A Different Concept of Instrumental Rationality
Eliezer Yudkowsky defines instrumental rationality as “systematically achieving your values” and goes on to say: “Instrumental rationality, on the other hand, is about steering reality—sending the future where you want it to go. It’s the art of choosing actions that lead to outcomes ranked higher in your preferences. I sometimes call this ‘winning.’” [1]
I agree with Yudkowsky’s concept of rationality as a method for systematised winning. It is why I decided to pursue rationality—that I may win. However, I personally disagree with the notion of “systematically achieving your values” simply because I think it is too vague. What are my values? Happiness and personal satisfaction? I find that you can maximise this by joining a religious organisation, in fact I think I was happiest in a time before I discovered the Way. But this isn’t the most relevant, maximising your values isn’t specific enough for my taste, it’s too vague for me.
“Likewise, decision theory defines what action I should take based on my beliefs. For any consistent set of beliefs and preferences I could have about Bob, there is a decision-theoretic answer to how I should then act in order to satisfy my preferences.” [2]
This implies that instrumental rationality is specific; from the above statement, I infer:
“For any decision problem to any rational agent with a specified psyche, there is only one correct choice to make.”
However, if we only seek to systematically achieve our values, I believe that instrumental rationality fails to be specific—it is possible that there’s more than one solution to a problem in which we merely seek to maximise or values. I cherish the specificity of rationality; there is a certain comfort, in knowing that there is a single correct solution to any problem, a right decision to make for any game—one merely need find it. As such, I sought a definition of rationality that I personally agree with; one that satisfies my criteria for specificity; one that satisfies my criteria for winning. The answer I arrived at was: “Rationality is systematically achieving your goals.”
I love the above definition; it is specific—gone is the vagueness and uncertainty of achieving values. It is simple—gone is the worry over whether value X should be an instrumental value or a terminal value. Above all, it is useful—I know whether or not I have achieved my goals, and I can motivate myself more to achieve them. Rather than thinking about vague values I think about my life in terms of goals:
“I have goal X how do I achieve it?”
If necessary, I can specify sub goals and sub goals for those sub goals. I find that thinking about your life in terms of goals to be achieved is a more conducive model for problem solving, a more efficient model—a useful model. I am many things, and above them all I am a utilitist—the worth of any entity is determined by its utility to me. I find the model of rationality as a goal enabler a more useful model.
Goals and values are not always aligned. For example, consider the problem below:
Jane is the captain of a boat full with 100 people. The ship is about to capsize and would, unless ten people are sacrificed. Jane’s goal is to save as many people as possible. Jane’s values hold human lives sacred. Sacrificing ten people has a 100% chance of saving 90 people, while sacrificing no one and going with plan delta has a 10% chance to save the 100, and a 90% chance for everyone to die.
The sanctity of human life is a terminal value for Jane. Jane when seeking to actualise her values, may well choose to go with plan delta, which has a 90% chance to prevent her from achieving her goals.
Values may be misaligned with goals, values may be inhibiting towards achieving our goals. Winning isn’t achieving your values; winning is achieving your goals.
Goals
I feel it is apt to define goals at this juncture, lest the definition be perverted and only goals aligned with values be considered “true/good goals”.
Goals are any objectives a self aware agent consciously assigns itself to accomplish.
There are no true goals, no false goals, no good goals, no bad goals, no worthy goals, no worthless goals; there are just goals.
I do not consider goals something that “exist to affirm/achieve values"—you may assign yourself goals that affirm your values, or goals that run contrary to them—the difference is irrelevant, we work to achieve those goals you have specified.
The Psyche
The Psyche is an objective map that describes a self-aware agent that functions as a decision maker—rational or not. The sum total of an individual’s beliefs—all knowledge is counted as belief—values and goals form their psyche. The psyche is unique to each individual. The psyche is not a subjective evaluation of an individual by themselves, but an objective evaluation of the individual as they would appear to an omniscient observer. An individual’s psyche includes the totality of their map. The psyche is— among other things—a map that describes a map so to speak.
When a decision problem is considered, the optimum solution to such a problem cannot be considered without considering the psyche of that individual. The values that individual holds, the goals they seek to achieve and their mental map of the world.
Eliezer Yudkowsky seems to believe that we have an extremely limited ability to alter our psyche. He posits, that we can’t choose to believe the sky is green at will. I never really bought this, and especially due to personal anecdotal evidence. Yet, I’ll come back to altering beliefs later.
Yudkowsky describes the human psyche as: “a lens that sees its own flaws”. [3] I personally would extend this definition; we are not merely “a lens that sees its own flaws”, we are also “a lens that corrects itself”—the self-aware AI that can alter its own code. The psyche can be altered at will—or so I argue.
I shall start with values. Values are neither permanent nor immutable. I’ve had a slew of values over the years; while Christian, I valued faith, now I adhere to Thomas Huxley’s maxim:
Scepticism is the highest of duties; blind faith the one unpardonable sin.
Another one: prior to my enlightenment I held emotional reasoning in high esteem, and could be persuaded by emotional arguments, after my enlightenment I upheld rational reasoning. Okay, that isn’t entirely true; my answer to the boat problem had always been to sacrifice the ten people, so that doesn’t exactly work, but I was more emotional then, and could be swayed by emotional arguments. Before I discovered the Way earlier this year (when I was fumbling around in the dark searching for rationality) I viewed all emotion as irrational, and my values held logic and reason above all. Back then, I was a true apath, and completely unfeeling. I later read arguments for the utility of emotions, and readjusted my values accordingly. I have readjusted my values several times along the journey of life; just recently, I repressed my values relating to pleasure from feeding—to aid my current routine of intermittent fasting. I similarly repressed my values of sexual arousal/pleasure—I felt it will make me more competent. Values can be altered, and I suspect many of us have done it at least once in our lives—we are the lens that corrects itself.
Getting back to belief (whether you can choose to believe the sky is green at will) I argue that you can, it is just a little more complicated than altering your values. Changing your beliefs—changing your actual anticipation controllers—truly redrawing the map, would require certain alterations to your psyche in order for it to retain a semblance of consistency. In order to be able to believe the sky is green, you would have to:
- Repress your values that make you desire true beliefs.
- Repress your values that make you give priority to empirical evidence.
- Repress your vales that make you sceptical.
- Create (or grow if you already have one) a new value that supports blind faith.
Optional:
- Repress your values that support curiosity.
- Create (or grow if you already have one) a new value that supports ignorance.
By the time, you’ve done the ‘edits’ listed above, you would be able to freely believe that the sky is green, or snow is black, or that the earth rests on the back of a giant turtle, or a teapot floats in the asteroid belt. I’m warning you though, by the time you’ve successfully accomplished the edits above, your psyche would be completely different from now, and you will be—I argue—a different person. If any of you were worried that the happiness of stupidity was forever closed to you, then fear not; it is open to you again—if you truly desire it. Be forewarned; the “you” that would embrace it would be different from the “you” now, and not one I’m sure I’d want to associate with. The psyche is alterable; we are the masters of our own mind—the lens that corrects itself.
I do not posit, that we can alter all of our psyche (I suspect that there are aspects of cognitive machinery that are unalterable; “hardcoded” so to speak. However, my neuroscience is non-existent—as such I shall leave this issue to those more equipped to comment on it.
Values as Tools
In my conception of instrumental rationality, values are no longer put on a pedestal, they are no longer sacred; there are no more terminal values anymore—only instrumental. Values aren’t the masters anymore, they’re slaves—they’re tools.
The notion of values as tools may seem disturbing for some, but I find it to be quite a useful model, and such I shall keep it.
Take the ship problem Jane was presented with above, had Jane deleted her value which held human life as sacred, she would have been able to make the decision with the highest probability of achieving her goals. She could even add a value that suppressed empathy, to assist her in similar situations—though some might feel that is overkill. I once asked a question on a particular subreddit:
“Is altruism rational?”
My reply was a quick and dismissive:
“Rationality doesn’t tell you what values to have, it only tells you how to achieve them.”
The answer was the standard textbook reply that anyone who had read the sequences or RAZ (Rationality: From AI to Zombies) would produce; I had read neither at the time. Nonetheless, I was reading HPMOR (Harry Potter and the Methods of Rationality), and that did sound like something Harry would say. After downloading my own copy of RAZ, I found that the answer was indeed correct—as long as I accepted Yudkowsky’s conception of instrumental rationality. Now that I reject it, and consider rationality as a tool to enable goals, I have a more apt response:
What are your goals?
If your goals are to have a net positive effect on the world (do good so to speak) then altruism may be a rational value to have. If your goals are far more selfish, then altruism may only serve as a hindrance.
The utility of “Values as Tools” isn’t just that some values may harm your goals, nay it does much more. The payoff of a decision is determined by two things:
- How much closer it brings you to the realisation of your goals?
- How much it aligns with your values?
Choosing values that are doubly correlated with your current goals (you actualise your values when you make goal conducive decisions, and you run opposite to your values when you make goal deleterious decisions) exaggerates the positive payoff of goal conducive decisions, and the negative payoff of goal deleterious decisions. This aggrandising of the payoffs of decisions serves as a strong motivator towards making goal conducive decisions— large rewards, large punishment—a perfect propulsion system so to speak.
The utility of the “Values as Tools” approach is that it serves as a strong motivator towards goal conducive decision making.
Conclusion
It has been brought to my attention that a life such as the one I describe may be “an unsatisfying life” and “a life not worth living”. I may reply that I do not seek to maximise happiness, but that may be dodging the issue; I first conceived rationality as a value decider when thinking about how I would design an AI—it goes without saying that humans are not computers.
I offer a suggestion: order your current values in a scale of preference. Note the value (or set thereof) utmost in your scale of preference. The value that if you had to achieve only one value, that you would choose. Pick a goal that is aligned with that value (or set thereof). That goal shall be called your “prime goal”.The moment you pick your prime goal, you fix it. From now on, you no longer change your goals to align with your values. You change your values to align with your goals. Your aim in life is to achieve your prime goal, and you pick values and subgoals that would help you achieve your prime goal.
References:
[1] Eliezer Yudkowsky, “Rationality: From AI to Zombies”, pg 7, 2015, MIRI, California.
[2] Eliezer Yudkowsky, “Rationality: From AI to Zombies”, pg 203, 2015, MIRI, California.
[3] Eliezer Yudkowsky, “Rationality: From AI to Zombies”, pg 40, 2015, MIRI, California.