Epistemic status: An idea I had a few days ago and shared with two or three friends.
Suppose you want to install a new habit. For example, gratitude journaling. You might try to give it a go, by pledging to write down 3 things you are grateful for every day right before you go to sleep. After a few days, you start skipping out on the habit, and after a month it’s completely gone. You aren’t too worried, remembering to beware of other(s)-optimizing (you), you say to yourself that maybe this thing just isn’t for you. A few weeks later, you get to hear again about how gratitude journaling is good so you give it another shot, which again, fizzles out after a few days.
It took me quite a few rounds this pattern (and a conversation about someone else’s habits) to notice it, and to try to think about a meta-level solution, which I call expected number of tries. Say there are N essentially different ways to do gratitude journaling, of which only n work for you. Then assuming you are randomly picking different methods, you should expect to try about e=N/n times before you find a way that works for you. Since you do gather both intuitive and logical while trying many times, you can probably do better than chance here, and e is only an upper bound on the expected number of tries. If we can get estimate for e, and a bound for how long it take to test one method, we can get a bound for how long should it take to get the habit installed. Moreover, a bound on the length of a test can be a useful period for checking up on the current attempt.
The hard part here is of course to estimate e.
Estimating N
Here are two methods I thought of for estimating N: 1. Asking ChatGPT How many unique HABIT habits you think you can suggest? 2. Googling N ways to do HABIT, for increasing N until there are no front page results which look fitting (at a glance).
For gratitude habits, ChatGPT suggested 35 different habits, while googling gave satisfactory results up until 200. For a different example, casual sports, ChatGPT gave 60 different suggestions. Googling was harder here, and gave results for 25 and 100 but nothing in between, and nothing more.
I think I’ll set my estimates closer to the number ChatGPT gives.
In both cases I think that the list of suggestion from google or ChatGPT could be used for inspiration, but that one should probably come up with their own ideas based on past experience.
Estimating n
This one seems to be harder to estimate. The worst case is n=1, and the best case is n=N, but I suspect these are far from the truth. The best thing I have been able to come up with is to ask ChatGPT again, via the prompt How many do you think might fit a particular person?. For the gratiude journal, it seems to estimate n=5, and for casual sports it estimates n somewhere between 5 and 10, giving an overall e≈7 for gratitude journaling and e≤12 for casual sports.
Improving the method
Currently, the biggest error probably comes from a lacking estimation of $n$. Ways of improving the estimates of each might be:
Creating a better prompt. I don’t know much about prompt optimising so there is probably a lot of room for improvement here.
Creating a large survey or updating database where people share what they tried and what worked for them.
Expected number of tries
Epistemic status: An idea I had a few days ago and shared with two or three friends.
Suppose you want to install a new habit. For example, gratitude journaling.
You might try to give it a go, by pledging to write down 3 things you are grateful for every day right before you go to sleep.
After a few days, you start skipping out on the habit, and after a month it’s completely gone.
You aren’t too worried, remembering to beware of other(s)-optimizing (you), you say to yourself that maybe this thing just isn’t for you.
A few weeks later, you get to hear again about how gratitude journaling is good so you give it another shot, which again, fizzles out after a few days.
It took me quite a few rounds this pattern (and a conversation about someone else’s habits) to notice it, and to try to think about a meta-level solution, which I call expected number of tries.
Say there are N essentially different ways to do gratitude journaling, of which only n work for you.
Then assuming you are randomly picking different methods, you should expect to try about e=N/n times before you find a way that works for you.
Since you do gather both intuitive and logical while trying many times, you can probably do better than chance here, and e is only an upper bound on the expected number of tries.
If we can get estimate for e, and a bound for how long it take to test one method, we can get a bound for how long should it take to get the habit installed.
Moreover, a bound on the length of a test can be a useful period for checking up on the current attempt.
The hard part here is of course to estimate e.
Estimating N
Here are two methods I thought of for estimating N:
1. Asking ChatGPT
How many unique HABIT habits you think you can suggest?2. Googling
N ways to do HABIT, for increasingNuntil there are no front page results which look fitting (at a glance).For gratitude habits, ChatGPT suggested 35 different habits, while googling gave satisfactory results up until 200.
For a different example, casual sports, ChatGPT gave 60 different suggestions. Googling was harder here, and gave results for 25 and 100 but nothing in between, and nothing more.
I think I’ll set my estimates closer to the number ChatGPT gives.
In both cases I think that the list of suggestion from google or ChatGPT could be used for inspiration, but that one should probably come up with their own ideas based on past experience.
Estimating n
This one seems to be harder to estimate.
The worst case is n=1, and the best case is n=N, but I suspect these are far from the truth.
The best thing I have been able to come up with is to ask ChatGPT again, via the prompt
How many do you think might fit a particular person?.For the gratiude journal, it seems to estimate n=5, and for casual sports it estimates n somewhere between 5 and 10, giving an overall e≈7 for gratitude journaling and e≤12 for casual sports.
Improving the method
Currently, the biggest error probably comes from a lacking estimation of $n$.
Ways of improving the estimates of each might be:
Creating a better prompt. I don’t know much about prompt optimising so there is probably a lot of room for improvement here.
Creating a large survey or updating database where people share what they tried and what worked for them.
I’d be happy to hear your thoughts!