This is the 2nd part of my personal and qualitative self-review of MIRI in 2013.
By “outreach” I refer to general outreach efforts, rather than e.g. outreach to specific researchers, which will be discussed in the post about MIRI’s 2013 research activities.
Outeach in 2013
- In early 2013, we decided to reduce our outreach efforts significantly, and we did.
- However, we learned throughout 2013 that some forms of “indirect” outreach tend to be pretty cost-effective for MIRI’s goals.1
- Therefore, we plan to put more effort into indirect outreach in 2014 than we did in 2013.
Reducing our outreach efforts
MIRI provisionally decided to reduce its outreach efforts in mid-2012, when we decided to enter negotiations with Singularity University (SU) in response to their offer to acquire the Singularity Summit, a deal finally closed in December 2012. The reduction in outreach efforts explained in our 2013 strategy post was largely a continuation of this earlier strategic decision, which had been conditional on the deal with SU closing successfully. (We would have continued to run the Singularity Summit had SU not acquired it.)
MIRI will co-produce the Singularity Summit for the next couple years, but this requires minimal effort on our part, and the scheduling of future Summits is in SU’s hands. SU has not yet scheduled the next Singularity Summit.
Without a Singularity Summit occurring in 2013, our direct outreach efforts this year were few. By “direct outreach efforts” I refer to efforts that attempt to directly expose people to MIRI’s mission focused on the superintelligence control problem. This is in contrast to “indirect outreach efforts,” which aim to expose people not to MIRI’s mission, but to other ideas that may begin to bridge the inferential distance that lies between common knowledge and the core ideas motivating MIRI’s mission — ideas such as effective altruism, applied rationality, global catastrophic risks, the security mindset, etc.
Our direct outreach efforts in 2013 were:
- A new website, a necessary part of our rebranding as the Machine Intelligence Research Institute, and some increased use of social media.
- A few general-audience talks about superintelligence control, given at universities and technology companies (Stanford, Oxford, Facebook, Quixey, and Heroku), and also at the Effective Altruism Summit.
- Two Bay Area events we hosted for our community: our April “relaunch as MIRI” event, and our September book launch event for James Barrat’s Our Final Invention. Two MIRI staffers also published reviews of Barrat’s book, here and here.
- Two new ebooks containing previously-written content: Facing the Intelligence Explosion and The Hanson-Yudkowsky AI-Foom Debate.
- We also attended several events for networking purposes, and had conversations with interested parties who contacted us, and thereby spoke to many people about superintelligence control.
- Various small writings and interviews, e.g. a short article for Quartz and an interview for Big Picture Science.
It is difficult to measure how much value these efforts are producing for our long-term goals. Some basic statistics on these efforts are recorded in a footnote.2
Overall, my own impressions on the value of these direct research efforts are as follows:
- The new domain name (intelligence.org) and website seem clearly worthwhile. Our limited use of social media also seems worthwhile, not to grow our community (because: too much inferential distance), but to keep our existing community engaged with MIRI’s work. (For many people, our Facebook updates are the easiest way to stay in touch with MIRI.)
- Our superintelligence control talks at the Effective Altruism Summit were worthwhile because there was less inferential distance with that audience. We also know that some donations came as a result of those talks. In contrast, we haven’t yet received any specific evidence that the other superintelligence control talks produced value toward our mission, other than by keeping us in touch with our existing community. We didn’t strongly expect to see such evidence so soon, but this still constitutes non-negligible evidence of absence of positive outreach effects.
- The two Bay Area events were valuable for staying in touch with our existing community, but they failed to produce major press stories even though we invited several major journalists to each. The events cost more staff time and attention than money. Our two book reviews of Our Final Invention are too recent for us to know whether they caused useful action (beyond merely purchasing the book) in their readers.
- The two ebooks of pre-existing content were inexpensive to produce, and they make our material more accessible. They are too recent for us to have specific evidence of positive effects, but I’d guess they were worth the expense.
- As far as I can tell, the networking events at which we spoke to people about superintelligence control (rather than effective altruism, rationality, etc.) seem to have produced negligible value. We experimented with several presentations of our material, but I think the inferential distance is just too great to engage new people who aren’t already familiar with effective altruism, applied rationality, global catastrophic risks, etc.
Indirect outreach efforts were more clearly successful
Our indirect outreach efforts in 2013 were also aimed at long-term effects:
- Eliezer Yudkowsky wrote new chapters of Harry Potter and the Methods of Rationality (HPMoR), and MIRI distributed 2500+ paperback copies of its first 17 chapters to schools, universities, and companies.3
- Carl Shulman contributed research and career counseling services to the Center for Effective Altruism (CEA).
- We produced several front-page posts for the Less Wrong community, e.g. a Decision Theory F.A.Q.4
The positive effects of our posts to Less Wrong are hard to measure, but our other indirect outreach efforts have produced clear evidence of their value more quickly than our direct outreach efforts have.
Carl’s work with CEA had measurable effects on the funding and attention going into long-term philanthropy (including the superintelligence control problem).
Finally, some of the evidence we’ve gotten about the indirect outreach value of HPMoR:
- When I observe someone doing something useful to MIRI’s mission (e.g. donation or research) for the first time, I typically ask how they came to be involved, and I trace the story back to their first contact with someone at MIRI, with one of MIRI’s close friends, or to something that MIRI or MIRI’s close friends produced. HPMoR is now the single most common “first contact” I encounter. (The 2nd most common first contact is with The Sequences.)
- The paperback version of HPMoR ends in the middle of chapter 17, meaning that readers coming to HPMoR.com after finishing the paperback should jump directly to chapter 17. According to Google Analytics, people visit chapter 17 significantly more often than any other chapter outside the first 10 chapters and newly released chapters. It’s hard to tell exactly, but my best guess is that 100-400 people per month are coming to hpmor.com as a result of first reading the paperback copy of the first 17 chapters.
These data fit with MIRI’s earlier experience of the effects of its direct and indirect outreach efforts. For example, as I said earlier: Eliezer Yudkowsky spent MIRI’s early years appealing directly to people about the importance of superintelligence control research. Some helpful people found his work, but the audience was being filtered for “interest in future technology” rather than “able to reason well,” and thus when Eliezer would make basic arguments for e.g. the orthogonality thesis or convergent instrumental values, the responses he would get were typically akin to the output of a free association task. So Eliezer wrote The Sequences and HPMoR, for which the primary audience filter is “interest in improving one’s reasoning.” This audience, in our experience, is more likely to do useful things when we present the case for effective altruism, for x-risk reduction, for FAI research, etc.
Lessons for 2014
Today I find myself even more confident than I was in April 2013 that certain kinds of indirect outreach by MIRI are more likely to produce value for superintelligence control than MIRI’s direct outreach efforts are. For this reason, I suspect that in 2014 we will continue to scale back our direct outreach efforts.5 Meanwhile, we are likely to increase our indirect outreach efforts, in particular by giving Eliezer time to write more HPMoR chapters, by distributing more HPMoR paperbacks, and by publishing an ebook of The Sequences.
We will also likely provide some consulting and other support for organizations who are doing the kinds of outreach that end up benefiting MIRI (and other organizations), for example CFAR (rationality) and CEA (effective altruism).
- We are currently gathering much more information about the effects of our direct and indirect outreach efforts, but these data will take several months to gather. It’s possible those data will overturn the basic conclusions in this post, but it’s more likely they will slightly adjust them. ↩
- Basic statistics on direct outreach efforts: (1) Our new intelligence.org domain name rose from PageRank 0 to PageRank 5, one rank below LessWrong.com, a site that has been heavily linked and active since 2009. (2) Since it launched in March 2013, intelligence.org has averaged 28,613 unique visitors per month. This is roughly double the traffic to our previous domain, singularity.org, which averaged 14,843 unique visitors per month during its operation (July 2012 through February 2013). (3) Our other “mini-sites” got negligible traffic, as expected. During their active months in 2013, here’s how many unique visitors per month our mini-sites got: IntelligenceExplosion.com (2,336/mo), Friendly-AI.com (50/mo), TheUncertainFuture.com (129/mo). (4) As of Jan. 2nd, MIRI’s Facebook page had ~5000 Likes, had been seen by ~19,500 unique people, and had “engaged” ~1100 unique people (meaning that they liked, commented on, or clicked on MIRI’s Facebook wall posts). MIRI’s Google+ page is more recent and was much less active: with ~230 +1s (analogous to Facebook Likes), and 17 re-shares. Note that our Facebook Likes count appears to be artificially high. From Oct. 9th to Nov. 9th, MIRI’s Facebook page received an unusual number of Likes. This began again on Dec. 22nd. We don’t know the cause. (5) Our newsletter subscription count grew only slightly, from 9,152 subscribers in January 2013 to 10,024 subscribers as of December 22nd. (6) Our April 2013 relaunch event was attended by ~65 people. The September 2013 book launch event was attended by ~75 people. In both cases, most attendees were close friends of MIRI. (7) Within one month of its publication, Luke’s review of Our Final Invention on KurzweilAI.net saw ~8,500 unique pageviews and was responsible for at least 20 purchases of the book, which describes MIRI’s mission and work in some detail. Meanwhile, Louie’s review of Our Final Invention for Singularity Hub was responsible for at least 60 purchases of the book as of Jan. 2nd, 2014. (8) From April-Nov 2013, Facing the Intelligence Explosion was purchased 1,379 times. Since September, The Hanson-Yudkowsky AI-Foom Debate has been downloaded at least 1,240 times, but this counts only downloads from people who clicked a link to the book from intelligence.org. (9) How many people attended our general-audience talks? By memory, I’d guess 40 at Stanford, 30 at Oxford, 20 at Facebook, 40 at Quixey, 15 at Heroku, and 50 at the Effective Altruism Summit. These were smaller, more targeted talks than e.g. our past talks at the Singularity Summits. ↩
- HPMoR is Yudkowsky’s private project, but MIRI sponsored this work in the sense that Yudkowsky remained on salary at MIRI while he took time off from “direct MIRI work” to write and release new chapters of HPMoR. ↩
- MIRI’s substantive, outreach-oriented, front-page posts for the Less Wrong community in 2013 were: (1) Decision Theory F.A.Q.; (2) Fermi Estimates; (3) Start Under the Streetlight, then Push into the Shadows; (4) Four Focus Areas of Effective Altruism; (5) How to Measure Anything; (6) MetaMed: Evidence-Based Healthcare. ↩
- There will probably be at least two major exceptions to this trend. First, in early 2014 we will publish an ebook called Smarter Than Us, which was written by Stuart Armstrong all the way back in late 2012. We just didn’t have staff time to turn it into an ebook until very recently. Second, we will probably publish something about the upcoming film Transcendence, because that is a unique opportunity to direct many eyeballs to our writings. Also, we will likely assist in the promotion of Nick Bostrom’s Superintelligence book, but this really qualifies as “outreach to researchers” rather than “general outreach” because Bostrom’s book is a scholarly monograph. ↩