When you procrastinate, you're probably not procrastinating because of the pain of working.
How do I know this? Because on a moment-to-moment basis, being in the middle of doing the work is usually less painful than being in the middle of procrastinating.
(Bolded because it's true, important, and nearly impossible to get your brain to remember - even though a few moments of reflection should convince you that it's true.)
So what is our brain flinching away from, if not the pain of doing the work?
I think it's flinching away from the pain of the decision to do the work - the momentary, immediate pain of (1) disengaging yourself from the (probably very small) flow of reinforcement that you're getting from reading a random unimportant Internet article, and (2) paying the energy cost for a prefrontal override to exert control of your own behavior and begin working.
Thanks to hyperbolic discounting (i.e., weighting values in inverse proportion to their temporal distance) the instant pain of disengaging from an Internet article and paying a prefrontal override cost, can outweigh the slightly more distant (minutes in the future, rather than seconds) pain of continuing to procrastinate, which is, once again, usually more painful than being in the middle of doing the work.
I think that hyperbolic discounting is far more ubiquitous as a failure mode than I once realized, because it's not just for commensurate-seeming tradeoffs like smoking a cigarette in a minute versus dying of lung cancer later.
When it comes to procrastinating, the obvious, salient, commensurate-seeming tradeoff, is between the (assumed) pleasure of reading a random Internet article now, versus the (assumed) pain of doing the work now. But this, as I said above, is not where I think the real tradeoff is; events that are five minutes away are too distant to dominate the thought process of a hyperbolic discounter like a human. Instead our thought processes are dominated by the prospective immediate pain of a thought, a cost that isn't even salient as something to be traded off. "Working" is an obvious, salient event, and "reading random articles" seems like an event. But "paying a small twinge of pain to make the decision to stop procrastinating now, exerting a bit of frontal override, and not getting to read the next paragraph of this random article" is so map-level that we don't even focus on it as a manipulable territory, a cost to be traded off; it is a transparent thought.
The real damage done by hyperbolic discounting is for thoughts that are only very slightly painful, and yet, these slight pains being immediate, they manage to dominate everything else in our calculation. And being transparent, we aren't even aware that's what's happening. "Beware of immediately trivially painful transparent thoughts", one might say.
Similarly, you may read a mediocre book for an hour, instead of a good book, because if you first spent a few minutes to search your library to obtain a better book, that would be an immediate cost - not that searching your library is all that unpleasant, but you'd have to pay an immediate activation cost to do that instead of taking the path of least resistance and grabbing the first thing in front of you. It's a hyperbolically discounted tradeoff that you make without realizing it, because the cost you're refusing to pay isn't commensurate enough with the payoff you're forgoing to be salient as an explicit tradeoff.
A related note that I might as well dump into this post: I'm starting to think that procrastination by reading random articles does not cause you to rest, that is, you do not regain mental energy from it. Success and happiness cause you to regain willpower; what you need to heal your mind from any damage sustained by working is not inactivity, but reliably solvable problems which reliably deliver experienced jolts of positive reinforcement. Putting in the effort to read a good book may do this; playing a good computer game may do this; reading random Internet articles, or playing bad games, probably won't. Literal mental exhaustion might mean that you don't have enough energy left to read a good book - or that you don't have enough energy left to pay the immediate cost of searching your library for good reading material instead of mediocre reading material - but in this case you shouldn't be reading random online articles. You should be sitting with your eyes closed listening to music, or possibly even napping; if dealing with a truly exhausted brain, reading random articles is probably too much effort.
If you don't feel good while reading a lot of forgettable online articles, and you don't feel renewed after doing so, your intuitive theory which says that this is how to rest is mistaken, and you need to look for other ways to rest instead - more active ways to regain willpower, less active ways to recover from immediate exhaustion. In general, poor performance often indicates poor models; if something seems incredibly difficult to predict or manipulate, it may be that you have mistaken beliefs about it, including transparent mistakes that are nonquestioned because they are nonsalient. This includes poor performance on the problem of resting.
Hopefully publishing this post will help me live up to it.
Exponential discounting is not sufficient to avoid exploitable dynamic inconsistencies.
For example, assume your decision theory's outputs are purely a function of:
In this case, the second criterion corresponds to exponential discounting (for k>1), yet it still leads the agent to make the same decision each period, meaning that a decision to defer now implies a perpetual decision to defer.
So despite the agent using exponential discounting, an agent is inconsistent to prefer both deferral of the work during the current period, and accomplishment of the work at some point.
But how would an unfriendly agent exploit that inconsistency?