In 2024, I did some analysis of changes over time in the LessWrong community. What I didn't do was compare the Rationalist community to the general population. I also historically haven't had enough charts to satisfy Ben Pace. Lets change that. 

I've run the Unofficial LessWrong Community Census for a couple years, and for various reasons the Astral Codex Ten survey is unusually relevant to my interests. I'm going to largely use the ACX survey here since it has a much sample size. One can reasonably ask whether there are population differences between people on LessWrong and people on ACX; I'm mostly going to shrug and consider either a reasonable approximation of "the rationalist community" unless curiosity and hunch compels me to go looking.

GSS data is from the General Social Survey explorer. ACX data is from the most recent ACX Survey. (It's labeled 2025, but was run in very early 2025.)

Also, I was working on this as part of a practicum led by David Gros on vibecoding datasets. This means it's not only largely vibecoded, but it's vibecoded by a relative vibecoding novice. You should be appropriately suspicious.

Comparisons

Age

• ACX 2025: mean 37.6 (median 40)

• GSS 2024: mean 50.0 (median 49).

  

  We skew younger. That's not especially surprising for a movement that took off in the late 2000s and has a strong showing among college students. 

Sex

• ACX 2025: Male 87.8%, Female 12.1%, Other 0.1%
• GSS 2024: Female 55.4%, Male 44.6%, Other not reported

Yeah, from eyeballing a lot of in-person meetups, a nine out of ten or four out of five rate of men sounds about right. What's surprising me is the GSS data. 5~10% isn't a ton, but is that a normal difference in general population?

Race

• ACX 2025: White 86.8%, Other 9.2%, Asian (East Asian) 3.4%, Black 0.6%

• GSS 2024: White 70.3%, Black 17.6%, Other 12.1%

  

The GSS data just has White, Black, and Other, so I mapped the ACX Asian responses to  Other. I'll flag that GSS just asks White, while ACX asks White (Non-Hispanic). Turns out, that makes it really easy to see what's demographically going on here.

Relationship status

• ACX: 44% married, 33% single, 24% in a relationship.
• GSS 2024: 41% married, 32% single (never-married), 16% divorced, 8% widowed, 3% separated.

Okay, I notice I'm confused.

We don't have Divorced, Separated, or Widowed in the ACX dataset. Fair enough. But unlike with race where I can make a pretty confident guess where a change went, I would be surprised if people who would say "relationship" on the ACX survey would say "Divorced" on the GSS survey. What gives?

Obviously the ACX people who got divorced now answer Single or Relationship or (re)Married. But it's surprising that this would result in almost exactly the same rates of saying Single or Married?

Overall I'm still confused here, but it looks like either I'm making a data processing error or ACX readers have basically the same luck in love as general population.

Work_status

• ACX: 62.7% employed, 9.3% self-employed, 10.6% students, 5.3% unemployed, 4.3% retired.
• GSS 2024: 55.8% employed, 24.1% retired, 8.9% homemaker, 5.1% unemployed, 2.7% students. 

We have more students and fewer retirees. Seems to basically follow from our age range.

Degree

• ACX: 10.2% high school or less, 2.1% associate, 37.1% bachelor’s, 50.6% graduate/professional
• GSS 2024: 54.8% high school or less, 9.0% associate, 21.8% bachelor’s, 14.3% graduate/professional

Okay, so GSS doesn't pick up Masters specifically. Once we've accepted that, this looks pretty sensible. We almost always have at least high school, and when we don't it's because we're still in college or we're too busy founding the field of AI Safety research. Me, I wanted to be on a cool sports team so I went to college^[1] but different strokes for different folks.

Income
 

Fun fact, one of the GSS fields for income topped out at 25k because it was designed for the 1970s. I'm a little suspicious of the GSS field I'm using saying there's nobody over 200k. Am I overestimating the prevalence of doctors and programmers in the general populati- oh, yeah, probably.

Political spectrum

Some compression had to happen since ACX uses 1-10 ranges and GSS uses 1-7.

After messing with it a bit, I took an LLM suggestion to do the statistically unsound thing. It said:

I reworked the ideology mapping with a percentile-based transformation: first compute each ACX respondent’s percentile rank within the ACX 1–10 scale, then project that percentile onto the GSS 1–7 range using 1 + 6·percentile, finally rounding and clipping to [1, 7]. This way the median ACX score now lands near 4 (the moderate midpoint), while the tails still align with the GSS endpoints.

Yeah that looks like a chart that's had terrible LLM clipping things happen to it.

Political affiliation

• ACX 2025: 65 % Democratic, 10 % Republican, 3 % Independent, 22 % in Other (which includes libertarians, socialists, neoreactionaries, etc.)
• GSS 2024: 28 % Democrat, 26 % Republican, 44 % Independent, 3 % Other.

Wait, 44% of the GSS answers are Independent? I think this is mostly down to having less formal party affiliations. If not, I'm confused.

Current Religion

• ACX 2025: 77 % identify as nonreligious (atheist/agnostic); ~20 % describe themselves as some variety of theist; ~3 % fall into other religious labels.
• GSS 2024: about 73 % theist (Protestant/Catholic/other faiths) versus 27 % reporting no religion.

Yep, that seems about as expected.

Religious background

• ACX 2025: Christian (39% Protestant, 27% Catholic), 33% report non-Christian or no religious background (11% Jewish, 23% none/other).
• GSS 2024: Christian (48% Protestant, 31% Catholic.) (1% Jewish, 20% none/other).

Note that the age sixteen part is less precise for the ACX group.

AI worry

• ACX 2025: 13 % “not at all worried,” 22 % “not very,” 22 % neutral, 26 % “somewhat,” and 17 % “extremely” worried.
• GSS 2024 (split-ballot, n≈1.5k responses): 40 % “somewhat worried” and 28 % “extremely” worried, while just 15 % fall into the low-worry buckets.

I'm sorry, am I reading that right? The ACX community is less concerned about AI risk than general population? 

That's because the GSS question is "Worried that AI will take over many jobs done by humans" while the ACX question was "How would you describe your opinion on existential risk from AI?" Now I'm not confused anymore.

Data Mapping

• Limited GSS analysis dataset to survey year `2024` (`year >= 2024`) to match the ACX 2025 survey (which was run in very early 2025.)
• Easy column mappings
  • `Age` (ACX) ↔ `age` (GSS)
  • `Sex` (ACX) ↔ `sex` (GSS) — map `1 → Male`, `2 → Female` after importing GSS codes.
  • `Race` (ACX) ↔ `race` (GSS) — harmonise categorical labels as needed.
  • `Country` (ACX) ↔ `country` (GSS).
  • `Children` (ACX) ↔ `children` (GSS) — both count number of children ever born.
  • `Degree` (ACX) ↔ `degree` (GSS) — convert GSS ordinal codes to text labels.
  • `Income` (ACX) ↔ `income` (GSS) — convert GSS income codes (1–23) to bracket labels; optionally derive continuous values from `realinc`.
• Fuzzy but reasonable column mappings
  • RelationshipStatus` ↔ `marital`
  • `WorkStatus` ↔ `wrkstat`
  • `ChildrenWanted` ↔ `chldnum` (ideal number of children)
  • `EducationComplete` ↔ `educ`
  • `PoliticalSpectrum` ↔ `polviews`
  • `PoliticalInterest` ↔ `polintrst`
  • `PoliticalAffiliation` ↔ `partyid`
  • `ReligiousViews` ↔ `relig`
  • `ReligiousDenomination` ↔ `denom`
  • `ReligiousBackground` ↔ `relig16`
• Tricky and untrustworthy mappings
  • AIxrisk` (ACX) ↔ `aiworry` (GSS)
    • Convert ACX responses to numeric 1–5 and treat `#NULL!` as missing.
    • GSS `aiworry` already uses 1–5 coding; apply shared label map `{1: "Not at all worried", 2: "Slightly worried", 3: "Moderately worried", 4: "Very worried", 5: "Extremely worried"}` for reporting consistency.
  • Geography
    • Keep `Country` (ACX) ↔ `country` (GSS) for geographic cross-checks.

Everything else got dropped.

Reminder, this analysis relied on some heavy vibecoding. Adjust your confidence accordingly.

 

1. ^{^}

   and joined Quidditch, obviously