Thoughts on the AI Impacts Site
Sent to AI safety researchers Paul Christiano and Katja Grace as they were planning AI Impacts — a public-facing research site for forecasting questions about advanced AI — with notes on how to design it so that it would actually attract returning readers and contributors, rather than going the way of the obscure Singularity FAQ Alyssa had co-written years earlier.
On the one hand, there clearly is a need for something like this. On the other, this unfortunately seems to be a domain where projects can fail for silly reasons, and where failure only becomes clear after projects have lots of time and effort invested in them. There are many examples, but, e.g., one that I’m familiar with is the Singularity FAQ written in… 2007? 2008? (I forget the exact date) by me and Kaj Sotala. We spent a few months coming up with questions, answering the questions, looking around for sources to support our arguments, and so on. Eventually, we published it, but as I remember it never got any significant web traffic. Looking back, I don’t think it helped save the world much.
My intuition is that, for websites, there is a sharp cutoff between ones that get visited a lot and ones that almost never do, because any given person can only remember so many sites at a time. So if you use the site a lot, you remember the URL and you keep coming back, but if you only use the site once or twice you don’t and you never come back (even if you’d in theory prefer to visit now and then). There seem to be two basic approaches: have enough high-quality, engaging content that people can read something, stop, and still feel like they want to come back the next day and read more; have the site content change frequently and visibly enough that people check back for updates.
Wikipedia and HPMOR follow approach #1, while news websites follow approach #2. I’d recommend approach #2 here, since the barrier to entry for #1 is generally fairly high (~thousands of pages of well-written and engaging content, which in this case is on a subject that is largely unexplored because few have historically been interested).
The simplest way to do #2 is with a blog, but while that can be quite successful at getting readers, it creates the problem that information is only sorted chronologically, so there is no easy way to explore items on any particular topic in rough order of importance or conceptual dependency. The successful finance website Mergers & Inquisitions solves this by having a feed of new material which is regularly updated, but by not displaying the date of a post and by having a menu where items are sorted by topic, so that it’s easy for people to find information on what interests them.
I think I like that strategy. However, it’s really only a strategy for effectively communicating the results of existing research, not for producing more research; to do that you also need a place where people can collaborate well. I’m not exactly sure how to do that effectively — perhaps we could ask some people in academia with lots of wide-audience-remote-collaboration experience — but my first guess is something like a hybrid blog/wiki system: have a network of wiki-like pages, with a core group of contributors who can edit, and with accounts available to the general public that can add comments; have some ( but not all ) edits to wiki pages and new pages be posted to a chronologically ordered feed, and there can be both a unified overall feed and sub-feeds segregated by topic; it’s very important that updates which go into a feed have high median quality , so that people will continue to pay attention; e.g. in a lot of systems, you get auto-subscribed to any comment about an article, and since a lot of comments are small, off-the-cuff, heavily rely on surrounding context to make sense, or are just generally uninteresting, people quickly learn to tune out the entire feed; you’d likely want to do this by enforcing minimum character count limits and likely other constraints; have a manually maintained front page and topic pages, which include both the auto-generated recent update feeds and a list of the most important pages within the topic, in order of conceptual dependence; this provides an easy path for people who are new to a particular topic to begin exploring it; for each user, have a publicly visible user page with their email address; when they write a comment, create a page, or add a block of text (as opposed to minor page edits), provide an overlay making it easy to click on the text and navigate to their user page, so it’s easy for you to contact someone if you’re interested in a particular thing they said; have a system of promoted comments, so that especially long or interesting comments get taken out of the comments section and (after editing) included in the page and the updates feed, and where people who make a lot of such comments automatically get editing powers; have limited email auto subscription, where if you comment on/edit/add a page in a given topic, there is opt-out subscription to the update feed for that topic and opt-in subscription to all comments about the topic or all comments on any particular page; once a given topic has gotten sufficiently popular (several dozen people contributing regularly), start having regularly scheduled hangouts or in-person meetups for discussion, although it’s also important to not try to force this if the audience isn’t yet big enough for people to be interested (this will just lead to getting demoralized).