CFAR Exercise Prize

From LessWrong
Jump to navigation Jump to search

The Exercise Prize is an open prize offer by the Center for Applied Rationality for exercises that teach a cognitive skill:

  • An immediate $50 for any suggestion which seems good enough that we actually test it, whether or not it works out.
  • An additional $500 for any suggestion adopted into an actual unit.

An archetypal case of a good suggestion for an exercise is Andrew Critch's Bayes game, as recounted here by Eliezer:

On Saturday I tested an 80-minute unit on Bayes's Rule. The audience were mostly either programmers or in other technical subjects, so I could go through the math fairly fast. Afterward I was worried that they hadn't really learned to apply Bayes's Rule and wished I had a small little pamphlet of practice problems to hand out.
On Wednesday, I attended Andrew Critch's course at Berkeley, LW-style cognitive-improvement material aimed at math students; and in this particular session, Critch introduced Bayes's Theorem, not as advanced math, but with the aim of getting the students to apply it to real life.
In particular, Critch demonstrated what he called the Really Getting Bayes game. He had Nisan (a local LWer) touch an object to the back of Critch's neck, a cellphone as it happened, while Critch faced in the other direction; this was "prior experience". Nisan said that the object was either a cellphone or a pen. Critch gave prior odds of 60% : 40% that the object was a cellphone vs. pen, based on his prior experience. Nisan then asked Critch how likely he thought it was that a cellphone or a pen would be RGB-colored, i.e., colored red, green, or blue. Critch didn't give exact numbers here, but said he thought a cellphone was more likely to be primary-colored, and drew some rectangles on the blackboard to illustrate the likelihood ratio. After being told that the object was in fact primary-colored (the cellphone was metallic blue), Critch gave posterior odds of 75% : 25% in favor of the cellphone, and then turned around to look.
Then Critch broke up the class into pairs and asked each pair to carry out a similar operation on each other: Pick two plausible objects and make sure you're holding at least one of them; touch the object to the other person while they face the other way; name prior odds; be told an additional fact; name the likelihood ratio; then state the posterior odds, and possibly adjust them.
This is the sort of in-person, hands-on, real-life, and social exercise that didn't occur to me, or Anna, or anyone else helping, while we were trying to design the Bayes's Theorem unit. Our brains just didn't go in that direction, though we recognized it as embarrassingly obvious in retrospect.

Critch's game was in fact adopted into the Bayes unit at Minicamp, and several campers said afterward that it was a key moment in comprehension.

The key idea about an exercise is that it should teach a useful skill by performance - not just illustrate an insight. In this case, the skill is to think about prior odds, think about likelihood ratios, multiply them, and check the result to see if it makes sense - in a highly experiential (sensory-motor) context which, we hope, makes it more likely that this skill will actually be performed in real life.

Prizes are payable via Paypal (or we can essay other payment methods if necessary). If you win, private-message Stephen Cole from the LessWrong account that won (so we know it's you) with your payment details. Considering the wide variety of skills we're trying to teach, there are no prespecified success metrics besides "Looks good enough to try" and "We decide to keep it".

The Center for Applied Rationality potentially has resources that can "build a simple computer game" or "come up with specific exercises along general pattern X", or even remote workers who can "research large numbers of trivia questions with known answers to use as tests". Suggestions of this type will have to look good in advance to be tested, since testing them requires resources, but shouldn't be eliminated from consideration. On the other hand, "Hire an author to write a novel that..." is too resource-intensive.

Suggestions that are intended to be performed at local Less Wrong meetups, rather than CfAR units, will also be considered eligible for prizes - $50 for a suggestion good enough to be tested, an additional $500 for a suggestion good enough to be adopted.

The prize also extends to unusually good suggestions for how to verify that a skill has been acquired.

Past requests for exercises, in chronological order from oldest to newest:

These are still open as requests, but except for the current newest post, they may be checked infrequently. Private message Eliezer_Yudkowsky to remind him to look at yours, if you're posting to an old thread.

Past winners of the Exercise Prize - suggestions that have been both tested ($50) and adopted ($500):

Miscellaneous $50 prizes have also been awarded for suggested tweaks to major suggestions, and ideas which were promising enough to test but didn't quite work out.

There is presently no system in place to accept fully general exercise suggestions. Feel free to PM or email Eliezer Yudkowsky or Anna Salamon if you have a burning insight that calls for this to be fixed sooner rather than later.