In January 2025, Mark Zuckerberg went on Joe Rogan’s podcast and said that the workplace has become “culturally neutered” and that he wants to bring “masculine energy” back to it. But what does “masculine energy” mean, exactly? It turns out you can measure it.
A lot of this project runs on AI embeddings — models that turn a piece of text into a point in a space with hundreds of dimensions, where text with similar meanings lands close together. But directions in that space carry meaning too. The classic party trick is that if you take the vector for “king,” subtract “man,” and add “woman,” you land near “queen.” A study from 2019 by Kozlowski, Taddy, and Evans called “The Geometry of Culture” turned that party trick into a measurement instrument: take many word pairs that differ only in gender — man/woman, he/she, father/mother — draw the arrow for each pair, and average them into a single, stable gender axis. Then any sentence can be projected onto that axis and scored: how strongly does this language associate with the masculine or feminine pole of the space?
This story does the same thing with the careers corpus. The axis is built from sixteen pure gender-term pairs — and, very importantly, there are no intuition words in the poles. I never told the model that “hardcore” or “relentless” is masculine. The axis is built entirely from words like he/she and father/mother; everything else is projected onto it and scored. If “hardcore” comes out masculine-coded, that’s the training data talking, not me.
The good news about biased training data is that it can reflect our biases back to us. When this story says “masculine-coded” or “feminine-coded,” that isn’t a personal judgment — not “direct communication is for men, therefore companies shouldn’t talk about direct communication.” It is a computed proximity to masculinity: if decades of English text have these concepts mapped to statements made by men, for men, or about men, the embedding shows that. In this particular study, gender bias is the feature, not a bug.
Every sentence of careers-page copy in this corpus is scored on that axis. One square per sentence.
Anduril
Basecamp
Ramp
SpaceX
Engine
Palantir
Netflix
Coinbase
Stripe
Shopify
HubSpot
NVIDIA
Amazon
Meta
Airbnb
GitHub
GitLab
Apple
Salesforce
Starbucks
Uber
Snap
feminine-coded (z ≤ −0.5) neutral band (−0.5 … +0.5) masculine-coded (z ≥ +0.5) one square = one sentence · deeper = further from neutral · hover for the sentence
Three notes on trusting this measurement. First, it passes the tests where the answer is known: occupation terms with heavily male-skewed workforces (infantry soldier, lumberjack) land cleanly on the masculine side, heavily female-skewed ones (kindergarten teacher, nurse) land on the feminine side, and genuinely neutral terms land near zero — that’s also why scores inside the ±0.5z band are shown as neutral rather than over-read. Second, the axis isn’t hostage to any particular word pair: splitting the sixteen pairs into two independent eight-pair axes ranks the corpus nearly the same way. Third — deliberately — there is no human-rating backstop here. Asking a human judge whether “bias for action” sounds masculine would re-import the very intuitions this method is designed to bypass; the coding lives in the corpus statistics, so validation is known-answer anchors and split-half reliability, not human agreement.
One caveat belongs next to the chart rather than in a footnote: the coding partly tracks who is doing the talking. Basecamp — famously anti-hustle — sits near the top of the ranking, because assertive founder-manifesto prose (“we don’t do X”) is masculine-coded independent of intensity, while HR-authored benefits copy pulls feminine. Companies whose pages are written in a founder’s voice read more masculine than companies whose pages are written by a communications team, so the fair comparisons in this story are within a register, not across registers. Relatedly, sentences that mention people by name score masculine for a boring reason — “Eric” is a gendered word, the same as “he” — and this story is trying to measure the register, not the roster, so the specimens quoted in the prose are free of names and pronouns. (One quote card below, Ramp’s letter, names both founders and earns part of its score that boring way; it stays because the letter itself is the genre’s purest specimen.)
If it sounds like it isn’t for you
Some careers pages say it almost outright. The register runs from SpaceX’s 2003 hiring criteria to Ramp’s 2026 letter persuading you not to apply — it predates the Netflix deck, which didn’t invent this language so much as make it quotable — and on the gender axis, its sharpest specimens are among the most masculine-coded sentences in the corpus. Coinbase’s homage to the Netflix severance line, “Unremarkable performance gets a generous severance package,” scores 1.8 standard deviations toward the masculine pole. Engine’s culture memo is a fountain of these: “Comfort is a signal to probe, not rest” (+1.7), “Silence is not humility here; it is abdication” (+1.6).
Meanwhile, the one nearly neutral card below is the plainest statement of the idea: Netflix’s “But Netflix is not for everyone, so please read on” scores almost exactly zero. It’s not the sentiment that’s masculine-coded. You can tell people your company isn’t for everyone in neutral language. It’s the performance of it — the hardcore, the hard mode, the not for the faint of heart — that carries the gender coding.
SpaceX · 2003 · +0.54z
“Out of necessity, SpaceX has exceptionally stringent hiring criteria.”
Apple · 2009 · +1.26z
“This isn't your cushy corporate nine-to-fiver.”
Netflix · 2009 · +1.20z
“Unlike many companies, we practice: adequate performance gets a generous severance package.”
Coinbase · 2024 · +1.11z
“You will be pushed beyond what you think you're capable of.”
Netflix · 2024 · -0.10z
“But Netflix is not for everyone, so please read on.”
Engine · 2025 · +1.33z
“Every seat must be earned, and underperformance is addressed quickly and respectfully.”
Ramp · 2026 · +1.82z
“Our two founders, Eric and Karim, thought it'd be a good idea to write a letter persuading you not to apply.”
That 2003 SpaceX page, the oldest artifact in this register, also contains the line “SpaceX does not discriminate on the basis of anything but skill.” Maybe they believed it. But the page is written in language that measurably sorts on something other than skill — it sorts on who feels spoken to.
The ideas companies share, ranked
It isn’t just companies that are gender-coded — the ideas they borrow from each other are. Clustering every sentence in the corpus into recurring concepts and keeping only the genuinely shared ones (at least three companies, two-plus sentences each, no company over 40%) puts the performance canon at the masculine pole — speed, flat structure, rejecting conventional norms — and the inclusion vocabulary at the feminine pole. Each concept is labeled with its own words.
0 = the average careers-page sentence (frozen 20-company baseline). Scores exclude sentences that explicitly mention gender. The axis passes a known-answer test: male-stereotyped occupations (infantry soldier, lumberjack) land on the masculine side, female-stereotyped ones (kindergarten teacher, nurse) on the feminine side, neutral terms near zero. Set aside by an LLM judge before ranking: 7 press-release / product marketing, 20 one company over the 40% cap, 4 diversity-reporting, 1 incoherent, 4 below the sharing floor, 4 navigation / boilerplate.
The two poles earn their coding differently, and the asymmetry is worth saying plainly. The masculine-coded concepts are style-driven: speed and iteration, flat organization, learning from difficulty rarely mention men at all. The feminine-coded concepts are often subject-driven: employee resource groups, pay equity, diversity-and-inclusion copy frequently talks explicitly about women. To separate the two, concept scores here exclude sentences with explicit gender mentions — and the feminine pole survives the exclusion. Employee resource groups barely move (−0.80 to −0.76), and diversity-and-inclusion culture holds at −0.56. Pay equity is the exception that shows the filter working: it softens from −0.95 to −0.66 once the sentences that literally say “women” are set aside, meaning a real share of its coding was lexical rather than stylistic.
One absence from this chart is deliberate: concepts dominated by a single company are excluded (twenty fell to the 40% cap). The most masculine-coded concept in the whole corpus — the severance-and-high-bar-hiring cluster — is one of them: it’s still 74% Netflix’s own voice, so it lives in the Netflix story rather than here, where the subject is the ideas companies actually share.
Masculine, or just not feminine?
“Masculine-coded” is one number, but it hides two different ways of earning it. The score is a difference: how close a sentence sits to the masculine pole of the axis, minus how close it sits to the feminine pole. That means a concept can score masculine by approach — actually living near the masculine pole — or by avoidance — sitting far from the feminine pole while not being especially close to the masculine one either.
Splitting the score back into its two halves shows both kinds exist in this corpus. Candor and boldness are masculine by approach: they’re personal, people-heavy registers that sit close to both poles, just closer to the masculine one. But concepts like effort-and-discipline, growth-and-scale, and communication-norms score masculine mostly by avoidance — impersonal process language whose distance from the feminine pole does the work. And the asymmetry runs one way: the strongly feminine-coded concepts all get there by approach. In this corpus, femininity is something careers copy moves toward; masculinity is often just what’s left when it moves away.
How this is measured: each pole is the average embedding of sixteen gender terms — woman, mother, sister, daughter, wife, queen… for the feminine pole, man, father, brother, son, husband, king… for the masculine — and a concept's position is its sentences' average cosine similarity to each pole, z-scored against all gender-mention-free sentences (0 = the average careers-page sentence). "Far from the feminine pole" means the concept's language is less similar to those feminine terms than the average sentence is. The ranking above scores each concept with one number — the difference between these two axes — so it can't distinguish the ringed concepts (masculine by avoidance) from the top-right ones (masculine by approach). The two pole similarities correlate 0.84, which is why position along the diagonal mostly measures how personal the language is at all.
Does any of this matter beyond the geometry? This corpus measures language, not outcomes — no promotion rates or pay gaps here. But the mechanism has been measured elsewhere: in a series of experiments published in 2011, people shown job ads that were identical except for gendered wording rated the masculine-coded jobs as less appealing and reported a lower sense of belonging — while their assessment of their own ability to do the job didn’t change. The language doesn’t convince anyone they can’t do the work. It convinces them they aren’t wanted. And self-selection is the cheapest, most deniable filter a company can run: nobody gets rejected, nothing shows up in the hiring statistics, people just quietly don’t apply.
So when a founder says he wants to bring masculine energy back to the workplace, the corpus has an answer: it never left. It’s been on SpaceX’s careers page since 2003, and it’s the water an entire generation of tech founders is swimming in. When your language sits at the far masculine end of the axis, neutral is what reads as neutered.