Responding To Overconfidence

Without Googling, what's your 95% confidence interval for the longest time spent in labor? Feel free to post it in the comments before reading on.

I just encountered this question on William MacAskill's Facebook wall. When I looked it up, after posting my answer, I discovered that my upper bound was off by a factor of 10.

I was reluctant to answer the question in the first place, but I didn't stop to examine why. It is now clear to me: When it's revealed that I'm extremely overconfident about something, my default response is shame and regret.

I used to respond much more strongly with shame and regret than I do now. I recall an occurrence of this reaction from about three years ago. The reaction was so strong that many vivid details of the context are readily available in memory. (Whether they're accurate is a separate question.) Robby and I were in the the pizzeria on Kirkwood sitting at a table by the door. It was raining. We were having bread sticks with cheese sauce, and he asked me, "What's your 90% confidence interval for when Reverend Bayes was born?"

Immediately I felt attacked and defensive. I did not know what a confidence interval was at the time, so he spent a few minutes explaining it to me. After that, he wanted me to answer, and I was so scared. I don't remember why, but I remember the feeling very well. I was terrified that I'd be way off, and that this test would reveal my embarrassing mistake. I answered anyway because I'd recently discovered Lesswrong, so I felt that this was a kind of question I was Supposed To answer if I wanted epistemic improvement. (The question was actually taken from a Lesswrong survey.) And I was definitely Supposed To know when Bayes was born!

Sure enough, I was way off. And sure enough, the shame flooded over me like a bucket of ice water. I felt terrible, and I regretted answering the question.

A lot of things have happened in the intervening time to lesson the shame response. I don't know what most of them are, but watching people I respect readily and casually test their predictions and reveal their mistakes has surely been part of it. Curing social anxiety also contributed, obviously.

It's not been enough, though. When I answered the labor question, I still felt enough of the shame to overshadow the recalibration going on beneath it. I did feel the recalibration as well, but it was subtle enough by comparison that if I hadn't been training mindfulness of this sort of mental motion in the recent past, I'd have missed it. And I certainly didn't catch the details.

That's a problem. I can't optimize a process I'm never aware of in real time. It doesn't matter how well I understand Bayesian updating when some other sensation is drowning out all my opportunities to apply my understanding.

You might think, "Ah, but those negative feelings are useful, because if you're punished for being overconfident then you might be less confident in the future!" What actually happens is that I'm less likely to put myself in situations that would reveal my overconfidence if it existed. Which shouldn't be surprising from a behavioral psychology perspective: The immediately preceding action was "answer calibration question", not "form a belief and establish a level of confidence in it".

If I were going to train a better reaction to calibration opportunities properly, I'd spend a few days studying my default reaction and becoming as mindful of it as possible. I'd also examine whether my default reaction suggested an emotional need of some sort that the optimal response ought to address, especially if my reaction were as strong now as it was three years ago. Only then would I begin considering possible interventions.

But this particular reaction seems to be in a class of habits that are very important but whose triggers are much too rare for the current version of the Tortoise Skills installation procedure. So instead, I'm going to try doing an abbreviated version of the procedure in case it turns out that I can get marginal improvements quickly from isolated cases like this.

My best guess at how I'd rather respond to discovering I'm extremely overconfident is about the same as the response I learned to have to failures. I'd like to feel nonchalant interest in my overconfidence. Further, I'd like that interest to inspire targeted curiosity about the cause of the overconfidence, and increased sensitivity to similar contexts or patterns of thought that might signal severe overconfidence if I encounter them when forming or considering other beliefs.

But the most important part is just letting go of the thing that drags me down into counterproductive emotions. A flavor of wu wei, maybe, of fluidity. My brain's pretty good, really, when I can keep from getting in its way. If I can just stop doing the stupid thing, I often don't need a brilliant solution on top of it.

I don't know how I do that "letting go of the dragging-downward" thing, but I do know that I've learned to do it at least once before. I'll plan to imagine Eliezer discovering overconfidence and his usual response, as a reminder that other responses are possible, in case I need some extra help.

So, here's the new trigger-action plan, which I will not train but will instead simply intend and await: If I notice that my overconfidence has been revealed, then I will loosen my grip on the downward-dragging sensations and direct my attention instead to even the tiniest sensation of reflective interest. If I have trouble with that, I will imagine how Eliezer would react in my place.