Mercor Hit $100 Million ARR Teaching Models to Hire. Humans Now Write the Exams.

Sarah Guo and Elad Gil open No Priors with a company that barely existed two years earlier. Mercor was founded in 2023 by three college dropouts and Thiel Fellows. By April 2025 it had raised $100 million, crossed a $100 million revenue run rate, and was working with the top AI labs. The work sounds circular until Foody states it plainly: Mercor trains models that predict how well someone will perform on a job, better than a human can.

The analogy is ordinary recruiting. A person reviews a résumé, runs an interview, and decides who to hire. Mercor automates those steps with language models. The loop is effective enough that frontier labs use it to hire thousands of people whose job is to train the next generation of models. Guo asks what skills those labs actually want. Foody's answer is almost a slogan: everything that is economically valuable.

Reinforcement learning, he says, has become so effective that once you create evals, models can learn the capability. That is why the shopping list runs from consultants and software engineers all the way to video-game hobbyists. Whatever the foundation-model companies or the application-layer companies decide to invest in, evals sit upstream of that decision. The hiring product started as a way to match talented people to jobs. The human-data vacuum is what made it a business at this scale.

Foody is careful about the origin story. Mercor did not start as a human-data company. The founders saw phenomenally talented people around the world who were not getting opportunities, and they thought language models could make matching them to jobs more efficient. Then they met a few customers and found a vacuum. The human-data market was changing shape faster than the old vendors could follow.

The old market was a crowdsourcing problem: how do you get a lot of low- and medium-skilled people writing barely grammatical sentences for early versions of ChatGPT. The new market is a vetting problem. Labs need some of the most capable people in the world, working directly with researchers, to push the frontier of model capabilities. Mercor kept the original DNA of hiring for roles, human-data and otherwise. Many customers now hire for both.

Gil asks whether all hiring, or at least all knowledge-work hiring, eventually moves to these systems. Foody does not hedge. On most of Mercor's evals, models are already better than human hiring managers at assessing talent, and this is still early innings. He thinks the world gets to a point where it is almost irrational not to listen to the model. Legal process may still require a human pressing the button. Trust, he says, will sit with the model's recommendation on who should do a given job.

Gil raises the 10x person: the coder, the physician, the investor who is wildly more productive than the average. Can Mercor's models find outliers? Foody lights up. The power-law nature of knowledge work, he says, is what frames the importance of performance prediction. If you can tell which engineers on a team will land in the 90th percentile, or that someone who costs half as much will still perform in the top quartile, you have named the value the company creates for customers.

People often talk about human performance as a bell curve. Foody says it is industry by industry. Investing is the most power-law thing imaginable: a handful of companies each decade matter a disproportionate amount, and the investors who win those bets capture it. Factory workers are closer to a commoditized skill set, with much less spread. Software engineering sits in between. It is definitely very power-law, he says, but not as extreme as the handful of best investors in the world.

His rule of thumb for where models pull ahead is blunt. Everything you can measure with text, the models are already superhuman at. If you can ask questions in an interview and read the transcript, they are better than people across more domains than he first expected — more domain-agnostic than he would have guessed. The lag is multimodal: how passionate is this person, how persuasive, how good at sales. Those signals will come. They will just take more time.

Volume is the other divider. Assess twenty people for the same job, hire them, and watch them perform, and it is easy to attribute a nuance in the interview or the résumé to the outcome. Give those twenty people twenty different jobs and attribution becomes a mess. High-volume, homogeneous processes get automated first. Guo asks what discovered features have surprised him. For engineering, Foody points at the public internet.

The best engineers leave signal everywhere: GitHub, personal projects on a website, a blog post written in college. Designers have Dribbble. Hiring managers do not have time to read it. Manual process is the bottleneck, so people under-index on what can be found online and over-index on what an interview can cover — passion, required skills — which humans are at least more practiced at today. Hidden signals show up in other professions too.

One finding that looks obvious only in hindsight: people based internationally who studied abroad in a Western country tend to work more collaboratively and communicate better. A human without full market context will not extract that. Intrinsic motivation is even more important, and it is not just who you hire but what they should work on. Hiring a biology PhD for biology problems is not the same as hiring the person whose thesis was on drug discovery to invent problems in that exact context. Matching still throws most of that information away.

The methodological rule is to stop testing proxies. If you care whether someone can build an MVP, the interview should be a scoped-down version of building an MVP. Proxies are for long-horizon tasks, and even then you should design them to carry as much signal as possible. Taste, Foody says, often shows up as softer signals: how someone approaches a problem, which nuances they notice in a product. Score it by putting people on a job-like task and comparing them to each other, including the thought process along the way.

Gil asks about the share of people who eventually get displaced once you can compare relative output. Foody does not offer a soothing timeline. Displacement in a lot of roles is going to happen very quickly, it is going to be very painful, and it will be a large political problem. He expects a big populist movement around it. The economic question underneath is how to respond: what should people now in customer support or recruiting do in a few years, and how do you reallocate wealth if the gains of superintelligence are themselves power-law.

Where do the displaced go? More of the physical world, he thinks, and more niche work. That can mean people creating robotics data, waiters, therapists — anything where people still want human interaction. Automation in the physical world will move much more slowly than in the digital one, because the virtual world has self-reinforcing gains and self-improvement that the physical world does not. Customer support and recruiting displacement is already being reported; a lot of it clusters in economic contractions, when companies get serious about the bottom line. The larger wave, he says, has not fully arrived yet.

On what skills to invest in, Foody quotes Sam Altman: optimize for being versatile, able to learn quickly and change what you do. Models get good at surprising things, fast. His own heuristic is verifiability. Math and student code, because they are verifiable, will get solved very quickly. You want a feedback loop the model can optimize against. Taste in a founder is much harder to automate, and the signal is sparse. He generally believes code-and-math reasoning will transfer, but you still need a reasonable amount of data to kick-start a new domain.

Counsel for parents of five-to-ten-year-olds: he would not push computer science, though he is not totally against it. Find something intellectually stimulating the child is passionate about, and let them learn general reasoning. He is skeptical that "people who can code" is the valuable thing in five years. More likely it is people with contrarian ideas about what is missing in markets, and the taste for the features that actually solve the problem. Guo's personal-assistant wish list makes the same point from the other side: models that ace math tests still cannot prioritize, sequence, and unblock a messy human calendar. Foody's diagnosis is that reasoning is not the missing piece. Evals and product are.

The so-called evaluation crisis is that models are clustering at the frontier and old academic, zero-shot test questions no longer tell them apart. Foody thinks the deeper mistake is what we eval. A software engineer at work does far more than open a pull request: coordinating with a product manager, fitting work into team priorities, turning all of that into output. The thing we actually need to measure is economically valuable work. He expects an immense amount of eval creation for agents, and calls that the largest barrier to automating most knowledge work.

Start with homogeneous tasks. Guo mentions Sierra's tau-bench as an early sample. Customer-support tickets are a good first target: one interface, a couple of tools, a relatively uniform job. Diverse work such as being a good software engineer — taste for how to approach a problem, a feel for products people enjoy — will be a years-long buildout even in verifiable domains. Gil asks how many people end up doing data collection. Foody thinks the volume is huge, because it all comes down to creating evals for everything in the economy. Some of that is current employees defining what good looks like so agents can learn. Some is contractors through a marketplace. It would not surprise him if that became the most common knowledge-work job in the world.

Elad offers a friend's analogy to the Nyquist sampling theorem: you can tell someone is smarter than you, but not how much smarter, because you cannot sample richly enough. Taken to superintelligence, Foody agrees that lands. His economic frame is simpler. Doing a task repeatedly is a variable cost; building an eval so a model can do it is a one-time fixed cost. Work should trend from the variable to the fixed. If models get so good so fast that humans are not needed to create evals, they will not be needed in many other parts of the economy either. Med-PaLM 2 is the cautionary case: physician panels already scored a Google health model as better than individual doctors. At some point human scoring should make the model worse. Foody's reply is that models will learn to ignore the mistakes and keep the insight. He also warns against reading SWE-Bench curves as a countdown to replacing software engineers. Coordination, tool use, and agent evals are a much longer build than a couple of years.

The technical hinge is reinforcement fine-tuning. Supervised fine-tuning never really took off: companies made a few hundred input-output pairs, scaled toward tens or hundreds of thousands, and still missed the capabilities they wanted. RFT instead defines the outcome. Sierra defines what a good support response looks like. Mercor defines the characteristics a candidate should show — passion in the interview, domain knowledge, a side project that proves the skill — and rewards the model for identifying them. The model then learns, in that environment, how to score well. Hundreds to thousands of examples, a tenable number for a mid-size business, is enough. That, Foody says, finally makes application-layer customization economic. Agents fill the roles employees currently fill; humans, in-house and on the marketplace, build the evals.

Asked for the biggest shift nobody is anticipating, Foody splits the answer. Most of the country, he thinks, has not grasped how fast jobs will be displaced, and governments need to stay proactive. Inside San Francisco, people under-index on agentic evals for non-verifiable domains, and they under-think the role humans will play in ten years because they are so focused on automating humans. Models should help figure out what people are passionate about, even if the project is not economically valuable. Gil notes that the economy already contains forms of de facto UBI: waste in government, administrative bloat in academia, high-pay low-output roles in big tech. Healthcare and education spend more every year with almost no gain in output, because there is no economic pressure. Foody expects better analytics on employee value to produce more layoffs, with political pushback, but he thinks the economics are too strong for companies in competitive sectors to refuse.

A quieter inversion: people assume AI will first be a great individual contributor. Foody thinks it may sooner become a great manager — taking a large problem, breaking it down, performance-managing people. Guo wants an assistant that, given enough context and objectives, prioritizes, tasks, and sequences the next three months so she is not the bottleneck. Today's models ace math tests and still fail that job. Foody does not think it is a reasoning gap. With a small amount of agent evals in various categories, the base model already has the reasoning. What it lacks is knowing when to use tools, how to synthesize their output, and what good looks like inside each company's knowledge base. That is post-training, which is why he is so bullish on RFT.

Mercor's next year, and the next five, come down to two bets. First, get the smartest people in the world onto the platform — Uber- and Airbnb-style supply-side network effects. Average labor marketplaces run about 50-to-1 supply to demand, which means most applicants talk to a friend who also applied and neither got a job. The way to scale to hundreds of millions of people is to give away AI mock interviews, career advice, and shareable profiles, funded by the monetization engine on the other side. Second, performance prediction: customer data on who is doing well and why, feeding a flywheel. The marketplace network effect is the obvious one. From early results, Foody thinks the data flywheel will matter more over time.

The largest inefficiency in labor markets, he argues, is fragmentation. A candidate applies to a dozen jobs. A San Francisco company considers a fraction of a percent of the world, because matching is still a manual process of résumés and interviews. Solve matching at the cost of software and you get a global unified labor market every candidate applies to and every company hires from. He calls that the largest economic opportunity in the world, and the most impactful one for finding people work they will be passionate about. Would it include AI agents competing alongside humans? He thinks so, because customers arrive with a problem to be solved, and the answer is usually some mix of the two. For startup hiring, the advice is specific. Early on, always index on quality over speed for those first employees, and be patient. As you scale, do not drop the bar, but get data-driven about which characteristics actually drive the outcomes you care about. The failure mode is vibes-based assessment, each manager inventing a standard that cannot be enforced. Gil's dichotomy: great companies either hire well or fire well early, rarely both. Google hired well and took years to clean people out. Facebook's early pool was mixed, but it removed underperformers quickly. Foody's romantic close is a Thiel-Fellow daydream: imagine Peter Thiel as a heuristic, interviewing everyone in the world at eighteen. He thinks that world is approaching quickly, and it will rewrite both labor markets and investing.