A bit about our last few months:
- We’ve been working on getting a simple clear mission and an organization that actually works. We think of our goal as analogous to the transition that the old Singularity Institute underwent under Lukeprog (during which chaos was replaced by a simple, intelligible structure that made it easier to turn effort into forward motion).
- As part of that, we’ll need to find a way to be intelligible.
- This is the first of several blog posts aimed at causing our new form to be visible from outside. (If you're in the Bay Area, you can also come meet us at tonight's open house.) (We'll be talking more about the causes of this mission-change; the extent to which it is in fact a change, etc. in an upcoming post.)
Here's a short explanation of our new mission:
-
We care a lot about AI Safety efforts in particular, and about otherwise increasing the odds that humanity reaches the stars.
-
Also, we[1] believe such efforts are bottlenecked more by our collective epistemology, than by the number of people who verbally endorse or act on "AI Safety", or any other "spreadable viewpoint" disconnected from its derivation.
-
Our aim is therefore to find ways of improving both individual thinking skill, and the modes of thinking and social fabric that allow people to think together. And to do this among the relatively small sets of people tackling existential risk.
Existential wins and AI safety
By an “existential win”, we mean humanity creates a stable, positive future. We care a heck of a lot about this one.
Our working model here accords roughly with the model in Nick Bostrom’s book Superintelligence. In particular, we believe that if general artificial intelligence is at some point invented, it will be an enormously big deal.
(Lately, AI Safety is being discussed by everyone from The Economist to Newsweek to Obama to an open letter from eight thousand. But we’ve been thinking on this, and backchaining partly from it, since before that.)
Who we’re focusing on, why
Our preliminary investigations agree with The Onion’s; despite some looking, we have found no ultra-competent group of people behind the scenes who have fully got things covered.
What we have found are:
- AI and machine learning graduate students, researchers, project-managers, etc. who care; who can think; and who are interested in thinking better;
- Students and others affiliated with the “Effective Altruism” movement, who are looking to direct their careers in ways that can do the most good;
- Rationality geeks, who are interested in seriously working to understand how the heck thinking works when it works, and how to make it work even in domains as confusing as AI safety.
These folks, we suspect, are the ones who can give humanity the most boost in its survival-odds per dollar of CFAR’s present efforts (which is a statement partly about us, but so it goes). We’ve been focusing on them.
(For the sake of everyone. Would you rather: (a) have bad rationality skills yourself; or (b) be killed by a scientist or policy-maker who also had bad rationality skills?)
Brier-boosting, not Signal-boosting
Everyone thinks they’re right. We do, too. So we have some temptation to take our own favorite current models of AI Safety strategy and to try to get everyone else to shut up about their models and believe ours instead.
This understandably popular activity is often called “signal boosting”, “raising awareness”, or doing “outreach”.
At CFAR, though, we force ourselves not to do “signal boosting” in this way. Our strategy is to spread general-purpose thinking skills, not our current opinions. It is important that we get the truth-seeking skills themselves to snowball across relevant players, because ultimately, creating a safe AI (or otherwise securing an existential win) is a research problem. Nobody, today, has copyable opinions that will get us there.
We like to call this “Brier boosting”, because a “Brier score” is a measure of predictive accuracy.
[1] By "We believe X", we do not mean to assert that every CFAR staff member individually believes X. (Similarly for "We care about Y). We mean rather that CFAR as an organization is planning/acting as though X is true. (Much as if CFAR promises you a rationality T-shirt, that isn't an individual promise from each of the individuals at CFAR; it is rather a promise from the organization as such.)
If we're going to build an art of rationality, we'll need to figure out how to create an organization where people can individually believe whatever the heck they end up actually believing as they chase the evidence, while also having the organization qua organization be predictable/intelligible.
ETA:
You may also want to check out two documents we posted in the days since this post:
- Further discussion of CFAR’s focus on AI safety, and the good things folks wanted from “cause neutrality”
- CFAR's mission statement (link post, linking to our website).
A few nitpicks on choice of "Brier-boosting" as a description of CFAR's approach:
Predictive power is maximized when Brier score is minimized
Brier score is the sum of differences between probabilities assigned to events and indicator variables that are are 1 or 0 according to whether the event did or did not occur. Good calibration therefore corresponds to minimizing Brier score rather than maximizing it, and "Brier-boosting" suggests maximization.
What's referred to as "quadratic score" is essentially the same as the negative of Brier score, and so maximizing quadratic score corresponds to maximizing predictive power.
Brier score fails to capture our intuitions about assignment of small probabilities
A more substantive point is that even though the Brier score is minimized by being well-calibrated, the way in which it varies with the probability assigned to an event does not correspond to our intuitions about how good a probabilistic prediction is. For example, suppose four observers A, B, C and D assigned probabilities 0.5, 0.4, 0.01 and 0.000001 (respectively) to an event E occurring and the event turns out to occur. Intuitively, B's prediction is only slightly worse than A's prediction, whereas D's prediction is much worse than C's prediction. But the difference between the increase in B's Brier score and A's Brier score is 0.36 - 0.25 = 0.11, which is much larger than corresponding difference for D and C, which is approximately 0.02.
Brier score is not constant across mathematically equivalent formulations of the same prediction
Suppose that a basketball player is to make three free throws, observer A predicts that the player makes each one with probability p and suppose that observer B accepts observer A's estimate and notes that this implies that the probability that the player makes all three free throws is p^3, and so makes that prediction.
Then if the player makes all three free throws, observer A's Brier score increases by
3*(1 - p)^2
while observer B's Brier score increases by
(1 - p^3)^2
But these two expressions are not equal in general, e.g. for p = 0.9 the first is 0.03 and the second is 0.073441. So changes to Brier score depend on the formulation of a prediction as opposed to the prediction itself.
======
The logarithmic scoring rule handles small probabilities well, and is invariant under changing the representation of a prediction, and so is preferred. I first learned of this from Eliezer's essay A Technical Explanation of a Technical Explanation.
Minimizing logarithmic score is equivalent to maximizing the likelihood function for logistic regression / binary classification. Unfortunately, the phrase "likelihood boosting" has one more syllable than "Brier boosting" and doesn't have same alliterative ring to it, so I don't have an actionable alternative suggestion :P.
I don't think the first problem is a big deal. No-one worries about "I boosted that from a Priority 3 to a Priority 1 bug".