Why experts writing AI evals is creating the fastest-growing companies in history | Brendan Foody

Brendan Foody, CEO and co-founder of Mercor, explains that the AI industry is entering the “era of evals.” This shift marks a pivotal moment where the primary bottleneck to improving AI models is no longer about sourcing massive amounts of pre-training data, but about creating high-quality evaluations to measure and guide model capabilities.

The “Era of Evals”: The New PRD for AI

The role of evaluations has become central to AI development. As Foody puts it, “If the model is the product, then the eval is the product requirement document (PRD).” Researchers run dozens of experiments daily, making small improvements against a specific eval set. Once a reliable evaluation exists, models can “hill climb” it effectively using reinforcement learning.

Key functions of evals include:

• A Measure of Success: Evals provide a systematic way to measure how well an AI can automate a core value chain, whether it’s generating architectural diagrams or performing investment banking analysis.
• A Guide for Research: They give researchers a clear target, transforming the abstract goal of “making the model better” into a concrete objective of improving performance on a specific, measurable task.
• Marketing and Sales Collateral: Evals are shifting from academic benchmarks (like Olympiad math problems) to practical, domain-specific capabilities. Labs and application-layer companies will increasingly use their performance on these custom evals to demonstrate the superiority and utility of their products.

The Shift in AI Training and Mercor’s Role

The evolution of LLMs has triggered a fundamental change in the human data market. The need has moved from low-skilled crowdsourcing for basic language tasks to a sourcing and vetting problem for high-caliber professionals. Top AI labs require experts—experienced software engineers, doctors, lawyers, and financial analysts—to evaluate and refine model capabilities in specialized domains.

This is where Mercor found its niche. The company provides a labor marketplace for these extraordinary professionals. These experts are not just labeling data; they are creating the foundation for reinforcement learning by defining what success looks like.

For example, a lawyer might be hired to create a detailed rubric for what constitutes a high-quality “redline” of a legal contract. This rubric then serves two purposes:

1. As the Eval: A benchmark to score the model’s performance.
2. As the Reward Signal: A basis for the reward function in a reinforcement learning environment.

This process is part of a broader market trend away from traditional Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) towards Reinforcement Learning from AI Feedback (RLAIF). In the RLAIF paradigm, a human expert defines the success criteria (the eval), and an AI system then uses this rubric to scalably generate rewards and penalties, making the training process far more data-efficient.

The Future of Work and In-Demand Skills

Contrary to the dominant narrative of job displacement, the “era of evals” is creating a new category of jobs. The market for human expertise is defined by “the amount of things where humans can do something that models can’t.” As long as this gap exists, there will be a need for experts to create training data. The entire economy may eventually become a large-scale reinforcement learning environment.

For individuals navigating their careers, Foody offers two key pieces of advice:

1. Focus on Jobs with “Elastic Demand”: Pursue careers where a 10x increase in productivity leads to a 10x or 100x increase in output and demand, rather than a reduction in the workforce. Software development is a prime example; if we can build software 100 times faster, we will simply build 100 times more software. Product management and other roles in the value chain of company-building also fit this description.
2. Master AI as a Tool: The most valuable skill is the ability to leverage AI to amplify your own capabilities. The adage holds true: “AI won’t replace you. People that are really good with AI will replace you.” Companies are already shifting their assessment methods, giving candidates access to all available AI tools during interviews and evaluating the quality of the product they can build in a short time.

On AGI and the Path Forward

Foody expresses skepticism about the imminent arrival of AGI or superintelligence. He believes the path forward is a long road paved with countless evals for every capability models currently lack, from basic tool use to long-horizon reasoning. Progress will come not from simply adding 10x more pre-training data, but from creating thoughtful, data-efficient post-training datasets built on the expertise that high-quality evals provide. The current landscape, with multiple competitive labs keeping each other in check, represents a healthy, stable state for innovation without the immediate risk of a single superintelligent entity emerging.

• Tags:
• Future of Work
• Large Language Models
• Reinforcement Learning
• Feature
• RLAIF
• AI Training
• AI Evals
• Mercor