Yes, the difference is that you are creating an external environment which rewards you for success and punishes you for failure. This is similar to taking a final exam, which is my inspiration.
The problem with committing to work rather than success is that you can always just rationalize something as “Oh I worked hard” or “I put in my best effort.” However, just as with a final exam, the only thing that will matter in the end is if you actually do what it takes to get the high score. This incentivizes good consequentialist thinking and disincentivizes rationalization.
I agree there are things out of your control, but the same is true with final exams. For instance, the test-maker could have put something on the test that you didn’t study much for. This encourages people to put extra effort into their assigned task to ensure robustness to outside forces.
I personally try to balance keeping myself honest by having some goal outside but also trusting myself enough to know when I should deprioritize the original goal in favor of something else.
For example, let’s say I set a goal to write a blog post about a topic I’m learning in 4 hours, and half-way through I realize I don’t understand one of the key underlying concepts related to the thing I intended to write about. During an actual test, the right thing to do would be to do my best given what I know already and finish as many questions as possible. But I’d argue that in the blog post case, I very well may be better off saying, “OK I’m going to go learn about this other thing until I understand it, even if I don’t end up finishing the post I wanted to write.”
The pithy way to say this is that tests are basically pure Goodhardt, and it’s dangerous to turn every real life task into a game of maximizing legible metrics.
For example, let’s say I set a goal to write a blog post about a topic I’m learning in 4 hours, and half-way through I realize I don’t understand one of the key underlying concepts related to the thing I intended to write about.
Interesting, this exact same thing just happened to me a few hours ago. I was testing my technique by writing a post on variational autoencoders. Halfway through I was very confused because I was trying to contrast them to GANs but didn’t have enough material or knowledge to know the advantages of either.
During an actual test, the right thing to do would be to do my best given what I know already and finish as many questions as possible. But I’d argue that in the blog post case, I very well may be better off saying, “OK I’m going to go learn about this other thing until I understand it, even if I don’t end up finishing the post I wanted to write.”
I agree that’s probably true. However, this creates a bad incentive where, at least in my case, I will slowly start making myself lazier during the testing phase because I know I can always just “give up” and learn the required concept afterwards.
At least in the case I described above I just moved onto a different topic, because I was kind of getting sick of variational autoencoders. However, I was able to do this because I didn’t have any external constraints, unlike the method I described in the parent comment.
The pithy way to say this is that tests are basically pure Goodhardt, and it’s dangerous to turn every real life task into a game of maximizing legible metrics.
That’s true, although perhaps one could devise a sufficiently complex test such that it matches perfectly with what we really want… well, I’m not saying that’s a solved problem in any sense.
Yes, the difference is that you are creating an external environment which rewards you for success and punishes you for failure. This is similar to taking a final exam, which is my inspiration.
The problem with committing to work rather than success is that you can always just rationalize something as “Oh I worked hard” or “I put in my best effort.” However, just as with a final exam, the only thing that will matter in the end is if you actually do what it takes to get the high score. This incentivizes good consequentialist thinking and disincentivizes rationalization.
I agree there are things out of your control, but the same is true with final exams. For instance, the test-maker could have put something on the test that you didn’t study much for. This encourages people to put extra effort into their assigned task to ensure robustness to outside forces.
I personally try to balance keeping myself honest by having some goal outside but also trusting myself enough to know when I should deprioritize the original goal in favor of something else.
For example, let’s say I set a goal to write a blog post about a topic I’m learning in 4 hours, and half-way through I realize I don’t understand one of the key underlying concepts related to the thing I intended to write about. During an actual test, the right thing to do would be to do my best given what I know already and finish as many questions as possible. But I’d argue that in the blog post case, I very well may be better off saying, “OK I’m going to go learn about this other thing until I understand it, even if I don’t end up finishing the post I wanted to write.”
The pithy way to say this is that tests are basically pure Goodhardt, and it’s dangerous to turn every real life task into a game of maximizing legible metrics.
Interesting, this exact same thing just happened to me a few hours ago. I was testing my technique by writing a post on variational autoencoders. Halfway through I was very confused because I was trying to contrast them to GANs but didn’t have enough material or knowledge to know the advantages of either.
I agree that’s probably true. However, this creates a bad incentive where, at least in my case, I will slowly start making myself lazier during the testing phase because I know I can always just “give up” and learn the required concept afterwards.
At least in the case I described above I just moved onto a different topic, because I was kind of getting sick of variational autoencoders. However, I was able to do this because I didn’t have any external constraints, unlike the method I described in the parent comment.
That’s true, although perhaps one could devise a sufficiently complex test such that it matches perfectly with what we really want… well, I’m not saying that’s a solved problem in any sense.