(Also posted on LessWrong)

Introduction

In EA orthodoxy, if you're really serious about EA, the three alternatives that people most often seem to talk about are

(1) “direct work” in a job that furthers a very important cause;

(2) “earning to give”;

(3) earning “career capital” that will help you do those things in the future, e.g. by getting a PhD or teaching yourself ML.

By contrast, there’s not much talk of:

(4) being in a job / situation where you have extra time and energy and freedom to explore things that seem interesting and important. 

But that last one is really important!

Examples

For example, here are a bunch of things off the top of my head that look like neither “direct work” nor “earning-to-give” nor “earning career capital”:

• David Denkenberger was a professor of mechanical engineering. As I understand it (see here), he got curious about food supplies during nuclear winter, and started looking into it in his free time. One thing led to another, and he now leads ALLFED, which is doing very important and irreplaceable work. (Denkenberger seems to have had no prior formal experience in this area.)
• I’m hazy on the details, but I believe that Eliezer Yudkowsky and Nick Bostrom developed much of their thinking about AGI & superintelligence via discussions on online mailing lists. I doubt they were being paid to do that!
• Meanwhile, Stuart Russell got really into AGI safety / alignment during a sabbatical.
• The precursor to GiveWell was a “charity club” started by Holden Karnofsky and Elie Hassenfeld, where they and other employees at their hedge fund “pooled in money and investigated the best charities to donate the money to” (source), presumably in their free time.
• I mean seriously, pretty much anytime anybody anywhere has ever started something really new, they were doing it in their free time before they were paid for it.

Three ingredients to a transformative hobby

Ingredient 1: Extra time / energy / slack

Honestly, I wasn’t really sure whether to put it on the list at all. Scott Alexander famously did some of his best writing during a medical residency—not exactly a stage of life where one has a lot of extra free time. (See his discussion here.) Another excellent blogger / thinker, Zvi Mowshowitz, has been squeezing his blogging / thinking into his life as a pre-launch startup founder and parent.

Or maybe those examples just illustrate that, within the “time / energy / slack” entry, “time” is a less important component than one might think. As they say, “if you want something done, ask a busy person to do it”. (Well, within limits—obviously, as free time approaches literally zero, hobbies approach zero as well.)

Note a surprising corollary to this ingredient: “direct work” (in the EA sense) and transformative hobbies can potentially work at cross-purposes! For example, at my last job, I was sometimes working on lidar for self-driving cars, and sometimes working on military navigation algorithms,^[1] and meanwhile I was working on AGI safety as a hobby (more on which below). Now, I really want there to be self-driving cars ASAP. I think they’re going to save lots of lives. They’ll certainly save me a lot of anguish as a parent! And we had a really great technical approach to automobile lidar—better than anything else out there, I still think. And (at certain times) I felt that the project would live or die depending on how hard I worked to come up with brilliant solutions to our various technical challenges. So during the periods when I was working on the lidar project, and I had extra time at night, or was thinking in the shower, I was thinking about lidar. And thus my AGI safety hobby progressed slower. By contrast—well, I have complicated opinions about military navigation algorithms, but let’s just say that they have never aroused in me a great passion. So during the periods when I was working on military navigation algorithms at my day job, and I had free time at night, or was thinking in the shower, I was thinking about AGI safety instead, and I made faster progress! (See Paul Graham’s essay “The Top Idea in your Mind”.)

Ingredient 2: ???

Here I’m referring to the fact that lots of people have extra time / energy / slack, and don’t use it for any world-changing hobbies. Instead, umm, I don’t know, maybe they watch a lot of TV, or argue about politics online, or host fancy parties, or build model ships, or whatever. (I'm not criticizing people who do those things. Your time is your time. Spend it as you wish!)

I don’t know what accounts for the difference. Certain types of motivation and skills and interests, I guess?

(I’m reminded of the book quote: “When you are older, you will learn that the first and foremost thing which any ordinary person does is nothing.”)

Ingredient 3: Willingness to pivot

As I understand it, Eliezer Yudkowsky was really into thinking about nanotechnology before deciding that actually thinking about AGI was a much better use of his free time, and then later pivoted again to thinking more specifically about AGI safety / alignment. Nate Soares likewise describes pivoting to AGI safety after a decade thinking about governance and institutions. I don’t know what David Denkenberger was thinking about in his free time before he came upon the question of feeding the world through nuclear winter, but I bet it was very interesting and different! For my part, AGI safety was the 5th (!!) long-term (i.e. multi-year) intense ambitious hobby of my life. (And now it’s my job.) 

I received the following comment on a draft version of this post:

There's also the danger that side projects also often have very unclear failure conditions. Because you're never trying that hard, lack of success feels fairly 'excusable', I can point to multiple examples of people who've devoted huge amounts of time+effort to side projects that were going nowhere, where I suspect that a full time 'put up or shut up' attempt would have caused them to (correctly) give up and try something else.

Here’s my message to “people who are devoting huge amounts of time+effort to side projects that are going nowhere”: Don’t do that, OK? I mean, jeez. One of the key selling points of hobbies is that if you think maybe you should pivot and start from scratch on something totally different, you can just up and do it! There’s nothing stopping you! You don’t owe anyone anything! You have to kill your darlings! Maybe you learned something along the way, or at least had fun, and life is long, go start something new, etc. etc. That’s my motivational speech, thanks for listening. Take it from me, a guy who spent 15 years building up a top-notch physics expertise that is now completely irrelevant for my life.

How can hobbies compete at all with jobs? My theory: compromises are terrible

Naïvely, if we’re going to compare “doing stuff as one’s job” versus “doing stuff as one’s hobby”, the former seems to have overwhelming advantages: jobs brings to bear heavy artillery in the form of time, energy, money to spend on resources, multiple people coordinating, and so on. And don’t get me wrong, jobs make the world go round. Nobody is manufacturing integrated circuits in their free time.

It’s no coincidence that David Denkenberger and Eliezer Yudkowsky and Holden Karnofsky (and me!) transitioned their hobbies into proper jobs approximately as soon as possible (more on which just below).

But we still have to explain how important things get done at first as hobbies, not jobs. I think the answer is: compromises are terrible.

If you need to simultaneously satisfy two criteria, i.e. “this is an important thing to do” and “this is something I can immediately get paid to do”, then there’s a tradeoff. You need to compromise on both.

I said above that jobs bring to bear “heavy artillery” that hobbies can’t. But hobbies can make up for that deficiency with better aim.

(For more on the theme of compromises when trying to satisfy multiple criteria at once, see for example Eliezer Yudkowsky’s Purchase Fuzzies and Utilons Separately or Alex Lawsen’s Know What You’re Optimizing For. For more on the advantages of doing things unconstrained by institutional incentives and inertia, see Paul Graham’s The Power of the Marginal, and Eliezer Yudkowsky’s book Inadequate Equilibria.)

Hobbies can eventually develop into jobs

A hopeful possibility, of course, is that the hobby can eventually lay the groundwork for the creation of a new job, one whose money- and status-related incentives are all perfectly aligned with the thing that you have now figured out is what you want to do. Then you get the best of both worlds—the aim of the hobby, and the heavy artillery (time, resources, etc.) of a job.

But you can’t get there until you know what you’re aiming for.

My own two stories

These are a couple anecdotes from my own life; feel free to skip down to the Conclusion if you're uninterested.

Story #1: Me & solar cells

Back in my innocent youth, I was passionate about inventing better solar cells. It was the reason I went into physics in the first place as an undergrad in 2003, and then I deliberately went to a physics grad school where there was an unusually high concentration of solar cell research happening.

How'd it go? I had some outputs in the "direct work" category (the thing I was “supposed to be doing” as a grad student), and also I had some outputs in the “slack / hobby” category (stuff I did informally, independent of my grad school advisor, unrelated to my eventual dissertation, etc.)

Shall we compare?

In the “direct work” category:

This category is where I was trying to compromise between

• “help develop better solar cell technology”, versus
• “do the normal grad student things”, e.g. “make progress towards my PhD” and “publish in high-profile journals” and “learn useful skills”.

Basically, I wound up doing decently at the latter and pretty much entirely failing at the former.

I worked in four labs:

• The first two labs were each doing interesting, trendy physics superficially related to solar cells, but their projects had no chance of actually making cheaper or better practical solar cells in the future. I wound up coauthoring a total of two papers in that category before switching labs. (The two papers now have 1400 and 500 citations respectively. It turns out that when you optimize for producing trendy high-profile papers, you can end up producing trendy high-profile papers. Who knew?)
• Then I switched to a lab that wasn’t working on solar cells at all, but I liked the professor. I figured, oh well, too bad about the solar cells, but at least I’m getting a physics education! As it turned out, I had my own external funding and thus wasn’t tied to any particular one of my professor’s grants, and in my free time I thought of a maybe-practically-useful solar cell project. I found some excellent collaborators, and we got a lovely paper out of it, but nothing came of it in practice, in part because our patent application got rejected for stupid reasons.
• I also spent some time as a visiting student in a lab that was doing plausibly-useful solar cell R&D. However, the nature of the work was a terrible match to my skills and inclinations. Well, I guess I learned something valuable about my skills and inclinations, although in hindsight I should have already known that much earlier.

In the “hobbies” category:

• I had written some optical simulation code related to solar cells (among many other things), and in my free time sometime during or after grad school I cleaned it up, documented it, and put it on GitHub. I’ve heard from many people over the years that they use my code to help understand and design solar cells, including people in the R&D departments of at least two solar cell companies. Great success!
• The theoretical physics underlying how solar cells work was basically all well-established long before I entered the field. But I did distribute some pedagogy on the topic. It’s hard to know what practical impact that had (if any), although I find it fun to open up a solar cell book, or watch a presentation, and hey, there’s my diagram, which I had put on Wikipedia years earlier!

Winner: Hobbies

I can say with some confidence that I have contributed a little bit to help bring cheaper and better solar cells into the world, exclusively via projects that I did in my free time, that my grad-school advisor didn’t even know about, and that probably would have done either nothing or almost-nothing to help me advance in academia.

Also, among my more standard legible grad-student-y projects, I had one near-miss which I think could have been practically useful (a priori). And wouldn’t you know it, it was the one project that I initiated and developed in my free time (before eventually looping in my advisor).

Story #2: Me & AGI safety

In 2019, having finished some other hobby, I decided my next hobby would be AGI safety. So I started reading and writing posts and comments on LessWrong, in my little bits of time squeezed between my full-time job and two young kids.

Of course, I found it very frustrating how little time I had, and I immediately started brainstorming how I could get much more time by doing AGI safety as a job.

The normal advice would be to apply for existing AGI safety jobs, or go back to grad school etc. But I had high living expenses and didn’t want to move out of Boston. Alternatively, I managed to come up with a handful of possible projects that I might do within my existing job, that seemed at least slightly relevant to AGI safety. For example, I could have tried to talk my way onto my coworkers’ existing DARPA-funded project related to machine learning uncertainty, or various other things. But I wound up dropping those ideas too.

So nothing came of any of that, and I’m glad it didn’t, because my current belief is that I did dramatically better by just blogging and commenting on LessWrong in my tiny amount of free time, than I would have done if I had successfully found a way to do AGI safety in an institutional setting right off the bat. I wound up going off in a weird intriguing direction, and then losing interest, and then going off in a different weird intriguing direction, which incidentally involved teaching myself neuroscience, and I wound up making enough progress to get a grant to do full-time independent research, and here I am! I think I’m doing something important and maybe a bit idiosyncratic, and it’s all thanks to having spent a significant amount of time just reading and writing about whatever the hell I wanted.

Conclusion

I was inspired to write this post right after talking to (what felt like) 3000 super-enthusiastic undergrads at EAGxBoston a couple months ago (I love you all!), who all wanted to get into AGI safety / alignment.

Among the people I talked to, there seemed to be an unspoken assumption that, for example, if they go to grad school at all, they should try to maneuver into doing AGI-safety-related projects with an AGI-safety-concerned advisor, and if they did so, victory!! Let’s be clear: I’m not opposed to doing that! Some people should definitely be doing that! Great work has been done that way, e.g. Alex Turner’s dissertation. But let’s not pretend that this isn’t a compromise. You’ll wind up in a project which is simultaneously optimizing for both advancing AGI safety and easy consumption by peer-reviewers, by your advisor’s future grant review committees, by your own dissertation committee, by future people reading your CV, etc.

(It’s no coincidence that the “weird sci-fi stuff” side of AGI safety is almost entirely absent from academia.)

Apart from grad school, other frequently-proposed plans were: joining existing AGI safety nonprofits, joining existing mentorship programs, and getting good at ML (or neuroscience or whatever) in normal legible ways like “taking online courses” or “getting jobs in that field”. Those are all good things too! I’m a huge fan of all those things!

But not many people brought up the idea of just having any job whatsoever, ideally a pleasant, invigorating job that you’re really good at and which has good work-life balance,^[2] and meanwhile making a plan in your free time for how to solve AGI safety. (Or if that’s too ambitious, maybe “write intelligent comments on people’s AGI-safety-related blog posts”, and/or “create AGI-safety-related pedagogy”, and work your way up from there—certainly that’s all I was hoping for at first!) And only then, when and if you come up with a plan, you can apply for funding, either for independent research, or starting a new organization, or joining up with other people who share your vision, etc.^[3]

Granted, that approach has its own issues. In some ways, it’s a terrible plan! But I think it at least deserves a modicum of consideration.

(As usual, consider reversing any advice!)

Thanks Alex Lawsen, Justis Mills, and Adam Shimi for critical comments on a draft.

1. ^{^}

   (among dozens of other projects, but these two can serve as prototypical examples)

2. ^{^}

   Or in the case of undergrads, I sometimes noticed perhaps a bit too much interest in filling up their résumé with the maximum number of research internships and good grades in hard classes etc., and perhaps a bit too little interest in having the mental space and energy to get sucked into thinking about interesting and important questions on their own time. [… Says the guy who spent his undergrad years filling up his résumé with the maximum number of research internships and good grades in hard classes etc. …]

3. ^{^}

   Or maybe you’ll wind up feeling that AGI technical safety research isn’t the right thing for you to do at all, and off you go in some other direction!