think a lot of the models, they're actually just not that good.

• 25:52Speaker 01

  Surprisingly so, like they're not that good at the different nuances of different languages or different dialects or different cultures. And so I think there actually still was a lot more progress to be made there. And then probably the biggest shift is just like the depth of expertise that a lot of the work requires. Like, yeah, you see the models winning IMO gold medals now, and you see them doing all these incredibly advanced tasks. And so you actually really, really need serious thinking power behind us. And so even today, some of the tasks that we do, they involve spending days or even weeks solving these really, really interesting problems. And so it's a very, very far cry from, you know, tasks five, 10 years ago where you might spend five seconds labeling a task.

• 26:29Lukas Biewald?

  So do you actually have people that like can solve like Olympiad level math problems, like creating those problems and then solving them?

• 26:36Speaker 01

  Yeah, yeah.

• 26:38Lukas Biewald?

  It's funny, this is kind of an aside, but I, I've been sort of surprised how at the, at the scores, the latest models are getting on these Olympiad problems. When, when I put in kind of more fun, you know, brain teaser problems that my friends pass around, they often can't do them. Do you have a sense for why there's that disconnect?

• 26:59Speaker 01

  Yeah. So I think a big problem with a lot of the frontier models today is that they've basically been benchmark hacked. So you have all these benchmarks out there. And they only have them are, a lot of them are just aren't very good. They're like overly academic or they're overly synthetic. And like, yeah, like a lot of these benchmarks, they have a single, going back to my point earlier, they have a single objective answer. And so the models have been narrowly constrained to be good at these very narrow objective problems. When, yeah, like a, you know, a math problem in the real world, or like the problem that you would ask as a frontier mathematician,

• 27:40Speaker 01

  they're not going to be closed-ended problems. They're going to be like open-ended explorations. So yeah, I, I think a lot of it stems from, from this kind of benchmark hacking that's, that's been going on.

• 27:50Lukas Biewald?

  So if you're going to build a benchmark today to compare frontier models, what, what kinds of things would it include?

• 27:57Speaker 01

  So what we always say is that the gold standard for evaluating models really is human evaluations, where you just can't fully automate it. Like, people try to build these leaderboards where they take kind of automatic verifiers. But automatic verifiers, they still kind of only work well in these very objective domains. Or they've tried building benchmarks like LMSys, which I think is an absolutely terrible, absolutely terrible leaderboard that has basically, I think, set the industry back by at least a year. So I can go into that in more detail. But I think a lot of the, a lot of the kind of benchmarks out there, they are flawed,

• 28:42Speaker 01

  either because you're built using low-quality data, or they are just overly synthetic, overly academic, overly objective in a way that the real world is not.

• 28:54Lukas Biewald?

  Okay, well, I mean, now I want to hear this. So why do you think the LMSys, you're saying that benchmark is not just bad, but it's set the industry back? Can you tell me more about that?

• 29:06Speaker 01

  Yeah. So, so basically, LMSys is this, if you don't know what it is, it's like this popular leaderboard of LLM models. And what happens is that the way it works is people go on to it, like literally anybody around the world. Like, you could be a, I mean, what we often hear is that it's already high schoolers, middle schoolers who can't access LMS any other way, and they're just going on to LLM Arena as their only mechanism. And so what happens is they go on to this chatbot arena, they'll enter a prompt, and then they see two model responses, and then they vote on which one's better. But think about it, they're not taking the time to read or evaluate these model responses at all. They're literally just like looking at responses for like two seconds and just picking whichever one just matches their fancy. And so the models could have made everything up. They could have completely hallucinated everything. They could have not followed your instructions at all.

• 29:50Speaker 01

  And they'll just vote on it because, okay, in order to progress, they need to vote on something. And oh, yeah, this model has emojis and has a lot of bold formatting, so it just looks really impressive. Like, one of the things that we've learned is that the easiest way to improve in this arena is simply to make your model responses a lot longer and to like double the number of emojis that you have. And if you actually think about some of the models that have been released in the past year, like think about Llama. Like, yeah, they, or at least the version of Llama 4 that was optimized for LMSys. Like if you actually looked at it, it exactly matched all of these patterns. And so one of the like, one of the phenomena that we often hear from researchers is they will tell us,

• 30:31Speaker 01

  they'll tell us, I'm only going to get promoted if like my BPS told me, I'm only going to get promoted if we advance or model by 10 points on the leaderboard. And they will look at the leaderboard data themselves and they'll see that, yeah, a lot of responses that are preferred by the, the LNS, uh, like in LNS data sets, like literally the, the model that has followed instructions worse, literally the model that has hallucinated everything. These are the responses that are getting preferred just because they have a lot of formatting and emojis. And so what like these researchers tell us is like, yeah, like I want to work on, um, improving fundamental capabilities of my model. Like I want to, I want to work on improving them in coding. I want to work on them, uh, affecting their hallucination abilities.

• 31:14Speaker 01

  But instead, if the only way I want to get promoted is by optimizing for this leaderboard, and the easiest way to optimize for this leaderboard is to make my model hallucinate, so that it just generates these crazy answers that are like, even though they're completely wrong, they're just like very, very compelling to, um, kind of like amateur, uh, amateur raiders who are only spending two seconds. Then yeah, that's, that's what they're going to do. And so we've often actually seen, uh, models, they've, um, like, uh, like if you look at actually some of the models that are, that are top spots on leaderboard today and compare them to how they were performing six months ago, you actually see that they're, they're, they're actually worse in, in many ways.

• 31:52Lukas Biewald?

  Do you think, so like another thing that kind of strikes me is like, you know, we have these different models made by totally different organizations, like, you know, you have like X, you know, like, you know, run by Elon and, and making Grok and, you know, you look at Anthropic and OpenAI, these clearly have different cultures and we've had different folks from these organizations on, on the podcast and, and yet. The models seem to have this kind of surprisingly consistent tone to me. They sort of like, you know, they feel like a little bit like annoyingly positive, like a little bit like, you know, boy scouty, like, yes, like, thank you, I will answer your question. And I, I sort of imagine that that must somehow be trained in at some point in the process. And I wonder if it's in like the, it's just the pre-processing, like, you know, is that just sort of like what

• 32:36Lukas Biewald?

  high-quality data on the web, you know, looks like? Or is it sort of intentionally inserted later in the process? Is somehow the annotation that you're doing, like contributing to that sort of consistent?

• 32:48Lukas Biewald?

  I feel like we'll look back on it as almost like this like LLM tone of, of 2025.

• 32:53Speaker 01

  Yep. I mean, I think it's kind of a combination of everything in that, like, for example, we actually don't teach the models explicitly to, or we generally don't teach the models to follow proper grammar. Like, we generally don't teach them all to use M dashes. But it's kind of just there in maybe in pre-training, where, where these kind of behaviors are baked in. I think in the future, the models will become more and more differentiated, where you will be able to tell that they have a noticeable difference in style. Like, actually, even today, I, I personally, maybe because I've looked at this data enough, I can actually just tell from reading the model responses themselves, which model is which. Like, some of them will just have certain preferences.

• 33:38Speaker 01

  Some of them will just use certain, certain words or certain stylistic patterns that I, I can actually generally tell.

• 33:46Lukas Biewald?

  Interesting. What's, can you, can you think of any tells? Is it more subconscious or could you cite some specific things that lets you know it's a particular model?

• 33:55Speaker 01

  Some of them would just use certain phrases like absolutely more often than others. Some of them would use emojis in a way that others wouldn't.

• 34:03Lukas Biewald?

  Lama, Lamy uses so many emojis. I totally agree with that.

• 34:07Speaker 01

  Some of them would like do a thing where they just like repeat three adjectives in a row. You know, some of them use a lot more markdown than others.

• 34:15Lukas Biewald?

  When you look at the way that these different organizations collect data, are there like striking differences? Or does it feel like everyone's sort of chasing the same types of data?

• 34:26Speaker 01

  I think there actually are very striking differences between the different companies. Like, every company almost has their own philosophy on the right way to both train and evaluate their models. So, yeah, I think it's actually been surprising how different they are.

• 34:45Lukas Biewald?

  Do you have like a point of view on like how you think it should be done? Like, if you're, like, if you're running an AJI company, do you, I guess, what would you do? Is there something you'd do differently than what you are asked to do?

• 34:59Speaker 01

  We had this very strong view that ROHF data was a lot more effective than SFT data.

• 35:09Lukas Biewald?

  Maybe you should just explain the difference here for the general.

• 35:13Speaker 01

  So the way SFT data works, just to give an example, would be.

• 35:16Lukas Biewald?

  And this is fine-tuned, fine-tuned data, supervised fine-tuning.

• 35:19Speaker 01

  Yes, so SFT stands for supervised fine-tuning. And so the way it works is, okay, so let's suppose that I wanted to train my model to become better at poetry. The way it would work would be, okay, I would, like, you as an hitter, you as a raider, would write a prompt. So the prompt might be, yeah, write a Shakespearean sonnet about cheese. And then you would literally write a Shakespearean sonnet about cheese, and that would be a demonstration to the model. So that's essentially supervised fine-tuning. In contrast, in ROHF, so ROHF stands for reinforcement learning with community feedback, the way it works is you write a prompt. So then you might take the same prompt, write a Shakespearean sonnet about cheese. And then you ask the model to generate two responses. So these might be two different models, like two different model checkpoints. But model A would spit on the answer, and model B would spit on the answer.

• 36:02Speaker 01

  And then what you would do is you would just rate, in some sense, which one is better. Like, okay, model A is better because, I don't know, it had emotion behind it, because it was a better written, because it actually followed the format of a Shakespearean sonnet. And so then you basically teach the model, like, okay, A is better than B. And so the benefit of this is, it's that you're, there are many benefits, but just to enumerate some of them. Like, one, it's a lot more efficient. Oh, yeah, like, even if you are a Nobel Prize laureate, it takes you a lot of time to write a poem about cheese. Second, it helps teach the model all these latent preferences, and it helps teach the model what's good, but also what's bad. And.

• 36:48Speaker 01

  Yeah, so basically, we had this very, very strong view that I mean, SFT is important sometimes. Like, one of the things you often tell our customers is that you often want to use a little bit of SFT data in order to more like efficiently bootstrap your models into a phase where RHF is useful. But we just had this view that RHF was a lot more effective. And so we actually kind of steered our customers into that direction. And if you actually read the Llama 2 paper, you'll see that one of the discoveries that they describe is like, yeah, like actually at the beginning, all of their researchers thought that SFT data would be more effective as well. But then they ran a bunch of experiments, they actually find that RHF was just so much more effective.

• 37:27Speaker 01

  Yeah, so that was one of our beliefs in the past.

• 37:30Lukas Biewald?

  You know, when you look into the future, like you roll things forward a few years,

• 37:37Lukas Biewald?

  what kinds of data do you think you're collecting now that you won't be collecting? And what new kinds of data do you think you'll be collecting?

• 37:47Speaker 01

  Yeah, so I actually don't think that we're going to stop collecting any of the data that we're still collecting right now. Like, one of the things that we've often found is that, I mean, the portions may change. Like, just in the same way that, sure, as you as you become a more sophisticated adult, you don't need to be trained as much on arithmetic that you, you know, you did when you were a little kid. But what we often found is that if you don't persist with a sum amount of this data, models often just regress in very surprising ways. Like, I won't name names here, but some of the frontier models, they just make the most bizarre mistakes these days. And I think it's just kind of a little bit wild.

• 38:29Speaker 01

  And that's just because the data measures have been tuned in very certain ways. But I'd say some of the things that we expect more and more,

• 38:39Speaker 01

  or some of the ways we expect the data trends to continue are

• 38:45Speaker 01

  in ways that they haven't before. It's, so I'll just list them out. So one is

• 38:52Speaker 01

  Even higher expertise problems than we have today. So like right now, there's a lot of work being done in kind of like I would call like graduate level STEM work. But if you actually want the models to make new scientific discoveries, you're going to need to progress beyond, you know, like what even like the average PhD can do. So I think there'll be kind of like super, I think, super STEM work that will, that will happen in the future. So that's one of them. Another will be.

• 39:19Lukas Biewald?

  Oh, wait, can I ask, so like super STEM work? Would that be like literally like, I'm gonna go out and discover some new like principle of chemistry just to train a model? Like, how would that even manifest?

• 39:33Speaker 01

  Oh, yeah, I mean, I think it actually is exactly that. Like, imagine that you are a Stanford professor and you're literally working on, I don't know, like your, your latest frontier research. What you want to be doing is, yeah, you want to be collaborating with an AI to kind of generate new app policies to help you test your experiments. And so like, it's this idea of a scientific collaborator. And it's in the same way that we have like a, you know, like a coding collaborator right now in the form of Claude Code and whatnot. So I think, um, like there, there will be things like that that are going to be increasing.

• 40:04Lukas Biewald?

  But how would you collect the data? Would you like get a Stanford professor and have it like just score responses? I guess, I guess the scoring response would be easier than just generating responses.

• 40:14Speaker 01

  Oh, yeah, I mean, it wouldn't necessarily be, I mean, so there's a lot of varieties, different types of data that we collect nowadays. So it's not just scoring. But yeah, in a sense of, yeah, it's like literally a Stanford professor or some other person who's capable for understanding frontier physics, who's collaborating with the models and then you see teaching them the, the right signals. Yeah, so certainly that's one domain. Another domain is just like longer and longer horizon tasks. Like today, the models can like, sure, they might be able to objectively perform tasks that a human could have done in, you know, five, 10, 15, 30 minutes, an hour. Some of their time horizons, I think, will just get very, very long. And so, yeah, well, what are types of things that you want the models to do that would take a couple of days or even weeks? So that's another domain.

• 40:58Speaker 01

  A third domain, almost going back to your point about the language that these models exhibit today. It's like, how do we actually make the models really, really good at kind of like long form creative writing? Like even today, it's kind of, kind of interesting. Like even today, I would say that I don't like any of the poetry or I don't like any of the short stories that the models can write. Even though they have like perfect prose, they're just not creative enough. Like they've almost collapsed to just like very, very vanilla generic dimension. And so I think there will be more work there. It's just that, yeah, like up and down, there can maybe be higher, higher economic value tasks that people want the models to do instead. I think it actually is an important thing for, for a multiple to do and to like exhibit this kind of creativity in a kind of like a non generic way.

• 41:43Speaker 01

  And then I guess a fourth one would just be more and more actions in a real world. Like agents have kind of percolated or I mean, probably wrong word, but agents have obviously exploded in, in popularity over the last few months. But just still at a very, very nascent level. And so just the concept of AI models that can plan and reflect on their actions and try new ideas and so on. And actually, like, again, like actually affect things in the world. I think that will be increasingly, increasingly important concept.

• 42:18Lukas Biewald?

  I guess one of the things that we might not have expected models to do a priori is put a lot of their reasoning into code. Like, I think, you know, we've kind of seen a trend where, you know, models use code to do the reasoning more and more. You can kind of see it like in the, you know, in the, in the chat window. Do you think that's the kind of an artifact of limitations of, of models today? Or do you kind of expect that, that trend to continue?

• 42:46Speaker 01

  I actually think it's right to continue. Like I do think that the code that they're writing is like very important for certain times, certain types of. Verifiability that you just might want the model to exhibit. So yeah, just in a way that if I could write a lot of code for my day to day life, I would. So yeah, I actually think it's very, very, very important capability for our mall stuff.

• 43:10Lukas Biewald?

  Do you have a sense, I guess I have this intuition that the amount of code used to train these models is kind of increasing as a fraction of the overall data. Do you think that's accurate? Like, are you generating more and more code for your customers than other stuff?

• 43:26Speaker 01

  Uh, yeah, yeah, we are.

• 43:31Lukas Biewald?

  I guess that's a lead into another question I have, which is, you know, do you think that the trend towards, is the trend towards reasoning models changing the types of data that you're collecting?

• 43:42Speaker 01

  Yeah, so definitely the reasoning models have led to a bunch of new types of collection methods. And so I think probably the biggest one is this concept of creating RL environments from scratch. Where essentially a lot of what we're doing is we're basically building these video game like universes with interesting tools and data sets that the models need to solve. So like, just to give you an example, like you can imagine a universe consisting of a simulated AI startup. And so in this environment or in this universe, you basically have a bunch of like Gmail messages and Slack messages and Jira tickets and GitHub PRs and code bases and so on. And so then what you need to do is like, you need to, like, you basically want this universe to be as rich and complex and like,

• 44:27Speaker 01

  high fidelity to challenging task in the real world as possible. So for example, you actually imagine in this universe, suddenly AWS goes down and suddenly Slack goes down. And so what do you, what do you do? What does the agent do? Like, I can't use Slack, so I need to make sure that knows how to do that. And he's like, figure out how to solve this problem. And yeah, I think there's like a lot of things that we're basically building into these, into these R environments.

• 44:51Lukas Biewald?

  How do you handle like the subjectivity of these models? Like I think one of the reasons that I spend a lot of time, I have a five year old. And, you know, she really likes to generate stories. So we spend a lot of time, I've gone like really deep trying to get these models to write, you know, interesting stories. And I like agree with you, I don't think they really write interesting stories. And I think a big part of it is, I think they're trying to maybe like average the preference of a lot of different people. But, you know, my five year old's taste is, is really different than, you know, my taste in stories. And, you know, we're actually like asking the model together. So maybe it's not even possible for it to make, you know, both of us kind of happy with the story, let alone, you know, the average person. So it seems like that might require like a different type of like labeling, if you

• 45:35Lukas Biewald?

  actually want to get like interesting art out of it. Because like good art probably shouldn't be, um, you know, like get thumbs up from like every single person that, that looks at it.

• 45:44Speaker 01

  Yep. Yeah, I mean, I think there's this concept of personalization that is actually very, very much unexplored with the models still. And there's like kind of like service level personalization where, yeah, the model knows that you live in Ohio versus you live in New York City. But there's this concept of like, there's just like a lot of unexplained latent preferences that you have that are kind of hard to articulate.

• 46:06Speaker 01

  And how do you get the model to somehow learn those from the data and for every single person and then apply those to, uh, to the response that they generate? I think that actually is like a really, really interesting concept that still hasn't been fully explored yet. Like, like just to give you an example where you are a mathematician, like here are the restaurants that you should try and pair. It's like sometimes the models are just like make these like weird little personalization generalizations. And so there, I think, I think that's all I worked at, that there remains to be done with personalization.

• 46:38Lukas Biewald?

  And is that, do you think that's because it's hard to collect good personalization data? Like, I'd imagine like to put yourself in the shoes of someone that you're not, the data is never going to be as accurate.

• 46:49Speaker 01

  I think that's part of it, but I mean, I think a big part of it is that all the frontier labs, only so many things that they can focus on. And so they just haven't quite doubled down on personalization yet.

• 46:59Lukas Biewald?

  Okay, so let's, I mean, let's talk about synthetic data. I think everybody's dream is always to be able to, you know, automatically, you know, generate useful data. We've seen, you know, a lot of people using LM as a judge for many applications, which is a kind of synthetic data. Like, do you, do you synthetic data generation? And how do you think about that?

• 47:20Speaker 01

  Yeah, so I actually do think synthetic data is really useful in some places. But I think a lot of people actually overestimate what synthetic data can do. So I can, I can give a couple of examples. So like right now, there are a bunch of models that have been trained really heavily on synthetic data. But similar to what you and I were saying before, it's like, that's partly why they're very good at these very academic homework style, benchmark style problems. And they're actually really, really terrible at real world use cases. So synthetic, but synthetic data has made models good at synthetic problems, but, but not real ones, like that, that kind of phenomenon. Yeah, like actually, I can think of an example now. So I remember, um, like maybe a year or so ago,

• 48:01Speaker 01

  we were running human evals for, um, one of the researchers that we work with. And our human evals showed that their models suddenly tanked. And when we dug into it, like when we talked to researcher, it turned out that they had just trained their model on like, 10 million or 20 million synthetic math problems. And so what they hadn't realized was that by training the model on these 10 to 20 million synthetic math problems, they, and they're like, synthetic math problems, like a very, very narrow domain of math. And what they hadn't realized was that how much that was making the model worse at. Basically, every other type of task, like, you know,

• 48:45Speaker 01

  there are only so many, like, you know, very closed-ended SAT-style math problems that you want the model, like, there's only so much you care about that. And so the model actually just became worse in basically every other domain. So, yeah, that's an example there. And then, I think, like, another thing we're actually often hear from companies is that they, they often tell us, yeah, I spent the past year training my models on synthetic data, but they've only now just realized all the problems that that's caused. And so they actually then spend, like, multiple months throwing a lot of it out. Like, a lot of them will actually tell us, they've thrown out 10, 20 million pieces of synthetic data because they found that even just a thousand pieces of really high-quality human data is, uh, is more useful.

• 49:25Speaker 01

  Like, the human data is both, uh, I think, more diverse, more creative, as opposed to, you know, kind of just getting 10 million pieces of the same thing over and over again. And so, actually, what ends up happening in practice is that, yeah, companies will try using synthetic data for certain problems for, like, six months, and then a lot of the work that we do is, ends up being cleaning up a synthetic data.

• 49:48Lukas Biewald?

  You know, you kind of have front-row seats to, you know, like, what most of the labs are doing. I mean, do you have a, um, do you have a point of view on if we're kind of at a moment where, performance is sort of, you know, stuck with, like, it's kind of needing, you know, new methods to get to the next level? Or if, um, you know, the, the current strategies of, you know, kind of, you know, pre-training and then, you know, like, reinforcement learning to learn reasoning will kind of get us to, um, a much better set of models?

• 50:21Speaker 01

  Yeah, so I definitely don't think we're stuck at all. Like, I think there are a lot of, um, new methods that are just appearing, and a lot of new data that hasn't been collected yet. And from some of these, I mean, even just like some of the early experiments that we run ourselves, I think that we are going to actually see some massive progress in some new domains very, very soon.

• 50:43Lukas Biewald?

  Can you describe what some of the new methods are at a high level?

• 50:46Speaker 01

  Yep, so, I mean, even just some of the methods that I was just describing in terms of RL environments and some of these new RL methods, I think people, uh, they're still kind of new to a lot of researchers in industry, and, um...

• 51:01Speaker 01

  It's basically this concept of exposing the models to all these new environments they haven't seen before. And

• 51:07Lukas Biewald?

  And what would be the example of a new environment that they haven't seen before?

• 51:11Speaker 01

  So

• 51:14Speaker 01

  So even even just the example I mentioned earlier, where you have like a simulated AI startup. And then suddenly, the the model, like the the environment, it loses access to Slack and it loses access to AWS. And so how does the agent continue operating in this environment and how does it like solve the problem? And it's something that just doesn't appear in some sense in any pre-existing data or any data that we've generated before. So it's just like this entirely new concept where the model needs to go through all of these different actions and needs to kind of reflect on what it's done and needs to just kind of like find new ways of solving a problem. And especially when you expose it to very, very messy type of data that it may try to retrieve, a very messy types of data that again, like it may not have

• 52:02Speaker 01

  encountered before.

• 52:07Speaker 01

  It's

• 52:10Speaker 01

  like just the models can can kind of fail and use like very, very unique ways. But I think in ways that it can be taught to to like progress through.

• 52:18Lukas Biewald?

  What do you think about the ARC benchmark? I mean, that's kind of an interesting one where it's it's it's like deceptively simple problems that I think you or I would have no no trouble with, certainly easier than, you know, Math Olympiad problems. There's sort of more visual pattern recognition. Like, do you think those are likely to get solved in in the next year or two?

• 52:38Speaker 01

  Yeah, so I will admit that even the ARC benchmark is very surprising to me as well. Like I, if you asked me without ever having seen ARC benchmark, if models could solve this, like I'll be like, yeah, absolutely. And so it actually is very, very certainly, there's something about them that I still don't quite understand in why models can't solve them. Yeah, it's actually very, very interesting because like right now, we are actually doing a lot of work to generate ARC style problems. So it'll be interesting to see how much that helps the models improve. But yeah, like I was saying, right now, I actually I don't have a good understanding of why models can't can't solve them.

• 53:17Lukas Biewald?

  Do you have an opinion on whether open source or closed source models are going to win in the long term?

• 53:25Speaker 01

  So my guess is that, at least with the current state of things, closed source models will continue winning. And

• 53:35Speaker 01

  part of that is because, I mean, LMS are just so valuable that

• 53:41Speaker 01

  if you ever try to build open source models, the way incentives currently work is that eventually you're going to be forced to close source them. And so if we want to build truly open source models that are really, really good, we almost need a different kind of incentive structure to make sure that that happens or remains in place. Otherwise, if you just look at the history of like other types of open source models, they have kind of gotten more closed over time.

• 54:08Lukas Biewald?

  What do you mean? Like, what are you referring to?

• 54:10Speaker 01

  So, I mean, if you even think about like Meta that is thinking about making their, uh, making VOMA closed stores. And just models are just so expensive to train. And people want to fully capture the value if you ever build a truly, truly good open source model that I think won't remain open, open source for very long. Again, unless you can change incentive structure in some way that I haven't figured out yet.

• 54:35Lukas Biewald?

  What do you think is like the ratio of spend on data to spend on compute in the, in the training of like a large model? And how do you expect that to change over time?

• 54:46Speaker 01

  So, I mean, I definitely think it should be a lot higher. Like, sometimes people will, I would say, they almost skimp on gathering data when...

• 54:55Lukas Biewald?

  What do you, what do you think it is today, like roughly?

• 54:59Speaker 01

  So I actually don't have a good sense myself. I think it actually varies highly depending on the labs, but it can be anywhere from like a couple percentage points of 1% to like 10%.

• 55:10Speaker 01

  But I think like one thing that we've often heard from some researchers is that there are some researchers who are really, really good at using human data. They do know how to come to us to gather it. They know how to like apply it in their own work. And what they often tell us is that like some of their counterparts, like they don't know how to use human data, and so they just move a lot slower. Um, like they may try these like, uh, kind of like somewhat wild ways to get around the fact that they don't have any human data. And it just ends up being this like kind of complex slowdown for them. So I, I hope it'll get higher in the future. But, um, yeah, I, I, I, I think there are a lot of ways in which people still underestimate these human data.

• 55:52Lukas Biewald?

  Awesome. I appreciate your time and congrats on building a fantastic business.

• 55:56Speaker 01

  Yeah, it was a great journey.

• 55:58Lukas Biewald?

  Great job. Thanks.

• 55:58Speaker 01

  Thanks much. Bye.

• 56:01Lukas Biewald?

  Thanks so much for listening to this episode of Grading Descent. Please stay tuned for future episodes.