Epistemic Status: Abstract claims, but grounded in data science... though the data science is somewhat stale. I wrote this on March 5th of 2026 based on memories of work I did and methods applied circa 2017, and pushed it out for publication after realizing that maybe there is appetite for it, after I saw this post.

Suppose we took a snapshot of each person in the US, and made a list of their "skills", as one might do with a D&D character.

I would like to report on what I expect would happen if this was attempted in real life, and why (until I get to the point that you understand the title of the essay about "Budgeting" skills being Important a little ways in).

I haven't done this recently, with modern data, but I felt that this was likely to be something in my brain that most people don't know about, and worth an article.

At the end of the essay there will be a call to action! I want to start at least one study group in the SF bay area to "Level Up Budgeting" so I could attend somewhere face to face and talk about books or essays or tools, and I'd be happy if various meetup communities around the world formed their own local study groups, so there can be cross-pollination and transient diversity and so on.

Object Level Skill Discussion

First, The Boring Skills

With a skill list for every person in the US, we would see a lot of lists for people whose chief skill is their ability to use a QWERTY keyboard and use a suite of Office Software to create spreadsheets, presentations, and prose.

It turns out you would see a lot of people who have a CPR certification, and almost everyone who gets that in modern times gets an AED certification as well.

(When I think of this, it reminds me of how rationalist!Harry went straight for a magical first aid kit on his first day of being able to buy magical gear. But also, it is interesting how "skills" blur very very swiftly into "certification of those skills" in the actual speech of actual people.)

In real life, writing down a skill list for every person would probably lead to variation in how people write.

Some people would write "Skills: word, cpr, aed, ..." and others might write "Skills & Certifications: MSOffice, AED, CPR, ..." and in some platonically perfect realm these would parse down and be applied to generate the same basic inference of the same basic capacities in the author of the list.

Languages already lump. To really dig into a skill, you don't just know "the name of the nameable skill" but actually can DO things.

There are probably subskills!

Some of the subskills might not even have names but still be something that can be transmitted and learned by someone saying "hold your hand <like this>... see?" and then "no, your pinky is getting in the way... more <like this>". We are interested in all of it, of course, from the microskills to the macro level.

Second, The List Of Skills

Most people in real life can't cast a "healing spell" (because magic isn't real, so far as I know).

And "krav maga" is pretty rare, and not what people think of first.

Fighting is not central to real life in the modern world... mostly (cops and soldiers and bouncers and bail bondsmen do make up a non-trivial chunk of the jobs though).

Usually people think of the civilian job market, in "real life", when they actually make lists of skills and post them online.

Even in the military, they have complex software helping with the task of killing people effectively to serve diplomatic ends (or whatever), but in practice they end up needing to model and optimize the logistics of hiring 18 year olds and turning just the right number of them into "fuel-truck drivers-license test proctors" as fast as possible. In a deep sense: a lot of skills are pretty prosaic.

The "reality" that is "out there" is nebulous, of course.

Depending on how much lumping and splitting you do, so far as I could tell back when I had access to the data, there are roughly 50k to 200k total skills that people will think are distinct enough to write out as distinct phrases.

Some of the phrases will point to having a specific security clearance level, but often someone might just write "Abilities: office, cpr/aed, ... security clearance, ..." as if those two words "security clearance" were a skill, or pointed to skills, or something?

(Metonymy is the name for what happens when people hit the limits of their words, and just name a thing for whatever thing happens to be nearby that they already know the name of.)

One could probably get the number of skills down around 35k if you put serious effect into de-duping the phrases that people use, and ignore hapax legomena.

But if you only de-dedupe the top 100 most common skills by hand based on semantics, and de-dupe the skill phrases that show up in skill lists algorithmically (reducing plural and single forms into one form), and keep the skill phrases that only ONE PERSON EVER thought deserved to be in a skill list, and yet also hunted far and wide to gather all the English language resumes you could (from the 1980s to 2026) I bet you could get as many as 300k skills, without much trouble.

And if you took each one of these and then traced it to the person, and asked them to teach someone this "officially listed skill" I bet it would turn out to often have 10 subskills that can only be described with a phrase or sentence or paragraph... which means that there are plausibly ~3M skills that are super granular and would take a "paragraph of description" to point to, potentially? Things like "the best way to hold a nail in your other hand when hammering nails".

When I estimate ~3M such skills, I'm being pretty sketch and rough... I would be surprised if it was less than 500,000 of them, and I would also be surprised if it was more than 10 million.

Third, The Serious Lumping Begins

A very reasonable person might think that this is crazy.

They might think "computers" is really "just one skill" and that "git" and "svn" are NOT meaningfully different skills.

They might assert that there is ONE "computer" skill, and that if you are super great at that good at the one skill (but better than a normie!) you could apply that skill in a tech support job, but then you might level that skill up until you could apply it to work in Computer Research.

This would be a very reasonable and pragmatic perspective, but in that case, there's a lot of prior classification work!

The reasonable people at the Bureau of Labor Statistics have already lumped things down to 832 narrow job categories fitting into 116 medial grained categories and just 21 basic categories.

Following them, seeking VERY lumpy lumps, you might reasonably say that the total number of "skills"...

...is either ~21 (some examples being "legal", "protection", "production", "management", "sales", and so on)...

...or else perhaps ~116 (some examples being "legal support skills", "fire fighting skills", "printing skills", "operational management skills", or "wholesale sales skills").

The BLS... is almost certainly too narrow.

Those aren't really skills, you know?

Those are really more like job categories... right?

But it does seem to be true that job categories sort of wink and nudge at the sense in which some skills might be more convergently useful than other skills, or have prerequisite skills, and be useful mostly only in concert with other skills.

Most people have played the piano at least a little, when killing time and kinda bored while waiting somewhere that a piano existed, and quite a few people have practiced specific piano skills (and there are many distinct such skills to practice separately (like just sitting properly, even)) but if we bow down to the lumpiest lumpers and their traditions, even using the full 832 BLS categories, all(!) of the piano skills (and so much more) are lumped into "27-2042  Musicians and Singers".

This is NOT GOOD ENOUGH to tell us what skills to spend 10 minutes practicing every day, or how to hire people who will pay the piano in the way that we really want to hear it played in a bar from just what can be found about them via online scraping of their data.

If we are going to HIRE for jobs, based on a gears level understanding of what cognitive or physical performance or capacity goes on in that job then we need something more granular than jobs to point at, or reason about, that "are the gears".

Fourth, Lumping The Skills With Math?!?

Here's a thought: maybe we take all the skill lists as is to create ~100k skills from some corpus (with regional choices, and choice of era, and some amount of de-duping) that someone ever mentioned.

We treat each unique skill skill as a node in a network (or as mathematicians might call it a "graph theoretic graph")... and then for each skill list in each resume, we draw lines between every pair of skills that occur in the the same list.

And then for the next list we do the same (making some lines +1 in strength if they were paired in a previous list, so AED and CPR are decently likely to have a three-weight-strong link after processing maybe a dozen random resumes).

This gives us a weighted graph, which comes up in a LOT of optimization problems like shortest path discovery and pagerank and so on.

I have done this before! That's why I am writing this essay ;-)

Lots of "clusters" of skills fall out of such an analysis in the form of dense cliques where many members of the clique are strongly linked to other members of the clique, and weakly linked to anything else. These cliques represent tasks that are demanded in the same job, or abilities taught in very common forms of general education, or sometimes tasks that show up in "standard career progresses" (where a drafting student becomes an engineering tech and eventually a licensed surveyor, and their skill list has things from that whole history).

Every so often you'll find skills that are very common, and that many cliques of skills ALL point to.

For example, there's a bunch of skills that teachers, in the field of education, are proud of, where, by mid career, almost everyone is bragging on their resume that they can do "curriculum design" (no matter whether they're a math teacher or a piano teacher or a kindergarten teacher).

There are other skills like "security clearance" that might show up in the protection area..

But then "cpr" is linked to both of these (and many others)!

This strongly suggests a high degree of convergent instrumental utility exists in skills like this, across a wide variety of fields, even though the skill is "narrowly a skill" that is more like "git reflogging" that can be taught and practiced and tested, but also "broadly a skill" in the sense that it comes up for practically everyone.

Fifth, Seeking "Betweenness" Centrality You'll Find "Budgeting"

There are many measures of "centrality" in graph theory.

One way to arrange the math of it might pinch out the thing at the very center of the very biggest clique (which itself is perhaps at the center of the biggest macro-clique and so on)... but that won't give us these insanely broadly applicable skills that can be taught and learned!

The thing we want, if we're looking for very very broadly valuable skills in almost any domain or any job is betweenness centrality.

The way this works (roughly) is that we pick a lot of pairs of nodes, and spend compute to find the shortest path between them... over and over... and every time we do this we add a point to all the nodes that were on this "shortest journey" from node to node.

At the end, we find the chokepoint... the master node, the node from which you can go ALMOST ANYWHERE very very quickly.

"Curriculum design" is more central than "math curriculum" because it is invoked by more kinds of people with more diversity of skills. And "CPR" has still more betweenness centrality than "curriculum design" (because CPR is useful for cops with cop skills, and life guards with swim instructor skills, and fire fighters, but also elementary school teachers, and summer camp instructors and so on).

Here is the punchline: the thing with the most betweenneess centrality out of all skills is "budgeting".

Which... uh... you know... makes sense? Maybe? <3

Sixth, Meditations On The Betweenness Centrality Of "Budgeting"

Here are some thoughts:

FIRST, consider the holy grail of "rationality" (in the sense of "verbally transmissible cognitive practices that conduce to higher chances of success at nearly any goal" is skill transfer).

Finding skills with high skill transfer and broad applicability lets you spend the least amount of time leveling up, and gives you the most benefit from them.

The skill of "budgeting" makes sense here because like... finitude is everywhere? Tradeoffs are everywhere? Also time is real and ubiquitous. And "making tradeoffs over time in the face of scarcity" is basically the essence of budgeting.

"Budgeting" arguably deserves some halo, because it makes sense, from first principles, that almost every agent would need this skill, if you actually think about agency itself in a first principles way.

SECOND, this comes up in practice in business because the "mangle" of organizations often leads to a handful of people on like..  "the budgeting committee" (which no one wants to be on because it SOUNDS SO BORING) wielding enormous organizational power, and needing to be able to justify their use of power when budgets cause weekly, monthly, quarterly, yearly adjustments in what the organization can and will choose to do.

This will show up, predictably, at the HEART of many moral mazes. We would kind of expect, then, like with the "security clearance" skill, that "budgeting" is a pointer to skills related to the use of language in clarifying AND politically obfuscating ways, depending on who a budgeter is talking to.

The skill might be anti-inductively hard to master because it is woven into bizlore (not academia) and it is very political by default.

THIRD, the lists of skills from which this kind of conclusion can be drawn often come from a wide variety of people, posting resumes online and such. Some of them are older, and more advanced in their careers.

There is a decent chance that if you start in almost any profession or organization, and are good at anything, eventually you'll rise to a role of teaching, guidance, planning, and in a word "leadership". And almost all leaders of non-trivial organizations have accounting and bookkeeping to handle... and also this idea that "the procedures of accounting" could be used to map and understand and control and generally guide a team?

So maybe the betweenness centrality of budgeting comes from a real pinch point, that shows up over and over again in MANY careers, where "budgeting" (about your expertise, on behalf of a group that collectively wields that expertise) becomes essential.

FOURTH, we can compare Biden's semi-famous quote "Don’t tell me what you value. Show me your budget—and I’ll tell you what you value" with a lot of common topics on LW like money itself (the unit of caring!) and kelly betting (essentially iterated bet budgeting), and the relation of money to VNM rational utility functions and bayesian beliefs.

If the skill of "budgeting" turned out to be a super power wouldn't actually be that surprising, if you have a sense that more is possible.

A Sort Of A Methods Section

This essay so far has been very high level. I haven't really substantiated it, except to casually suggest what MIGHT be observed, if someone hypothetically looked at certain ways of describing the world in an intensely operational way. Maybe you just don't think it makes sense?

I generally don't like talking about myself, but... uh...

I applied to work at Google in 2013 when I believed that they had the Mandate Of Heaven with respect to the Singularity.

(For several reasons. One big one was that I had previously thought The Singularity Institute had the Mandate Of Heaven earlier than that, but then they sold their banner and trademark and social outreach apparatus to Kurzweil in 2012, who worked at Google. This was before OpenAI was founded in 2014. If I had waited another year, I might have joined OpenAI, instead.)

By 2018 I had given up on Google, because it is a moral maze, and was clearly not going to save the world, and I moved on to working on a blockchain project with friends that was not much related to the Singularity or to AGI in any obvious ways, but it gave me an opportunity to design decentralized utopias and dystopias for decent pay, and that was just super super hard to turn down.

In the intervening time, from 2013 to 2017, I first worked inside the Google Brain "cultural silo" (because I had been told I could work on deep learning aimed at comprehensive proactive benevolence) but after about a year of working on bullshit (like trying to get people to watch longer YouTube content that ads could fit in the middle of) I changed to work on something that actually fucking matters.

So I hopped over into the "PeopleOps" silo by 2015, and instead of increasing the size of YouTube's "ad inventory" I helped build the job search engine that is one of many topic-focused engines existing inside the overall Google Machine for that overall technological macro object to wield according to high level metrics.

(Job search seemed like the closest thing I would be able to find to a sword of good, that The Google Machine would have a hard time abusing too egregiously. And it seemed like it could plausibly produce tens of thousands of dollars in consumer surplus for people who used "free job search" (supported merely by ad money, that is) to find better jobs.)

For about 18 months, between 2015-2017, I was doing data science on the theory and practice of searching for jobs specifically through the lens of "skills and education".

I didn't have the ability to contact people in the outside world about it (because we didn't want to leak that we were working on job search and so on) but my team had access to "all the resumes" and "all the job postings"... in English anyway (and it turns out that the language used in different job markets around the US was regionally dialectical, and so we optimized at the beginning for just a handful of cities, and when we tried to use machine translation to generalize it to Telugu or Japanese the relevance numbers just totally cratered because there is a lot of nuance in the euphemisms people use to talk about hiring and firing and such).

We could play with ways to parse and analyze them using now-old-fashioned pre-LLM NLP techniques. So for a bit more than a year, I could, in fact, do the kind of analyses I described above, but with real data.

And "budgeting" did, in fact, fall out of that actual data, back in ~2017, in the way described <3

A Call To Action: Study Groups

Does anyone else want to do this?

I hope that people in the comments chime in in lots of ways because I still feel like "budgeting" is something I'm STILL learning to do well. I would love to hear good textbooks. I would love to hear war stories. I would love to hear ways to spend an hour a day practicing something for a week and become really good at "budgeting".

Does anyone know how to hire the best CFO who is a total wizard at the "budgeting" skill? I don't! I think that's an important thing to be able to do, and I can't reliably do it.

I would be open to driving to Berkeley once a week to talk about some shared reading, or report on homework we assigned ourselves, and hear about other people's challenges and growth at "budgeting, the most central of skills" <3