Debugging my apparent 2016 stagnation

If you look at my quarterly reviews over the past few quarters (here, here, and here) it does look like there’s been stagnation in a number of respects, including production and consumption. As somebody who’s frequently raised the issue of stagnation in others’ output, I’m concerned about my own stagnation.

Stagnation has been seen in the following respects:

  • Reduction in the number of Wikipedia pages created: For instance, August and September 2016 were two consecutive months where I didn’t create any Wikipedia pages at all, after a track record of having created at least one page a month for the past two years.
  • Reduction in total pageviews for the Wikipedia pages I’ve created: This has two aspects: first, total pageviews for Wikipedia pages I’ve created over my lifetime fell year-over-year in 2016. Second, the new pages I’ve created in 2016 have received much fewer pageviews in 2016 compared to what pages I created in 2015 received in 2015 (even on a per-page basis, i.e., even after controlling for the lower volume of page creation). And both are way less than the corresponding counts for 2014. To be concrete, the 24 pages I created in 2016 got 4,235 desktop pageviews in September 2016, compared to 36,940 for the 61 pages I created in 2015 in September 2015, and 79,562 for the 155 pages I created in 2014 in September 2014.
  • Reduction in the total number of blog posts published: In total, across various fora (including Open Borders: The Case, the Effective Altruism Forum, and LessWrong), I’ve published very few posts this year. The respective numbers for the first nine months of 2016 are 2, 2, and 0. The corresponding numbers for the entirety of 2015 were 18, 4 (3 of them in December), and 1.
  • Reduction in other public online activity: My rate of Facebook posting has declined significantly. I’ve become basically inactive on Quora.
  • No significant increase in number of WikiHow articles: I’ve created 5 pages in 2016, compared to 6 in 2015. My activity on WikiHow began in May 2015, so considering the lengths of the respective time periods, activity in 2016 seems lower than that in 2015.

This is interesting to me because observing stagnation in other things often causes me some concern and puzzlement. Seeing the same (apparent) phenomenon in myself gives me a unique opportunity to dig deep and understand what’s going on. That can help me better guess what might be going on in other contexts where I don’t have inside information.

I believe there are a few different things going on. I describe them below for a better understanding of the phenomenon of (apparent) stagnation.

Note: I talk a lot about people I pay to create content on Wikipedia. Although I use the first-person singular (“I”), I should clarify that a lot of the management of the editors is done by Issa Rice, so “we” could be more appropriate. However, since there is a lot of comparison with the past, and Issa’s involvement in managing editors has been significant only recently, I’ve stuck to first-person singular to make the comparisons easier.

Significant shift to producing longer and much more thoroughly researched content

One way I’ve shifted gradually since 2014 is toward producing longer pieces of content. There are three levels of hierarchy:

  • At the highest level is the overall project, e.g., “understand 19th century United States migration” or “understand website traffic analysis”. At this level, the project is essentially unending.
  • At the mid-level is an original research/synthesis deliverable, usually in the form of a blog post or WikiHow article. The piece could be several thousand words long, include many pictures in some cases, and combine many topics into a coherent whole. An example under “understand 19th century United States migration” is the blog post How did we get here? Chinese Exclusion Act buildup (1848-1872). An example under “understand website traffic analysis” is the WikiHow article How to Understand Your Website Traffic Variation with Time.
  • At the lowest level is mini-contributions, such as Wikipedia articles or Facebook posts, that fall out as a “trail” of the research. Note that Wikipedia articles don’t make sense in all cases, but they do make sense in cases where Wikipedia’s notability criteria are satisfied. For instance, in the process of writing up the post High-skilled hacks: a (very) bref overview of H-1Bs (more to follow), I wrote a number of Wikipedia articles related to the subject (as summarized in a paragraph in the post itself). The same is true of my Chinese Exclusion Act post.

One of the implications of this is that the Wikipedia articles I create have a higher ratio of (research complexity) to (expected pageviews) than the ones I used to create. There are two aspects to this phenomenon.

  • The topics themselves are more obscure, and less likely to be of interest to many people. So regardless of the content of the article, they are going to get less pageviews.
  • My goal isn’t just “create a Wikipedia page that’s good enough for casual readers of the topic” but “create a Wikipedia page and in the process research the topic throughly enough (even if all details don’t make it to Wikipedia)” because I want it to be a building block to my mid-level goal.

For instance, the Wikipedia pages on Consular nonreviewability and Systematic Alien Verification for Entitlements took me a lot more time than they would if my goal was just to create something that would pass muster with Wikipedia. The latter in particular took me a lot of time as I was pulling in from many diverse sources.

The mid-level content I produce generally signals its own thoroughness reasonably well, but the extent to which quality can be signaled does not grow in direct proportion to the effort expended. I could probably have spent about half the time I did on my WikiHow articles and still produced something that, to a casual reader, would look approximately as good. Of course, if they were looking for the specific additional things that came out of the extra time spent, it would affect their impression more. But a casual reader just surveying the article may not fully see what exactly the return-on-investment was on the effort spent. Part of this can be combated by looking at the GitHub revision history, that helps better explain the step-by-step addition of content.

Here are a couple of examples of comparisons between posts I wrote in 2016 and 2015 that were part of the same big project, so you can understand more clearly the differences in thoroughness:

  • The project of understanding migration history actually began in late 2014, with a blog post back in January 2015: How Did We Get Here? Chinese Exclusion Act — Implementation (1882-1910). That’s actually illustrative for contrast because, viewed in isolation, it seems like a decent, fairly well-researched post. However, my April 2016 post How did we get here? Chinese Exclusion Act buildup (1848-1872) took more than double that time to write, and is much more heavily researched. And yet, it’s hard to say that most readers will be able to judge the difference in isolation, without a careful side-by-side comparison of the posts.
  • The project of understanding high-skilled migration began late in 2014. My early 2015 post High-skilled hacks: the case of Optional Practical Training was fairly well-researched, but mostly piggybacked on my existing knowledge rather than involving fresh research. It also had a number of minor errors and omissions that I fixed as I understood the topic better over subsequent months of studying it. In contrast, my June 2016 post High-skilled hacks: a (very) brief overview of H-1Bs (more to follow) required well over double the research effort, and again it’s unclear that a casual reader would experience a significantly different affect on the second post than the first.

Net effect: Fewer articles created and fewer pageviews per article created.

Heavier outsourcing of “pop” content creation and thematic content creation

I’ve cut down quite a bit on creating “pop” content on Wikipedia, i.e., creating articles about popular topics. I still believe they are valuable to create, but they don’t require my skills and expertise. Part of my sponsored Wikipedia editing focuses on such content. For instance, last week I asked Alex K. Chen to make the Wikipedia page for Michael Seibel, which he did.

There is another category of content that I have successfully outsourced, which is “thematic” content creation. Here, we identify a particular type of article, with some variable parameters (for instance, timeline of healthcare in country X, where X can vary over the 200+ countries). I closely supervise and provide feedback on the first few articles with that theme, but then the person creating those articles basically knows the procedure and can keep doing it for other countries. I’ve had success using this approach with Jesse Clifton (who’s made pages on animal welfare and rights by country), Sebastian Sanchez (who’s made pages on timeline of healthcare by country, and timelines of diseases), and Alex K. Chen (who has covered timelines for tech companies and has recently expanded to other companies).

Even with thematic content creation, there’s a difference between my focus areas in 2015 and my focus areas in 2016. In 2015, when I was trying to establish the tractability of thematic content creation, my focus was on topics I expected to be more popular. Hence, I focused more on timelines for technology companies, with timelines created by me (in 2014) and later by Alex K. Chen (starting 2015). In 2016, I picked more exotic themes, such as animal welfare and rights by country, timeline of healthcare by country, and timeline of diseases. Why the switch? Partly, the topics selected in 2016 were ones that formed part of a bigger picture of interest. Specifically, healthcare and diseases are important because a lot of people believe that health interventions offer very high value-for-philanthropic-money. I wanted to get deeper background on healthcare and diseases, and funding Sebastian’s work on these timelines is helping with that process. Animal welfare and rights by country was similarly motivated. As a result, if you look at total pageviews for thematic content creation I’ve funded in 2016, it doesn’t come out that different than 2015, even though the number of such pages funded has increased significantly.

Net effect: The changes from 2014 to 2015 shifted article creation away from me to the people I funded. The changes from 2015 to 2016 caused a decline in the incremental pageviews for new articles, despite an increase in the number of articles created.

Outsourcing of some other low-level mini-contributions

After Issa Rice created the timeline of global health, he and I identified a number of missing Wikipedia articles on related topics. Some of these articles were subsequently created by Issa. Others were given out to other contract workers. Many of them are interesting mini-projects for whoever chooses to do them. The goal is to improve our understanding of global health, for reasons discussed earlier in this post: namely, a lot of people thing health interventions are very good value-for-money in philanthropy.

Other areas where we outsourced some content creation include philanthropic foundations and the United States welfare system. You can go through the GitHub revision history of the new article pool to see how articles entered the pool, were reserved for specific editors, and then got removed as the article got completed.

Net effect: Even adding in articles whose creation I funded does not boost pageviews significantly, because many of these were low-level mini-contributions to grander projects. They require effort to create and don’t get a lot of pageviews.

Switch away from social media and public solicitation

This year, I’ve switched a lot to paying people to do stuff. In previous years, my model was more along the lines of doing stuff myself (or paying a couple of others, specifically Issa Rice and Alex K. Chen). I relied on public solicitation of feedback and thoughts, plus discussion with Alex and Issa and occasionally relevant subject-matter experts.

Now, a lot of the discussion happens privately. Issa and I discuss the topic and potential sources with the editor in question. We might sometimes seek thoughts from the general public (through Facebook) but that’s often not the case. This is mostly because the topics are more obscure, and we don’t expect most people to know things about them that are hard to Google.

Net effect: Less activity on social media, giving appearance of less work.

Some other measurement confounders

In January 2016, I switched counting of Wikipedia pageviews to the Wikimedia API, excluding bot views. Views till 2015 had included bot views. The transition to no bots led to a decline in estimated pageviews. Also, since it affects pages with fewer overall pageviews hardest, it makes the performance of my 2016 pages look worse compared to 2015. With that said, even adjusting for this still shows much less traction achieved in 2016 by pages I created.

Overall, are the changes “good” changes? Are they working out well?

For the most part, the apparent decline is more a story of a change in strategy than an actual decline. But is the change in strategy an improvement? Unfortunately, at this point it’s hard to be conclusive. My intuition is that a lot of these changes are improvements. However, you can differ in terms of values and empirics. For instance, if all pageviews are weighted equally, then the switch to the sort of pages I’ve created and funded recently is worse than sticking to pages like the old ones. On the empirics side, you might have a low opinion of the quality of the content that I’ve created and funded of late, or at least, think it is not of better quality than past, much more quickly produced, content.

Was my past approach wrong for that era?

Although I believe that my present approach to content creation makes more sense now, I do not mean to say that my past approach to content creation was wrong. In fact, I’d say that my original approach to content creation was probably a good fit for that stage of things. It provided a proof of concept in a relatively low-stakes environment (one-off page creations without undertaking a larger project). Also, by selecting more popular pages, I was able to achieve pageview counts that would come across as reasonably impressive to casual observers. But the current approach makes more sense as part of a plan to drive long-term value. Or so it seems to me, right now.

Implications for understanding apparent stagnation in others

One lesson is that increasing depth somewhat can result in a substantial reduction in apparent output, without the increases in quality being blatantly obvious to a casual observer. Whether the increases in quality are still worthwhile can depend on context, but casual observation won’t be a very reliable guide to that.

I believe one implication is to look more closely at revealed preference, especially preference in terms of financial investment and depth of engagement. The fact that I’m now paying significantly more (funding others’ work in addition to doing stuff on my own) is some evidence that I have good grounds for believing I am making progress. Similarly, if evaluating others’ progress, the extent to which they are able to draw people to work for and engage with them, and the extent to which they are able to raise money, can be stronger indicators than just looking at raw numeric counts of the amount produced or consumed. I’m not too satisfied with this answer, primarily because there is also a flip side: optimism bias pushes people to incorrectly believe that things are improving and that they are getting better, regardless of whether they actually are (and that’s a bias I might be under the influence of, too).