2024 Unofficial LessWrong Survey Results

Issues with the dutch book beyond the marginal value of money:

It's not as clear as it should that the LLM IQ loss question is talking about a permanent loss (I may have read it as temporary when answering)
Although the LLM IQ drop question does say "your IQ" there's an assumption that that sort of thing is a statistical average - and I think the way I use LLMs, for example, is much less likely to drop my IQ than the average person's usage.
I think is that the LessWrong subscription question is implictly asking about the marginal value of LessWrong given the existence of other resources while the relative LessWrong/LLM value question is implicitly leaning more towards non-marginal value obtained, which might be very many times more

impact: these issues increase LLM/IQ and (Lesswrong/LLM relative to LessWrong/$), which cause errors in the same direction in the LLM/IQ/$/Lesswrong/LLM cycle, potentially by a very large multiplier.

Marginal value due to the high IQ gain of 5 lowers $/IQ which increases IQ/$. This also acts in the same direction.

(That's my excuse anyway. I suspected the cycle when answering and was fairly confident, without actually checking, that I was going to be way off from a "consistent" value. I gave my excuse as a comment in the survey itself that I was being hasty, but on reflection I still endorse an "inconsistent" result here, modulo the fact that I likely misread at least one question).

[-]Screwtape7mo40

If I try this again next year I plan to use the exact same text and values on both sides, which hopefully will clear up most of that kind of issue. It doesn't really fix marginal value, but I'm not sure that's fatal to this kind of analysis- I can quote a reasonable price for an apple even though my marginal value of apples drops very fast by the time I hit three digits of apples. I could try and fix this by picking things I think people value vaguely the same but then we miss out on catching scope insensitivity.

[-]SarahNibs7mo74

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?

[-]Screwtape7mo20

Yeah, I should probably add a bit at the start or end of that section that everything in it is potentially selection effect. I don't know how to look at the thing I'm curious about without that.

Thinking out loud: If you get a random selection of people from the Pushup Club and count how many pushups they can do, then do the same for general population, the difference could be selection effect. People who like doing pushups are more likely to go to pushup club in the first place, and more likely to stick with it. But I can't realistically pay a bunch of Mechanical Turkers to hang out on LessWrong for six years and watch what happens. Presumably there's some approach actual scientists have here, but I don't know what it is. Suggestions welcome.

In the mean time I'm going to add a bit towards the start of the section warning of potential selection effects.

[-]Garrett Baker7mo51

One answer is to promote it more, but December is generally kind of busy.

An alternative is to do it in the summer!

[-]jenn7mo41

I put four questions into the survey that formed a loop.

what omg this is the coolest thing ever. kudos!

[-]Screwtape7mo30

Thank you! I felt quite clever setting it up.

[-]lsusr7mo30

"The #1 career of people on this website is Computers."

"What's #2?"

"More computers."

[-]mako yass7mo20

I'd be surprised if LLM use for therapy or sumarization is that low irl, and I'd expect people would've just forgot to mention those usecases. Hope they'll be in the option list this year.

Hmm I wonder if a lot of trends are drastically underestimated because surveyers are getting essentially false statistics from the Other gutter.

[-]Screwtape7mo20

If Other is larger than I expect, I think of that as a reason to try and figure out what the parts of Other are. Amusingly enough for the question, I'm optimistic about solving this by letting people do more free response and having an LLM sift through the responses.

[-]mako yass7mo20

Contemplating an argument that free response rarely gets more accurate results for questions like this because listing the most common answers as checkboxes helps respondents to remember all of the answers that're true for of them.

[-]Garrett Baker7mo23

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

There is no “normal” calibration, its all relative to the questions you have, eg you get like 0.2 if you answer 50% to everything, which can be pretty horrible if the questions were all “will the sun rise tomorrow?” but pretty awesome if the questions were all something like “Will there be a plague in 2020?” (asked in 2016) such that everyone puts 1% or something

[-]Screwtape7mo20

Hrm. I guess what would be helpful here would be a sense of the range; the average briers floated around .20 to .23, and I don't have a sense of whether that's a tight clustering with a bit of noise or a meaningful difference. To use running a mile as a comparison, differences of seconds mostly aren't important (except at high levels) but differences of minutes are, right?

[-]Garrett Baker7mo30

Yeah, I don't know the answer here, but I will also say that one flaw of the brier score is that its not even clear that these sorts of differences will be even all that meaningful. Like, what you actually want to know is, how much more information does one group here give over the other groups here, and how much credence should we assign to each of the groups (acting as if they were each hypotheses in a Bayes update) given their predictions on the data we have. And for that, you can just run the bayes update.

The brier score was chosen for forecasters as far as I can tell because its more fun than scoring yourself based on log-odds (equivalent to the bayes update thing). Its less sensitive to horribly bad predictions, and it has a bounded "how bad can you be". Its also easier to explain and think about, and has a different incentive landscape for those trying to maximize their scores, which may be useful if you're trying to elicit good predictions.

But if you're trying to determine who you should listen to (ie in what proportion you should update your model given so-and-so says such-and-such) you can't do better than a Bayes update (given the constraints), so just use that!

[-]Garrett Baker7mo20

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.

what? We have , and similarly for confident.

[-]Screwtape7mo20

11.2% is if I remove the CFAR attendees. 36.8% is if I remove the non-attendees. Possibly this is a backwards way of setting things up but I think it's right?

Say I have a general population and I know how many pushups they can do on average (call this Everyone Average), and I remove everyone who goes to the gym and see how many pushups those remaining can do on average (Call this Gym-Removed Average) and then I go back to the general population again this time removing everyone who doesn't go to the gym (Call this No-Gym-Removed Average.)

This is a confusing label scheme but I don't immediately know what the better one is.

If No-Gym-Removed Average < Gym-Removed Average, then it looks like the gym helps.

(Totally possible I'm screwing something up here still)

[-]Garrett Baker7mo20

Ok first, when naming things I think you should do everything you can to not use double-negatives. So you should say "gym average" or "no gym average". Its shorter, and much less confusing.

Second, I'm still confused. Translating what you said, we'd have "no gym removed average" -> "gym average" (since you remove everyone who doesn't go to the gym meaning the only people remaining go to the gym), and "gym removed average" -> "no gym average" (since we're removing everyone who goes to the gym meaning the only remaining people don't go to the gym).

Therefore we have,

gym average = no gym removed average < gym removed average = no gym average

So it looks like the gym doesn't help, since those who don't go to the gym have a higher average number of pushups they can do than those who go to the gym.

[-]Sherrinford7mo20

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.

[-]Screwtape7mo20

No, I think that's correct.

There's 107 people who answered above 200, 21 who answered exactly 200, and 113 people who answered below 200. The second quartile (aka the median) is 200. But nobody guessed a negative number, so the people who guessed low aren't pulling the mean down that much. Meanwhile 33 people guessed 1000 or higher, and they can yank the mean a lot without doing that much to the median. If you're asking people to generate numbers, you tend to get whole number quartiles because nobody guesses there's 100.5 stations.

Imagine a the set [1,1,1,2,2,2,2,100,100]. The average is ~23.444, but the median is 2.

Or have I misunderstood the thing that you think needs to be corrected?

[-]Sherrinford7mo20

Sorry, I confused something (not about the median, but about the 400). Thanks for illustrating.

[-]Screwtape7mo20

Future Survey Discussion thread

[-]Screwtape7mo20

A Screwtape Point (and upvotes) to whoever can tell me how to fix the table of contents.

[-]dirk7mo10

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

[-]Screwtape7mo20

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.

[-]dirk7mo10

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).

[-]Alice Blair7mo11

When taking the survey, I figured that there was something fishy going on with the conjunction fallacy questions, but predicted that it was instead about sensitivity to subtle changes in the wording of questions.

I figured there was something going on with the various questions about IQ changes, but I instead predicted that you were working for big adult intelligence enhancement, and I completely failed to notice the dutch book.

Regarding the dutch book numbers: it seems like, for each of the individual-question presentations of that data, you removed the outliers. When performing the dutch book calculations, however, it seems like you keep the outliers in. This may be part of why the numbers reflect on our dutch book resistance so poorly (although not the whole reason).

[-]Screwtape7mo30

but predicted that it was instead about sensitivity to subtle changes in the wording of questions.

If I try this again next year I'm inclined to keep the wording the same instead of trying to be subtle.

Regarding the dutch book numbers: it seems like, for each of the individual-question presentations of that data, you removed the outliers. When performing the dutch book calculations, however, it seems like you keep the outliers in.

Yep. Well, in the individual reports I reported the version with the outliers, and then sometimes did another pass without outliers. I kept all of the entries that answered all the questions for the dutch book calculations, even if they were outliers. I think this is the correct move: if someone's valuations are wild outliers from everyone elses but in a way that multiplies out and gets them back to a 1:1 ratio, then being an outlier isn't a problem.

(Imagine someone who values a laptop at one million bikes, and a bike at equal to one car, and a car at one millionth of a of a laptop. They're almost certainly a wild outlier, and I'm confused as heck, but they are consistent in their values!)

[-]Alice Blair7mo31

I think your reasoning-as-stated there is true and I'm glad that you showed the full data. I suggested removing outliers for dutch book calculations because I suspected that the people who were wild outliers on at least one of their answers were more likely to be wild outliers on their ability to resist dutch books; I predict that the thing that causes someone to say they value a laptop at one million bikes is pretty often just going to be "they're unusually bad at assigning numeric values to things."

The actual origin of my confusion was "huh, those dutch book numbers look really high relative to my expectations, this reminds me of earlier in the post when the other outliers made numbers really high."

I'd be interested to see the outlier-less numbers here, but I respect if you don't have the spoons for that given that the designated census processing time is already over.

Use Case	Approx. Count	% of ~190
Writing or debugging code	~125	~65%
Rubber ducking	~95–100	~50–55%
Giving me angles on research	~85–90	~45–50%
Editing my writing	~70–75	~35–40%
Generating new writing that I edit	~60–65	~30–35%
Translation between languages	~60–65	~30–35%
Search	~25	~10–15%
Any “Other” use (therapy, fun, etc.)	~60–70	~30–40%

“Other” Subcategory	Approx. Count	% of 190
Emotional support / therapy (e.g. “Therapy,” “emotional processing,” “talking through problems”)	~10	~5%
Brainstorming / idea generation (casual or creative idea‑bouncing, Q&A beyond “research angles”)	~20	~10%
Data extraction / summarizing (OCR, content moderation, doc simplification, info extraction, etc.)	~15	~8%
Fun / entertainment / roleplay (poems, music, “just messing around,” “testing capabilities”)	~20	~10%
Recipes / cooking suggestions (meal planning, recipe ideas)	~5	~3%

110

2024 Unofficial LessWrong Survey Results

110

110

The Data

0. Population

I. Demographics

II. Sex, Gender, and Relationships

III. Work and Education

IV. Politics and Religion

V. Numbers Which Attempt To Measure Intellect

VI. LessWrong, the basics

VII. LessWrong, the community

VIII. Probabilities

IX. Traditional LessWrong Census Questions

X. LessWrong Team Questions

XI. Adjacent Community Questions

XII. Indulging My Curiosity

XIII. Detailed Miscellaneous Questions

XIV: Bonus Political Questions

XV: Wrapup

Fishing Expeditions

Meetup ComparisonsHow good are they?

What’s the community overlap like anyway?

Skill Issues

What might influence how rational someone is?

How might you measure how rational someone is?

The Census Over Time

LW Age

Degree

Time in Community

Time on LW

Attends Meetups

Income

Race

Sequences

Political

Referrals

Singularity

Conclusions

About the respondents

About the survey

The public data

Meetup Comparisons
How good are they?