[PROLOGUE – EVERYBODY WANTS A ROCK]
[PART I – THERMOSTAT]
[PART II – MOTIVATION]
[PART III – PERSONALITY AND INDIVIDUAL DIFFERENCES]
[PART IV – LEARNING]
[PART V – DEPRESSION AND OTHER DIAGNOSES]
[PART VI – CONFLICT AND OSCILLATION]
[PART VII – NO REALLY, SERIOUSLY, WHAT IS GOING ON?]
[INTERLUDE – I LOVE YOU FOR PSYCHOLOGICAL REASONS]
[PART VIII – ARTIFICIAL INTELLIGENCE]
[PART IX – ANIMAL WELFARE]
[PART X – DYNAMIC METHODS]

---

There’s a fascinating little paper called Physiological responses to maximal eating in men. 

The researchers recruited fourteen men (mean age: 28 years old) and invited them back to the lab to eat “a homogenous mixed-macronutrient meal (pizza)”. The authors note that “this study was open to males and females but no females signed up.” 

They invited each man to visit the lab two separate times. On one occasion, the man was asked to eat pizza until “comfortably full”. The other time, he was asked to eat pizza until he “could not eat another bite”.

When asked to eat until “comfortably full”, the men ate an average of about 1500 calories of pizza. But when asked to eat until they “could not eat another bite”, the men ate an average of more than 3000 calories. 

The authors view this as a study about nutrition, but we saw it and immediately went, “Aha! Pizza psychology!”

While this isn’t a lot of data — only fourteen men, and they only tried the challenges one time each — it shows some promise as a first step towards a personality measure of hunger and satiety, because it measures both how hungry these boys are, and also how much they can eat before they have to stop.

When asked to aim for “could not eat another bite”, the men could on average eat about twice as much pizza compared to when they were asked to aim for “comfortably full”. But there was quite a lot of variation in this ratio for different men:  

All the men ate more when they were asked to eat as much as they could, than when they were asked to eat as much as they liked. But there’s a lot of diversity in the ratio between those two values. When instructed to eat until they “could not eat another bite”, some men ate only a little bit more than they ate ad libitum. But one man ate almost three times as much when he was told to go as hard as he can. 

People have some emotions that drive them to eat (collectively known as hunger), and other emotions that drive them to stop eating (collectively known as satiety). While these pizza measurements are very rough, they suggest something about the relationship between these two sets of drives in these men. If nothing else, it’s reassuring to see that for each individual, the “could not eat another bite” number is always higher. 

It’s a little early to start using this as a personality measure, but with a little legwork to make it reliable, we might find something interesting. It could be the case, for example, that there are some men with very little daylight between “comfortably full” and “could not eat another bite”, and other men for whom these two occasions are like day and night. That would suggest that some men’s hunger governor(s) are quite strong compared to their satiety governor(s), and other men’s are relatively weak. 

The general principle of personality in cybernetic psychology is “some drives are stronger than others”. So for personality, we want to invent methods that can get at the question of how strong different drives are, and how they stack up against each other. Get in loser, we’re making a tier list of the emotions. 

We may not be able to look at a drive and say exactly how strong it is, since we don’t yet know how to measure the strength of a drive. We don’t even know the units. When this is eventually discovered, it will probably come from an unexpected place, like how John Dalton’s work in meteorology gave him the idea for the atomic theory. 

But we can still get a decent sense of how strong one drive is compared to another drive. This is possible whenever we can take two drives and make them fight. 

Some drives are naturally in opposition — this pizza study is a good example. The satiety governor(s) exist specifically to check the hunger governor(s). Hunger was invented to start eating, and satiety was invented to make it stop. So it’s easy to set up a situation where the two of them are in conflict. 

Or somewhat easy. We think it’s more accurate to model the pizza study as the interaction between three (groups of) emotions. When asked to eat until “comfortably full”, the hunger governor voted for “eat pizza” until its error was close to zero, then it stopped voting for “eat pizza”, so the man stopped. That condition was simple and mainly involved just the one governor.

The other condition was more complex. When asked to eat until they “could not eat another bite”, the hunger governor first voted for “eat pizza” until its error was close to zero. Then, some kind of “please the researchers” governor(s) kept voting for “eat pizza” to please the researchers. 

At some point this started running up against the satiety governor. The satiety governor tracks something like how full you are, so as the man started to get too full, the satiety governor started voting against “eat pizza”. The man kept eating until the vote from the “please the researchers” governor(s) was just as strong as the vote from the satiety governor, at which point the two votes cancel out and the man could not eat another bite. 

This reveals the problem. In one sense, hunger and satiety are naturally in opposition. Hunger tries to make you eat enough and satiety tries to make sure you don’t eat enough too much. But in a healthy person, there’s plenty of daylight between the set points of these two drives, and they don’t come into conflict. 

Same thing with hot and cold — the drive that tries to keep you warm is in some sense “in opposition” to the drive that tries to keep you from overheating, but they don’t normally fight. If you have a sane and normal mind, you don’t put on 20 sweaters, then overheat, then in a fit of revenge take off all of your clothes and jump in a snowbank, etc. These drives oppose each other along a single axis, but when they are working correctly, they keep the variable they care about in a range that they agree on. Hunger and satiety, and all the paired governors, are more often allies than enemies. 

But any two drives can come into conflict when the things they want to do become mutually exclusive, or even just trade off against each other. Even if you can do everything you want, the drives will still need to argue about who gets to go first. Take something you want, anything at all, and put it next to a tiger. Congratulations, fear is now in conflict with that original desire. 

Many people experience this conflict almost every morning:

This is actually a more complicated situation, where the governors have formed factions. The pee governor wants to let loose on your bladder. But your hygiene governor votes against wetting the bed. Together they settle on a compromise where you get up and pee in the toilet instead, since this satisfies both of their goals (bladder relief + hygienic). 

But the governor that keeps you warm, the sleep governor (who wants to drift back into unconsciousness), and any other governors with an interest in being cozy, strenuously oppose this policy. They want you to stay in your warm, comfy bed. So you are at an impasse until the bladder governor eventually has such a strong error signal — you have to take a leak so bad — that it has the votes to overrule the cozy coalition and motivate you to get up and go to the bathroom. 

The point is, the bladder governor, warmth governor, and sleep governor don’t fundamentally have anything to do with each other. They all care about very different things. But when you have to pee in the middle of the night, their interests happen to be opposed. They draw up into factions, and this leads to a power struggle — one so universal that there are memes about it. And as is always the case in politics, a power struggle is a good chance to get a sense of the relative strength of the factions involved.

If you met someone who said they didn’t relate to this — they always get up in the middle of the night to pee without any hesitation or inner struggle — this would suggest that their bladder governor is very strong, or that their warmth and/or sleep governors are unusually weak. Whatever the case, their bladder governor wins such disagreements so quickly that there doesn’t even appear to be a dispute. 

In contrast, if your friend confesses that they have such a hard time getting up that they sometimes wet the bed, this suggests that their bladder governor, and probably their hygiene governor, are unusually weak compared to the governors voting for them to stay in bed. 

To understand these methods, we have to understand the difference between two kinds of “strength”. 

In general when we say that a drive is strong, we mean that it can meet its goals, it can vote for the actions it wants. This is why we can learn something about the relative strength of two drives by letting them fight — we can present the organism with mutually exclusive options (truth or dare?) and see which option it picks. If we have some reasonable idea which drive would pick which option, we know which drive is stronger from which option is picked. 

However! Another way a drive can be strong is that it can have a big error signal in that moment. If you are ravenously hungry, you will eat before anything else. If you are in excruciating pain, you will pull your hand off the stove before doing anything else. This kind of urgency tells us that the current error is big, but it doesn’t tell us much about the governor. 

A drive does get a stronger vote when its variable is further off target. But it’s also true that for a given person, some drives seem stronger in all situations. 

The normal sense of strength gets at the fact that a governor can be stronger or weaker for a given error. Some people can go to sleep hungry without any problem. For other people, even the slightest hint of appetite will keep them awake. When we talk about someone being aggressive, we mean that they will drop other concerns if they see a chance to dominate someone; if we talk about someone being meek, we mean the opposite. 

The current strength of any drive is a function of the size of its current error signal and the overall strength or “weight” of the governor. Unfortunately, we don’t know what that function is. Also, it might be a function of more than just those two things. Uh-oh! 

Ideally, what we would do is hold the size of the error constant. If we could make sure that the error on the salt governor is 10 units, and the error on the sweet governor is 10 units, then we could figure out which governor is stronger by seeing which the person would choose first, skittles or olives. This is based on the assumption that the strength of the vote for each option is a combination of the size of the errors and the strength of the governor itself. Since in this hypothetical we know that the strength of the errors is exactly the same, the difference in choice should be entirely the result of the difference in the strength of the governors.

Unfortunately we don’t know how to do that either. We don’t know how to measure the errors directly, let alone how to hold the size of the errors constant. 

But we can use techniques that should make the size of some error approximately constant, and base our research on that. The closer the approximation, the better. 

The important insight here is that even when we can’t make measurements in absolute terms, we can often make ordinal comparisons. “How strong is this drive” is an impossible question to answer until we know more about how strength is implemented mechanically, but we can make very reasonable guesses about which of two drives is stronger, what order their strengths are in, i.e. ordinal measurements. 

We can do this two ways: we can compare one of your drives to everyone else’s version of that same drive, or we can compare one of your drives to your other drives.

Compare One of Your Drives to Everyone Else’s Version of that Same Drive

The first is that we can compare one of a person’s drives to the same drive in other people. 

It’s reasonable to ask if your hunger, fear, pain, or shame drive is stronger or weaker than average. To do this, we can look at two or more individuals and ask if the drive is stronger for one of them or for the other. 

This will offer a personality measure like: your salt governor is stronger than 98% of people. You a salty boy.

Again, to get a measure of strength, we need to make everyone’s errors approximately constant. One way we can make errors approximately constant is by fully satisfying the drive. So if we identify a drive, like the drive for salt, we can exhaust the drive by letting people eat as much salt or salty food(s) as they want. Now all their errors should be close to zero. Then we can see how long it takes before they go eat something salty again. If someone goes to get salty foods sooner, then other things being equal, this is a sign that their salt governor is unusually strong.

This won’t be perfectly the same, and other things will not be perfectly equal. Some people’s salt error may increase more quickly than others’, like maybe they metabolize salt faster, or something. So after 5 hours without salty foods, some people’s error may be much bigger than others’. But it should be approximately equal, and certainly we would learn something important if we saw one guy who couldn’t go 10 minutes without eating something salty, and someone else who literally never seemed to seek it out. 

When we say things like, “Johnnie is a very social person. If he has to spend even 30 minutes by himself he gets very lonely, so he’s always out and spending time with people. But Suzie will go weeks or even months without seeing anyone,” this is a casual version of the same reasoning, and we think it’s justified. It may not get exactly at the true nature of personality, but it’s a start. 

When we figure out what the targets are for some governors, we’ll be able to do one better. For example, let’s imagine that we find out that thirst is the error for a governor that controls blood osmolality, and through careful experimentation, we find out that almost everyone’s target for blood osmolality is 280 mOsm/kg. Given the opportunity, behavior drives blood osmolality to 280 mOsm/kg and then stops.

If we measure people’s blood osmolality, we can dehydrate them to the point where their blood osmolality is precisely 275 mOsm/kg. We know that this will be an error of 5 mOsm/kg, because that’s 5 units less than the target. Then we would know almost exactly what their error is, and we could estimate the relative strength of their thirst governor by measuring how hard they fight to get a drink of water. 

On that note, it’s possible that a better measure than time would be effort. For example, you could take a bunch of rats and figure out the ideal cage temperature for each of them. Separately, you teach them that pushing a lever will raise the temperature of their cage by a small amount each time they press it. 

Then, you set the cage temperature 5 degrees colder than they prefer. This should give them all errors of similar magnitude — they are all about 5 degrees colder than they’d like. Then you give them the same lever they were trained on. But this time, it’s disconnected. You count how many times they press the lever before they give up. This will presumably give you a rough measure of how much each rat is bothered by being 5 degrees below target, and so presumably an estimate of the strength of that governor. If nothing else, you should observe some kind of individual difference. 

Compare One of Your Drives to Your Other Drives

The second approach is to ask how your drives compare to each other, basically a ranking. We can look at a single person and ask, in this person, is drive A stronger than drive B? 

The main way to do this is to give the person a forced choice between two options, one choice that satisfies governor A, and the other that satisfies governor B. This doesn’t have to be cruel — you can let them take both options, you just have to just make them choose which they want to do first.

This would offer a personality measure like: you are more driven by cleanliness than by loneliness, which is why you keep blowing off all your friends to stay in and scrub your toilet.

There are some drives that make us want to be comfortable and other drives that make us want to be fashionable; there are at least some tradeoffs between comfort and fashion; if you reflect on each of the people in your life, it’s likely that you already know which coalition of drives tends to be stronger in each person.  

Every time you see someone skip work to play videogames, refuse to shower even when it ruins all their friendships, blow up their life to have an affair with the 23-year-old at the office, or stay up late memorizing digits of pi, you are making this kind of personality judgment implicitly. People have all kinds of different drives, and you can learn a lot about which ones are strongest by seeing which drives are totally neglected, and which drives lead people to blithely sacrifice all other concerns, as though they’re blind to the consequences.

The Bene Gesserit, a sect of eugenicist, utopian nuns from the Dune universe, use a simplified version of this method in their famous human awareness test, better known as the gom jabbar. Candidates are subjected to extreme pain and ordered not to pull away, at penalty of taking a poisoned needle in the neck. In his success, Paul demonstrates that some kind of self-control governor is much stronger than his pain governor, even when his pain error is turned way up.

But no shade to the Bene Gesserit, this is not a very precise measure. By turning the pain governor’s error extremely high, they can show that a candidate has exceptional self-control. But this doesn’t let them see if self-control is in general stronger than pain, because the error gets so huge. To compare the strength of governors, you ideally want the error signals to be as similar as possible.

As before, the best way to get at strength is to take two drives, try to make their errors as similar as possible, and then see which drive gets priority. Other things being equal, that drive must be stronger. 

When we were trying to compare personality between people, this was relatively easy. If nothing else, we were at least looking at the same error. We can’t get an exact measure of the error, but we could at least say, both of these people have gone 10 hours without eating, or 20 hours without sleep, or are ten degrees hotter than they find comfortable. These are the same kinds of things and they are equal for both people.

But to compare two governors within a single person, we are comparing two different errors, and we have no idea what the units are. So it may be hard to demonstrate differences between the strength of the governors when those differences are small. If one error is ten times stronger than the other, then we assume that the governor behind that error will win nearly all competitions between the two of them. If one error is 1.05 times stronger than the other, that governor has an edge, but will often get sidelined when there are other forces at play.

But like the common-sense examples above, it should be possible to make some comparisons, especially when differences are clear. For example, if we deprive a person of both sleep and food for 48 hours (with their consent of course), then offer them a forced choice between food and sleep, and they take the food, that suggests that their drive to eat may be stronger than their drive to sleep. This is especially true if we see that other people in the same situation take the option to sleep instead. 

If we deprive the person of sleep for 48 hours and food for only 4 hours, and they still choose the food over sleep, that is even better evidence that their drive to eat is stronger than their drive to sleep, probably a lot stronger. 

While these methods are designed to discover something inside an individual person, they might also shed some light on personality differences between people. For example, we might find that in most people, the sugar governor is stronger than the salt governor. But maybe for you, your salt governor is much stronger than your sugar governor. That tells us something about your personality in isolation (that one drive is stronger than another), and also tells us something about your personality compared to other people (you have an uncommon ordering of drives). 

Return to Pizza Study

The pizza study is interesting because it kind of combines these techniques.

Each person was compared on two tasks — “comfortably full” and “could not eat another bite”, which gives us a very rough sense of how strong their hunger and satiety governors are. If you ate 10 slices to get to “comfortably full” and only 12 slices to get to “could not eat another bite”, your satiety governor is probably pretty strong, since it kicks in not long after you ate as much as you need. (There could be other interpretations, but you get the gist.) 

In addition, each person can be compared to all the other people. Some men could eat only a little more when they were asked to get to “could not eat another bite”. But one man ate almost three times as much as his “comfortably full”. This man’s satiety governor is probably weaker than average. There are certainly other factors involved, but it still took a long time before that governor forced him to stop eating, suggesting it is weak. 

A final note on strength. The strength of a governor is probably somewhat innate. But it may also be somewhat the result of experience. If someone is more motivated by safety than by other drives, some of that may be genetic, but some of that may be learned. It would not be ridiculous to think that your mind might be able to tune things so that if you have been very unsafe in your life, you will pay more attention to safety in the future.

Even the part that’s genetic (or otherwise innate) still has to be implemented in some specific way. When one of your governors is unusually strong, does that governor have a stronger connection to the selector? Does it have the same connection as usual, but it can shout louder? Does it shout as loud as normal, but it can shout twice as often? We don’t know the details yet, but keep in mind that all of this will be implemented in biology and will include all kinds of gritty details. 

Deeper Questions

People can differ in more ways than just having some of their drives be stronger than others. For example, some people are more active than other people in general, more active for every kind of drive. They do more things every single day. 

Some people seem to get more happiness from the same level of accomplishment. For some people, cooking dinner is a celebration. For others, routine is routine. 

Some people seem more anxious by default. Even a small thing will make them nervous. 

These seem like they might be other dimensions on which people can differ, and they don’t seem like they are linked to specific governors. 

Studying the strength of the governors is nice because the governors are all built on basically the same blueprint, so the logic needed to puzzle out one of them should mostly work to puzzle out any of the others. The methods used to study one governor should work to study all of them, only minor tweaks required. If you find techniques to measure the strength of one governor, you should be able to use those techniques to measure the strength of any governor.

But other ways in which people differ seem more idiosyncratic. They are probably the result of different parameters that tune features that are more global, each of which interacts with the whole system in a unique and different way. So we will probably need to invent new methods for each of them. 

That means we can’t yet write a section on the different methods that will be useful. These methods still need to be invented. And we might only get to these methods once we have learned most of what there is to know about the differences in strength between the governors, and have to track down the remaining unexplained differences between people. But we can give a few examples to illustrate what some of these questions and methods might look like.

Learning

Every governor has to have some way of learning which behaviors increase/decrease their errors. We don’t know exactly how this learning works yet, but we can point to a few questions that we think will be fruitful.

For example, is learning “both ways”? 

The hot governor (keeps you from getting too hot) and the cold governor (keeps you from getting too cold) both care about the same variable, body temperature. Certainly if you are too cold and you turn on a gas fireplace, your cold governor will notice that this corrects its error and will learn that turning on the gas fireplace is a good option. So when you get too cold in the future, that governor will sometimes vote for “turn on the gas fireplace”.

But what if you are too hot and you turn on the gas fireplace? Well, your hot governor will notice that this increases its error, and will learn that this is a bad option, which it will vote against if you’re in danger of getting too hot. 

What does your cold governor learn in this situation? Maybe it learns the same thing your hot governor does — that the gas fireplace increases temperature. The hot governor thinks that’s a bad outcome, but the cold governor thinks it’s a good outcome. If so, then next time you are cold, the cold governor might vote for you to turn on the gas fireplace. 

But maybe a governor only learns when its error is changed. After all, each governor only really cares about the error it’s trying to send to zero. And if that error isn’t changed, maybe the governor doesn’t pay attention. If the error is very small, maybe that governor more or less turns off, and stops paying attention, to conserve energy. Then it might not do any learning at all. 

If this were the case, the cold governor shouldn’t learn from any actions you take when you’re too hot, even when these actions influence your body temperature. And the hot governor shouldn’t learn from anything you do when you’re too cold, same deal. 

You could test this by putting a mouse in a cage that is uncomfortably hot, and that contains a number of switches. Each switch will either temporarily increase or temporarily decrease the temperature of the cage. With this setup, the mouse should quickly learn which switches to trip (makes the cage cooler) and which switches to avoid (makes the cage even more uncomfortably hot). 

Once the mouse has completely learned the switches, then you make the cage uncomfortably cold instead, and see what happens. If the cold governor has also been learning, then the mouse should simply invert its choice of switches, and will be just as good at regulating the cage temperature as before. 

But if the cold governor wasn’t paying close attention to the hot governor’s mistakes, then the mouse will have to do some learning to catch up. If the cold governor wasn’t learning from the hot governor’s mistakes at all, then the mouse will be back at square one, and might even have to re-learn all the switches through trial and error.

We definitely might expect the former outcome, but you have to admit that the latter outcome would be pretty interesting. 

The Model of Happiness

Or consider the possibility that happiness might drive learning.

This would explain why happiness exists in the first place. It’s not just pleasant, it’s a signal to flag successful behavior and make sure that it’s recorded. When something makes you happy, that signals some system to record the link between the recent action and the error correction.

This would also explain why it often feels like we are motivated by happiness as a reward. We aren’t actually motivated by happiness itself, but when something has made us happy, we tend to do it more often in the future. 

Previously we said that happiness is equal to the change in an error. In short, when you correct one of your errors, that creates a proportional amount of happiness. This happiness sticks around for a while but slowly decays over time. 

That’s a fine model as a starting point, but it’s very simple. Here’s a slightly more complicated model of happiness, which may be more accurate than the model we suggested earlier. Maybe happiness is equal to the reduction in error times the total sum of all errors, like so:

happiness = delta_error * sum_errors

If happiness is just the result of the correction of an error, then you get the same amount of happiness from correcting that error in any circumstance. But that seems a little naïve. A drink of water in the morning after a night at a five-star hotel is an accomplishment, but the same drink of water drawn while hungry and in pain, lost in the wilderness, is a much greater feat. Remembering the strategy that led to that success might be more important. 

If you multiply the correction by the total amount of error, then correcting an error when you are in a rough situation overall leads to a much greater reward, which would encourage the governors to put a greater weight on successes that are pulled off in difficult situations. If you correct an error when all your other errors are near zero, you will get some happiness. But if you are more out of alignment generally — more tired, cold, lonely, or whatever — you get more happiness from the same correction.

This might explain fetishes. Why do so many sexual fetishes include things that cause fear, pain, disgust, or embarrassment? Surely the fear, pain, disgust, and embarrassment governors would vote against these things. 

We have to assume that the horny governor is voting for these things. The question is, why would it vote for anything more than getting your rocks off? Why would an orgasm plus embarrassment be in any way superior to an orgasm in isolation?

If learning is based on happiness rather than raw reduction in error, then governors will learn to vote for things that have caused past happiness.

And if happiness is a function of total error, not just correction in the error they care about, governors will sometimes vote for things that increase the total error just before their own error is corrected. 

The point is, if happiness is a function of total error, governors will actually prefer to reduce their errors in a state of greater disequilibrium. This doesn’t decrease their error any more than in a state of general calm, but it does lead to more happiness, greater learning, and so they learn to perform that action more often. And in some cases they will actually vote to increase the errors of other governors, when they can get the votes.

The horny governor only cares about you having an orgasm. But since it learns from happiness, not from the raw correction in its error, it has learned to vote for you to become afraid and embarrassed just before the moment of climax, because that increases your total error, which increases happiness. And since the horny governor has the votes, it overrules the governors who would vote against those things.

We don’t know how to quantify any of the factors involved, so we can’t test precise models. There are probably constants in these equations, but we can’t figure those out either, at least not yet.

But we can still make reasonable tests of general classes of models. We can make very decent guesses about whether or not something is a function of something else, and we can probably figure out if these relationships are sums or products, whether relationships are linear or exponential, and so on. For example:

happiness = delta_error

This is the original model we proposed, and it’s the most simple. In this case, happiness is caused when an organism corrects any error, and the amount of happiness produced is a direct function of how big of an error was corrected. Eating a cheeseburger makes you happy because, assuming you are hungry, it corrects that error signal. The cheeseburger error. 

Not shown in that equation is the kind of relationship. Maybe it’s linear, but maybe it’s exponential. Does eating two cheeseburgers cause more than twice as much happiness as eating one?

This very simple model has the virtue of being very simple. And it seems like it lines up with the basic facts — eating, sleeping, drinking, and fucking do tend to make us happy, especially if we are quite hungry, tired, thirsty, or horny. 

But we should also think about more complex models and see if any of them are any better. For example:

happiness = delta_error * product_errors

In this case, the correction in an error is multiplied not by the sum, but by the product of all other errors. So eating a cheeseburger while tired and lonely will be much more pleasurable than eating a cheeseburger while merely tired or merely lonely. 

This seems pretty unlikely just from first glance. If happiness were dependent on the product of your other errors, that seems like it would be pretty noticeable, because the difference between correcting an error while largely satisfied and largely unsatisfied would be huge and thus obvious. But this is also something that you could test empirically and maybe there could be some kind of truth to it. 

Is this a better model? Not entirely clear, but it certainly makes predictions that can be compared to parts of life we’re familiar with, and it can be tested empirically. That’s a pretty good start.

Or another example:

happiness = delta_error / sum_errors

Instead of multiplying the correction to produce happiness, this time we tried dividing it. In this case, happiness is smaller when the total amount of error is bigger. So correcting the same error leads to less happiness if you’re more out of alignment. 

This one seems right out. The joy we get from a cup of hot chocolate is greater when we are lonely, not less. Living in extremis seems like it should only magnify the satisfaction of our experiences. It’s possible that this doesn’t stand up to closer inspection, but people certainly find the idea intuitive:

Finally, one more example. You remember this equation from the learning and memory section above: 

Another model of happiness is that happiness is proportional to the TD error in the equation above, or the equivalent in whatever system our brain really uses. The TD error is the difference between the current and projected outcome of the action and the expected outcome of the action. So in this model, we get happiness when something corrects an error by more than the governor expects. 

Having an especially great sandwich for the first time feels great. This is because you didn’t know how good it would be. But having the same sandwich for the 100th time isn’t as good, even if it corrects the same amount of error. This is because you anticipated it would be that good, so there’s no TD error. In fact, if the sandwich hits the spot less than usual, you’ll be disappointed, even if it’s still pretty good. 

In this model, you’d expect that doing the same enjoyable stuff over and over wouldn’t keep you happy for very long. You’d have to mix it up and try new things that correct your errors.

This model does seem to capture something important. But that said, in real life correcting a big enough error usually creates some happiness. So happiness doesn’t seem like it could be entirely based on how unexpected the correction is. Some amount of happiness seems to come from any correction. But it does seem like more unexpected corrections usually make us more happy. 

So this is an example of how we can test general models, even before we can make precise measurements. We can think about classes of models, bring them to their limits, ask how the implications of these models compare to other things we already know about life and happiness, things we experience every day.

Just thinking of these questions mechanically, thinking of them as models, prompts us to ask questions like — What is the minimum amount of happiness? Can happiness only go down to zero, or can there be negative happiness? Is there a maximum amount of happiness? Even if a maximum wasn’t designed intentionally, surely there is some kind of limit to the value the hardware can represent? Can you get happiness overflow errors? What is the quantum of happiness? What are the units? — questions that psychologists wouldn’t normally ask. 

---

[Next: HELP WANTED]

---