Mercor CEO Brendan Foody: working with 6 of the Mag 7, 45% month-over-month growth, and why AI training now needs real-world professionals not academics

in the studio breaking it down. Ultra lightning round. Ultra lightning round. We're going to rip through these really quickly. Uh thank you so much for your patience, Brendon. Great to hear from you.

Uh start uh kick us off with a quick introduction of how things are going at Merkore because uh it seems like you've been on an absolute tear. Things just keep the dominoes keep falling in your direction. Give it to us. What's the latest update from you? Of course.

Well, thanks first of all for having me back on, but it's been an exciting few months since we spoke. Uh we're now working with six out of the mag seven all of the top five that have a lot that's good. Um and have averaged 45% month- over-month growth for the last 12 months.

Um and so it's which of the mag seven are missing from the six out of seven. I wonder sorry that's me not you. Yeah, it's been it's been an exciting time u especially the wake of of all the news with scale of course um and seeing a huge amount of of customer interest um after that. Yeah. Okay.

We were talking to Clem about this earlier from hugging face uh where where's the biggest demand? Uh people used to think of what you do as uh and and just human data generation generally as uh kind of work that anyone could do.

Mechanical Turk work um you know just just you know checking for hallucinations and factchecking and uh you know is this a dangerous prompt? Is this saying a bad word? Now it's moved into PhD level work. Uh where's the state-of-the-art? Where do you see it going? Totally.

So I think the key development in the market is that reinforcement learning is becoming so effective that once we have evals for something the models can saturate them and so really anything that we want to create agents or LMS that are capable of doing we need to build out evals and RL environments in those domains and so actually while some of it is PhD level work there's a huge push into all of the professional non-academic domains moving away from things that that were academic of how do we find the consultants, doctors, lawyers, bankers that can evaluate and teach models how to do the things that we would want those professionals to do on the job.

Uh, and so that's been a really exciting growth area that we've been leaning significantly into. What's the solution to booking a flight? Like, do we need to get travel agents to define the computer use workflows, create some sort of a environment?

It seems like I should be able to go to an chat app or Siri and just say, uh, get me to New York tomorrow. Remember my preferences. I don't like connections. I'm willing to pay this much. I like to fly at these times. Um, and yet it's a very human problem that I don't think people are fully comfortable delegating to AI.

Yeah, it's a fascinating dichotomy that we're simultaneously talking about PhD level reasoning and Olympian math yet it can't do a lot of the very basic things like booking a flight.

Um, and I think the key is that really one of the primary barriers to research is anything that the model can't do, we need an effective way to measure. That way we can experiment with all the different data sets to uh help achieve those capabilities.

So we need people that otherwise could book those flights to create eval how agents can do that to measure what success versus failure looks like in in all of those cases.

And there's just that huge buildout happening across all of the hyperscalers um with respect to everything from simple tool use uh in how we book things or buy things online all the way to super high complexity reasoning over how all of these things interact with really complex knowledge bases.

What are you seeing in uh video specifically uh and like is there a role for humans in the loop in the training cycle or data generation around uh these video models? V3 is incredible. Feels like it's a beneficiary of YouTube. Um what are you seeing there?

Yeah, I think expansion to multimodality has definitely been something that's uh really exciting to see. I I think that uh there's definitely a lot of rich content online for video. Um but ultimately how we measure what those good videos look like and sample that um has been really important.

In particular, there's a significant amount of eval buildout in retrieval over videos and long context of like how can we watch an hourong YouTube lecture and understand uh and obviously that's just an example for uh one video case uh and understand like what are the elements of that content that really matter to users and how can um we more than just build useful models build useful products um that all of these companies can distribute to everyone.

Yeah. Um, yeah, that makes sense in this in the context of a lecture. I I guess I'm wondering like for V4, V3 is incredible.

Uh, the physics are remarkable and I'm wondering if it's important to have a human in the loop describing I mean I I generate a lot of like Michael Bay knockoff videos honestly and and there's things where it gets the physics perfectly and I'm wondering if there's if that's a beneficiary of not just a YouTube video of a Michael Bay trailer but actually having someone sit there and describe exactly what's happening in the Michael Bay trailer in plain text that then can be fed into the system and if we need to go further with the tagging and additional like metadata or transcription, not just transcription of what's said, but actually what's happening in the in the actual videos.

Is that an important step to get us to V4 where there's even less hallucinations? Yeah, we're definitely seeing a meaningful amount of that in that people want to create all of the tags using a combination of LMS and uh human expert data.

uh around what's happening in the video so that they can effectively work backwards from there to use those texts for for generation.

Um and especially for all of the really high complexity stuff at the frontier of what the models can't yet do or or might require a significant amount of reasoning, but we tend to do all of all of that highest complexity work, the stuff that is very difficult um and and high skilled to produce. Anything else, Jordy?

No, this is great. We're going through a lightning round. I know it was quick. We'll have you back soon to talk more. I'm sure there will be more news. It's the uh it's the hottest industry in the world. I just want to see the actual revenue ramp chart. It's going to be dropping soon.

You can trust me with the revenue ramp. Yeah. I won't tweet it out before you. No, we will. We will respect embargos. If you have an upcoming milestone, I'm sure there's something coming up. So, congratulations on all the success. Fantastic positioning in the market. Get some sleep. We'll talk to you soon. I'm the best.

See you guys. Have a great one. Bye. Up next, we have our next guest in the studio already. Sam from Super Dial. Welcome to the stream. How