Note: I'm writing every day in November, see my blog for disclaimers.

It’s hard to be correct, especially if you want to be correct at something that’s non-trivial. And as you attempt trickier and trickier things, you become less and less likely to be correct, with no clear way to improve your chances. Despite this, it’s often possible to bias your attempts such that if you fail, you’ll fail in a way that’s preferable to you for whatever reason.

As a practical example, consider a robot trying to crack an egg. The robot has to exert just enough force to break the egg. This (for a sufficiently dumb robot) is a hard thing to do. But importantly, the failure modes are completely different depending on whether the robot uses too much force or too little: too much force will break the egg and likely splatter the yolk & white all over the kitchen, too little force will just not break the egg. In this scenario it’s clearly better to use too little force rather than too much force, so the robot should start with a lower-estimate of the force required to break the egg, and gradually increase the force until the egg cracks nicely.

This also appears in non-physical contexts. This idea is already prevalent in safety related discussions: it’s usually far worse to underestimate a risk than it is to overestimate a risk (e.g. the risk of a novel pathogen, the risk of AI capabilities, the risk of infohazards).

Looking at more day-to-day scenarios, students regularly consider whether it’s worth voicing their uncertainty “I don’t understand equation 3” or just keeping quiet about it and trying to figure out the uncertainty later. But I’d argue that in these cases it’s worthwhile having a bias towards asking rather than not asking, because in the long-run this will lead to you learning more, faster.

Salary negotiation is another example, in which you have uncertainty about exactly what amount your potential employer would be happy to pay you, but in the long-run it’ll serve you well to overestimate rather than underestimate. Also, you should really read patio11’s Salary Negotiation essay if you or a friend is going through a salary negotiation.

You see similar asymmetric penalties with reaching out to people who you don’t know, asking for introductions, or otherwise trying to get to know new people who might be able to help you. It’s hard to know what the “right” amount of cold emails to send is, but I’d certainly rather be accused of sending too many than feel the problems of having sent too few.

This idea is a slippery one, but I’ve found that it applies to nearly all hard decisions in which I don’t know the right amount of something to do. While I can’t figure out the precise amount, often I have strong preferences about doing too much or too little, and this makes the precise amount matter less. I give my best guess, update somewhat towards the direction I’d prefer to fail, and then commit to the decision.