Repeating complaints from last year:

So in 2012 we started asking for SAT and ACT scores, which are known to correlate well with IQ and are much harder to get wrong. These scores confirmed the 139 IQ result on the 2012 test.

The 2012 estimate from SATs was about 128, since the 1994 renorming destroyed the old relationship between the SAT and IQ. Our average SAT (on 1600) was again about 1470, which again maps to less than 130, but not by much. (And, again, self-reported average probably overestimates actual population average.)

Last year I complained about horrible performance on calibration questions, but we all decided it was probably just a fluke caused by a particularly weird question. This year's results suggest that was no fluke and that we haven't even learned to overcome the one bias that we can measure super-well and which is most easily trained away. Disappointment!

I still think you're asking this question in a way that's particularly hard for people to get right. (The issue isn't the fact you ask about, but what sort of answers you look for.)

You've clearly got an error in your calibration chart; you can't have 2 out of 3 elite LWers be right in the [95,100] category but 100% of typical LWers are right in that category. Or are you not including the elite LWers in typical LWers? Regardless, the person who gave a calibration of 99% and the two people who gave calibrations of 100% aren't elite LWers (karmas of 0, 0, and 4; two 25% of the sequences and one 50%).

With few exceptions you were very overconfident.

The calibration chart doesn't make clear the impact of frequency. If most people are providing probabilities of 20%, and they're about 20% right, then most people are getting it right- and the 2-3 people who provided a probability of 60% don't matter.

There are a handful of ways to depict this. One I haven't seen before, which is probably ugly, is to scale the width of the points by the frequency. Instead, here's a flat graph of the proportion of survey respondents who gave each calibration bracket:

Significant is that if you add together the 10, 20, and 30 brackets (the ones around the correct baseline probability of ~20% of getting it right) you get 50% for typical LWers and 60% for elite LWers; so most people were fairly close to correctly calibrated, and the people who thought they had more skill on the whole dramatically overestimated how much more skill they had.

(I put down 70% probability, but was answering the wrong question; I got the population of the EU almost exactly right, which I knew from GDP and per-capita comparisons to the US. Oops.)