So it looks like CFAR and the Guild both increase comfort in these skills. There’s two giant reasons not to trust this. First, this is self reported comfort levels, aka we’re basically measuring a vibe. Second, my sample size of CFAR goers and Guild of the Rose identifiers is like, a dozen people in the Yes category.
Zeroth, did they increase comfort or select for those already comfortable?
729.1 + 3546.2 (90, 200, 400) [n=241]
About the Tokyo estimates: I assume that the (90, 200, 400) need to be corrected. But I misunderstand something.
A relatively easy solution (which would, unfortunately, mess with your formatting; not sure if there's a better one that doesn't do that) might be to convert everything you don't want in there to paragraph style instead of heading 1/2/3
Wouldn't that get rid of all of the table of contents?
Ideally I'd have a hierarchy of headings. I think what's happening is it picks up some (but not all) lines that are entirely bold, and treats those as a sort of Heading 4.
Sorry, I meant to change only the headings you didn't want (but that won't work for text that's already paragraph-style, so I suppose that wouldn't fix the bold issue in any case; I apologize for mixing things up!).
Testing it out in a draft, it seems like having paragraph breaks before and after a single line of bold text might be what triggers index inclusion? In which case you can likely remove the offending entries by replacing the preceding or subsequent paragraph break with a shift-enter (still hacky, but at least addressing the right problem this time XD).
Thanks to everyone who took the Unofficial 2024 LessWrong Survey. For the results, check out the data below.
The Data
0. Population
There were two hundred and seventy nine respondents over thirty three days. Previous surveys have been run over the last decade and a half.
2009: 166
2011: 1090
2012: 1195
2013: 1636
2014: 1503
2016: 3083
2017: "About 300"
2020: 61
2022: 186
2023: 558
2024: 279
That’s an annoying drop. I put a bit less oomph into spreading the good word of the census this year largely due to a packed December, but I’d hoped to get above a thousand anyway. Hope is a strategy, but it’s not a very good strategy.
Out of curiosity, I checked the Slate Star Codex/ Astral Codex Ten survey numbers.
2014: 649
2017: 5500
2018: 8077
2019: 8171
2020: 8043
2022: 7341
2024: 5982
That comparison makes me think about the ways in which LessWrong and Astral Codex Ten are different. Everyone who subscribes to ACX gets the ACX survey email, and everyone who subscribes to ACX is at least aware of who Scott Alexander is. LessWrong is a bit more spread out. I tune out most of the AI posts for instance, and I’m sure there’s someone around who tunes out everything but the AI posts. Those two people see a website with rather different content.
Enough musing, let's see some numbers.
Previous Surveys
Have you taken previous incarnations of the Less Wrong Census/Survey?
Yes: 140, 51.7%
No: 119, 43.9%
Prefer not to answer: 12, 4.4%
A note on formatting: For questions like this where there’s obvious categories, I show the question in bold, the categories, then the total number of people who picked that category, then the percent. For instance, 140 people said yes they had taken previous LW surveys, which is about 51.7% of everyone who answered.
I habitually rounded to the nearest tenth of a percent, and often binned any answer with less than three other people in it to Other.
I. Demographics
Age: 32 + 10.1 (25, 31, 37) [n=274]
This is the other common format for the numbers. Here, I show the mean answer (32), plus the standard deviation (10.1), then the first, second, and third quartiles. (25, 31, and 37.) Lastly, I show the number of people who answered the question. (274.)
Country:
United States of America: 135, 49.3%
United Kingdom: 19, 6.9%
Canada: 18, 6.6%
Germany: 16, 5.8%
France: 12, 4.4%
Australia: 7, 2.6%
Russia: 7, 2.6%
New Zealand: 6, 2.2%
Finland: 4, 1.5%
Netherlands: 4, 1.5%
Poland: 4, 1.5%
Denmark: 3, 1.1%
India: 3, 1.1%
Israel: 3, 1.1%
Italy: 3, 1.1%
Other: 30, 10.9%
I binned everyone who had less than three other people in their country to the Other category.
Race
With what race do you most identify?
White (non-Hispanic): 211, 77.9%
Asian (East Asian): 13, 4.8%
Asian (Indian subcontinent): 10, 3.7%
White (Hispanic): 10, 3.7%
Middle Eastern: 7, 2.6%
Black: 4, 1.5%
Jewish: 3, 1.1%
Other: 13, 4.8%
II. Sex, Gender, and Relationships
Sex
What sex were you assigned at birth?
Male: 251, 91.6%
Female: 23, 8.4%
Gender
With what gender do you primarily identify?
Male (Cisgender): 220, 80.6%
Female (Cisgender): 18, 6.6%
Female (Transgender m->f): 14, 5.1%
Non-binary: 10, 3.7%
Male (Transgender f->m): 2, 0.7%
Other: 9, 3.3%
Sexual Orientation
What is your sexual orientation?
Heterosexual: 188, 69.9%
Bisexual: 48, 17.8%
Asexual: 12, 4.5%
Homosexual: 12, 4.5%
Other: 9, 3.3%
Relationship Style
What is your preferred relationship style?
Prefer monogamous: 151, 55.1%
Uncertain / no preference: 67, 24.5%
Prefer polyamorous: 56, 20.4%
Number of Current Partners
0: 116, 43.0%
1: 139, 51.5%
2: 7, 2.6%
3: 7, 2.6%
4: 1, 0.4%
Relationship Goals
. . . and currently not looking for more relationship partners: 176, 66.2%
. . . and currently looking for more relationship partners: 90, 33.8%
Relationship Status
What is your current relationship status?
Single: 122, 44.9%
Relationship: 77, 28.3%
Married: 73, 26.8%
Living With
Who do you currently live with most of the time?
With partner/spouse: 109, 39.9%
Alone: 63, 23.1%
With roommates: 52, 19.0%
With parents/siblings: 41, 15.0%
Other: 8, 2.9%
Children
How many children do you have? (Do not count sperm-donations)
0: 227, 83.2%
1: 17, 6.2%
2: 21, 7.7%
3: 5, 1.8%
4: 2, 0.7%
5: 1, 0.4%
More Children
Are you planning on having more children? Answer yes if you don't have children but want some, or if you do have children but want more.
No: 107, 39.3%
Yes: 91, 33.5%
Uncertain: 74, 27.2%
III. Work and Education
Work Status
What do you currently do?
For-profit work: 110, 40.0%
Student: 52, 18.9%
Unemployed: 29, 10.5%
Non-profit work: 27, 9.8%
Self-employed: 27, 9.8%
Government work: 9, 3.3%
Academics (on the teaching side): 8, 2.9%
Independently wealthy: 8, 2.9%
Other: 5, 1.8%
Profession
In what field do you currently work or study? If more than one, please choose most important.
Computers (practical: IT, programming, etc): 98, 36.7%
Computers (AI): 41, 15.4%
Finance / Economics: 17, 6.4%
Mathematics: 16, 6.0%
Computers (other academic, computer science): 15, 5.6%
Engineering: 11, 4.1%
Business: 7, 2.6%
Law: 7, 2.6%
Medicine: 7, 2.6%
Art: 5, 1.9%
Philosophy: 4, 1.5%
Physics: 4, 1.5%
Other: 35, 13.1%
We’ve got all kinds here, programmers, AI researchers, and computer scientists. I joke, but it looks like that’s fully half of the responses.
Also, I’ve been reusing these categories from previous years but at some point AI work starts becoming practical computer work. Probably that point has already happened.
What is your highest degree earned?
None: 13, 4.7%
High school: 63, 23.0%
2-year degree: 5, 1.8%
Bachelor's: 107, 39.1%
Master's: 52, 19.0%
PhD: 24, 8.8%
MD/JD/Other professional degree: 10, 3.6%
IV. Politics and Religion
Political
Given that no label can completely describe a person's political views, with which of these labels do you MOST identify?
Liberal: 94, 36.4%
Libertarian: 71, 27.5%
Social democrat: 46, 17.8%
Conservative: 13, 5.0%
Green-party: 12, 4.7%
Socialist: 7, 2.7%
Anarchist: 6, 2.3%
Communist: 3, 1.2%
Reactionary: 3, 1.2%
Syndicalist: 2, 0.8%
Social Darwinist: 1, 0.4%
Write-in: What political label do you most identify with?
This was a write in, and therefore a pain to do any kind of statistics to. Eyeballing it, Classical Liberal and Social Liberal were common answers. The Georgists lost ground this year from 4 to 2, but the Pirate party stayed strong at 2. Gray Tribe and dath ilani answers are appreciated - truly, my people - but the one that made me laugh was “specifically me-led monarchist.” I admire the honesty!
Religious Views
How would you describe your religious views?
Atheist and not spiritual: 177, 64.4%
Atheist but spiritual: 37, 13.5%
Agnostic: 30, 10.9%
Lukewarm theist: 14, 5.1%
Committed theist: 11, 4.0%
Deist/Pantheist/etc.: 6, 2.2%
Religious Denomination
If you are religious, which best describes the religion you practice? If atheist, please skip this question.
Not religious, but I wanted to click a button: 83, 65.4%
[Yep, that continues to be the most popular option. Let's remove those answers and see what happens.]
Christian (Protestant): 10, 22.7%
Jewish: 7, 15.9%
Buddhist: 6, 13.6%
Christian (Other non-Protestant, eg Eastern Orthodox): 5:=, 11.4%
Christian (Catholic): 3, 6.8%
Mixed: 3, 6.8%
Unitarian Universalism or similar: 3, 6.8%
Other: 7, 5.5%
We did get one answer of “AI” here, so I would like to take this moment to point out that even if worshipping Cthulhu means he eats you first, it is still not advisable to call up that which you cannot put down.
Family Religion
How would you describe the religious views of your family when you were growing up?
Lukewarm theist: 90, 33.0%
Committed theist: 81, 29.7%
Atheist and not spiritual: 49, 17.9%
Mixed: 18, 6.6%
Agnostic: 16, 5.9%
Atheist but spiritual: 13, 4.8%
Other: 6, 2.2%
We got one person who put “Communist.” I think historical communists were fairly anti-religion (understatement) but I can appreciate the similarities.
Religious Background
What is your family's religious background, as of the last time your family practiced a religion?
Christian (Protestant): 97, 36.1%
Christian (Catholic): 77, 28.6%
Jewish: 25, 9.3%
Not religious in living memory: 18, 6.7%
Christian (Other non-Protestant, eg Eastern Orthodox): 16, 5.9%
Mixed: 15, 5.6%
Hindu: 8, 3.0%
Muslim: 5, 1.9%
Christian (Mormon): 3, 1.1%
Other: 5, 1.9%
Moral Views
With which of these moral philosophies do you MOST identify?
Accept / lean toward consequentialism: 165, 61.8%
Other / no answer: 42, 15.7%
Accept / lean toward virtue ethics: 39, 14.6%
Accept / lean toward deontology: 21, 7.9%
V. Numbers Which Attempt To Measure Intellect
IQ: 137.4 + 11.4 (130, 135, 141.75) [n=66]
IQ Age: 18.1 + 7.7 (11.5, 18, 23.5) [n=67]
SAT scores out of 1600: 1493.4 + 89.7 (1440, 1520, 1560) [n=73]
SAT scores out of 2400: 2257.6 + 107.3 (2240, 2285, 2325) [n=34]
ACT score out of 36: 33.4 + 2.6 (32, 34, 35) [n=41]
[American students take the SATs before going to college. From 2005 to 2016, the test was scored out of 2400, in other years it was out of 1600. IQ Age was the age at which the IQ test was taken.]
Bonus: IQ Exchange
How much would you pay in USD for an extra five points of IQ? (Assume it's not purely an artifact of the test, but genuinely lifts your overall intellect.)
82542 + 167514.3 (5000, 20000, 100000) [n=186]
Any time the standard deviation is bigger than the mean, my bucket finger gets itchy.
>0 to <=10: 2, 1.0%
>10 to <=100: 3, 1.6%
>100 to <=1k: 18, 9.4%
>1k to <=10k: 57, 29.7%
>10k to <=100k: 76, 39.6%
>100k to <=1M: 31, 16.1%
>1M to <=10M: 0, 0.0%
Nobody gave a negative number, though a few people gave answers that were not numbers and I deleted those. I’m going to come back to this later.
VI. LessWrong, the basics
Referrals
How did you find out about Less Wrong?
Referred by Harry Potter and the Methods of Rationality: 57, 21.9%
Referred by a link on Slate Star Codex / Astral Codex Ten: 45, 17.3%
Referred by a friend: 33, 12.7%
Referred by a link on another blog: 27, 10.4%
Referred by a search engine: 22, 8.5%
Been here since it was started in the Overcoming Bias days: 20, 7.7%
Referred by an Effective Altruism site: 13, 5.0%
Other: 43, 16.5%
HPMOR continues to be the largest source of referrals, and it’s been ten years since the last update. Last year I checked and that was still true if I only paid attention to people who had been in the community for five years. How about if I only pay attention to people who have been around for a year or less?
Referred by a link on Slate Star Codex / Astral Codex Ten: 9, 21.4%
Referred by Harry Potter and the Methods of Rationality: 8, 19.0%
Referred by a link on another blog: 7, 16.7%
Referred by a search engine: 3, 7.1%
Referred by an Effective Altruism site: 3, 7.1%
Referred by a friend: 2, 4.8%
Other: 10 23.8%
19% of the people who joined the community in the last year were referred via HPMOR. That’s some seriously impressive outreach. To the newcomers: Welcome! I also got here via HPMOR, so you’re in good company.
And hey, if you’d like more company and weren’t aware, there’s HPMOR Anniversary Parties in the works. Check it out, or if you’re reading this after March 16th or so you can probably also find HPMOR fans at your local meetup.
Blog Referrals
If you were referred here by another blog (other than Overcoming Bias or SSC/ACX) or website, please name the site. Otherwise, please leave this blank.
Common Sense Atheism: 2, 11.1%
Gwern.net: 2, 11.1%
Hacker News: 2, 11.1%
Wait But Why: 2, 11.1%
A Quora post: 1, 5.6%
Don't remember, but it wasn't either of those: 1, 5.6%
https://habr.com AND multiple prior references: 1, 5.6%
LessOnline: 1, 5.6%
Marginal Revolution: 1, 5.6%
Marginal Revolution, EA blogs: 1, 5.6%
Rationally Speaking: 1, 5.6%
reddit: 1, 5.6%
Ted talk of Yudkowsky: 1, 5.6%
The sequences epub: 1, 5.6%
Less Wrong Use
How do you use Less Wrong?
I lurk, but never registered an account: 45, 16.5%
I've registered an account, but never posted: 45, 16.5%
I've posted a comment, but never my own post: 72, 26.4%
I've made my own post, but never gotten a front page tag: 37, 13.6%
At least one post I made got a front page tag: 74, 27.1%
I defy the data here, 27% of LessWrong users have not wound up with a post on the frontpage. That's an obvious selection effect from the kind of person who takes the survey.
Sequences
About how much of the Sequences have you read?
All or nearly all of the Sequences: 111, 41.1%
About 75% of the Sequences: 38, 14.1%
About 50% of the Sequences: 25, 9.3%
About 25% of the Sequences: 14, 5.2%
Some, but less than 25%: 43, 15.9%
All or nearly all of the highlights: 13, 4.8%
Know they existed, but never looked at them: 18, 6.7%
Never even knew they existed until this moment: 8, 3.0%
Time in Community
How long, in years, have you been in the Overcoming Bias/Less Wrong community?
6.6 + 5.2 (2.46, 5, 10) [n=255]
I was curious, so I broke this out into sections.
>=0 to <=0: 3, 1.2%
>0 to <=1: 23, 8.9%
>1 to <=5: 116, 45.1%
>5 to <=10: 59, 23.0%
>10 to <=15: 39, 15.2%
>15 to <=100: 17, 6.6%
[I did a tiny bit of data cleaning here. One person said they’d been in the community for 1.5 hours, which I went ahead and converted to 0.00017 years. To the people who have been here for exactly zero years, welcome, I’m glad to have you!]
Time on LW
How long, in number of minutes, do you spend on Less Wrong in the average day?
20.1 + 31.4 (5, 10, 20) [n=247]
>=0 to <=0: 27, 10.9%
>0 to <=5: 53, 21.5%
>5 to <=15: 74, 30.0%
>15 to <=30: 63, 25.5%
>30 to <=60: 20, 8.1%
>60 to <=120: 6, 2.4%
>120 to <=480: 4, 1.6%
Apropos of nothing, did you know the LessWrong Core Team is four people?
LW Karma
If you have no account, your karma score is 0.
1161.3 + 3520.3 (0, 89, 500) [n=207]
>=0 to <=0: 59, 28.5%
>0 to <=100: 55, 26.6%
>100 to <=1k: 61, 29.5%
>1k to <=10k: 27, 13.0%
>10k to <=100k: 5, 2.4%
>100k to <=1M: 0, 0.0%
LessWrong Subscription
Imagine LessWrong was offered as a monthly subscription, like a newspaper or a software fee. How much in USD would that subscription be worth to you?
80.2 + 719.8 (1.995, 5, 20) [n=195]
If you ever see a standard deviation that’s about ten times your mean, get a bucket.
>-10 to <0: 2, 1.0%
>=0 to <=0: 31, 15.9%
>0 to <=10: 93, 47.7%
>10 to <=100: 61, 31.3%
>100 to <=1k: 7, 3.6%
>1k to <=10k: 1, 0.5%
>10k to <=100k: 0, 0.0%
We’ll come back to this, but for the moment, if you’re thinking “LessWrong is really worth a lot to me” consider that the LessWrong team would appreciate your donations.
(In the interests of transparency, Lightcone (the organization that runs LessWrong) has in the past paid me to do things like work on big events they run, and they also help out ACX meetups in various ways.)
VII. LessWrong, the community
Adjacent Community Identity
Do you consider yourself a member of the following communities?
[Effective Altruism]
Yes: 69, 26.0%
Sorta: 81, 30.6%
No: 115, 43.4%
[Slate Star Codex / Astral Codex Ten]
Yes: 87, 33.1%
Sorta: 100, 38.0%
No: 76, 28.9%
[Post-Rationality]
Yes: 8, 3.1%
Sorta: 23, 8.9%
No: 226, 87.9%
[TPOT / Vibecamp]
Yes: 10, 3.9%
Sorta: 34, 13.2%
No: 214, 82.9%
[Guild of the Rose]
Yes: 9, 3.5%
Sorta: 20, 7.8%
No: 228, 88.7%
[Glowfic]
Yes: 10, 3.9%
Sorta: 32, 12.4%
No: 216, 83.7%
[Forecasting]
Yes: 28, 10.7%
Sorta: 72, 27.5%
No: 162, 61.8%
[AI Alignment]
Yes: 62, 23.8%
Sorta: 75, 28.8%
No: 123, 47.3%
[R/Rational]
Yes: 20, 7.7%
Sorta: 55, 21.2%
No: 184, 71.0%
[The Bayesian Conspiracy]
Yes: 35, 13.6%
Sorta: 34, 13.2%
No: 189, 73.3%
[LessWrong]
Yes: 120, 45.5%
Sorta: 112, 42.4%
No: 32, 12.1%
Adjacent Community Other
If there's another community you think is obviously adjacent which you belong to, feel free to list it below.
This was a write in, I don’t think any community got more than two submissions.
Attending Meetups
Do you attend Less Wrong meetups?
No: 175, 65.1%
Yes, once or a few times: 50, 18.6%
Yes, regularly: 44, 16.4%
Have you ever attended a rationalist-style Secular Solstice?
No: 207, 77.5%
Yes: 60, 22.5%
CFAR Workshops
Have you ever attended a CFAR style workshop?
No: 245, 91.1%
I have been to at least one CFAR class, but not a full (3+ day) workshop: 6, 2.2%
Yes, I have been to a full (3+ day) workshop: 18, 6.7%
Organizing Meetups
Do you organize Less Wrong meetups?
No: 235, 87.7%
Yes, once or a few times: 20, 7.5%
Yes, regularly: 13, 4.9%
(If you organize ACX meetups or you want to, and you’re reading this before March 23rd, consider volunteering for ACX Everywhere. Ahem. Look, when the ACX Meetups Czar runs your census and nobody asks him not to you get a little bit of meetup encouragement.)
Jobs
Have you ever gotten a job or found an employee through the Less Wrong community?
No: 217, 81.0%
Maybe, it was one of several factors or I'm not sure: 17, 6.3%
Yes, once: 22, 8.2%
Yes, repeatedly: 12, 4.5%
Romance
Have you ever been in a romantic relationship with someone you met through the Less Wrong community?
No: 231, 86.5%
I didn't meet them through the community, but they're part of the community now: 12, 4.5%
Yes: 24, 9.0%
VIII. Probabilities
Instructions
Do you understand the instructions above?
Yes, I understand I should answer in percentages between 0 and 100: 257, 98.5%
No, I don't read instructions and am going to make the data look weird unless someone does the work of fixing my answers: 4, 1.5%
Test Question To See If You're Paying Attention
Please give the obvious answer to this question, so I can automatically throw away all surveys that don't follow the rules: What is the probability of a fair coin coming up heads?
50: 255, 97.7%
0.5: 4, 1.5%
49.99: 1, 0.4%
49.9999999999999: 1, 0.4%
I can’t tell whether to be annoyed or appreciative. On the one hand, c’mon. On the other hand, you can’t fault them for honesty.
Wait, no, I can fault them. Someone said they understood, and then immediately screwed up and used 50% instead of 50. That person I do fault. And are those 49.99s imagining a fair coin lands on the edge sometimes or what?
Reminder: the result format used is "Question: Mean + standard deviation (1st quartile, 2nd quartile, 3rd quartile) [n= number responding]"
P(Many Worlds): 48.5 + 32.4 (20, 50, 75) [n=218]
P(Aliens): 67.8 + 33.8 (50, 80, 99) [n=241]
P(Aliens) 2: 32.4 + 34.5 (2, 20, 60) [n=238]
P(Trump Tariffs): 36.4 + 20.6 (20, 30, 50) [n=102]
P(Supernatural): 11.6 + 25.8 (0, 0.1, 5) [n=238]
P(God): 14.2 + 26.5 (0, 1, 20) [n=236]
P(Religion): 8.4 + 23.3 (0, 0.0001, 1) [n=241]
P(Cryonics): 21.7 + 27.1 (3, 10, 30) [n=241]
P(Trump Tariffs and DOGE Cuts): 25.9 + 18.9 (11, 20, 33) [n=121]
P(Anti-Agathics): 37.1 + 35.3 (4, 25, 70) [n=242]
P(Flat Earth): 0.1 + 0.4 (0, 0, 0.0000000775) [n=246]
P(Simulation): 25.8 + 29.4 (2, 15, 40) [n=219]
P(Warming): 88.5 + 20.7 (90, 95, 99) [n=240]
P(Global catastrophic risk, fixed): 38.7 + 31.1 (10, 30, 65.25) [n=236]
P(DOGE CUTS): 23.1 + 18.9 (10, 20, 30) [n=101]
I had a bit of fun with this year’s probability questions.
First off, Flat Earth is this year’s Lizardman Constant trap. This was a restrained year for lizards, with the highest probability being 4.2%, and only four respondents >1% (my threshold for ‘okay, you’re not keeping a healthy skepticism, you’re just trolling or wrong.’)
Second was the conjunction fallacy test. I split people up randomly (by whether their birth date was even or odd- there are statistical differences between people born earlier in the year vs later in the year, there are differences in what letters people’s names start with, but I’ve never heard of a reason people born on the 1st or 3rd of a month should differ from the ones born on the 2nd or 4th) and asked them different questions.
Half got these two as separate questions:
P(Trump Tariffs): 36.4 + 20.6 (20, 30, 50) [n=102]
P(DOGE CUTS): 23.1 + 18.9 (10, 20, 30) [n=101]
Half got this as one question:
P(Trump Tariffs and DOGE Cuts): 25.9 + 18.9 (11, 20, 33) [n=121]
Here’s how the Conjunction Fallacy works: if I tell you that Linda cares a lot about racial equality and was involved in anti-nuclear demonstrations, then ask you if she’s a bank teller, many people will give low odds. If instead I ask you if she’s a bank teller and also a feminist, many people will give higher odds – not necessarily high, but higher than the first group. This is wrong, since P(A) strictly cannot be higher than P(A & B). If B was “the sun will rise tomorrow” then P(A&B) should basically be equal to P(A) unless your AI timelines are very short indeed.
It’s hard to prove any one person is making this mistake, since if you ask them both questions they tend to figure out you’re up to something, so you want to split people up. It’s one of those straightforward mistakes brains make.
I’m going to come back to this, but basically the more ‘rational’ people are, the lower we expect the combined question to be relative to the separate questions.
IX. Traditional LessWrong Census Questions
Singularity
By what year do you think the Singularity will occur?
2676.9 + 7054.6 (2032, 2040, 2060) [n=201]
[Ah, outliers, my old friends. I arbitrarily removed everyone with an answer that was in the past or higher than 3000 and got the following:]
2062.4 + 67.6 (2032, 2040, 2060) [n=193]
[You should be aware that there was a cluster at 2500 I saw before dropping outliers, who got included. This kind of eyeballing method of checking outliers is statistically sketchy and I could have nudged the numbers around a bit, though less than previous years.]
Cryonics Status
Are you signed up for cryonics?
No - still considering it: 95, 36.3%
No - and do not want to sign up for cryonics: 78, 29.8%
No - would like to sign up but haven't gotten around to it: 51, 19.5%
No - would like to sign up but unavailable in my area: 15, 5.7%
Yes - signed up or just finishing up paperwork: 13, 5.0%
Never thought about it / don't understand: 10, 3.8%
Calibration IQ
What do you think is the probability that the IQ you gave earlier in the survey is greater than the IQ of over 50% of survey respondents?
54.2 + 25.8 (40, 57, 73.75) [n=62]
Calibration Question
Without checking a source, please give your best guess for the number of interconnected rail stations in the Tokyo Metropolis (according to Wikipedia's "Transport in Greater Tokyo" article):
103,306,511.2 + 1,607,060,819.5 (92.5, 200, 400) [n=242]
[Wait, what? Okay, I looked at the data and one person said 25000000000. Removing that, we get:]
729.1 + 3546.2 (90, 200, 400) [n=241]
The correct answer per wikipedia is 882. I was worried this wasn’t as good a question as we usually use; too much reasonable uncertainty in what counts as “greater Tokyo” or what counts as an “interconnected rail station.” We wound up vaguely close (a mean of 729 is off, but it’s the right order of magnitude) but zero people were within 10%, so in hindsight I should have been more worried and used something else for calibration. Oops.
Calibration Answer
Without checking a source, estimate the probability that the answer you just gave is within 10% of the true value. Same rules for percents apply as above.
19.7 + 16.3 (10, 15, 26.25) [n=240]
The calibration questions haven’t been a big part of the survey over the years, but they are a consistent part. Good calibration is another spot where it’s hard to spot if someone’s making big mistakes from one question, but where given a lot of questions you’d expect to be able to sort the rational from the irrational. I’m going to come back to this later.
X. LessWrong Team Questions
LLM Frequency
About how often do you use LLMs like ChatGPT while active? If you only use them for work, answer based on your workdays.
Daily: 123, 46.9%
Weekly: 64, 24.4%
Hourly or more: 31, 11.8%
I've tried them once or twice, but not regularly: 21, 8.0%
Maybe a once or twice a month?: 15, 5.7%
Never used one directly: 8, 3.1%
LLM Usecases
What do you use LLMs for, if you use them?
Editing my writing: 91, 38.4%
Generating new writing that I edit: 77, 32.5%
Writing or debugging code: 149, 62.9%
Giving me angles on research: 108, 45.6%
Rubber ducking: 111, 46.8%
Translation between languages: 74, 31.2%
Count: 5, 2.1%
Remember, this was a checkbox answer so people could say they used it for editing their writing and for writing or debugging code. There’s a lot of Other answers. Everything shown except Counting was one of the listed answers, which explains the big jump. Google Sheets formatted this in an annoying way, and I didn’t want to read through the Other answers to clean them up and group them.
. . . So I made an LLM do it.
I asked it to break down “Other” a bit more.
Emotional support / therapy
(e.g. “Therapy,” “emotional processing,” “talking through problems”)
Brainstorming / idea generation
(casual or creative idea‑bouncing, Q&A beyond “research angles”)
Data extraction / summarizing
(OCR, content moderation, doc simplification, info extraction, etc.)
Fun / entertainment / roleplay
(poems, music, “just messing around,” “testing capabilities”)
Recipes / cooking suggestions
(meal planning, recipe ideas)
LLM Brainpower:
Imagine it was conclusively proved that regular LLM use made your mind less sharp. How many points of IQ loss would be 'worth' using LLMs for the next year? E.g., 0.5 if half a point of IQ, 5 if five points of IQ.
1.5 + 3.5 (0.1, 0.65, 2) [n=212]
Bring out the buckets.
>-40 to <0: 2, 0.9%
>=0 to <=0: 47, 22.2%
>0 to <=1: 91, 42.9%
>1 to <=3: 38, 17.9%
>3 to <=5: 20, 9.4%
>5 to <=10: 14, 6.6%
>10 to <=15: 0, 0.0%
There’s one negative thirty-five as an outlier. The only other negative number was 0.3. I removed the -35 and got the following:
1.7 + 2.4 (0.1, 0.8, 2) [n=211]
This was the question I didn’t get from the LessWrong team. It fit in with the other LLM questions, and I wanted it for a fishing expedition I was going on.
Best LW Thing
What's the best thing about LessWrong for you?
Worst LW Thing
What's the worst thing about LessWrong for you?
These were write-ins, and as such hard to do any statistics to. However, you know what I can do? That’s right, I can feed them to an LLM, tell it to do ‘some statistics to this’ and then cut and paste some of what it spits back.
Start of LLM content:
Best Aspects
Worst Aspects
Other complaints include brand/reputation problems, Bay Area–centric perspectives, and minor user interface frustrations. Overall, respondents find the rationalist community and thoughtful content most valuable, while they worry about perceived insularity, AI overload, and accessibility issues.
End LLM content.
LW Recommendation
How likely are you to recommend LessWrong to a friend or a colleague?
6 + 2.9 (3, 7, 8) [n=237]
Holy Buckets Batman!
10: 33, 13.9%
9: 17, 7.2%
8: 37, 15.6%
7: 39, 16.5%
6: 18, 7.6%
5: 15, 6.3%
4: 16, 6.8%
3: 22, 9.3%
2: 25, 10.5%
1: 15, 6.3%
LW Donation:
Have you ever donated to Lightcone? (The organization that runs the LessWrong website as well as other things.)
No: 202, 82.1%
Yes, a little bit: 31, 12.6%
Yes, a significant amount: 13, 5.3%
Hamming Question 1
What is the most important problem in your field?
This is another write-in. There’s an unsurprising amount of AI Alignment answers.
Hamming Question 2
Why aren't you working on it? (If you are working on it, please put "I am".)
The “I am”s have around 45%. That is way higher than I expected.
XI. Adjacent Community Questions
Dojo Organizations (Guild of the Rose question 2)
What organizations are you aware of that are providing some kind of rationality dojo format (courses focused on improving the skill of rationality)?
Guild of the Rose got at least 20%, and CFAR got at least 14%. I say “at least” because splitting out text answers is a pain in the butt and I don’t think the exact percentage changes any of my plans. Someone who wrote in “Guild of the Rose, CFAR” would count for neither of those, because I’m doing primitive text matching.
SPARC came up a few times. Raemon’s Feedback Loop Rationality came up a few times. Clearer Thinking came up a few times. Then there’s a handful of things only one person put down like Manifold, University of Bayes, Lsusr’s Dojo, and Ricki Heicklen’s thing.
Lots of people seemed under the impression CFAR was dead in the water. CFAR, you’re welcome to chime in and assure us all that rumours of your death have been greatly exaggerated.
(I’m not sure how to think about that myself. The last couple years of my own life have involved trying to help the global rationalist community and pushing for better decisionmaking tools, and in the process I notice I spend a lot of time working on plumbing under the surface that people don’t notice. It wouldn’t wildly surprise me if CFAR had a much more advanced case of the same kind of thing going on.)
Optimizing for Fun 1 (Vibecamp Question 1)
Have you tried optimizing for group fun? ["No, never tried", "Theorized about it", "Experimented once or twice", "Yes, in detail"]"
No, never tried: 126, 68.9%
Experimented once or twice: 23, 12.6%
Theorized about it: 21, 11.5%
I wish I’d thought to ask if people had tried optimizing for fun, in full generality, because this answer looks like nonsense to me. I’m assuming that basically everyone likes to have fun, and basically everyone has fun at some point, and basically everyone who has had fun more than twice looks at the kinds of fun they’ve had and goes ‘okay, the second was better than the first, lets do more of that.’
Is it the group part? Do people think ‘optimizing’ implies more rigor? Am I just really high on my empiricist supply and most people taking the LessWrong general census do not try stuff like this?
Optimizing for Fun 2 (Vibecamp Question 2)
If the answer to the last question was "Experimented once or twice" or "Yes, in detail" then what did you learn?
I’m delighted by the answers here, and if you are too then you should read the answers in the dataset. Pulling out a few highlights, entirely based on what stood out to me:
(Whoever that last person is, I’d be interested in talking to you or seeing a writeup of your results! Making meetups better is relevant to my interests these days.)
Oh, and a special shoutout to whoever put this:
I’d also be interested in talking to you, because I’m a giant tabletop RPG nerd.
Which rationalist podcast do you listen to the most, if any? (Bayesian Conspiracy Question 1)
Bayesian Conspiracy: 43, 21.0%
Clearer Thinking: 24, 11.7%
Rationally Speaking: 9, 4.4%
The Mindkiller: 5, 2.4%
After that, it trails off into some with one or two responses.
After the standout showing the Bayesian Conspiracy made in last year’s census, I proactively reached out to Eneasz and Steven about what they’d want in a question. If I recall correctly, I went back and forth a little about making this a radio button or a checkbox.
Regarding that podcast: What type of episodes have been most valuable to you? (Bayesian Conspiracy Question 2)
“Deep dives into rationality concepts
Personal growth/productivity discussions
Discussions of specific skills or insights
Current events analysis through a rationalist lens
Guest interviews
Casual conversations about community topics”
XII. Indulging My Curiosity
Appreciating Users
Please name a LessWrong user who is not yourself, and write something you appreciate about them.This question will be separated from the rest of your responses, though still shared anonymously if you were okay with any of your responses being shared.
Some people answered this question, but had put that they were okay being in summaries but not having their row shared. Defaulting to caution, I removed their answers. I’m going to pick two of the remaining to pull out, one by randomness and one by blatant personal bias.
“Raemon is one of my dearest role models. I deeply resonate with his approach to developing applied rationality, and he seems to embody the ideal rationalist as I imagined it when I joined the community three years ago. In September he helped me out process stuckness around work and motivation at Lighthaven, and it was one of the most moving discussions of my life!”
“Screwtape - idk if you know who that is. Runs the census, rationality meetups, etc. Every time we've talked, I've walked away with some new mental tool in my toolbox that I use. Usually it's a small tool, but a tool nonetheless.”
I’ll let you guess which one was random and which was biased. You can view the rest of the appreciations here.
Websites Smarter Than Me
About how valuable do you find a month of LessWrong in terms of a month of LLMs like ChatGPT? E.g. if LessWrong is twice as valuable to you as LLMs put 2, if LessWrong is a quarter as valuable put 0.25.
6840.5 + 81917.7 (0.15, 0.9, 4) [n=149]
This survey has more buckets than a homestuck convention.
>0 to <=0.01: 7, 4.7%
>0.01 to <=0.1: 28, 18.8%
>0.1 to <=0.9: 40, 26.8%
>0.9 to <=1.1: 13, 8.7%
>1.1 to <=10: 42, 28.2%
>10 to <=100: 10, 6.7%
>100 to <=1,000,000: 9, 6.0%
Okay, the 1,000,000 answer is an obvious outlier. I’m going to make an unprincipled decision to toss out every answer over a hundred or under a hundredth, but you should be advised that this is super arbitrary and there wasn’t an obvious cutoff point. (There were no answers under a hundredth, but that’s the symmetrical side of things.)
4.9 + 15.1 (0.115, 0.75, 3) [n=140]
Huh. That’s not what I was expecting. I love LessWrong but ChatGPT is much quicker about answering weird statistics questions at 2am.
Pulling out a few answers which were not numbers and therefore get the census-giver glare:
“100 (but I wouldn;t pa 100x what I'm currently paying for ChatGPT and Claude).”
Sure, and you don’t have to. I am going to take a moment to ask why though.
“wow in retrospect my answers to some of the previous questions were wildly inconsistent. in that case I do not trust myself to answer this one either.”
Apropos of nothing, I’ve been taking a singing class recently and one of the warmups involves doing an evil villain laugh. It’s been good to get the practice in. I’ll see you in a fishing expedition later.
Internalized Lessons
Which of the following techniques or skills do you feel comfortable and confident that you can personally use?
[Calibration]
Very comfortable: 42, 21.4%
Sorta comfortable: 103, 52.6%
Not at all: 51, 26.0%
[Units of Exchange]
Very comfortable: 36, 19.1%
Sorta comfortable: 77, 41.0%
Not at all: 75, 39.9%
[Goal-factoring]
Very comfortable: 26, 13.9%
Sorta comfortable: 76, 40.6%
Not at all: 85, 45.5%
[Noticing you are confused]
Very comfortable: 108, 52.7%
Sorta comfortable: 84, 41.0%
Not at all: 13, 6.3%
[Speaking truth against social pressure]
Very comfortable: 74, 36.3%
Sorta comfortable: 102, 50.0%
Not at all: 28, 13.7%
First off, I defy the data here. I do not believe that 15% more people are comfortable and confident in speaking against social pressure than calibration. Am I typical minding super hard? Maybe.
Noticing when you’re confused and goal-factoring make for a good sanity check. The latter is a specific technique with multiple steps, referenced mainly from CFAR materials. The former is referenced at least obliquely throughout The Sequences, was one of the common write-ins from previous census questions about rationalist techniques, and at least for me it’s one singular mental motion. As predicted, a lot more people are confident and comfortable with Noticing they’re confused than goal-factoring.
Related: Not bad, CFAR managed to take a technique, introduce it to the community, and get something like 13% of people community wide to be comfortable and confident with it! What happens if I remove everyone who hasn’t been to CFAR?
Goal-factoring, Removed everyone who hadn’t been to CFAR
Very comfortable: 7, 36.8%
Sorta comfortable: 8, 42.1%
Not at all: 4, 21.1%
Goal-factoring, Removed everyone had been to CFAR
Very comfortable: 19, 11.2%
Sorta comfortable: 68, 40.2%
Not at all: 82, 48.5%
Going to CFAR looks like it increases comfort and confidence with Goal-factoring. If we accept that goal-factoring is a useful thing, then by self-report CFAR helps. How much do I trust self-assessment? I don’t know how else to test goal-factoring via a census like this, goal-factoring is just a really open ended technique. I’m going to come back to this later for some of the less open ended techniques: I can’t quickly use statistics to tell if you suck at goal-factoring, but theoretically at least I can tell if you suck at calibration >:)
The list of techniques I checked are somewhat drawn from past survey responses, somewhat drawn from my own experiences trying to practice rationality. If you have a strong case to make for what five skills an aspiring rationalist should know, please make it in the comments!
Rationality Quote
What's your favourite rationality quote that's only one or two sentences long?
Where were you when I was looking for Rationality Quotes last fall?
The litanies of Gendlin and Taski came up a lot. You can look at this list of quotes people gave, but I’ll grab a few that I personally liked:
“‘The first rule of infohazard club is to immediately tell everyone else’ ~ Eve, in the garden of Eden”
“The optimal amount of cringe is not zero.”
“But if there’s one thing I’ve learned in life, it’s that the important things are accomplished not by those best suited to do them, or by those who ought to be responsible for doing them, but by whoever actually shows up. “
XKCD Scorecard: Stayed
In your life, when you thought you were not really happy where you were and maybe this is the best you can expect and you'd regret giving it up, how many times did it turn out you should have stayed?:
1.3 + 5.6 (0, 0, 1) [n=105]
There’s one outlier at 57, and everything else is 5 or below. Removing that outlier. . .
0.8 + 1.1 (0, 0, 1) [n=104]
XKCD Scorecard: Left
And how many times did it turn out you should have left sooner?:
9.5 + 50.7 (1, 2, 4) [n=103]
No obvious outliers here. Or, rather, there’s some obviously unusual answers (a 500, a 100) but it’s not as obvious. There’s a few in the mid teens, some tens, a bunch of sixes and sevens, and then a mostly smooth progression downward. I’m going to be super arbitrary and get rid of everything above twenty.
2.8 + 3.3 (0.875, 2, 4) [n=100]
I have a reference to this comic on my office wall, and have ever since reading Heads Or Tails: The Impact Of A Coin Toss On Major Life Decisions by Steven Levitt. It’s a paper that asked participants to make a major or minor life change (like whether to break up with their partner) based on a coin toss, then followed up six months later and found people who had followed the coin to make a change were happier.
Imagine for a moment that the study was true, no statistical trickery involved. Wouldn’t that be the ultimate refutation of the whole rationality project? A coin toss is literally the archetype of a random process. If we can’t make better decisions than random processes, why bother?
My position is that this reflects a cognitive bias. We stay too long in the status quo, and should make changes more often. It’s like a generic version of the Monty Hall problem; once you know where you are, and what the replacement rate is like, often you should make a switch. Or it’s like the planning fallacy, where once you realize that “how long will this task take me to finish?” is a question where you are usually going to be overconfident, and you are almost always well served by nudging that number up. This is a (small, self-reported) further nudge in that direction.
Fight 1:
Who would win in a fight, Eliezer Yudkowsky or Scott Alexander?
Eliezer: 53, 27.7%
Scott: 71, 37.2%
Wait, what's this? It's Robin Hanson with a steel chair!: 67, 35.1%
Look, the questions can’t all be deep.
(I totally claim the majority is wrong though, Scott’s got the killing intent of a bowl of wet grapes.)
XIII. Detailed Miscellaneous Questions
Income
What is your approximate annual income in US dollars?
127,570.1 + 247,962.8 (20,000, 65,000, 140,000) [n=153]
>=0 to <=25,000: 43, 28.1%
>25,000 to <=50,000: 23, 15.0%
>50,000 to <=75,000: 18, 11.8%
>75,000 to <=100,000: 21, 13.7%
>100,000 to <=150,000: 15, 9.8%
>150,000 to <=300,000: 21, 13.7%
>300,000 to <=2,000,000: 12, 7.8%
Charity
How much money, in number of dollars, have you donated to charity over the past year?
7,085 + 21,102.7 (0, 300, 5000) [n=160]
>=0 to <=2,500: 113, 70.6%
>2,500 to <=5,000: 11, 6.9%
>5,000 to <=7,500: 9, 5.6%
>7,500 to <=10,000: 7, 4.4%
>10,000 to <=15,000: 3, 1.9%
>15,000 to <=30,000: 10, 6.3%
>30,000 to <=200,000: 7, 4.4%
Height
What is your height in number of centimeters?
177.2 + 13.4 (173, 179, 183) [n=193]
Blood
Have you ever donated blood?
No: 135, 56.7%
Yes: 103, 43.3%
Primary Language
What is your primary language?
English: 171, 71.0%
Russian: 10, 4.1%
French: 9, 3.7%
German: 7, 2.9%
Finnish: 4, 1.7%
Dutch: 3, 1.2%
Italian: 3, 1.2%
Polish: 3, 1.2%
Spanish: 3, 1.2%
XIV: Bonus Political Questions
Political Interest
How would you describe your level of interest in politics?
3.3 + 1.1 (2, 3, 4) [n=238]
5: 32, 13.4%
4: 79, 33.2%
3: 61, 25.6%
2: 51, 21.4%
1: 15, 6.3%
Voting
Did you vote in your country's last major national election? If you were ineligible to vote for some reason, the answer is No.
Yes: 191, 78.3%
No: 52, 21.3%
My country does not have elections: 1, 0.4%
Political Centralization:
Should power be more centralized, or more dispersed? (1 centralized, 5 dispersed):
3.7 + 0.9 (3, 4, 4) [n=215]
5: 45, 20.9%
4: 86, 40.0%
3: 66, 30.7%
2: 15, 7.0%
1: 3, 1.4%
Politically Unempathetic
Which of these political identities do you find the hardest to empathize with?
Communist: 48, 22.3%
Reactionary: 47, 21.9%
Social Darwinist: 37, 17.2%
Anarchist: 26, 12.1%
Conservative: 22, 10.2%
Green-party: 12, 5.6%
Syndicalist: 10, 4.7%
Liberal: 7, 3.3%
Libertarian: 4, 1.9%
Social democrat: 2, 0.9%
It seems like there would be something fruitful in comparison against the political identification question. Remember, those answers were:
Liberal: 94, 36.4%
Libertarian: 71, 27.5%
Social democrat: 46, 17.8%
Conservative: 13, 5.0%
Green-party: 12, 4.7%
Socialist: 7, 2.7%
Anarchist: 6, 2.3%
Communist: 3, 1.2%
Reactionary: 3, 1.2%
Syndicalist: 2, 0.8%
Social Darwinist: 1, 0.4%
Earmarking this as a place I might follow up on in a future analysis. I’d be interested to check if there’s an obvious pattern in individual rows, e.g. Libertarians have a hard time empathizing with communists and vice versa.
Politics Open Section
In as much detail as you care to give, how would you describe your political views?
This one is a write-in. A few entries I would like to curate for your attention:
“TESCREAL with Reactionary Characteristics. . . ”
“perpetually confused”
“There's one rightful king, and ought to be me. . . “
XV: Wrapup
Census Length
How long was this survey? (1 is I’m happy answering more, 5 is way too long):
3 + 1 (3, 3, 4) [n=248]
5: 12, 4.8%
4: 62, 25.0%
3: 114, 46.0%
2: 38, 15.3%
1: 22, 8.9%
Huh. Movement in the right direction, last year we were at a mean of 3.2. At a glance it looks like the movement was mostly from 4s becoming 3s. On the other hand, I would have expected cutting about a third of the questions to move the needle a bit more.
Have you answered this survey honestly and accurately?
Yes: 258, 100.0%
No: 0, 0.0%
This question continues to make me laugh. Social darwinists and altruists who donate six figures to charity, people who have never used an LLM and people who think the singularity happened last year, people who’ve been here for a couple of decades and people who’ve been here for a couple of hours, we all have this that unites us. Men, women, non-binaries, and people who put “goth” in their gender, every year the census shows that LessWrong can speak in unison, our answers ringing out with one confident voice.
Everyone answers the survey honestly and accurately.
Fishing Expeditions
Meetup Comparisons
How good are they?
I’m the ACX Meetup Czar. What that means is a little broad, but basically it involves trying to keep the Astral Codex Ten in-person community running smoothly. LessWrong is obviously adjacent, and closely intertwined enough that I sometimes have to remind myself the venn diagram is only ~2/3rds overlapped. So, how are meetups going?
Remember, the LessWrong answers were
Attending Meetups
Do you attend Less Wrong meetups?
No: 175, 65.1%
Yes, once or a few times: 50, 18.6%
Yes, regularly: 44, 16.4%
Last year I spent a bunch of time making charts of how things changed over time. I’m not going to do them in general, but meetups are relevant to my interests so sure.
Fewer once-off attendees, more regular attendees or people who have never gone.
Out of curiosity, I grabbed the ACX 2025 survey data. (Run in late December/early January, and so loosely equivalent to LessWrong’s 2024. Though I bet someone’s going to forget that date mismatch in a couple years.)
Meetup (2025)
Have you ever been to an Astral Codex Ten meetup?
No, I don't want to: 1701, 30.7%
No, I want to but I just haven't made it there yet: 1461, 26.3%
No, I don't know of any, or can't make it to any: 1010, 18.2%
Yes, I attended one meetup: 544, 9.8%
Yes, I've attended a few meetups: 585, 10.6%
Yes, I attend meetups regularly: 244, 4.4%
Meetup Approval (2025)
For people who went to ACX meetups: how much did you enjoy it?
5: 312, 22.7%
4: 572, 41.7%
3: 335, 24.4%
2: 119, 8.7%
1: 34, 2.5%
While I’m grabbing survey data, what were the ACX numbers for 2022?
Meetup (2022)
Have you ever been to an Astral Codex Ten meetup?
No, I don't want to: 2221, 32.1%
No, I want to but I just haven't made it there yet: 1884, 27.3%
No, I don't know of any, or can't make it to any: 1337, 19.3%
Yes, I attended one meetup: 630, 9.1%
Yes, I've attended a few meetups: 605, 8.8%
Yes, I attend meetups regularly: 233, 3.4%
Meetup Approval (2022)
For people who went to ACX meetups: how much did you enjoy it?
5: 330, 22.4%
4: 639, 43.4%
3: 359, 24.4%
2: 120, 8.2%
1: 23, 1.6%
Overall that looks like noise, possibly meetups are getting a tiny bit less attended and a tiny bit worse. That's discouraging. I took over as ACX Meetup Czar at the start of 2023, so this is the range where I might have impacted things and these are obvious questions to check if I'm changing anything. Meaningfully changing these numbers is on the one hand ambitious (the in-person community is big and only a small piece of it is gathered in any one place) but I'm an ambitious person.
Looking back, there's not much I did that I would have expected to move the needle on whether someone attends a meetup. ACX Everywhere (I think) the main source of new attendees, but that was going on for years before I started. There's a bunch of things I've done to try and upgrade how much people enjoy ACX meetups though, and none of that's showing up. Worthwhile for me to think about as I'm planning this year I guess.
What’s the community overlap like anyway?
I’m the ACX Meetup Czar. I’m not the LessWrong Meetup Czar. I’m definitely not the EA Meetup Czar. On the other hand, that’s a venn diagram with a lot of overlap. How much overlap?
Here’s a few questions from the LW Census.
Adjacent Community Identity
Do you consider yourself a member of the following communities?
[Effective Altruism]
Yes: 69, 26.0%
Sorta: 81, 30.6%
No: 115, 43.4%
[Slate Star Codex / Astral Codex Ten]
Yes: 87, 33.1%
Sorta: 100, 38.0%
No: 76, 28.9%
[LessWrong]
Yes: 120, 45.5%
Sorta: 112, 42.4%
No: 32, 12.1%
Here’s a few questions from the ACX Survey.
LW ID
Do you identify as a Less Wronger or "rationalist"?
Yes: 700, 12.7%
Sorta: 2396, 43.6%
No: 2397, 43.6%
EA ID
Do you identify as an effective altruist?
Yes: 656, 11.9%
Sorta: 1851, 33.5%
No: 3023, 54.7%
Sadly I can’t find an EA survey that asks a close enough variation of the same question, and the ACX Survey didn’t ask if people identified as a member of the ACX community. I really want the latter. Maybe ACX has more casual readers who keep their identity separate from whatever blogs they read? Reasonable if so, but that’s a wild guess.
I should keep an eye on this.
Skill Issues
What might influence how rational someone is?
I got into the rationality community looking for a martial art of rationality, a way to practice making better decisions. I’m not pulling that out of my own idiosyncrasies, here’s the first text after a Piet Hein poem on the About page for LessWrong:
“LessWrong is an online forum and community dedicated to improving human reasoning and decision-making. We seek to hold true beliefs and to be effective at accomplishing our goals. Each day, we aim to be less wrong about the world than the day before.”
If I may quote from Patio11:
“You know, have a plan, write it down, measure the success of the plan with numbers, cease doing the thing which is ineffective and try other things that would be more effective, etc etc. Play the game like you're playing it to actually win."
Emphasis mine.
What might influence how rational someone is? Well, here’s a few things that might.
Sequences
About how much of the Sequences have you read?
All or nearly all of the Sequences: 111, 41.1%
About 75% of the Sequences: 38, 14.1%
About 50% of the Sequences: 25, 9.3%
About 25% of the Sequences: 14, 5.2%
Some, but less than 25%: 43, 15.9%
All or nearly all of the highlights: 13, 4.8%
Know they existed, but never looked at them: 18, 6.7%
Never even knew they existed until this moment: 8, 3.0%
Time in Community
How long, in years, have you been in the Overcoming Bias/Less Wrong community?
6.6 + 5.2 (2.46, 5, 10) [n=255]
>=0 to <=0: 3, 1.2%
>0 to <=1: 23, 8.9%
>1 to <=5: 116, 45.1%
>5 to <=10: 59, 23.0%
>10 to <=15: 39, 15.2%
>15 to <=100: 17, 6.6%
Do you consider yourself a member of the following communities? [Guild of the Rose]
No: 228, 88.7%
Sorta: 20, 7.8%
Yes: 9, 3.5%"
Attending Meetups
Do you attend Less Wrong meetups?
No: 175, 65.1%
Yes, once or a few times: 50, 18.6%
Yes, regularly: 44, 16.4%
CFAR Workshops
Have you ever attended a CFAR style workshop?
No: 245, 91.1%
Yes, I have been to a full (3+ day) workshop: 18, 6.7%
I have been to at least one CFAR class, but not a full (3+ day) workshop: 6, 2.2%
Organizing Meetups
Do you organize Less Wrong meetups?
No: 235, 87.7%
Yes, once or a few times: 20, 7.5%
Yes, regularly: 13, 4.9%
The sequences were written to create rationalists, so if people read it maybe they get more rational. The LessWrong website is about improving reasoning, so maybe if people spend time on the website they reason better. The Guild of the Rose is a rationalist (adjacent?) organization focused on self improvement, so maybe if people are a member of the Guild they’ll have improved reasoning. Less Wrong community is about improving reasoning, so maybe if people attend meetups they’ll reason better. CFAR is trying to develop clear thinking and run (ran?) workshops, so maybe if people go to workshops they’ll have clearer thinking. And lastly, we might hope that whatever improved reasoning LessWrong is about, the people running the meetups might have it.
(If I could have surveyed CFAR workshop instructors I absolutely would have. That would have probably been deanonymizing though.)
Before you read on, I want you to contemplate the litany of Tarski with me. Today, the litany goes like this:
I’m about to throw a bunch of stuff at the wall and see what sticks. I’m not a trained data scientist, I’m not a beisutsukai. I’m here because I want the art, not because I have it. If this section is of particular interest to you, I suspect you’ll get something extra out of it if you write down your guesses on what each test will find before you read what I found.
If you look at this section and think “this isn’t how you test rationalist skills, obviously none of those variables would influence rationalist skill levels” then I really want you to tell me what tests and what variables you think matter. That’s not rhetoric, I’m trying to figure this out here.
One more note: This is very much a fishing expedition. I didn’t preregister anything, other than having ideas in mind when I added some of the questions. This section is going to be rambly, because I’m going to try and describe what I tried to look for even when it doesn’t turn up anything interesting or it turns out not to work. Ideally, that minimizes some of the Garden of Forking Paths style nonsense I could have gotten up to, but ultimately I got to play around in the data for a couple months before writing this and there’s still going to be stuff I intuited or checked and forgot to write down.
How might you measure how rational someone is?
Self-assessment
How about pure self assessment to start?
CFAR ran workshops. CFAR workshops taught different things at different times, but two of the things in their handbook were Goal Factoring and Units of Exchange. Calibration wasn’t specifically in the handbook, but it’s a pretty obvious skill.
The Guild of the Rose runs workshops. The Guild of the Rose workshops teach different things at different times, but one of the things in their workshops is calibration.
Units of Exchange, only people who hadn’t been to any CFAR
Not at all: 71, 42.0%
Sorta comfortable: 66, 39.1%
Very comfortable: 32, 18.9%
Units of Exchange, only people who had been to a full CFAR workshop
Not at all: 3, 21.4%
Sorta comfortable: 7, 50.0%
Very comfortable: 4, 28.6%
~20% more people who had been to CFAR were at least sorta comfortable with Units of Exchange. About ten percent more were Very Comfortable.
Calibration, only people who hadn’t been to any CFAR
Not at all: 81, 48.2%
Sorta comfortable: 68, 40.5%
Very comfortable: 19, 11.3%
Calibration, only people who had been to a full CFAR workshop
Not at all: 2, 14.3%
Sorta comfortable: 8, 57.1%
Very comfortable: 4, 28.6%
~30% more people who had been to CFAR were at least sorta comfortable with calibration. About ten percent more were Very Comfortable.
Calibration, only people who identify Yes Guild of the Rose
Not at all: 48, 27.1%
Sorta comfortable: 93, 52.5%
Very comfortable: 36, 20.3%
Calibration, only people who identify Yes Guild of the Rose
Not at all: 2, 14.3%
Sorta comfortable: 6, 42.9%
Very comfortable: 6, 42.9%
~10% more people who identify as part of the Guild of the Rose are at least sorta comfortable with calibration. About twenty percent more were Very Comfortable.
So it looks like CFAR and the Guild both increase comfort in these skills. There’s two giant reasons not to trust this. First, this is self reported comfort levels, aka we’re basically measuring a vibe. Second, my sample size of CFAR goers and Guild of the Rose identifiers is like, a dozen people in the Yes category.
Calibration
Self assessment is fuzzy. Calibration exercises are probably the most objective test for any skill in the rationalist canon that we have. If I ask you a hundred True/False questions and ask you how confident you are that they’re True, you say 90% True on all of them, and it turns out half the time the answer is True and half the time the answer is False, you’re badly calibrated.
On the other hand, I don’t have a hundred True/False questions on the survey. What do I have? I have half a dozen calibration questions from last year that have pretty clear answers, and a few hundred people.
The data from this section is taken from the 2023 LessWrong Census.
What do I have to work with?
P(GPT-5 Release)
What is the probability that OpenAI will release GPT-5 before the end of 2024? "Release" means that a random member of the public can use it, possibly paid.
P(Dow Jones)
What is the probability that the Dow Jones Industrial Average price will be higher on July 1st, 2024, than it was on December 1st, 2023? ($36,245.50)
P(LLM Hallucination) or The 1st Manifold Question
Will there exist by May 1, 2024 an LLM, that is at least as otherwise capable as GPT-4, that hallucinates in typical conversations on questions where human experts exist at most (about as) often as human experts hallucinate when asked similar questions? Don't look at Manifold when you're about to answer this question.
P(Hamas Rules Gaza) or The 2nd Manifold Question
Will Hamas still rule Gaza by June 1st, 2024? Hamas must be the ruling party in Gaza, operating freely to supply public services and collect taxes. Don't look at Manifold when you're about to answer this question.
Calibration IQ
What do you think is the probability that the IQ you gave earlier in the survey is greater than the IQ of over 50% of survey respondents? If you left that blank, also leave this blank.
Calibration Question
Without checking a source, please give your best guess for the weight of the Opportunity rover in kg (according to Wikipedia's "Opportunity (Rover)" article)
Calibration Answer
Without checking a source, estimate the probability that the answer you just gave is within 10% of the true value. Same rules for percents apply as above.
I removed anyone who didn’t answer all of the questions. Put another way, these are the numbers for people who answered all the probability questions. Then I computed people’s brier scores- a way to average prediction accuracy. In brier, lower is better.
The average brier score is as follows:
0.2129 + 0.0747 (0.16, 0.2012, 0.2473) [n=80]
Selecting only people who read All or Nearly All of the Sequences:
0.2118 + 0.075 (0.155, 0.2029, 0.2269) [n=35]
Selecting only people who have been in the community >5 years:
0.2017 + 0.0591 (0.1629, 0.1996, 0.2244) [n=39]
Selecting only people who answered Yes to whether they identify as Guild of the Rose:
0.2323 + 0.0583 (0.2042, 0.2283, 0.2832) [n=5]
Selecting only people who attend meetups regularly:
0.2138 + 0.0513 (0.1907, 0.2158, 0.2259) [n=12]
Selecting only people who organize meetups regularly:
0.2192 + 0.0131 (0.2146, 0.225, 0.2267) [n=3]
Is that significant? Someone at Metaculus or Fatebook should maybe publish what “normal” calibration looks like, though then again the kind of person who uses Metaculus isn’t exactly normal. I’m tempted to pay for a Mechanical Turk survey or something for a better control.
The best score was .0876, the worst score was .4431.
The best scorer had read some but less than 25% of the sequences, had been in the community less than two years, wasn’t a guild of the rose member and had been to meetups once or a few times but never organized any. The worst scorer had read all or nearly all the sequences, had been in the community a little longer but not by much, and was otherwise the same on those questions.
If I took these numbers as significant, then I’d suggest that being in the LessWrong community longer is good for your calibration and identifying as a Guild of the Rose member is bad for your calibration. I do not actually trust these numbers to be significant.
Conjunction Fallacy
Remember what I said back in the Probability section?
I split people up randomly (by whether their birth date was even or odd- there are statistical differences between people born earlier in the year vs later in the year, there are differences in what letters people’s names start with, but I’ve never heard of a reason people born on the 1st or 3rd of a month should differ from the ones born on the 2nd or 4th) and asked them different questions.
Half got these two as separate questions:
P(Trump Tariffs): 36.4 + 20.6 (20, 30, 50) [n=102]
P(DOGE CUTS): 23.1 + 18.9 (10, 20, 30) [n=101]
Half got this as one question:
P(Trump Tariffs and DOGE Cuts): 25.9 + 18.9 (11, 20, 33) [n=121]
Here’s how the Conjunction Fallacy works: if I tell you that Linda cares a lot about racial equality and was involved in anti-nuclear demonstrations, then ask you if she’s a bank teller, many people will give low odds. If instead I ask you if she’s a bank teller and also a feminist, many people will give higher odds – not necessarily high, but higher than the first group. This is wrong, since P(A) strictly cannot be higher than P(A & B). If B was “the sun will rise tomorrow” then P(A&B) should basically be equal to P(A) unless your AI timelines are very short indeed.
I don’t need to wait and find out how the tariffs or doge cuts work out. If the half of people who got the separate version gave lower odds than the half of people who got the combined version — that is, if people think DOGE is less likely than Tarriffs & DOGE — then I have good reason to suspect people are being irrational. Any variable that makes this happen less often indicates . . . well, realistically it could be any number of things including selection effects, but it might be because that variable makes people more rational.
The normal numbers:
P(Trump Tariffs): 36.4 + 20.6 (20, 30, 50) [n=102]
P(DOGE CUTS): 23.1 + 18.9 (10, 20, 30) [n=101]
P(Trump Tariffs and DOGE Cuts): 25.9 + 18.9 (11, 20, 33) [n=121]
The combined form is above the DOGE Cuts numbers. I suspect there’s some Conjunction Fallacy going on.
The Read All Or Nearly All of the Sequences numbers:
P(Trump Tariffs):36.04 + 17.54 (25, 33, 50) [n=45]
P(DOGE CUTS):22.66 + 17.71 (10, 20, 30) [n=45]
P(Trump Tariffs and DOGE Cuts): 19.09 + 15.4 (10, 15, 21.25) [n=48]
The combined form is below both independent versions. Tentatively it looks like reading the sequences is correlated with beating the conjunction fallacy.
The Time in Community >5 years numbers:
P(Trump Tariffs): 33 + 17.37 (20, 30, 50) [n=38]
P(DOGE CUTS): 17.06 + 14.67 (8, 10.5, 25) [n=38]
P(Trump Tariffs and DOGE Cuts): 22.22 + 18.23 (10, 19, 27.5) [n=59]
The combined form is above the DOGE Cuts numbers. I suspect there’s some Conjunction Fallacy going on.
Attends meetups regularly numbers:
P(Trump Tariffs): 36.44 + 19.66 (23.75, 31.5, 50) [n=16]
P(DOGE CUTS): 31.4 + 24.57 (15, 30, 45) [n=15]
P(Trump Tariffs and DOGE Cuts): 20.05 + 16.43 (9.25, 14.5, 25) [n=22]
The combined form is below both independent versions. Tentatively, it looks like attending meetups is correlated with beating the conjunction fallacy.
Organizes meetups regularly numbers:
P(Trump Tariffs): 25.75 + 5.38 (23.75, 25, 27) [n=4]
P(DOGE CUTS): 22.5 + 15.2 (15.5, 26.5, 33.5) [n=4]
P(Trump Tariffs and DOGE Cuts): 11.33 + 5.65 (8.25, 10.5, 14.25) [n=6]
The combined form is below both independent versions. Organizing meetups looks correlated with beating the conjunction fallacy, though on top of everything else that might go wrong with this approach my sample sizes are single digit.
Yes identifies as Guild of the Rose numbers:
P(Trump Tariffs): 32 + 16.05 (25, 30, 45) [n=5]
P(DOGE CUTS): 11.6 + 10.69 (2, 10, 20) [n=5]
P(Trump Tariffs and DOGE Cuts): 5.17 + 4.75 (2.75, 5, 7.5) [n=3]
The combined form is below both independent versions. My sample size is again super small, but tentatively it looks like Guild of the Rose identification is correlated with beating the conjunction fallacy.
Attended a full CFAR workshop numbers:
P(Trump Tariffs): 30.83 + 22.23 (13.75, 25, 43.75) [n=6]
P(DOGE CUTS): 13.6 + 5.9 (10, 10, 20) [n=5]
P(Trump Tariffs and DOGE Cuts): 27.3 + 23.8 (12.75, 20, 25) [n=10]
The combined form is above the DOGE Cuts. I suspect there’s Conjunction Fallacy going on, though again my sample sizes are really small.
Based on this house of cards, if you want to beat the conjunction fallacy you should read the sequences. Which doesn’t totally surprise me; that’s the only one of these where I know for sure that the Conjunction Fallacy even comes up, and it comes up by name. This also suggests you should go to rationalist meetups and run them yourself. That is such a self-serving conclusion for the ACX Meetup Czar to reach, a conclusion with such a tentative causal mechanism, that I’m frankly suspicious of myself already and someone should double check that.
Someone mentioned in the comments of the survey that they noticed the ROT13 was repeating chunks. That alone might have been enough to make someone wary and pay more attention to what I was up to. Wanting to run this test on people in a sneakier way was almost enough to make me use Airtable instead of Google Forms, so if I decide this kind of split test is a good line to pursue I might want to look into that again. Alternatively, the second saint of rationalist surveys uses twitter and Guided Track. Maybe I should look into that.
Values and Dutch Booking
Have you ever read The Cambist and Lord Iron? If you haven’t, I recommend it, great story.
A dutch book is a set of bets that ensures a guaranteed loss. You can also express these bets as trades. The way I usually put it is, if you consider trading two bicycles for a laptop fair in both directions, and two laptops for an old car fair in both directions, I can’t be sure you’re being irrational. If you’d then go on to trade two old cars for a bicycle and would do that in either direction, then I can be sure you’re making a mistake, since someone could trade you a bicycle for two old cars, an old car for two laptops, a laptop for two bicycles, and you’d wind up strictly worse off than if you’d never traded. I made a whole game out of the idea, but the basic concept has a lot in common with the CFAR class Units of Exchange.
I like this exercise a lot because it forces you to put numbers on fuzzy intuitions, and then checks how well you’re doing with some not-at-all fuzzy math. It’s the next best thing to calibration practice I’ve found.
I put four questions into the survey that formed a loop.
LessWrong Subscription
Imagine LessWrong was offered as a monthly subscription, like a newspaper or a software fee. How much in USD would that subscription be worth to you?
80.2 + 719.8 (1.995, 5, 20) [n=195]
Bonus: IQ Exchange
How much would you pay in USD for an extra five points of IQ? (Assume it's not purely an artifact of the test, but genuinely lifts your overall intellect.)
82542 + 167514.3 (5000, 20000, 100000) [n=186]
LLM Brainpower:
Imagine it was conclusively proved that regular LLM use made your mind less sharp. How many points of IQ loss would be 'worth' using LLMs for the next year? E.g., 0.5 if half a point of IQ, 5 if five points of IQ.
1.5 + 3.5 (0.1, 0.65, 2) [n=212]
Websites Smarter Than Me
About how valuable do you find a month of LessWrong in terms of a month of LLMs like ChatGPT? E.g. if LessWrong is twice as valuable to you as LLMs put 2, if LessWrong is a quarter as valuable put 0.25.
6840.5 + 81917.7 (0.15, 0.9, 4) [n=149]
LessWrong use -> USD -> IQ -> LLM use -> implied LessWrong use.
There’s a bunch of reasons not to put too much weight on this; marginal value is one (If I have no money, I value my first buck higher than my billionth dollar) and the fact that the questions aren’t worded quite as closely as the bicycles and laptops is another. I’m going to try anyway.
First,I dropped anyone who didn’t answer all four questions with a number. Then I converted the questions units (e.g. LLM use for a month is equal to LLM use for a year times 12.) Then I went through the cycle to get an implied value of LessWrong use for a month. Last I checked the ratio of the implied value vs the stated value from LessWrong Subscription, getting a ratio of 1:X. If X is 1, your ration is 1:1 and your values are consistent. Lower is better.
(Thanks to NoSignalNoNoise for helping me set this up in a spreadsheet to do it in bulk.)
Everyone: 2991.6 + 19737.5 (8, 33.33, 167) [n=73] (Best score was 1.33)
Read all or nearly all of the sequences: 5744.2 + 28024.2 (5.16, 38.25, 375) [n=36]. (Best score was 1.33)
Been in the community >5 years: 5618.7 + 28046.5 (6.5, 20.83, 51) [n=36] (Best score was 1.33)
Everyone who regularly attends meetups: 2000.9 + 7143.8 (25, 41.67, 333) [n=17] (Best score is 4.17)
Regularly organizes meetups: 385.2 + 932.7 (26.79, 41.67, 46) [n=7] (Best score was 9.38)
Everyone who’s been to a full CFAR workshop: 20888 + 58903.6 (13.33, 19.79, 115) [n=8] (Best score was 2.5)
Identifies yes as Guild of the Rose: 19.1 + 22.3 (11.21, 19.09, 27) [n=2] (Best score was 3.33)
Small sample size means I don’t trust my Guild of the Rose, CFAR, or Organizes Meetups numbers. It would be reasonable to point out that all my sample sizes are pretty small. Which is a shame because I was pretty excited to see this in action. Oh well. Still, if you want to extrapolate from this the Guild of the Rose comes out looking really good, as does organizing meetups which I was not expecting. I was guessing CFAR was going to be the winner here.
Next year I might try the same thing, but with less attempt to be sneaky about what I’m up to and a more apples to apples comparison.
Conclusions
About the respondents
According to the census, the median LessWronger is a 32 year old white man from the U.S.A. They’re monogamous, straight, and not looking for a partner. They do some kind of work with computers. They grew up Christian, but these days they’re an atheist, and they’ve been a member of the LessWrong community for about six years. They consider themself at least sorta a member of the Effective Altruism and Astral Codex Ten communities.
That profile doesn’t describe every person. It’s what in the software industry I usually call a Persona, a story about the kind of person who uses your product. You don’t arrange everything to fit the common persona, but you do assume that kind of person is going to show up a lot.
(I checked, we don’t have a perfectly median user this year.)
About the survey
Last year there was an indication that people thought the survey was too long, and this year I shrunk the survey to about two thirds of its previous size. It’s currently about a hundred questions long. I think I could get it down to seventy-five questions and still get the information I want out of it, maybe fifty if I made harder cuts. Worth nudging downward a little, but I don’t think we’re too bad.
We got less responses than I wanted. One answer is to promote it more, but December is generally kind of busy. Another concern is overlap; the first place I go when I want as many people in the rationalist community to see something as I can get is to ask Scott Alexander to put it in ACX. ACX already has a survey though, and the ACX and LessWrong communities are maybe half overlapping with half not overlapping. That generalizes for everywhere I’d advertise it off of LessWrong; one of the reasons I suspect the Guild of the Rose and Bayesian Conspiracy show up in the numbers they do is those places pitch the survey.
Increasingly I think there are shards of the community that descend from but don’t reference LessWrong or Overcoming Bias. How they want to relate to LW and vice versa is going to be different in different places. My position is I’m pretty happy to use the LessWrong Community Census as a collective point for the rationalist diaspora and filter in or out as needed based on the Adjacent Identity questions. I’m also delighted to add questions from adjacent communities. On the other hand, some branches have their own surveys (ACX does, Thing of Things does.)
I feel good about the evaluation questions and the whole Skill Issues fishing expedition. In 2014, Scott Alexander wrote:
In 2022, Ben Pace wrote:
I don’t think I’m the appointed LW Census Czar. I am tentatively planning to run another of these in 2025, mainly checking the usual demographic information but also trying to test more specific hypotheses about improving human rationality. I’ll take discussion of future surveys into the comments, and there might also be a separate post about my hypotheses.
The public data
For the public dataset, I made the following changes:
The public data is accessible here. If you do your own analysis and get different answers than I did, it’s likely either because the respondents who didn’t want to be in the public dataset mattered, or because I made a mistake.