Unpredictability of AI Models
This was my third project with Mercor, and each one continues to be a learning experience.
One of the surprises has been the unpredictability of AI models. They can handle very complex prompts with ease, yet fail on seemingly simple ones. That tension—what works well versus what breaks down—reminds me why our role matters. We’re here to uncover blind spots, highlight weaknesses, and help guide improvement.
The paradox is that we often overestimate AI and underestimate it at the same time. We assume it can do more than it can in some areas, while overlooking its potential in others.
How AI Training Works
The AI training process is both rigorous and intellectually stimulating. Every task a physician writes is reviewed by another physician, then checked again by a super reviewer (also a physician), with an additional layer of quality control on top.
Now that I serve as a reviewer/super-reviewer myself, I see just how important this structure is. I’ve learned a lot—not only about AI, but also about medicine, as reviewing requires thinking critically about accuracy, clarity, and depth in every task.
We’re expected to commit at least 10 hours a week and, just as importantly, stay responsive. If feedback comes back from a reviewer, the expectation is to address it within a day—not a week. That keeps the projects on track and ensures consistency.
Compensation and Bonuses
Compensation isn’t always straightforward. While projects often have a base hourly rate, some also include performance bonuses, which can make the upside much higher than expected. There are many factors that go into the bonus, so it’s best not to expect it—but when it does come through, it’s a happy surprise.
Mercor’s Growth Trajectory
For context, Mercor itself has grown at a remarkable pace—reaching a $450M annual run rate in under two years and now aiming for a valuation above $10 billion. That makes it one of the fastest-growing companies in history, and a reflection of just how central AI training and evaluation have become to the future of large language models.
A Common Question
There’s debate about whether physicians should even participate in training these systems. My perspective is that AI will advance regardless. The real question is whether we want to help shape it responsibly, or leave those decisions entirely to others.
My prediction is that AI will not replace physicians. Rather, it will unlock new innovations, open doors to different ways of working, and expand what’s possible in medicine. Like every major technological shift before, it will likely create new roles and opportunities that we can’t yet fully imagine—while also raising new challenges we’ll need to address.
Final Reflection
For me, being part of this work has been both eye-opening and rewarding. Up until this point, AI has largely been a toy in high-stakes settings—interesting to experiment with, but not something physicians could reliably use for decision-making.
The real test is ahead: whether this kind of rigorous training and evaluation will make the tools actually useful in clinical and professional settings for physicians. And it’s just the beginning of what we’ll see as AI continues to evolve in healthcare and beyond.