Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

Comment author: Jiro 08 July 2015 02:32:26PM 6 points [-]

Okay, a summary of my attitude towards EA is that EA rationally follows from a set of weird premises that are not shared by most people and certainly not by me. I do not have any desire to maximize utility in a way that considers utility for every human being equally. I prefer increasing utility for myself, my family, friends, countrymen, and people like me. Every time I pay for electricity for my computer rather than sending the money to a third world peasant is, according to EA, a failure to maximize utility.

Also, I believe that most cases of EA producing very counterintuitive results are just examples of cases where the weirdness of EA becomes obvious.

Comment author: benkuhn 12 July 2015 03:23:49AM 6 points [-]

Every time I pay for electricity for my computer rather than sending the money to a third world peasant is, according to EA, a failure to maximize utility.

I'm sad that people still think EAers endorse such a naive and short-time-horizon type of optimizing utility. It would obviously not optimize any reasonable utility function over a reasonable timeframe for you to stop paying for electricity for your computer.

More generally, I think most EAers have a much more sophisticated understanding of their values, and the psychology of optimizing them, than you give them credit for. As far as I know, nobody who identifies with EA routinely makes individual decisions between personal purchases and donating. Instead, most people allocate a "charity budget" periodically and make sure they feel ok about both the charity budget and the amount they spend on themselves. Very few people, if any, cut personal spending to the point where they have to worry about, e.g., electricity bills.

Comment author: ChaosMote 04 June 2015 04:34:28AM 1 point [-]

Your argument assumes that the algorithm and the prisons have access to the same data. This need not be the case - in particular, if a prison bribes a judge to over-convict, the algorithm will be (incorrectly) relying on said conviction as data, skewing the predicted recidivism measure.

That said, the perverse incentive you mentioned is absolutely in play as well.

Comment author: benkuhn 04 June 2015 07:07:57AM 0 points [-]

Yes, I glossed over the possibility of prisons bribing judges to screw up the data set. That's because the extremely small influence of marginal data points and the cost of bribing judges would make such a strategy incredibly expensive.

Comment author: Davidmanheim 03 June 2015 04:31:22AM 8 points [-]

The incentive to try "high volatility" methods seems like an advantage; if many prisons try them, 20% of them would succeed, and we'd learn how to rehabilitate better.

Comment author: benkuhn 03 June 2015 10:49:53PM 0 points [-]

Yep. Concretely, if you take one year to decide that each negative reform has been negative, the 20-80 trade that the OP posts is a net positive to society if you expect the improvement to stay around for 4 years.

Comment author: ChaosMote 03 June 2015 01:44:15AM 20 points [-]

Great suggestion! That said, in light of your first paragraph, I'd like to point out a couple of issues. I came up with most of these by asking the questions "What exactly are you trying to encourage? What exactly are you incentivising? What differences are there between the two, and what would make those difference significant?"

You are trying to encourage prisons to rehabilitate their inmates. If, for a given prisoner, we use p to represent their propensity towards recidivism and a to represent their actual recidivism, rehabilitation is represented by p-a. Of course, we can't actually measure these values, so we use proxies; anticipated recidivism according to your algorithm and re-conviction rate (we'll call these p' and a', respectively).

With this incentive scheme, our prisons have three incentives: increasing p'-p, increasing p-a, and increasing a-a'. The first and last can lead to some problematic incentives.

To increase p'-p, prisons need to incarcerate prisoners which are less prone to recidivism than predicted. Given that past criminality is an excellent predictor of future criminality, this leads to a perverse incentive towards incarcerating those who were unfairly convicted (wrongly convicted innocents or over-convinced lesser offenders). If said prisons can influence the judges supplying their inmates, this may lead to judges being bribed to aggressively convict edge-cases or even outright innocents, and to convict lesser offenses of crimes more correlated with recidivism. (Counterpoint: We already have this problem, so this perverse incentive might not be making things much worse than they already are.)

To increase a-a', prisons need to reduce the probability of re-conviction relative to recidivism. At the comically amoral end, this can lead to prisons teaching inmates "how not to get caught." Even if that doesn't happen, I can see prisons handing out their lawyer's business cards to released inmates. "We are invested in making you a contributing member of society. If you are ever in trouble, let us know - we might be able to help you get back on track." (Counterpoint: Some of these tactics are likely to be too expensive to be worthwhile, even ignoring morality issues.)

Also, since you are incentivising improvement but not disincentivizing regression, prisons who are below-average are encouraged to try high-volatility reforms even if they would yield negative expected improvement. For example, if a reform has a 20% chance of making things much better but a 80% chance of making things equally worse, it is still a good business decision (since the latter consequence does not carry any costs).

Comment author: benkuhn 03 June 2015 10:48:52PM 2 points [-]

To increase p'-p, prisons need to incarcerate prisoners which are less prone to recidivism than predicted. Given that past criminality is an excellent predictor of future criminality, this leads to a perverse incentive towards incarcerating those who were unfairly convicted (wrongly convicted innocents or over-convinced lesser offenders).

If past criminality is a predictor of future criminality, then it should be included in the state's predictive model of recidivism, which would fix the predictions. The actual perverse incentive here is for the prisons to reverse-engineer the predicted model, figure out where it's consistently wrong, and then lobby to incarcerate (relatively) more of those people. Given that (a) data science is not the core competency of prison operators; (b) prisons will make it obvious when they find vulnerabilities in the model; and (c) the model can be re-trained faster than the prison lobbying cycle, it doesn't seem like this perverse incentive is actually that bad.

Comment author: othercriteria 14 May 2015 04:38:11AM 4 points [-]

Given that at least 25% of respondents listed $0 in charity, the offset you add to the charity ($1 if I understand log1p correctly) seems like it could have a large effect on your conclusions. You may want to do some sensitivity checks by raising the offset to, say, $10 or $100 or something else where a respondent might round their giving down to $0 and see if anything changes.

Comment author: benkuhn 19 May 2015 04:16:56AM 5 points [-]

Gwern has a point that it's pretty trivial to run this robustness check yourself if you're worried. I ran it. Changing the $1 to $100 reduces the coefficient of EA from about 1.8 to 1.0 (1.3 sigma), and moving to $1000 reduces it from 1.0 to 0.5 (about two sigma). The coefficient remains highly significant in all cases, and in fact becomes more significant with the higher constant in the log.

Comment author: gwern 16 May 2015 04:00:55PM 0 points [-]

It's a nonlinear transformation to turn nonlinear totals back into something which is linear, and it does so very well, as you can see by comparing the log graph with an unlogged graph. Again, I'm not seeing what the problem here is. What do you think this changes? Ordering is preserved, zeros are preserved, and dollar amounts become linear which avoids a lot of potential problems with the usual statistical machinery.

Comment author: benkuhn 18 May 2015 12:00:41AM 0 points [-]

What do you mean by "dollar amounts become linear"? I haven't seen a random variable referred to as "linear" before (on its own, without reference to another variable as in "y is linear in x").

Comment author: Baisius 18 April 2015 05:16:33AM 0 points [-]

Also, as I understand, it's actually better not to cancel the cards you sign up for (unless they have an annual fee), because "average age of credit line" is a factor in the FICO score. Snip them up, set up auto-pay and fraud alerts and forget about them, but don't cancel them.

It does not seem like the expected value of the probability of something slipping through the cracks would pay for the marginal increase in the credit score.

Comment author: benkuhn 18 April 2015 09:06:26PM 0 points [-]

For people who would otherwise not have multiple credit cards, the increase in credit score can be fairly substantial.

In addition to Dorikka's comment, you are not liable for fraudulent charges; usually the intermediating bank is.

Comment author: benkuhn 18 April 2015 02:39:15AM 0 points [-]

If you don't want to bother signing up for a bunch of cards, the US Bank Cash+ card gives 5% cash back for charitable donations, up to I think $2000 per quarter. This is a worse percentage but lower-effort and does not ding your credit (as long as you don't miss payments, obvs).

Also, as I understand, it's actually better not to cancel the cards you sign up for (unless they have an annual fee), because "average age of credit line" is a factor in the FICO score. Snip them up, set up auto-pay and fraud alerts and forget about them, but don't cancel them.

Comment author: vbuterin 01 April 2015 05:45:59AM 1 point [-]

Thanks, I think this might actually be the argument I was looking for.

Whether the Bitcoin markets are efficient enough to worry about this is an open question

Right, so now the question is one of, does this idea of adverse selection actually apply?

I suppose one reformulation of the point made in the article is: if I believe X will happen with probability 5%, then I do not necessarily want to bet on X at 4.99% and bet against X at 5.01%, because it could be that my confidence is low enough that the very fact that someone wants to bet for or against me will shift my estimation of X in either direction outside that range.

So a safety factor is necessary. Question is, how large? The current markets are willing to bet on the proposition at 0.7% (as a first approximation; in reality the rectangle of $34000 * 5% is only part of the probability distribution so it's probably more like 0.2%). I'm not sure that many people are willing to bet against it at 0.7%; my hunch is that the people shorting it now would disappear once some threshold is passed (eg. the old $1242 all-time high) and are merely going on short and medium-term technicals.

In general, I'm hypothesizing that the Bitcoin markets have an inefficiency in that many people who are in them are already in them deeply, and so marginal additional investment even at positive expected value is a bad idea for them because in those worlds where BTC goes up a lot they would already be very rich and so they would rather optimize the remainder of their portfolio for the worlds where that doesn't happen; essentially limitations due to risk.

A claim that would significantly work against my hypothesis is the BTC price not going up by much or at all over the next year, as Bitcoin ETFs for mainstream investors are now available.

For instance, a bitcoin detractor could argue that the reference class should also include Beanie Babies, Dutch tulips, and other similar stores of value.

True, I hadn't thought of those. Of course, the case of Beanie Babies is more comparable to Dogecoin than Bitcoin, and the Dutch tulip story has in reality been quite significantly overblown (see http://en.wikipedia.org/wiki/Tulip_mania#Modern_views , scrolling down to "Legal Changes"). But then I suppose the reference class of "highly unique things" will necessarily include things each of which has unique properties... :)

Comment author: benkuhn 01 April 2015 05:59:50PM 1 point [-]

Of course, the case of Beanie Babies is more comparable to Dogecoin than Bitcoin, and the Dutch tulip story has in reality been quite significantly overblown (see http://en.wikipedia.org/wiki/Tulip_mania#Modern_views , scrolling down to "Legal Changes"). But then I suppose the reference class of "highly unique things" will necessarily include things each of which has unique properties... :)

I think the way to go here is to assemble a larger set of potentially comparable cases. If you keep finding yourself citing different idiosyncratic distinctions (e.g. Bitcoin was the only member to be not-overblown AND have a hard cap on its supply AND get over 3B market cap AND ...), this suggests that you need to be more inclusive about your reference class in order to get a good estimate.

Comment author: seer 01 April 2015 05:29:36AM 7 points [-]

For instance, a bitcoin detractor could argue that the reference class should also include Beanie Babies, Dutch tulips, and other similar stores of value.

The difference is that it's easy to make more tulips or Beanie Babies, but the maximum number of Bitcoins is fixed.

Comment author: benkuhn 01 April 2015 04:43:45PM *  2 points [-]

The difference is that it's easy to make more tulips or Beanie Babies, but the maximum number of Bitcoins is fixed.

Yes, this is what I mean by reference class tennis :)

Actually, according to Wikipedia, it's hypothesized that part of the reason that tulip prices rose as quickly as they did was that it took 7-12 years to grow new tulip bulbs (and many new bulb varieties had only a few bulbs in existence). And the Beanie Baby supply was controlled by a single company. So the lines are not that sharp here, though I agree they exist.

View more: Next