How to be Miserable
A Mathematical Model I Find Useful
By default, mental agents are negative. That is, they guide us away, rather than towards, certain behaviors. I find this to be the case empirically, but it also makes sense given the strongest agents typically form in response to bad experiences. Let's create a simple mathematical model. The end goal is, as always, is representing actions as points in high dimensional space interfacing with multiple very complex semi-correlated functions. But I want to develop some intuition in this case, so let's start simple. Let's say you're just picking a point from $\mathbb{R}$, and feeding it into a single function with range $[0, 1]$.
Let's call this function $1-0.9^{|\lfloor \log_w(|x|) \rfloor - 5|}$1, or "dislikes when you have too many or too few powers of w." Maybe your parents yelled at you when you were a 17 digit number as a child. Who knows. The point is, it's very easy to be happy as this function; just pick a number around $w^5$, and you output zero. In this case, the domain represents action space, and the range is your misery score; the farther from zero, the more miserable the action makes you. Now let's add a second function: $\left|\frac{x \bmod 10}{5} - 1\right|$, or "dislikes when you don't end in something close to 5." This function notably has no correlation with the first (for most w), so the entire range of $[0, 1]^2$ is achievable. Also note that because misery is euclidean distance from 0, $(0, 1)$ is a bit worse than $(0.5, 0.5)$ — this effect is, of course, vastly exacerbated in higher dimensions. I claim this is realistic; egregious violation of a single agent is, in most cases, far more uncomfortable than a slight violation of many.
Let's say function three cares about unique prime factors under ten of $\lfloor |x| \rfloor$, incentivising fewer of them. Suddenly, the best choice is not so obvious. The point is, this function is negatively correlated with function 2; it is no longer possible to have a misery value of zero (your misery coordinate can also no longer be $(1, 1, 1)$, but given this system assumes perfect play, this is far less relevant). The question posed by this system is not one of behavior, but one of limitation. The nature of agents is one of framework, and as such, this model sheds light on the limitations of how happy you can be assuming perfect action given a set framework, rather than investigating what perfect action looks like. Of course, it is possible to change agents over time, and so this is, of course, an imperfect map. But it seems to me that this incongruity is decently tangential to the point of the investigation in this case (as of yet), so I intend to ignore it.
To add a bit of nuance, let's replace $\mathbb{R}$ with $\mathbb{R}^2$. There are a few different ways to treat this, each of which carries its own interpretation. I will put forward two options: simple, and good. The easiest answer is just using multivariable functions as our agent representations. This allows the domain to act not only as an action space, but also as a situation space (i.e. points now act as all actions across all situations). This unfortunately loses the nice intuition of a single best point in the action space that we look for. We are instead left with an area-under-the-curve situation. This leaves us with overvaluing ridiculous situations, which can be fixed by discounting points far from zero in the action space, but it's not so clear how to effectively implement this. If you prefer this model, I think it's best to sweep the math under the rug a bit, and accept the easy intuition it gives: functions that take all dimensions as input are more universal, while isolated ones are specific. For example, say our first function uses distance from 0 in action space. This is a decent analog for a general fear of acting out, being judged, being different, etc, and acts as a general incentive towards more normal behavior. Meanwhile, the "end in 5" function might only address the first coordinate, making it analogous to a more specific preference (e.g. you dislike when a food is particularly sour).
The complex model is, strictly speaking, isomorphic. Here we have two inputs: situation space, and action space. Our agents are now represented by function pairs: our misery functions, each from action space to $[0, 1]$, and weight functions, each from situation space to $[0, 1]$, which determine how active the corresponding misery agent is in that specific situation. The two outputs are multiplied to get the final agent misery score. This model is also imperfect — these two functions are definitely not independent; certain actions in an otherwise benign situation might trigger otherwise dormant agents. Still, I think in the vast majority of situations, the model behaves quite well. It also lends itself to a much more concrete area under the curve notion: speaking nebulously, situation space can be organized placing more outlandish scenarios further from the origin, allowing us to integrate the minimum misery scores of our situation surface in a way that makes some sense. Black boxing, we have a function that, for a given set of agents, takes a situation and outputs a misery score in a way that should (?)2 be continuous if done correctly. Taking $\mathbb{R}^2$ as an example, we lift it into $\mathbb{R}^3$ with $z = \left(\frac{1}{2}\right)^{\|(x, y)\|}$, then scale by our continuous misery surface, and integrate to get a final misery value of our agent set. Obviously this is very reductive, and the numbers would need a bit of tweaking to make sense, but to me this seems to act as a decent model for very simply evaluating agent systems. Returning to our intuition check from earlier, function one, the "don't act out" agent, would just have a situation function of $f(x,\ldots) = 1$, meaning it's universal (though in reality it may have slight exceptions), while the "don't like sour food" rule may have a situation function of $f(x,\ldots, y, \ldots, z) = 0.5y$, where y indicates whether the situation pertains to eating.
So what set of agents makes you as miserable as possible?
The antithesis, what makes you as happy as possible, is surprisingly boring. The best system is one without agents. This obviously isn't quite right — it relies on the assumption that agents are far more negative by default. While I do think there's a truth behind this — that the fewer restrictions you place on your own happiness, the better your odds are — it's obviously not quite right. I oriented the model in this direction on purpose — I think it's much easier to be miserable all the time than to be happy all the time. I find not being miserable is a much bigger part of being happy than not being happy is of being miserable.
Just make two agents have a correlation of -1. Actually, this isn't possible, as the range is $[0, 1]$. But let's preserve that idea, and have two agents be in direct opposition and always sum to 1. Let's also make them as universal as possible — in the simple model, make them include every dimension, in the complex model make the mean situation function output ~1. Even assuming every other agent is satisfied, this gives you a minimum misery score of ~0.7 for every single situation. Put differently, you are constantly miserable. Congrats! The analog of this is that every choice feels wrong for one of two reasons. Note that the least miserable point, the 0.7, is where both agents are at 0.5, representing the middle road, being a little miserable in both directions. I think of this as inaction.
But surely it's not so easy to find a universal agent, along with its antithesis? Well, Id is pretty universal. If anything, it's as universal as an agent could possibly be. There's a culture I've seen emerge recently, particularly in SF, that idolizes self control. The ideal is someone with complete dominance over Id, to the point that wanting to do something is a sufficient reason not to do it. Setting the superego as the active antithesis of Id like this is really bad. Given Freud is the ~simplest possible agent system, this is, as far as I can tell, the easiest way to be unhappy.
The much simpler example is burnout. The cycle of burnout is fundamental a pendulum between Id and superego — to use the complex framing, as one passes through the cycle of burnout, the strength of the Id agent and the superego agent wax and wane as they are starved and sated respectively, pushing the cycle forwards. To keep pushing forwards along the cycle is actually to keep walking the line of minimum misery at any given time. The state of self justifying fatigue and inaction is a state of $(0.5, 0.5)$, trying to keep either agent from cutting too deep. Incidentally, this is also why the cure to burnout is taking some time off without hating yourself for it — this allows the situation functions of the two agents to reset to a healthy balance. The idea that both overwork and underwork set you spiraling also fits into this notion nicely.
I do think that the culture in SF is uniquely bad on this front. Between overwork and underwork, there exists a stable state where both agents are somewhat sated. The size of this stable state, however, depends on how diametrically opposed your id and superego actually are. One major factor is how much you expect of yourself, another is how much the culture expects of you. This being the case, I think certain environments in SF, particularly among some of the smartest kids in the country, have no stable state. As a result, practically every kid is burnt out, cycling, and decently miserable.
This also explains why so many people working on AIS are so miserable — orienting yourself around needing to save the world leaves no room for Id without guilt, and thus no room for a healthy stable state.
While I have written no solutions here, I find this framework helpful in rationalizing an internal process I have long found enigmatic.