On Defecting On Yourself

post by David Udell · 2022-03-18T02:21:01.572Z · LW · GW · 6 comments


  Fooling Yourself
  Trusting Yourself
  The Comparative Advantages of Your Time-Slices

Sometimes I get into an anxious frame of mind where I try to prep for something as if my memory, my past self, cannot be trusted. This feels like an epistemically virtuous way to prepare for the world: double-checking all your past work and preparations as if they were made by someone you don't fully trust sure feels like it'll make all the difference in catching sloppy errors. If you have something to protect [LW · GW], then this is good, no? Doing without a second- or third-pass seems reckless or epistemically immodest [LW · GW], as anyone can make mistakes, and plenty of people do.

Fooling Yourself

One time my brother was telling me about his mindset running track. He mentioned that one of his track tricks is to get himself into a frame of mind where he swears that he'll stop running … as soon as he gets to that lamppost over there. He pushes through the pain, knowing that the end is in sight … and then finds himself thinking that, "Well, I can stop after I get to the next nice round stopping-point." Rinse and repeat until his run is completed.

This didn't feel right to me then. My inarticulate thought at the time was something like, "But you have to be able to trust your past self. Only with that intertemporal trust can you expect to get anything done."

In George Ainslie's Breakdown of Will, Ainslie tries to explain the thought process of a sober alcoholic about to relapse, weighing his alternatives. On the one hand, he desperately wants to never touch that stuff again, and would much prefer a future where he never touches the sauce again to one where he does. On the other hand, there is always a looming local temptation: an even better future would be one where he has one final binge today and cleans up beginning tomorrow. This would harvest just about all the value of getting clean while still letting him have a non-miserable time now. The trouble is that once he defects on his future self by binging now, the thought that his future self will clean up tomorrow becomes less plausible. Your future self will behave more or less like you. And that thought becomes all the less likely when he drinks despite being at a natural stopping point; if he started drinking again early in the new year and threw away the fresh start the year had granted him, what're the odds that he'll successfully stop drinking on just any old Tuesday? The only equilibria are to swear off the sauce forevermore, because of his long-run interest in doing so, or to always defect and follow his local interests -- never getting clean, but never having to suffer through a failed attempt to go clean when all hope is already gone. In summary, Ainslie's theory is that addiction is an intertemporal Prisoner's Dilemma.

Trusting Yourself

Imagine yourself as a series of time-slices. There's you-an-hour-ago, you-right-now, you-a-year-hence, and so on. Your past selves are more or less like you, and your future selves will be too. It might feel appropriately epistemically modest to assume away your own historical competence for a moment and treat all your past preparations as if they were made by a stranger, but doing that systematically underestimates your past preparations!

Not trusting your past self also seems like an intertemporal Prisoner's Dilemma. If I have a crippling fear of eating poisoned food, then locally it's always in my interest to test my meal for poisoning. This has some relatively small cost in hassle. That small cost would add up, though, over a lifetime of neuroticism. If the base rates of poisoning are pretty low in my environment, then I might prefer never worrying to always worrying. But, what's better is the option of worrying just this once and then after that never worrying. If I go for that, though, I'll destroy my credibility to my future self. The only equilibria are to go for full neuroticism or no neuroticism.

Say that I'm working on something and punch-out time rolls around. In this case, the cooperate option is defer to your past self's planning and spend your evening recuperating. The defect option is keep working now, then resume sticking to your past self's plan. Having tried it both ways, I've found I'm more productive when I know I can trust my future and past selves to play their part in our harmonious intertemporal-pact. It's really important to me to preserve my inherited credibility as a member of our pact! If I squander that credibility, the best my future selves will be able to do is to blindly follow their local interests, and the global interest we'd all prefer to see realized might become irreversibly inaccessible. So long as I preserve that credibility, I can and should think of my past selves as handing off a competently prepared state of affairs to me, in pursuit of our global goals.

The Comparative Advantages of Your Time-Slices

One of the most profound insights in economics is Ricardo's trade theorem. It states that two trading actors with different productive abilities can produce a larger collective-pie-of-goods by each working on what they're least worst at, relative to the other actor. 'Least worst' here means "minimizes the number of gizmos you forgo when you produce widgets." If Britain and Spain are trading, and Britain produces both wine and cloth quicker than Spain does, owing to its more industrialized economy, it still makes sense for Britain and Spain to individually specialize and rely on one another. Britain should throw all its resources at whichever good it produces with the lowest cost in forgone gizmos (wine, cloth, whatever), and so should Spain. The collective-pie-of-goods will strictly grow if both countries move from making everything in-house as generalists to specializing and fully exploiting their respective strengths.

And the same is true for individual people, of course! A large society of specialists will be far more productive overall than its counterpart society of generalists. Plausibly, this is why we have professions at all.

What do your various time-slices forgo when they opt for one activity? When you're sleepy and choose to work, you forgo falling asleep quickly and leave that sleep-quota to your future self (who might be have to give up more to complete it). Your intertemporal pact will do better if you specialize in recuperating when you're worn down, and in doing research when you're alert. Let that principle guide your allocation of weekly time to your tasks, and then trust that planning by not defecting when it seems like you can incrementally get ahead just this once. You(-right-now) are not in this alone! Plan around having a backup team in the past that you trust to hand you good work, and around having a team in the future that will intelligently continue whatever partially completed tasks you send on to them.


Comments sorted by top scores.

comment by Richard_Kennaway · 2022-03-18T09:53:50.272Z · LW(p) · GW(p)

The way I frame these situations to myself is to consider myself not as a succession of agents figuring out how to cooperate, but as a single four-dimensional entity [LW(p) · GW(p)] that coheres in time as it does in space. Myself tomorrow is as surely a part of myself as my right hand is today.

I go to the gym today because that's what I planned yesterday, if there is no new information prompting a reconsideration. I make sure all the dishes are washed before retiring at night, because that's one of the rules that I have. Whatever decisions I make are what I do when the time comes, because that is what a decision is.

I cannot claim to do this perfectly, any more than I can claim any other perfection, but that is the frame of mind.

Replies from: Gunnar_Zarncke, David Udell
comment by Gunnar_Zarncke · 2022-03-18T12:38:25.078Z · LW(p) · GW(p)

You can do that only if you are a coherent entity, to begin with - or at least close enough to it that this model works. 

Replies from: Richard_Kennaway
comment by Richard_Kennaway · 2022-03-18T12:49:17.065Z · LW(p) · GW(p)

All beginnings are paradoxical, for how can a thing begin, without its seeds having already been present? Yet things do begin, notwithstanding philosophers proving that the arrow does not move, that learning is but remembrance of forgotten knowledge, that an omnipotent, self-caused being must have started this universe, that none can do good of their own volition, or that we are machines that cannot do at all.

If you're screwed enough, you're screwed, as the saying is, but if not then one can learn to do such things. Sometimes an outside stimulus may help one to start.

comment by David Udell · 2022-03-18T17:15:35.229Z · LW(p) · GW(p)

I think Ainslie's theory of willpower excels at explaining instances of weakness-of-will -- I badly want to be the sort of person who regularly goes to the gym, but it wouldn't hurt much to hold off starting till next week

On your theory, isn't it a little mysterious that you have conflicting local and global interests like this? Or do you think that you basically don't have this kind of goal-conflict?

Replies from: Richard_Kennaway
comment by Richard_Kennaway · 2022-03-18T18:07:26.353Z · LW(p) · GW(p)

"Not today, but tomorrow I really will!" is an obvious failure mode. Do not do obvious failure modes. Today's cross-section and tomorrow's should no more conflict than your left hand and your right.

"I badly want to be the sort of person who regularly goes to the gym, but it wouldn't hurt much to hold off starting till next week" — on the contrary, it does hurt to hold off until next week, by a week of not being the person you want to be. You can be that person right now by going to the gym. If you want a thing, and you can have it right now, take it. When you grok this, action is effortless.

comment by Gunnar_Zarncke · 2022-03-18T12:43:50.505Z · LW(p) · GW(p)

It's really important to me to preserve my inherited credibility as a member of our pact! If I squander that credibility, the best my future selves will be able to do is to blindly follow their local interests, and the global interest we'd all prefer to see realized might become irreversibly inaccessible. 

Arguably that describes the function of the self and maybe (partly) why it evolved: To better coordinate over time. There is probably overlap with modeling other agents in this way too.

This suggests sharp criteria for which animals have selves in this sense.