On futurism as a community practice
A speculative note sent to Oliver Habryka the day after a New Year’s Eve conversation with Valentine Smith (a CFAR instructor) about whether the rationality community could develop a sharper version of the scientific method. The institutions named at the start — MIRI (the Machine Intelligence Research Institute), FHI (Oxford’s Future of Humanity Institute), FLI (the Future of Life Institute), and CSER (Cambridge’s Centre for the Study of Existential Risk) — were all then trying to make rigorous predictions about long-run technological change. Holden Karnofsky, of GiveWell, is the “Holden” mentioned below; Tetlock is the political-science forecaster of Superforecasting .
(This is an inferential chain, so the later steps will be less certain than the earlier steps.)
Suppose we did develop some kind of better scientific method, which lets people reach the truth more easily. We’d want to practice discovering specific new truths, both to hone our skills and to show that our stuff really works. But how could we do that?
If Alice is studying biology, she might use the tools of Science++ to discover cool new biological things. But that wouldn’t be very good to base a community around, because biology has tons of background knowledge that you’d have to absorb before you could even read Alice’s papers. To make it work effectively, you’d have to convert a good chunk of the whole biological community to Science++… which is not at all a bad idea, come to think of it, and Holden is working on that kind of thing right now, but it’s not what CFAR is really about.
So instead, we could use historical examples — get some older tools, and try to re-discover existing things like cells and genetics and bacteria. But then you run into the problem of hugely varying amount of knowledge. For biology and pretty much every other major field, the knowledge level of people we know varies all the way from “total noob” to “experienced professor”. So it would be hard to base a real community around that, either.
So what’s a) important, b) uncertain, and c) doesn’t have a large body of existing academic knowledge that’s accessible only to specialists? Futurism. For a), well, all of MIRI and FHI and FLI and CSER are based on the ability to make predictions about the long-term future that are 1) non-obvious and 2) might allow people to steer it. For b), the future is so poorly understood that even fairly obvious things catch almost everyone by surprise. And for c), Tetlock’s existing work shows that mostly-ordinary people can outperform the CIA with only basic training.
Of course, some people we know already talk about the future, but there’s no real means for collaboration; people just post stuff on blogs, and then other people might read it (or not). The 17th century’s solution to this problem was the academic journal. We could just copy that. But we should be able to do better, nowadays, with all our software and computation abilities…
So we could make a dedicated website for this, with each person able to create and share predictions. (Me and other MIRI interns under Anna Salamon more-or-less did this back in 2008.) But not just predictions for a certain year; those aren’t very useful. Instead, we could have predictions that come with pre-requisites: I think thing A will happen when specific things X, Y, and Z have happened. The “intelligence explosion” hypothesis is one such prediction — once AI gets to a certain point, it will very quickly spiral out of control and create superintelligence, and that’s very good to know even if we can’t attach a specific year.
In addition, the software can help solve the problem of focusing on just one particular thing (global warming! terrorism! Moore’s Law!) and ignoring all the other stuff. We could have support for trendlines — linear, polynomial, exponential — and have all the trendlines operating simultaneously , so any given year can have a snapshot of “the whole future” as extrapolated from today. We could set up a range of default trendlines, based on what we think is most plausible, and then let users have their own modifications or additions. This would help tie mid-range strategy into long-range strategy in a much more concrete way.
Hmmmm… if I wrote this software myself, is there any chance that anybody would pay me for it, so I could keep paying rent and stuff?