Math kids vs humanities kids in tech: the split I stopped believing

I used to sort résumés with a private prejudice I would not have said in a hiring meeting. Physics, CS, math olympiad — “can think.” English, history, philosophy, a bootcamp after a classics degree — “will write the docs.” I hired both. I was wrong about which label predicted who shipped.

The split is a campfire story the industry tells to make taste look like a filter. After a decade of 1:1s, performance packets, and incident reviews, the people who shipped had a different habit in common, and it did not live on a transcript.

What I thought I was buying

From the math-shaped candidates I expected comfort with invariants, a stomach for formal systems, less fear of a proof or a probability. Sometimes I got that. Sometimes I got a person who could talk about Big-O and could not tell me why last Tuesday’s payout job double-sent. The olympiad glow does not automatically include “I read the logs.”

From the word-shaped candidates I expected communication, product sense, empathy with a confused user. Sometimes I got that. Sometimes I got a person who wrote beautiful RFCs and would not close the loop on a failing test because the test was “too pedantic.” Prose is not ownership.

The worst hires in both piles shared a trait: they optimized for looking like the archetype. The math kid performed rigor in design review and disappeared when the staging data was dirty. The word kid performed collaboration in Slack and disappeared when the API contract needed a hard no. Archetypes are costumes. Costumes do well in interviews.

What actually showed up in the work

The people who shipped — a CS master’s, a history BA, a self-taught former support lead, a stats PhD — did a small number of unromantic things on a schedule.

• They finished. Not “I have a branch.” Finished: in production, with a rollback, with a note in the PR about the thing that could still go wrong.
• They could hold a system in their head well enough to predict a failure. The history hire who asked “what happens if the webhook is two days late” was doing the same job as the math hire who wrote the state machine. Different vocabulary, same move.
• They wrote the sentence other people needed. Sometimes that sentence was SQL. Sometimes it was a paragraph in the runbook. The medium was not the talent. The habit of leaving a next person a trail was.
• They changed their mind in public when the data said so. I have worked with brilliant formal thinkers who treated a wrong model as a personal insult. I have worked with “soft” majors who treated a wrong model as Tuesday. Guess who unblocked the release.

I stopped asking “are they a math person.” I started asking “do they finish, predict, leave a trail, and update.” Those are observable in a work sample and a reference. They are not observable in a degree line.

Where the labels still leak into process

Our loops still punish word people and flatter math people in dumb ways. A leetcode screen is a math-shaped ritual even when the job is a Spring service and a lot of Hibernate. I still run a small algorithmic screen when the role is actually algorithmic. I do not run it as a class filter. When I did, I filtered for people who had time to grind, which correlates with a certain kind of university more than with on-call judgment.

System design interviews can go the other way. A confident talker will draw boxes. A quieter engineer will ask about the data. I used to score the boxes. I now score the questions. The history major who asked about the write path’s idempotency beat the CS major who invented a clever consistent-hash ring we did not need.

Take-homes punish people with kids and reward people who enjoy homework. That is not a math/humanities split. It is a life split we pretend is a merit split. I prefer a pairing session on our repo or a sanitized slice. I watch whether they ask about the existing tests. That ask has come from both piles. It has also been missing from both piles.

The skills I was mis-attributing

Reading a dense RFC or a paper is not a humanities exclusive. Writing a careful email to a regulator is not either. Quantitative comfort is not a math exclusive — I have had English majors who became the person everyone asked about a funnel because they were the only one who opened the warehouse.

What I will say, narrowly: if the job is a compiler, a quantitative library, or a research-shaped ML eval, I want evidence of that craft, and a degree can be one kind of evidence. If the job is a product team that ships a B2B workflow, I want evidence of shipping. I have wasted calendar time treating the first kind of evidence as a proxy for the second.

Domain language is the other mis-attribution. “Word people will talk to the business.” Some will. Some math people will too, if you do not train them that talking to the business is a status demotion. I have done that training by accident, by praising only the clever design. That is on me, not on their major.

Two people I keep using as a correction

Mara had a topology background and a habit of rewriting other people’s tickets into a more formal language. For three months I thought I had hired a staff-shaped thinker. Then I noticed her PRs never included the rollback. She treated production as a corollary. When a migration locked a table, she had a correct proof that it should not have, and a team that needed a working down. She got better when we made rollback part of the definition of done. The degree did not prevent the miss. The process caught it.

Jonah had studied comparative literature and came through a support-to-eng bridge. He wrote slowly. He asked questions that sounded naive in Slack. He also shipped the refunds tool that finance had been begging for, because he had been the person who used to paste those refunds by hand. He knew the ugly cases. He encoded them. I almost screened him out for “not technical enough” on a phone screen where he refused to pretend he remembered a red-black tree. I still think about that phone screen when I catch myself warming to a fluent algorithm recitation.

I do not tell these stories to invert the prejudice — humanities good, math suspect. I tell them because my first-pass story about each person was a cartoon, and the cartoon was expensive. Mara became excellent once we stopped praising only the formalization. Jonah was excellent as soon as we let him near the actual work. The split would have had me over-index on one and under-index on the other.

How I hire now

I read the degree after I have a work sample, not before I decide if the sample is worth doing. I ask for a story of a production miss and what they changed. I ask them to explain a system to me as if I will own it next week. I do not care if the explanation starts with a metaphor or a formula. I care if I could page them from the explanation.

References: I ask “when did they last finish something ugly.” I do not ask “are they technical.” Technical is a compliment people give to people who sound like them.

On the team I try to mix the vocabularies on purpose. A spec that only a math-shaped person can read will be implemented wrong by everyone else, including other math-shaped people who were tired. A spec that is only a narrative will miss the invariant. The mix is the point. The caste system is not.

What I tell leads who still believe the split

If your best engineer has a physics degree, that is a person, not a pipeline. If your worst incident was caused by someone with a literature degree, that is a person, not a warning about English departments. Base rates in your last company are not a science.

Believe the work. Believe the trail they leave. Believe whether they update. Unbelieve the campfire. I hired both “math” and “word” people. What predicted who shipped was whether they treated the system as something that could hurt a user on a Thursday, not whether they had taken real analysis.

I still like people who can do real analysis. I like them more when they also write the runbook. I like the history hire more when they also read the query plan. The split I stopped believing is the idea that I have to choose. The job already chose: it wants both moves in one skull, or in a team that does not rank one skull above the other.

If you need a practical rule for Monday: hide the school line until after the work sample. I have done this with a coordinator who redacts the education block. It is a small trick. It has saved me from talking myself into a fluent interview and talking myself out of a quieter finisher. That is the only split I still believe in — fluent versus finished — and it cuts across every department I have hired from.