What does a Data Scientist really do?

Eugene Yan has hired data scientists, managed them, and fielded enough LinkedIn messages to notice a pattern: almost everyone outside the field pictures the job as one long neural-network bake-off. Stack a PhD, some olympiad-grade math, and a GitHub full of transformer experiments, and you're set. Yan says that picture is mostly wrong, and he's got the receipts from years of actually doing the work.

His theory for why the myth persists is simple availability bias. The data scientists who make headlines are Turing Award winners cited in Forbes, or DeepMind researchers whose team bios read like a PhD directory. Coursera courses reinforce it further, loading up on backpropagation and loss functions because that's the sexy, teachable part. So the public builds its mental model of 'data scientist' from the 1% who spend 80% of their day training the same architecture over and over, hoping for convergence, and generalizes it to everyone else.

Yan splits the field into two rough camps — Type A, the analysts who behave a lot like statisticians, and Type B, the builders who write production code and ship models that actually run somewhere. He writes mostly about Type B, and the list of what that work involves is longer than most job postings admit: framing the problem, negotiating what data even exists, wrangling nulls and outliers, joining messy sources without quietly corrupting them, building validation frameworks that don't leak information, running experiments, and then — only then — worrying about which model architecture to pick. After the model exists, someone still has to wrap it in an API, schedule it, monitor drift, plan for rollback when it breaks, and explain to a room of skeptical executives why they should trust a number a machine spat out.

When Yan surveyed Chief Data Officers and data science leads about what separates a great data scientist from a mediocre one, he expected an answer rooted in math depth or algorithmic cleverness. Instead nearly everyone said some version of the same thing: the best ones use data to deliver measurable value. Not a technique you can drill with more LeetCode or textbook chapters — a habit of judgment that's much harder to teach and, by his own admission, much harder to define. He leaves that mystery for a follow-up post, but the shift in framing alone undercuts the hiring pipelines and course curricula built around the flashier, narrower skill set.