One of the most useful concepts I have learned recently is the distinction between actions which directly improve the world, and actions which indirectly improve the world.
Suppose that you go onto Mechanical Turk, open an account, and spend a hundred hours transcribing audio. At current market rates, you'd get paid around $100 for your labor. By taking this action, you have made yourself $100 wealthier. This is an example of what I'd call a Level 1 or object-level action: something that directly moves the world from a less desirable state into a more desirable state.
On the other hand, suppose you take a typing class, which teaches you to type twice as fast. On the object level, this doesn't move the world into a better state- nothing about the world has changed, other than you. However, the typing class can still be very useful, because every Level 1 project you tackle later which involves typing will go better- you'll be able to do it more efficiently, and you'll get a higher return on your time. This is what I'd call a Level 2 or meta-level action, because it doesn't make the world better directly - it makes the world better indirectly, by improving the effectiveness of Level 1 actions. There are also Level 3 (meta-meta-level) actions, Level 4 (meta-meta-meta-level actions), and so on.
The most important difference between Level 1 and Level 2 actions is that Level 1 actions tend to be additive, while Level 2 actions tend to be multiplicative. If you do ten hours of work at McDonald's, you'll get paid ten times as much as if you did one hour; the benefits of the hours add together. However, if you take ten typing classes, each one of which improves your ability by 20%, you'll be 1.2^10 = 6.2 times better at the end than at the beginning: the benefits of the classes multiply (assuming independence).
One result is that spending time on Level 2 actions can have a much greater return than spending time on Level 1 actions. If your labor is worth $20 an hour, and you can't change that, then the amount of money you can earn in a year has a fairly hard upper bound- no matter how you slice it, there are only 168 hours in a week. If you spend that year trying to increase the value of your labor, on the other hand, the upper bound on your performance is both a lot higher (because you can then make more money every year for the next three decades), and a lot more fuzzy. It's a lot more fuzzy because, while everyone has the same number of hours in a week, how effective Level 2 actions are depends a lot on your intelligence, what methods you use, and lots of other stuff. Most Americans spend too little time on higher-level actions, like being strategic - doing a quick analysis of what your goals are, and which Level 1 or Level 2 actions would best accomplish those goals. Witness the hordes of lawyers who spend thirty years on the Level 1 action of working at a law firm, three years on the Level 2 action of getting a law degree, and three minutes on the Level 3 action of deciding what to do after college. (Being strategic is one level up from whichever actions you're being strategic about.)
It is also possible to have the opposite problem, of under-valuing Level 1, and I suspect that quite a few people in the nerdier communities do. People sometimes fall into the trap of noticing that the higher levels are (when applied properly) far more useful on the margin than Level 1, and then reacting by giving blind praise to the meta level at the expense of the object level. One cultural example is the ancient Greeks- who, though they were good thinkers for their day, didn't invent science. Science involved actually going out and looking at the world, and that was manual labor and manual labor was for slaves. The ultimate extreme of this is Aristotle, who got philosophy off to an unfortunate beginning by starting his Metaphysics with the assumption that the most noble knowledge would be the most useless.
The problem there is that, because Level 2 actions are multiplicative and not additive, you still need at least some Level 1 actions to multiply by. It doesn't matter how high the value of one's labor is, if one never actually goes out and does labor. A very large number, multiplied by zero, is still zero. If one just does Level 2 actions, without any Level 1 actions, it is a failure to do something instead of nothing. Taking only meta-level actions accomplishes less, in the end, than the ten-year-old who just mowed the neighbor's lawn for a dollar.
On a societal level, one can run into this problem even more easily, because having a large society allows one to build up more meta-levels. For the most part, people's day-to-day labor is made up of Level 1 actions- the stuff that directly improves the world. Engineering technology that helps improve people's productivity is then a Level 2 action. Doing science that helps with engineering is then a Level 3 action (meta-meta), and doing math that helps with science is a Level 4 action (meta-meta-meta). In order for working on Level 4 to be effective, there have to be three steps chaining back to Level 1: in this example, from math to science, from science to engineering, and from engineering to productivity. If any one of these steps fails - if the math isn't useful for science, if the science is in a different field than the engineering, or if the engineered devices aren't used effectively - then working on Level 4 won't accomplish anything.
Going meta can be very powerful, for the reasons outlined above- each action taken on Level N + 1 makes it easier to do lots of things on Level N. The invention of science, which is almost always Level 2 or higher, changed the world and created modern civilization, and science can still radically improve your life today. However, in order for the higher levels to be useful within a specific project, that project has to incorporate all the steps from the meta level back down to the object level, and this becomes much more difficult with each meta level added. The Manhattan Project managed to pull it off with two meta-levels- science to engineering and engineering to real-world effects- but the Manhattan Project had dozens of world-class scientists and hundreds of top-notch engineers working on it. Attempting too many meta levels without having the infrastructure to support the attempt will wind up like filing an IPO for a one-man beer pong business.
How should we counter this, while still getting the benefits of the higher levels? One suggestion that gets talked about a lot is simply to always do something directly useful - do something instead of nothing, which is the first step towards accomplishing any goal. On the small scale, we are all familiar with this. If you take the garbage out, the garbage can becomes empty; if you don't, the room starts to stink. If you do the dishes, you have clean dishes to eat off of; if not, you wind up eating off of dirty dishes or a table. If you go and buy food, you have lots of food to eat; if not, you go hungry, or pay a lot for food at a restaurant.
I think that, collectively, we are all living in a world that's isomorphic to a pigsty house, where no one ever does the dishes and no one ever takes out the trash. We're doing better now than we have in the past. But the world of today is nothing, compared to the shining, sparkling utopia that could exist if people simply did more things, using the decision algorithms that they already know how to execute. So many things ; to name one particular field, there are many examples of science which is obviously ridiculously underfunded. Yet, in all these cases, with the exception of a tiny handful of part-time volunteers, nothing ever gets done.
It is, of course, also important to choose effective actions, in addition to simply choosing to act. Preventing existential risk is a more important goal by far than coming up with a more effective way to do biotech research. However, the first and most important step is to just do something instead of nothing. If, at the end of the day, what you actually wind up doing is nothing instead of X, it doesn't matter that there are ten other things you could do which would be a hundred times more effective than X. Those don't show up on the bottom line. Your final score is still zero, instead of at least being positive.
Even if we consciously agree that doing X could be useful and is unlikely to result in serious harm, I strongly suspect that there are still barriers to acting: unconscious ones. Most of the machinery of our brain operates below the level of deliberate reasoning. While typing this sentence, I remembered to breathe oxygen, match the words I was thinking to the motions of my hands (so as to make them appear on the screen), remain sitting upright, blink every so often, look around to make sure my stuff hadn't gotten lost, and move around a bit so my legs don't go numb, all without paying any deliberate attention at all. Yet, each of these things are quite difficult to do, in the sense that it would be a lot of work to build a robot that can do them. Just as we frequently breathe without noticing, I think we all refrain from doing useful things, out of unjustified fear, without noticing. Why is that?
My current guess is that it's because of the increasing institutionalization of society, which is caused by economic growth. When your tribe is made up of a hundred people, you can model each person in high detail when you interact with them - taking into account their personality, their strengths and weaknesses, their past interactions with you, and so on. However, in a corporation with a hundred thousand people, the CEO doesn't have time to construct complex models of each worker, and yet he must ensure that all the workers cooperate effectively. How does he do that? By making each worker simple to model - by constructing a set of rules which governs each worker's behavior, and constrains them to behave in simple, easily understandable ways. The net effect of this is that large institutions train people to be afraid of taking actions they aren't explicitly told to take, because that would make life more complicated for the managers.
One possible solution to this is to lower your general level of inhibition, by practicing doing things that you feel inhibited about. Studies have shown that there is a significant positive correlation between alcohol use, and income in life. Why would that be? Drinking alcohol doesn't make you smarter. Nor does it make you work harder, or become more skilled, or gain additional knowledge. I think the reason is that alcohol is disinhibitory. People do things while drunk that they wouldn't do otherwise, and even though a lot of them are stupid and destructive, some of them are useful (like meeting new people). And the world counts the good things and forgets about the bad things, like how everyone forgot about George W. Bush being branded with a coat hanger by his fraternity.
There is also a very interesting way, which I highly recommend, of getting both the benefits of Level 1 and Level 2 actions. One does this by going out and doing Level 1 things that one hasn't done before - for, to paraphrase Eliezer, if you want to find a better route to work, you must necessarily explore a different route to work.
Ordinarily, going to the grocery store is a Level 1 action. But what if you've never been to a grocery store? Then, going to the store is actually both a Level 1 and a Level 2 action. By going to the grocery store, you acquire food. And you also learn lots of useful things about how grocery stores work, which will help you on all of your subsequent trips.
The downside of this is that, when you consider things purely as Level 1 actions, it might be less worthwhile to do something new than something you're already familiar with. If you've never ridden a bus before, it might be faster to walk than to take the risk of getting lost. But, in most cases, this tends not to be a very big deal. If you do it badly, it's no biggie, you can just try again later when you're more skilled.
In the ancestral environment, the range of skills one could acquire was fairly limited. Hence, we humans evolved to employ a two-part strategy: try new things during your childhood, and then when you mature (at age 14 or so), forget about trying new things, and concentrate only on Level 1 actions. Now that the range of possible skills is so much larger, this is terribly suboptimal- but humans have this thing about continuing to do stuff, like eating chocolate, that has long since lost its utility.
The key benefit of doing lots of new, unfamiliar Level 1 actions is that world is an extremely complicated place, and as a general rule, no matter how much you read and learn about something (Level 2 actions), there's always some sort of surprise when you actually go and do it; something that the authors of the stuff you read didn't notice, or forgot to write down. In computer programming, we have the general principle of humility regarding bugs: even if you can't think of anything you did wrong when writing software, you had better go and test it before releasing it, because the odds are pretty darn good that you made a mistake somewhere. The analogous principle is, never assume that you can do something (even if it seems simple) unless you've actually done it before, because there will probably be some sort of hidden surprise.
And the more things you've done before, the fewer hidden surprises there will be. The counterpart point is that if you do go out and do things that you haven't done before, you'll be able to pass over mostly-invisible barriers that other people will mysteriously smack into. Suppose you try to construct a ten-step plan, where each of the ten steps is something seemingly simple, but is still something that you haven't done before. Will it work? Probably not. Even if the probability of success on each step is 90%, the probability of the whole plan working is only 0.9^10 = 0.35. You can go over each and every one of the steps, analyze them, figure out that they're very likely to succeed individually- and still fail, because without a strong assurance of success on each step, executing all of them in the proper sequence becomes very unlikely.
On the other hand, if you have lots of experience doing something Level 1, you can understand it well enough to actually make the probability of failure arbitrarily small, not just smallish-seeming. How narrow or wide your probability distribution is for X is a function of how much information you have about X. And doing X offers the possibility of gaining arbitrarily complete information about X, not just the sort of information that one can communicate effectively in words. I think the evidence is fairly conclusive that someone who hasn't experienced, for example, love, war, or torture can't really have complete information about it, because there are parts of the brain which are only wired to receive information directly from the external environment. (Of course, the reverse is also true, which is why people who don't read a lot can be very capable in their domains of expertise, but still terrible at abstract thought.)
Making the probability of failure arbitrarily small, then, allows you to construct long chains of sequential events which will work reliably. An average human can take a step forward ten thousand times in a row without falling down once, because they (the mostly subconscious systems in their brain that handle movement) understand it so well. And, once a particular chain of Level 1 events has become reliable enough, you can then use it as a building block to construct new, higher-level chains of Level 1 events, which is itself a Level 2 action. This is how society can accomplish extremely complicated things, like taking a company public, with any reliability at all- by having people build larger chains out of smaller building blocks that they already understand, and can execute many times without slipping up.