How Should I Spend My Time?

Back in 2013, I was in my final year of college and wanted to think seriously about what I do. I drafted up 11 initial options (though I now feel like my initial list was kind of shortsighted), narrowed it down to 5 options that I more thoroughly assessed, compared them on 11 different traits (too many, in retrospect), talked with Satvik, talked with Holden, ended up deciding to enter software engineering and didn’t regret it one year later.

Then, I made a decision to get a day job working to earn to give, but also spend a significant amount of time outside of work on direct EA activities. I told myself that I’d continue to re-evaluate as I got more information. While I offered some considerations in my last review, I feel like I never really assessed things formally with an eye toward what is best. Until now.

Right now, I’d say that I spent 2016 allocating time among four activities: (a) a day job as a data scientist, (b) volunteering for Charity Science Health, (c) doing work on cause prioritization (such as blogging, doing veg research, and being on the ACE board), and (d) volunteering for Hillary Clinton. I also wanted to work more on starting a tech blog and doing fundraising, but didn’t really do much of either.

I wanted to look back at all seven of those activities to assess what their value was or could be as well as consider some additional possible paths forward. Also, while back-of-the-envelope calculations are not everything and should not be taken literally, I wanted to construct these calculations for each activity to try to inform my intuitions with more precise numbers and to make my thinking more clear for others.

Day Job

I work as a data scientist at an advertising company in Chicago.

I think the utilitarian value of it comes from (a) providing a salary from which to donate, (b) providing the opportunity to grow my salary and donate more in the future, and (c) providing skills that could be useful for direct EA work.

My guess is that (c) is not special here and could be arrived at without working where I do, though because I work at an advertising company, I’ve learned some stuff that is relevant to veg advocacy (since that is mostly about advertising) and I’ve learned some stuff that is relevant to my work on Charity Science Health, such as A-B testing.

 

To make this easier to estimate, though, I’d like to value my day job solely on (a) and (b). I’ve historically been bad at estimating my own salary growth, but I made a new attempt at guessing how much I would donate in the future that tries to adjust for this. This suggests I’d donate about $160K over 2016-2019 (95% interval: $100K to $260K).

After over a year of time tracking, it’s clear that my job takes up an average of 21 hours a week of time-tracked work (does not include lunch breaks, bathroom breaks, socializing with co-workers, doing non-work things at work, commuting, etc.). So if I could be expected to work 4380 hours over 2016-2019, earn $660K (95%: $580K to $860K) and donate $160K, that’s an expected earnings of $150.68 per hour worked. While not all of that is donated, the part of it that isn’t donated can allow me to work additional hours for free on EA projects and not draw money off of the EA movement, so I consider my entire earnings to be the altruistic value of this project.

Charity Science Health

 

Charity Science Health is an experiment to create a new GiveWell top charity. We’re using SMS to remind mothers in India to get their children vaccinated, which we find to be highly cost-effective with a high strength of evidence and good ease of testability. I’ve been a long-time board member and volunteer to Charity Science Health, and this is where I put a significant portion of my not day job time.

To look at the value of working at Charity Science Health, I wrote up an in-depth analysis with an accompanying model which suggested that each full-time equivalent staff member (FTE) could be expected per year to generate the equivalent of a $400K donation (95%: $220K/yr - $720K/yr). Adjusting for “senior” staff status (which I have) boosts the expected value to $4.1M per FTE year (95%: $550K/yr to $25M/yr). Converting this to a “per hour” rate for comparison would be $320 per hour in altruistic value using the first model and $3280 per hour using the second model, assuming 1250 time-tracked hours worked per year by a FTE.

Cause Prioritization

I’ve been spending more time recently working broadly on cause prioritization, which I take to be the broader theme behind all my EA forum blogging (including this post), all my work on veg advocacy research, and all my work with Animal Charity Evaluators.

A lot of this work is scattered and it’s very unlikely that each hour would have the same returns. Therefore, I might someday want to be mindful of this and do some cause prioritization prioritization. Until then, I’m suggesting a simpler model for how to value cause prioritization -- how better can you make decisions and how many people can you influence?

Through cause prioritization, I can obviously influence my own giving, which I project to be about $160K over 2016-2019 (95% interval: $100K to $260K). I also can influence my marginal additional time and the time and money of others.

But the other question of “how better can you make decisions?” also matters, as the value of cause prioritization is only equal to the value of information for each thing we learn. The only way to actually create value is to stumble upon a decision that actually changes where I and others target our efforts. So far this has happened once, when two full-time years were invested into the Charity Entrepreneurship Project, which shifted our Charity Science team from doing fundraising to doing Charity Science Health. I think another two full-time years doing more broad research could potentially yield similar gains.

Putting all these assumptions into another highly uncertain Guesstimate model, looking at the resources I could potentially influence and the expectation for improvement yields the assumption that investing two full-time years into cause prioritization would produce an expected value of $11M over the four years after project completion (95% interval: $620K to $53M). Assuming a full-time year is 1250 hours of solid, honest time-tracked work, that means we could generate an expected value of $4500/hr (95% interval: $250/hr to $21K/hr).

Campaigning for Hillary

I spent a few days volunteering for the Hillary campaign outside of my day job. I told people to vote and ask their friends to vote. I told people again to ask their friends and I asked all my friends. I phone banked and placed 274 calls.

I’d guess my efforts lead to 1-3 additional votes to Clinton in swing states. 80,000 Hours sketches out some notes on the value of a vote that could potentially be construed as an estimate of $1M in US social value per vote, though this estimate is incredibly rough and very likely to be biased toward a larger number. (Earlier I had interpreted this number more literally, but Ben helpfully guided my analysis with this comment, leading to these revisions.)

This number could be potentially converted to a comparative GiveWell scale by suggesting it is ~100x better to help someone in the developing world than someone in the US (due to comparative consumption and income) and that furthermore interventions like AMF and SCI are ~5x better than pure cash value. This would re-scale the value of a vote to ~$2000.

I'd then further double the value of the vote to readjust for the fact that the US does spend government funds on certain things (e.g., science, some military exercises, foreign aid) that are of value to the rest of the developing world, and that these can change if government policy changes.

Lastly, I’d be more comfortable, relatively speaking, also reducing the figure 5x to account for the figure being potentially biased upward (though I’m still not comfortable, generally speaking, with this figure). Thus, when adding in the developing world discount and uncertainty discount, I’d clock my efforts at generating $800-$2400 in expected value from 1-3 votes.

Since I spent about nine total hours on campaigning for Clinton, that’s an estimated $89/hr to 267/hr in expected altruistic value. However, the error bars on this estimate would be immensely wide and include negative numbers, since it certainly is possible Hillary Clinton would have been a worse president overall than Donald Trump.

Overall, it bears re-emphasizing for a third time that I’m significantly uncertain about how heavily I should penalize the value of a vote or how skeptical I should be that I actually was able to get people to vote who otherwise wouldn’t have. But, fortunately or fortunately, I don’t feel the need to worry about this more since the election is over. It’s possible future political interventions could be high value, but I’d expect the opportunities to be much less great outside of a presidential election.

 

Start a technical blog about data science

Another thing I’ve considered doing but haven’t put much time into yet is starting a technical data science blog. A technical blog could help sharpen and advertise my skills to my current and future employers. Building a tech following from my blog would also enhance my networking skills a lot. Maybe there’s a 25% chance the blog would amount to nothing, a 50% chance the blog would garner the equivalent of a $5K/yr raise, and a 25% chance it would garner the equivalent of a $20K/yr raise. I’ll assume the raise matters for five years, a $5K/yr raise for five years has a $20K present value and a $25K/yr raise for five years has a $105K present value (assuming a 6% discount rate).

If I spend 5 hours on each post and post 50 times a year, that’s 250 hours of work. The present expected value of the blog would be $36,250, or $145/hr.

 

Fundraising

 

Over the past few years I’ve spent a few times fundraising, most notably through Charity Science Outreach (now scaled down) and through running a birthday fundraiser. I wrote a separate post looking back at the past and trying to project some more to the future, and I was able to estimate that future fundraising work could be expected to generate between $40/hr and $4000/hr, with the most reasonable estimate being around $140/hr, though admitting significant problems with steeply diminishing marginal returns on the highest value opportunities.

Another fundraising method I haven’t tried yet, talking to friends about taking the GWWC pledge, could offer much higher returns -- according to an estimate by those involved, the potential is to secure ~0.6 GWWC pledge each hour. Since GWWC estimates the value of an individual pledge at $73K each, that is $43.8K per hour. However, when I adjust the estimate for the expected income of my friend group and add skepticism about how long people will stay with the pledge, I get a value of $3900/hr, or roughly the same as some other fundraising work.

 

Train to Become an AI Safety Researcher

AI Safety research seems potentially very important, though I remain unsure about how to properly evaluate it and compare it to other causes. Given that my experience so far has been in engineering and machine learning, I feel like I could potentially be a good fit for an AI safety researcher, but I would have to spend a lot of time overcoming my poor abilities in mathematics. In short, I’d try to “pull a Soares”. I haven’t really worked at this much at all, besides by working through a calculus textbook for a bit. I put some assumptions about how long training would take and how much value I would realize in this Guesstimate model, getting a wildly uncertain, hugely wide average of $13/hr to $23K/hr with an average of $2K/hr.

Spreading Mobile Money

 

Inspired by several EA friends moving to Wave (most notably Jeff Kaufman), an EA startup working to improve cash transfers. Following Jeff’s analysis, I modeled the impact of spreading mobile money (and summarized my disagreements and confusions with Jeff). Based on this model, the direct impact of working at Wave would be about $200/hr (95% interval: $50/hr to $511/hr). If I had to take a 50% paycut like Jeff did to work at Wave (assuming I got hired by Wave), I’d expect to realize an additional $75/hr from donating the salary, for a total of $275/hr.

Notably, Jeff has more of an inside view than I do. When I update my model with more of his original inputs, the value of the direct work comes out to $383/hr (95%: $145/hr to $834/hr). Using my model to define the lower and mean values, my model with Jeff’s inputs to define the high value, and adding half of my day job value to each, gives me a range between $116/hr and $932/hr with a mean of $350/hr.

All options compared

I was not expecting my day job numbers to come out lower than all my other options (except arguably tech blogging and fundraising), but I was also not expecting my day job numbers to come out as high as they did.

However, as I said before, back-of-the-envelope calculations are not everything and should not be taken literally. While I think these numbers are somewhat useful, I worry about focusing on them too much, especially when there’s differences in rigor in the estimates, high chances of systematic model error, tricky philosophical problems like double counting and counterfactual impact, plus the estimates could be off in very outlandish ways I don’t currently understand (e.g., $20K/hr and $21K/hr both sound far too good to be true).

I also care about how doing the option builds long term success rather than just being the best thing to do in the moment -- I don’t want to get caught in a local optimum. For example, having concrete opportunities to build flexible career capital (e.g., learning generally applicable skills like management, building professional networking contacts) or opportunities to stumble upon something new. Also, personal enjoyment is important.

So when I re-look at my options, I prefer making a grid like this:

 

Looking at it this way from a few different angles, I feel better about how things stand. This exercise leads me to two concrete changes:

First, I will focus on the three highest performing activities -- my day job, cause prioritization, and Charity Science Health. I will cancel my plans to start a tech blog, avoid seeking out or expanding new opportunities to get into fundraising, avoid expanding or seeking out new opportunities to get into politics, and avoid spending a lot of time learning a lot about AI.

Second, among the three activities I am focusing on, I will prioritize them in the following waterfall -- first, I will seek to do whatever I need to do for my day job, then I will seek to do whatever is needed for Charity Science Health, and then I will seek to do whatever I can do to follow my cause prioritization agenda.

Also, Wave looks like a cool idea and it may be worth looking into that further.

In Defense of Multiprojecting

The most important accusation I think that could be leveled at me is why work on three activities instead of one or two?

I think having at least two projects makes sense to me given that I want to stay in my day job because I do not think I should focus solely on my day job. My day job works a bit backwards from my other activities in so far as I think I get a higher impact from doing less of my day job. This is because my salary will remain mostly the same regardless of how much work I do, as long as I complete the baseline required, and I win by driving this baseline down and collecting the same amount of money divided by a smaller number of hours.

This is especially true because my bonus is not tied to individual performance. The only reason that I would have to work more on my day job beyond the simple baseline is that it could get me to a raise more quickly. However, I have not seen any indication yet that working more than I currently do moves me any faster toward a raise (I already am seen as going “above and beyond” at work) and the actual size of the raise would be very minuscule compared to what I could accomplish on my other projects.

That means the real question is why not just stop at working at Charity Science Health? The answer to this is that I find that sometimes I can complete the work I’ve set out to do faster than I can get new work assigned, which leads dead time that I would otherwise like to fill up. The waterfall I’ve suggested above would address this.

Second, I’m worried somewhat about the diminishing marginal returns of work at CSH not being as important as further cause prioritization work.

Thirdly, I find that my interests between different activities can change and I find myself strongly preferring to work on cause prioritization research even if it may not be the highest impact choice. I find that I often have separate energies and that I can be refreshed from switching tasks and draw from time that otherwise would not have been used productively.

Lastly, since I’m still highly uncertain about what the best choice is, there is a large amount of value to continuing to explore the field so that I can reassess among many possible options rather than just a very limited selection. Sometimes, the only way to find out the effectiveness of an activity is to do the activity and see how it turns out.

 

-

 

Update 16 Jan: I revised the figure on the value of the vote downward by 30x, following Ben Todd's helpful comment.

Update 16 May: I revised the figure on the value of the vote downward from 30x to 50x, following Toby Ord's analysis that the developing world gets about ~100x more value per money than the developed world. I also reduced my uncertainty adjustment from 10x to 5x. The net effect of this was ~0.