Mercor hits $500M ARR run rate in 17 months: Brendan Foody on AI talent and the future of expert data

next guest, Brendan from Meror coming in the studio.

Brendan, welcome to the stream. How you doing?

I'm doing great. How are you?

I'm great. Sorry we couldn't uh link up on Tuesday. You have some massive news. Uh uh, introduce yourself. break it down for us. What's the latest and greatest?

Yeah. I'm glad you finally, you know, it's good you get to the $500 million revenue mark and you decide, all right, I'll do some podcast.

Yes,

exactly. Yeah. Um, so I'm Brendan, the CEO and co-founder of Meror. Started the company with my best friends from high school when we were 19. Um, scaled up a little bit to a million dollar revenue run rate. Dropped out of college. uh started working with all of the AI labs and scaled from 1 to 500 million in revenue run rate in 17 months.

Congratulations.

Which has been pretty wild. Uh yeah. No. And super super exciting and surreal as you can imagine. Um so uh very excited about all of the um progress with the business and the best is yet to come.

What was the first task you did? Like what was the first data you labeled? What was the first project to get from go to how do you go from zero to one basically?

Yeah, name every piece of data.

Well, so I remember the very first oh let's see there there were a few but the first meeting that really jumped out uh was when we were hiring Olympiad medalists uh because everyone was interested in how the models could become superhuman at Olympiad math. And so we turned around 25 Olympiad medalists in 24 hours. A and I I bring up this example cuz I wasn't completing the task successfully. I was uh obviously haven't don't have an Olympiad gold medal in math. But it was uh it was incredible to just see like how capable the models are. uh a and this indication of the huge trend underway in the market away from low and medium skilled talent towards this super high-skilled sourcing and vetting paradigm ranging from Olympiad math all the way to the fang software engineers and top investment bankers and consultants that helped to push the frontier of models.

What was the what was the onboarding process for those 25 medalists? Was this just like cold outreach or something like how did you actually meet these folks? largely through referrals because we have a big pool of people that we've already hired on the platform. And so, uh, the largest sourcing channel by far is that people we've previously worked with, we'll send the link to their friends and we'll pay them 250 bucks as for a successful referral to help grow the talent network.

It feels like most of the labs are I mean, we saw this with uh, OpenAI. Mark Chen was just posting that they basically dominated at every hard programming and math competition this year. Um what are the labs interested in next? Yeah, I I think the largest transition is and we'll share more about this in one of our product releases soon, but it's away from academic evals like Olympiad math or GBQA for PhD level reasoning uh and moving towards all of these professional domains of how do we measure what it means to be a great software engineer to build products? How do we measure what it means to be an investment banker that can do thoughtful financial analysis or a consultant that can help to uh you know segment a market and these endto-end evals over all of the professional capabilities I think will be one of the largest most exciting trends in the market over the next year or two.

What's the shape of that task then? I mean it's so vague. It's it's so much like less quantifiable than what we what your score was on the IMO, although that's incredibly impressive if someone can do that level of math. Um, go build good software feels really unverifiable, feels really broad.

Exactly. And so that's why you need humans to define the stasis points. It's so much more difficult to measure. And so one way of doing it could be to build a rubric where the model can use the criteria to score the deliverable that's being produced. Like imagine you want the model to be really good at building a web app that looks beautiful. You could have rubric criteria for you know all of the different elements of of said web app or or whatever um you're ultimately building. And so having humans define the success criteria and using that those verifiers as part of in our environment to train models iteratively to learn how to optimize for those criteria is one of the enormous trends that we're seeing across all of the frontier labs. How do you how do you sort of plan with the team when you have overwhelming demand for a current product yet simultaneously need to try to predict future demand in a way that I think normally when companies are doing traditional demand planning it's like very clear of like just how many customers can we reach with our current set of products and future products but in your business it's not always entirely obvious what the needs are going to look like you know even a couple years out.

Totally. I think that the most important thing is always working on the frontier and understanding what are the leading indicators of what the entire economy is going to be doing soon and adopting soon. And emphasizing that frontier in all of our product development and all of our investments is one of the most important decisions that we've made historically as sort of a a framework for resource allocation. That makes sense.

Uh we were just watching a video of a of a French company, Wandercraft, uh making a humanoid robot. People were kind of joking about it. I thought it was pretty impressive. I haven't seen any humanoid robots out of uh out of France. Um but uh can you talk about like what the data collection process for humanoid robots looks like now? Yeah, specifically a lot of, you know, we were at Meta Connect yesterday talking with Zach and Bos and the team and a lot of people saw the announcement and they said, "Okay, a lot more people are going to have cameras on their faces soon." Yeah. How valuable is this? Should should frontline factory workers be, you know,

uh be collecting data today or or is that not necessary?

Yeah, it's interesting. The key thing for the models to learn is having a clearly defined reward. And so I'll give like a couple of examples of that and sort of the role that humans can play. The first one actually without humans is that uh I was at the um this like robotics office where they had robots that were folding laundry and then they would have a vision model look at the laundry to see if it was folded properly as the reward. So right having that stasis point where you can have a 100 different model trajectories see which five trajectories are right and then reward those trajectories so that the model increases its probability of doing that correctly in the future is very powerful. All the way to another example where models are proposing scientific experiments and then we need humans that we hire as contractors to run those experiments in the physical world report on the results and say how they did. And so I very much believe that models will learn from their experience and interacting with the real economy. But so much of that experience similar to the way that you or I learn is curated by humans in the way that you know we help models with running the experiment in the physical world and giving feedback etc.

You tweeted the letters IPO a while back. What did you mean by that? Well, I did I did comment in parenthesis below that kidding. Uh

oh, okay. I missed that part. I missed that part. Okay, that

it's all good. Well, people people took it seriously because I remember that

I'm sure you got a lot of like frantic calls from bankers being like, Brandon, you told me you'd tell me when you're ready.

Yeah. Well, it's funny because last year when we were a seedstage company, I tweeted IPO by end of year and everyone thought I was, you know, a little bit crazy because we'd raised our our $40 million seed round. Uh or sorry, I tweeted Yeah, Unicorn by end of year. And uh we, you know, ended up making it happen. And so people thought maybe this time was was real, but I was I was just kidding about the IPO.

Got to keep

What's the uh what what's the shape of the business now? Obviously there's a huge amount of focus on on these expert networks and and these really highskilled specialized talent uh getting data from that and working through different problems and all the examples that you gave. Uh is there still a need uh from big labs for just the more uh traditional RHF? Is this good? Is this bad? thumbs up, thumbs down or has that been completely absorbed by uh just

users of the of the

the users or also just the models themselves like is is are GPT5 or or you know any of the other models the frontier models are they able to deliver the the the thumbs up thumbs down if you need to do some sort of fine-tuning on a specific uh problem?

Yeah, it depends on the lab. There's definitely still large investments that are happening in RHF. However, it seems like it's more efficient to collect those via data flywheels in the real world. Uh, and where you really need expert human involvement that's incredibly valuable is someone that will think about a problem for 5 hours and come up with this like very structured framework for how to evaluate model success in a way that it's difficult to expect uh users uh of products to do in a reliable way. Is that the is that like a reasonable way to think about a task a single task like five hours of

that might be one way but one thing I'm very excited about is that the time horizons of agentic trajectories will go up dramatically right and so I think a lot of people initially think about AI in the context of what can they see on their screen on chat GBT at any given time but over time it's and maybe using one tool with like online research uh but over time we're going to have the models working on problems that would take a human 30 days to do or 90 days to do that are using 10 different tools are interacting with various employees in the companies and we need environments for all of that right we need ways to eval to measure success to define the rewards and that's going to be a very exciting problem space to continue pushing the frontier of

uh how does how do you or just humans generally fit into solving the problem of like booking a flight or uh or you know ordering Door Dash we've heard about these uh simulated environments RL environments verifiable rewards like uh is it is it mostly designing the environment with the reward uh or is there actually a process for um someone who's just a fantastic travel agent to uh you know create a a rubric or create a uh or just actually do a ton of tasks to generate data. Yeah, it could be. So, one way you would do it for those kinds of browser use workflows is that you could have a simulated application and then a unit test that measures if the model effectively, you know, completed the task. It changed the state and like booking the applica the flight or whatever the action is. And ultimately, you do need a human expert to help write what is that unit test. Um, but my guess is that computer use will be solved relatively quickly in the next like or at least in the next like two or three years and then there's going to be this much longer tale of sort of the broad space of knowledge work and everything that we want to do of how do we get the model to build a startup or uh help help uh prep for a podcast episode or or whatever the workflow is.

Yeah. How do you think about uh solving problems that take like decades? Like I I always go back to like you know health there's certain things where you know the the FDA does a lot of work to try and understand uh you know if you're consuming this particular ingredient for 50 years how does that affect you? It feels like until we can simulate the entire human body and and run it at a faster clock rate, like that seems like something that you just can't really short circuit. We're already getting to like longer and longer rollouts, and that feels like that might be some sort of like damping function on how quickly we can compound. But are there any promising uh strategy that you've heard for dealing with uh problems that just take a long time to actually understand the reward or understand the did the past fail?

Yeah, the concern I would have with that scenario is it's so difficult to perfectly simulate like the human body and how that would play forward. And so my guess is that for a very very long time models will more so look at empirical analysis. However, models might survey people that took certain uh you know, vitamins or or whatever drugs uh when they were a certain age and see the impact that that's had in their analysis, but it'll it'll be difficult to simulate in that case. There are other cases where it's like a well scoped physics simulation of how well uh does you know this like ball roll down the plane or or whatever uh we're modeling out. that'll be easier to simulate uh and therefore for models to have an accurate understanding of how things will play out in the real world.

What's uh are you guys uh naturally 996? Do you do you I always get this question but but I but I feel like you guys are probably more on like a

like 997 like six is almost for the week. you guys are 20 21 like what what else do you have to do besides uh besides this?

Well, it's funny because we have actually never really mandated ours the company. It was just

that's I figured but there's like if you're ramping revenue from one to 500 million and

it's a lot of work. Yeah. And a little bit of time.

Exactly. So I think it the way it started was our initial core team was working like seven days a week. um and everyone was in the office like super late, all this stuff. And so we initially gave that rough guidance because we wanted people to have a little bit more balance and going home earlier, etc. Uh but obviously as the companies developed, I think it's important to be able to hire people that have families and even though they'll work hard on the weekends, they might still be at home. And so there's um there's some of that as well. Still emphasizing a lot of intensity and you know moving mountains for customers but at the same time not necessarily being as input oriented and and much more focusing on outputs.

How big is the team now? Like the full-time core not the network

relatively large. We're 250 people now.

Wow. Congratulations.

Fast. Yeah exactly. Um across the US and India. um and then a little bit across Latin America and the UK as well. But um yeah, it's it's been exciting. Certainly a crazy feeling to start having all these people in our our new SF office and and new faces that I have to meet. So, um lots of

What kind of predictions are you going to make? Uh uh any numbers you're throwing out? You you said unicorn by end of year last year, but what about what about going forward? anything you're willing to I'm willing to to take

we could say duckorn by end of year this year so

feels like you might uh might have already had it in the bag but

yeah we'll have to call R for rock to get this congratulations we'll talk to you soon on Bye

next we have Darren Mauy from Google Cloud coming in big Google announcement today big event uh over at Google in the uh global startups at Google Cloud