We start with a single painted particle moving upwards at a decreasing velocity, with a trajectory that is slowly rotating at a constant velocity. At each step, there is some small probability for the particle to “split” into two particles, one of which begins rotating in the opposite direction. And if a particle’s trajectory is ever about to start turning downwards, it instead reverses its angular velocity and begins turning upwards.

Those are all the rules. So what sort of shapes does this incredibly simple algorithm generate? Take a look!

It might be unbelievable that such a simple algorithm could generate such complex and organic shapes, but you can try it for yourself: the images above are produced by only 25 lines of code! And of course, it makes one very curious about whether there is a real connection between this algorithm and the way that plants actually grow.

If you’ve watched some of the popular movies out there about the 2008 financial crisis, chances are that you’ve been misled about one or two things. (I’m looking at you, Big Short.) For example:

Entertaining? Sure! Accurate? No, very much not so. This analogy is very much off the mark, as you’ll see in a minute.

Here’s a quote from Inside Job, often described as the most rigorous and well-researched of the popular movies on the crisis:

In the early 2000s, there was a huge increase in the riskiest loans, called subprime. But when thousands of subprime loans were combined to create CDOs, many of them still received AAA ratings.

The tone that this is stated in is one of disbelief at the idea that by combining subprime loans you can create extremely safe loans. And maybe this idea does sound pretty crazy if you haven’t studied much finance! But it’s actually correct. You can, by combining subprime loans, generate enormously safe investments, and thus the central conceit of a CDO is actually entirely feasible.

The overall attitude taken by many of these movies is that the financial industry in the early 2000s devoted itself to the scamming of investors for short-term profits through creation of complicated financial instruments like CDOs. As these movies describe, the premise of a CDO is that by combining a bunch of risky loans and slicing-and-dicing them a bit, you can produce a mixture of new investment opportunities including many that are extremely safe. This is all described in a tone that is supposed to convey a sense that this premise is self-evidently absurd.

I want to convince you that the premise of CDOs is not self-evidently absurd, and that in fact it is totally possible to pool risky mortgages to generate extremely safe investments.

So, why think that it should be possible to pool risky investments and decrease overall risk? Well first of all, that’s just what happens when you pool assets! Risk always decreases when you pool assets, with the only exception being the case where the assets are all perfectly correlated (which never happens in real life anyway).

As an example, imagine that we have two independent and identical bets, each giving a 90% chance of a $1000 return and a 10% chance of nothing.

Now put these two together, and split the pool into two new bets, each an average of the original two:

Take a look at what we’ve obtained. Now we have only a 1% chance of getting nothing (because both bets have to fail for this to happen). We do, however, have only a 81% chance of getting $1000, as opposed to the 90% we had earlier. But what about risk? Are we higher or lower risk than before?

The usual way of measuring risk is to look at standard deviations. So what are the standard deviations of the original bet and the new one?

Initial Bet
Mean = 90% ($1000) + 10% ($0) = $900
Variance = 90% (100^2) + 10% (900^2) = 90,000
Standard deviation = $300

New Bet
Mean = 81% ($1000) + 18% ($500) + 10% ($0) = $900
Variance = 81% (100^2) + 18% (400^2) + 1% (900^2) = 45,000
Standard deviation = $216.13

And look at what we see: risk has dropped, and fairly dramatically so, just by pooling independent bets! This concept is one of the core lessons of financial theory, and it goes by the name of diversification. The more bets we pool, the further the risk goes down, and in the limit of infinite independent bets, the risk goes to zero.

So if you’re watching a movie and somebody says something about combining risky assets to produce safe assets as if that idea is self-evidently absurd, you know that they have no understanding of basic financial concepts, and especially not a complex financial instrument like a CDO.

In fact, let’s move on to CDOs now. The setup I described above of simply pooling bets does decrease risk, but it’s not how CDOs work. At least, not entirely. CDOs still take advantage of diversification, but they also incorporate an additional clever trick to distribute risk.

The idea is that you take a pool of risky assets, and you create from it a new pool of non-identical assets with a spectrum of risk profiles. Previously all of the assets that we generated were identical to each other, but what we’ll do now with the CDO is that we’ll split up our assets into non-identical assets in such a way as to allocate the risk, so that some of the assets that we get will have very high risk (they’ll have more of the risk allocated to them), and some of them will have very little risk.

Alright, so that’s the idea: from a pool of equally risky assets, you can get a new pool of assets that have some variation in riskiness. Some of them are actually very safe, and some of them are very, very risky. How do we do this? Well let’s go back to our starting example where we had two identical bets, each with 90% chance of paying out $1000, and put them together in a pool. But this time, instead of creating two new identical bets, we are going to differentiate the two bets by placing an order priority of payout on them. In other words, one bet will be called the “senior tranche”, and will be paid first. And the other bet will be called the “junior tranche”, and will be paid only if there is still money left over after the senior tranche has been paid. What do the payouts for these two new bets look like?

The senior tranche gets paid as long as at least one of the two bets pays out, which happens with 99% probability. Remember, we started with only a 90% probability of paying out. This is a dramatic change! In terms of standard deviation, this is $99.49, less than a third of what we started with!

And what about the junior tranche? Its probability of getting paid is just the probability that both people don’t default, which is 81%. And its risk has gone up, with a standard deviation of $392.30. So essentially, all we’ve done is split up our risk. We originally had 90%/90%, and now we have 99%/81%. In the process, what we’ve done is we’ve created a very very safe bet and a very very risky bet.

The important thing is that these two bets have to both be sold. You can’t just sell the senior tranche to people who want safe things (pensions funds), you have to also sell the junior tranche. So how do you do that? Well, you just lower its price! A higher rate of return awaits the taker of the junior tranche in exchange for taking on more risk.

Now if you think about it, this new lower level risk we’ve obtained, this 1% chance of defaulting that we got out of two bets that had a 10% chance of defaulting each, that’s a real thing! There really is a 1% chance that both bets default if they are independent, and so the senior tranche really can expect to get paid 99% of the time! There isn’t a lie or a con here, a pension funds that gets sold these senior tranches of CDOs is actually getting a safe bet! It’s just a clever way of divvying up risk among two assets.

I think the idea of a CDO is cool enough by itself, but I think that the especially cool thing about CDOs is that they open up the market to new customers. Previously, if you wanted to get a mortgage, then you had to find basically a bank that was willing to accept your level of risk, whatever it happens to be. And it could be that if you’re too high risk, then nobody wants to give you a mortgage, and you’d just be out of luck. Even prior to CDOs, when you mortgage pooling but no payment priority, you have to have investors that are interested in the level of risk of your pool. The novelty of CDOS is in allowing you to alter the risk profile of your pool of mortgages at will.

:Let’s say that you have 100 risky loans, and there’s only enough demand for you to sell 50 of them. What you can do is create a CDO with 50 safe loans and 50 risky loans. Now you get to not only sell your risky loans, but you can also sell your safe loans to interested customers like pension funds! This is the primary benefit of the new financial technology of CDOs: it allows banks to generate tailor-made risk levels for the set of investors that are interested in buying, so that they can sell more mortgage-backed securities and get more people homes. And if everything is done exactly as I described it, then everything should work out fine.

But of course, things weren’t done exactly as I described them. The risk levels of individual mortgages were given increasingly optimistic ratings with stated-income loans, no-down-payment loans, and no-income no-asset loans. CDOs were complex and their risk level was often difficult to assess, resulting in less transparency and more ability for banks to over-report their safety. And crucially, the different tranches of any given CDO are highly dependent on each other, even after they’ve been sold to investors that have nothing to do with each other.

Let’s go back to our simple example of the two $1000 bets for illustration. Suppose that one of the two bets doesn’t pay out (which could correspond to one home-owner defaulting on their monthly payment). Now the senior tranche owner’s payment is entirely dependent on how the other bet performs. The senior tranche owner will get $1000 only if that remaining bet pays out, which happens with just 90% probability. So his chance of getting $1000 has dropped from 99% to 90%.

That’s a simple example of a more general point: that in a CDO, once the riskier tranches fail, the originally safe tranches suddenly become a lot riskier (so that what was originally AA is now maybe BBB). This helps to explain why once the housing bubble had popped, all levels of CDOs began losing value, not just the junior levels. Ordinary mortgage backed securities don’t behave this way! A AA-rated mortgage is rated that way because of some actual underlying fact about the reliability of the homeowner, which doesn’t necessarily change when less reliable homeowners start defaulting. A AA-rated CDO tranche might be rated that way entirely because it has payment priority, even though all the mortgages in its pool are risky.

Another way to say this: An ordinary mortgage backed security decreased risk just because of diversification (many mortgages pooled together make for a less risky bet than a single mortgage). But a CDO gets decreased risk because of both diversification and (in the upper tranches) the order priority (getting paid first). In both cases, as some of the mortgages in the pool fail, you lose some of the diversification benefit. But in the CDO case, you also lose the order priority benefit in the upper tranches (because, for example, if it takes 75 defaults in your pool for you to lose your money and 50 have already failed, then you are at a much higher risk of losing your money than if none of them have failed). Thus there is more loss of value in safe CDOs than in safe MBSs as default rates rise.

Consequentialism is a family of moral theories that say that an act is moral or immoral based on its consequences. If an act has overall good consequences then it is moral, and if it has bad consequences then it is immoral. What precisely counts as a “good” or “bad” consequence is what distinguishes one consequentialist theory from another. For instance, act utilitarians say that the only morally relevant feature of the consequences of our actions is the aggregate happiness and suffering produced, while preference utilitarians say that the relevant feature of the consequences is the number and strength of desires satisfied. Another form of consequentialism might strike a balance between aggregate happiness and social equality.

What all these different consequentialist theories have in common is that the ultimate criteria being used to evaluate the moral status of an action is only a function of the consequences of that action, as opposed to, say, the intentions behind the action, or whether the action is an instance of a universalizable Kantian rule.

In this essay, we’ll explore some puzzles in consequentialist theories that force us to take a more nuanced and subtle view of consequentialism. These puzzles are all adapted from Derek Parfit’s Reasons and Persons, with very minor changes.

First, we’ll consider a simple puzzle regarding how exactly to evaluate the consequences of one’s actions, when one is part of a collective that jointly accomplishes some good.

Case 1: There are 100 miners stuck in a mineshaft with flood waters rising. These men can be brought to the surface in a lift raised by weights on long levers. The leverage is such that just four people can stand on a platform and provide sufficient weight to raise the lift and save the lives of the hundred men. But if any fewer than four people stand on the platform, it will not be enough to raise the lift. As it happens, you and three other people happen to be standing there. The four of you stand on the platform, raising the lift and saving the lives of the hundred men.

The question for us to consider is, how many lives did you save by standing on the platform? The answer to this question matters, because to be a good consequentialist, each individual needs to be able to compare their contribution here to the contribution they might make by going elsewhere. As a first thought, we might say that you saved 100 lives by standing on the platform. But the other three people were in the same position as you, and it seems a little strange to say that all four of you saved 100 lives each (since there weren’t 400 lives saved total). So perhaps we want to say that each of you saved one quarter of the total: 25 lives each.

Parfit calls this the Share-of-the-Total View. We can characterize this view as saying that in general, if you are part of a collective of N people who jointly save M lives, then your share of lives saved is M/N.

There are some big problems with this view. To see this, let’s amend Case 1 slightly by adding an opportunity cost.

Case 2: Just as before, there are 100 miners stuck in a mineshaft with flood waters rising, and they can be saved by four or more people standing on a platform. This time though, you and four other people happen to be standing there. The other four are going to stand on the platform no matter what you do. Your choice is either to stand on the platform, or to go elsewhere to save 10 lives. What should you do?

The correct answer here is obviously that you should leave to save the 10 lives. The 100 miners will be saved whether you stay or leave, and the 10 lives will be lost if you stick around. But let’s consider what the Share-of-the-Total View says. According to this view, if you stand on the platform, your share of the lives saved is 100/5 = 20. And if you leave to go elsewhere, you only save 10 lives. So you save more lives by staying and standing on the platform!

This is a reductio of the Share-of-the-Total View. We must revise this view to get a sensible consequentialist theory. Parfit’s suggestion is that we say that when you join others who are doing good, the good that you do is not just your own share of the total benefit. You should also add to your share the change that you caused in the shares of the benefits produced by each other by joining. On their own, the four would each have a share of 25 lives. So by joining, you have a share of 20 lives, minus the 5 lives that have been reduced from the share of each of the other four. In other words, by joining, you have saved 20 – 5(4) lives, in other words, 0 lives. And of course, this is the right answer, because you have done nothing at all by stepping onto the platform!

Applying our revised view to Case 1, we see that if you hadn’t stepped onto the platform, zero lives would be saved. By stepping onto the platform, 100 lives are saved. So your share of those lives is 25, plus 25 lives for each of the others that would have had zero without you. So your share is actually 100 lives! The same applies to the others, so in our revised view, each of the four is responsible for saving all 100 lives. Perhaps on reflection this is not so unintuitive; after all, it’s true for each of them that if they change their behavior, 100 lives are lost.

Case 3: Just as in Case 2, there are 100 miners stuck in a mineshaft. You and four others are standing on the platform while the miners are slowly being raised up. Each of you know of an opportunity to save 10 lives elsewhere (a different 10 lives for each of you), but to successfully save the lives you have to leave immediately, before the miners are rescued. The five of you have to make your decision right away, without communicating with each other.

We might think that if each of the five of you reasons as before, each of you will go off and save the other 10 lives (as by staying, they see that they are saving zero lives). In the end, 50 lives will be saved and 100 lost. This is not good! But in fact, it’s not totally clear that this is the fault of our revised view. The problem here is lack of information. If each of the five knew what the other four planned on doing, then they would make the best decision (if all four planned to stay then the fifth would leave, and if one of the other four planned to leave then the fifth would stay). As things stand, perhaps the best outcome would be that all five stay on the platform (losing the opportunity to save 10 extra lives, but ensuring the safety of the 100). If they can use a randomized strategy, then the optimal strategy is to each stay on the platform with probability 97.2848% (saving an expected 100.66 lives)

Let’s move on to another type of scenario.

Case 4: X and Y simultaneously shoot and kill me. Either shot, by itself, would have killed.

The consequence of X’s action is not that I die, because if X had not shot, I would have died by Y’s bullet. And the same goes for Y. So if we’re evaluating the morality of X or Y’s action based on its consequences, it seems that we have to say that neither one did anything immoral. But of course, the two of them collectively did do something immoral by killing me. What this tells us that the consequentialist’s creed cannot be “an act is immoral if its consequences are bad”, as an act can also be immoral if it is part of a set of acts whose collective consequences are bad.

Inheriting immorality from group membership has some problems, though. X and Y collectively did something immoral. But what about the group X, Y, and Barack Obama, who was napping at home when this happened? The collective consequences of their actions were bad as well. So did Obama do something immoral too? No. We need to restrict our claim to the following:

“When some group together harm or benefit other people, this group is the smallest group of whom it is true that, if they had all acted differently, the other people would not have been harmed, or benefited.” -Parfit

A final scenario involves the morality of actions that produce imperceptible consequences.

Case 5: One million torturers stand in front of one million buttons. Each button, if pushed, induces a tiny stretch in each of a million racks, each of which has a victim on it. The stretch induced by a single press of the button is so minuscule that it is imperceptible. But the stretch induced by a million button presses produces terrible pain in all the victims.

Clearly we want to say that each torturer is acting immorally. But the problem is that the consequences of each individual torturer’s action are imperceptible! It’s only when enough of the torturers press the button that the consequence becomes perceptible. So what we seem to be saying is that it’s possible to act immorally, even though your action produces no perceptible change in anybody’s conscious experience, if your action is part of a collection of actions that together produce negative changes in conscious experiences.

This is already unintuitive. But we can make it even worse.

Case 6: Consider the final torturer of the million. At the time that he pushes his button, the victims are all in terrible agony, and his press doesn’t make their pain any perceptibly worse. Now, imagine that instead of there being 999,999 other torturers, there are zero. There is just the one torturer, and the victims have all awoken this morning in immense pain, caused by nobody in particular. The torturer presses the button, causing no perceptible change in the victims’ conditions. Has the torturer done something wrong?

It seems like we have to say the same thing about the torturer in Case 6 as we did in Case 5. The only change is that Nature has done the rest of the harm instead of other human beings, but this can’t matter for the morality of the torturer’s action. But if we believe this, then the scope of our moral concerns is greatly expanded, to a point that seems nonsensical. My temptation here is to say “all the worse for consequentialism, then!” and move to a theory that inherently values intentions, but I am curious if there is a way to make a consequentialist theory workable in light of these problems.

I devoured science fiction as a kid. One of my favorite books was Asimov’s Foundation, which told the story of a splintering Galactic Empire and the seeds of a new civilization that would grow from its ashes. I marveled and was filled with joy at the notion that in our far future humanity might spread across the galaxy, settling countless new planets and bringing civilization to every corner of it.

Then I went to college and learned physics, especially special relativity and cosmology, and learned that there are actually some hugely significant barriers in the way of these visions ever being manifest in reality. Ever since, it has seemed obvious to me that the idea of a galactic civilization, or of humans “colonizing the universe” are purely fantasy. But I am continuously surprised when I hear those interested in futurism throw around these terms casually, as if their occurrence were an inevitability so long as humans stay around long enough. This is deeply puzzling to me. Perhaps these people are just being loose with their language, and when they refer to a galactic civilization, they mean something much more modest, like a loosely-connected civilization spread across a few nearby stars. Or maybe there is a fundamental misunderstanding of both the limitations imposed on us by physics and the the enormity of the scales in question. Either way, I want to write a post explaining exactly why these science fiction ideas are almost certainly never going to become reality.

My argument in brief:

We are really slow.

The galaxy is really big.

The universe is even bigger.

Back in 1977, humans launched the Voyager 1 spacecraft, designed to explore the outer solar system and then to head for the stars. Its trajectory was coordinated to slingshot around Jupiter and then Saturn, picking up speed at each stage, and then finally to launch itself out of the solar system into the great beyond. 36 years later, in 2012, it has finally left the Sun’s heliopause, marking the first steps into interstellar space. It is now 11.7 billion miles from Earth, which sounds great until you realize that this distance is still less than two-tenths of a single percent of one light year.

Compare this to, say, the distance to the nearest star Alpha Centauri, we find that it has traveled less than .05% of the distance. At this rate, it would take another 80,000 years to make contact with Alpha Centauri (if it were aimed in that direction)! On the scale of distances between stars, our furthest current exploration has gotten virtually nowhere. It’s the equivalent of somebody who started in the center of the Earth hoping to burrow up all the way to the surface, and takes 5 days to travel a single meter. Over 42 years, they would have travelled just under 3000 meters.

OK, you say, but that’s unfair. The Voyager was designed in the 70s! Surely we could do better with modern space shuttle design. To which my reply is, sure we can do better now, but not better enough. The fastest thing humans have ever designed is the Helios 2 shuttle, which hit a speed of 157,078 mph at its fastest (.023% of the speed of light). If we packed some people in this shuttle (which we couldn’t) and sent it off towards the nearest star (assuming that it stayed at this max speed the entire journey), guess how long it would take? Over 18 thousand years.

And keep in mind, this is only talking about the nearest star, let alone spreading across the galaxy! It should be totally evident to everybody that the technology we would need to be able to reach the stars is still far far away. But let’s put aside the limits of our current technology. We are still in our infancy as a space-faring species in many ways, and it is totally unfair to treat our current technology level as if it’s something set in stone. Let’s give the futurists the full benefit of the doubt, and imagine that in the future humans will be able to harness incredible quantities of energy that vastly surpass anything we have today. If we keep increasing our energy capacity more and more, is there any limitation on how fast we could get a spacecraft going? Well, yes there is! There is a fundamental cosmic speed limit built into the of the universe, which is of course the speed of light. The speed-vs-energy curve has a vertical asymptote at this value; no finite amount of energy can get you past this speed.

So much for arbitrarily high speeds. What can we do with spacecraft traveling near the speed of light? It turns out, a whole lot more! Suppose we can travel at 0.9 times the speed of light, and grant also that the time to accelerate up to this speed is insignificant. Now it only takes us 4.7 years to get to the nearest star! The next closest star takes us 6.6 years. Next is 12.3 years. This is not too bad! Humans in the past have made years-long journeys to discover new lands. Surely we could do it again to settle the stars.

But now let’s consider the trip to the center of the galaxy. At 90% the speed of light, this journey would take 29,000 years. As you can see, there’s a massive difference between jumping to nearby stars and attempting to actually traverse the galaxy. The trouble is that most people don’t realize just how significant this change in distance scale is. When you hear “distance to Alpha Centauri” and “distance to the center of the Milky Way”, you are probably not intuitively grasping how hugely different these two quantities are. Even if we had a shuttle traveling at 99.99% times the speed of light, it would take over 100,000 years to travel the diameter of the Milky Way.

You might be thinking “Ok, that is quite a long time. But so what! Surely there will be some intrepid explorers that are willing to leave behind the safety of settled planets to bring civilization to brand new worlds. After all, taking into account time dilation, a 1000 year journey from Earth’s perspective only takes 14 years to pass from the perspective of a passenger on a ship traveling at 99.99% percent of the speed of light.”

And I accept that! It doesn’t seem crazy that if humans develop shuttle that travel a significant percentage of the speed of light, we could end up with human cities scattered all across the galaxy. But now the issue is with the idea of a galactic CIVILIZATION. Any such civilization faces a massive problem, which is the issue of communication. Even having a shared society between Earth and a planet around our nearest star would be enormously tricky, being that any information sent from one planet to the other would take more than four years to arrive. The two societies would be perpetually four years behind each other, and this raises some serious issues for any central state attempting to govern both. And it’s safe to say that no trade can exist between two planets that have a 10,000 year delay in communication. Nor can any diplomacy, or common leadership, or shared culture or technology, or ANY of the necessary prerequisites for a civilization.

I think that for these reasons, it’s evident that the idea of a Galactic Empire like Asimov’s could never come into being. The idea of such a widespread civilization must rely on an assumption that humans will at some point learn to send faster than light messages. Which, sure, we can’t rule out! But everything in physics teaches us that we should be betting very heavily against it. Our attitude towards the idea of an eventual real life Galactic Empire should be similar to our attitude towards perpetual motion machines, as both rely on a total rewriting of the laws of physics as we know them.

But people don’t stop at a Galactic Empire. Not naming names, but I hear people that I respect throwing around the idea that humans can settle the universe. This is total madness. The jump in distance scale from diameters of galaxies to the size of the observable universe, is 40 times larger than the jump in distance scale we previously saw from nearby stars to galactic diameters. The universe is really really big, and as it expands, every second more of it is vanishing from our horizon of observable events.

Ok, so let’s take the claim to be something much more modest than the literal ‘settle galaxies all across the observable universe’. Let’s take it that they just mean that humans will spread across our nearest neighborhood of galaxies. The problem with this is again that the distance scale in question is just unimaginably large. Clearly we can’t have a civilization across galaxies (as we can’t even have a civilization across a single galaxy). But now even making the trip is a seemingly insurmountable hurdle. The Andromeda Galaxy (the nearest spiral galaxy to the Milky Way) is 2 million light years away. Even traveling at 99.99% of the speed of light, this would be a 28,000 year journey for those within the ship. From one edge of the Virgo supercluster (our cluster of neighbor galaxies) to the other takes 110 million years at the speed of light. To make this trip doable in a single lifetime requires a speed of 99.999999999999% of c (which would result in the trip taking 16 years inside the spacecraft). The kinetic energy required to get a mass m to these speeds is given by KE = (γ – 1) mc^{2}. If our shuttle and all the people on it have a mass of, say, 1000 kg, the required energy ends up being approximately the entire energy output of the Sun per second.

Again, it’s possible that if humans survive long enough and somehow get our hands on the truly enormous amounts of energy required to get this close to the speed of light, then we could eventually have human societies springing up in nearby galaxies, in total isolation from one another. But (1) it’s far from obvious that we will ever be capable of mastering such enormous amounts of energy, and (2) I think that this is not what futurists are visualizing when they talk about settling the universe.

Let me just say that I still love science fiction, love futurism, and stand in awe when I think of all the incredible things the march of technological progress will likely bring us in the future. But this seems to be one area where the way that our future prospects are discussed is really far off the mark, and where many people would do well to significantly adjust their estimates of the plausibility of the future trajectories they are imagining. Though science and future technology may bring us many incredible new abilities and accomplishments as a species, a galactic or intergalactic civilization is almost certainly not one of them.

Merry Christmas! In the spirit of the season, let’s talk about altruism.

Anybody familiar with the effective altruism movement knows about the concept of earning to give. The idea is that for some people, the ideal altruistic career path might involve them making lots of money in a non-charitable role and then donating a significant fraction of that to effective charities. Estimates from GiveWell for the current lowest cost to save a life put it around $2,300, which indicates that choosing a soulless corporate job that allows you to donate $150,000 a year could actually be better than choosing an altruistic career in which you are on average saving 65 lives per year, or more than a life per week!

What I’m curious about lately is the concept of earning to save up to give. At the time of writing, the rate of return on US treasury bills is 1.53%. US treasury bills are considered to be practically riskless – you get a return on your investment no matter what happens to the economy. Assuming that this rate of return and the $2,300 figure both stay constant, what this means is that by holding off on donating your $150,000 for five years, you expect to be able to donate about $11,830 more, which corresponds to saving 5 extra lives. And if you hold off for twenty years, you will be able to save 23 more lives!

If you choose to take on bigger risk, you can do even better than this. Instead of investing in treasury bills, you could put your money into a diversified portfolio of stocks and expect to get a rate of return of around 7% (the average annualized total return of the S&P 500 for the past 90 years, adjusted for inflation). Now you run the risk of the stock market being in a slump at the end of five years, but even if that happens you can just wait it out longer until the market rises again. In general, as your time horizon for when you’re willing to withdraw your investment expands, your risk drops. If you invest your $150,000 in stocks and withdraw after five years, you expect to save an extra 26 lives than you would by donating right away!

In general, you can choose your desired level of risk and get the best available rate of return by investing in an appropriate linear combination of treasury bills and stocks. And if you want a higher rate of return than 7%, you can take on more risk by leveraging the market (short selling the risk-free asset and using the money to invest more in stocks). In this model, the plan for maximizing charitable giving would be to continuously invest your income at your chosen level of risk, donating nothing until some later date at which time you make an enormous contribution to your choice of top effective charities. In an extreme case, you could hold off on donations your entire life and then finally make the donation in your will!

Alright, so let’s discuss some factors in favor and against this plan.

Factors in favor
Compounding interest
Decreasing moral and factual uncertainty

Factors against
Shrinking frontier of low hanging fruit
Personal moral regression
Bad PR
Future discounting
Infinite procrastination

Compounding interest and vanishing low hanging fruits

First, and most obviously, waiting to give means more money, and more money means more lives. And since your investment grows exponentially, being sufficiently patient can mean large increases in amount donated and lives saved. There’s a wrinkle in this argument though. While your money grows with time, it might be that the effective cost of improving the world grows even quicker. This could result from the steady improvement of the world: problems are getting solved and low hanging altruistic fruits are being taken one at a time, leaving us with a set of problems that take more money to solve, making them less effective in terms of impact per dollar donated. If the benefit of waiting is a 5% annual return on your investment, but the cost is a 6% decrease in the effective number of lives saved per dollar donated, then it is best to donate as soon as possible, so as to ensure that your dollars have the greatest impact.

Estimates of the trends in effectiveness of top charities are hard to come by, but crucially important for deciding when to give and when to wait.

Moral and factual uncertainty

Say that today you donate 50,000 dollars to Charity X, and tomorrow an exposé comes out revealing that this charity is actually significantly less effective than previously estimated. That money is gone now, and you can’t redirect it to a better charity in response to this new information. But if you had waited to donate, you would be able to respond and more accurately target your money to the most effective charities. The longer you wait to donate, the more time there is to gather the relevant data, analyze it, and come to accurate conclusions about the effectiveness of charities. This is a huge benefit of waiting to give.

That was an example of factual uncertainty. You might also be concerned about moral uncertainty, and think that in the future you will have better values than today. For instance, you may think that your future values, after having had more time to reflect and consider new arguments, will be more rational and coherent than your current values. This jumps straight into tricky meta-ethical territory; if you are a moral realist then it makes sense to talk about improving your values, but as a non-realist this is harder to make sense of. Regardless, there certainly is a very common intuition that individuals can make moral progress, and that we can “improve our values” by thinking deeply about ethics.

William MacAskill has talked about a related concept, applied on a broader civilizational level. Here’s a quote from his 80,000 Hours interview:

Different people have different sets of values. They might have very different views for what an optimal future looks like. What we really want ideally is a convergent goal between different sorts of values so that we can all say, “Look, this is the thing that we’re all getting behind that we’re trying to ensure that humanity…” Kind of like this is the purpose of civilization. The issue, if you think about purpose of civilization, is just so much disagreement. Maybe there’s something we can aim for that all sorts of different value systems will agree is good. Then, that means we can really get coordination in aiming for that.

I think there is an answer. I call it the long reflection, which is you get to a state where existential risks or extinction risks have been reduced to basically zero. It’s also a position of far greater technological power than we have now, such that we have basically vast intelligence compared to what we have now, amazing empirical understanding of the world, and secondly tens of thousands of years to not really do anything with respect to moving to the stars or really trying to actually build civilization in one particular way, but instead just to engage in this research project of what actually is a value. What actually is the meaning of life? And have, maybe it’s 10 billion people, debating and working on these issues for 10,000 years because the importance is just so great. Humanity, or post-humanity, may be around for billions of years. In which case spending a mere 10,000 is actually absolutely nothing.

Personal moral regression

On the other side of the issue of changing values over time, we have the problem of personal moral regression. If I’m planning to save up for an eventual donation decades down the line, I might need to seriously worry about the possibility that when the time comes to donate I have become a more selfish person or have lost interest in effective altruism. Plausibly, as you age you might become more attached to your money or expect a higher standard of living than when you were younger. This is another of these factors that is hard to estimate, and depends a lot on the individual.

Bad PR

Earning to give is already a concept that draws criticism in some quarters, and I think that waiting to give may look worse in some ways. I could easily see the mainstream revile the idea of a community of self-proclaimed altruists that mostly sit around and build up wealth with the promise to donate it some time into the future. Tied in with this is the concern that by removing the signaling value of your form of altruism, some of the motivation to actually be altruistic in the first place is lost.

Future discounting

Economists commonly talk about temporal discounting, the idea of weighting current value higher than future value. Give somebody a choice between an ice cream today and two ice creams in a month, and they will likely choose the ice cream today. This indicates that there is some discount rate on future value to make somebody indifferent to a nominally identical current value. This discount rate is often thought of purely descriptively, as a way to model a particular aspect of human psychology, but it is also sometimes factored into recommendations for policy.

For instance, some economists talk about a social discount rate, which represents the idea of valuing future generations less than the current generation. This discount rate actually also factors importantly into the calculation of the appropriate value of a carbon tax. Most major calculations of the carbon tax explicitly use a non-zero social discount rate, meaning that they assume that future generations matter less than current generations. For instance, William Nordhaus’s hugely influential work on carbon pricing used a “3 percent social discount rate that slowly declines to 1 percent in 300 years.”

I don’t think that this makes much moral sense. If you apply a constant discount rate to the future, you can end up saying things like “I’d rather get a nickel today than save the entire planet a thousand years from now.” It seems to me that this form of discounting is simply prejudice against those people are most remote from us: those that are in our future and as such do not exist yet. This paper by Tyler Cowen and Derek Parfit argues against a social discount rate. From the paper:

Remoteness in time roughly correlates with a whole range of morally important facts. So does remoteness in space. Those to whom we have the greatest obligation, our own family, often live with us in the same building. We often live close to those to whom we have other special obligations, such as our clients, pupils, or patients. Most of our fellow citizens live closer to us than most aliens. But no one suggests that, because there are such correlations, we should adopt a spatial discount rate. No one thinks that we would be morally justified if we cared less about the long-range effects of our acts, at some rate of n percent per yard. The temporal discount rate is, we believe, as little justified.

Infinite procrastination

There’s one final consideration I want to bring up, which is more abstract than the previous ones. Earlier I imagined somebody who decides to save up their entire life and make their donation in their will. But we can naturally ask: why not set up your will to wait another twenty years and then donate to the top-rated charities from your estimate of the most reliable charity evaluator? And then why stop at just twenty years? Why not keep on investing, letting your money build up more and more, all to the end of making some huge future donation and makes an absolutely enormous benefit to some future generation? The concern is that this line of reasoning might never terminate.

While this is a fun thought experiment, I think there are a few fairly easy holes that can be poked in it. In reality, you will eventually run up against decreasing marginal value of your money. At some point, the extra money you get by waiting another year actually doesn’t go far enough to make up for the human suffering you could have prevented. Additionally, the issue of vanishing low hanging fruits will become more and more pressing, pushing you towards

Summing up

We can neatly sum up the previous considerations with the following formula.

V = Q − p(1 − d)(Q + I − F)(R − p)
If V > 0, then giving now is preferable to giving in a year. If V < 0, then you should wait for at least a year.

Q = Current lives saved per dollar
R = Rate of return on investment
I = Reflection factor: yearly increase in lives saved per dollar from better information and longer reflection
p = Regression factor: expected percentage less you are willing to give in a year
F = Low hanging fruit factor: yearly decrease in lives saved per dollar
d = Temporal discount factor: percentage less that lives are valued each year

The ideal setting for waiting to give is where F is near zero (the world’s issues are only very slowly being sorted out), I is large (time for reflection is sorely needed to bring empirical data and moral clarity), p is near zero (no moral regression), R is large (you get a high return on your investment), and d is near zero (you have little to no moral preference for helping current people over future people).

“The only free lunch in finance is diversification”

Harry Markowitz

Suppose you have two assets that you can invest in, and $1000 dollars total to split between them. By the end of year 1, Asset A doubles in price and Asset B halves in price. And by the end of year 2, A halves and B doubles (so that they each return to their starting point).

If you had initially invested all $1000 in Asset A, then after year 1 you would have $2000 total. But after year 2, that $2000 is cut in half and you end up back at $1000. If you had invested the $1000 in Asset B, then you would go down to $500 after year 1 and then back up to $1000 after year 2. Either way, you don’t get any profit.

Now, the question is: can you find some way to distribute the $1000 dollars across Assets A and B such that by the end of year 2 you have made a profit? For fairness sake, you cannot change your distribution at the end of year 1 (as that would allow you to take advantage of the advance knowledge of how the prices will change). Whatever weights you choose initially for Assets A and B, at the end of year 1 you must move around your money so that you ensure it’s still distributed with the exact same weights as before.

So what do you think? Does it seem impossible to make a profit without changing your distribution of money between years 1 and 2? Amazingly, the answer is that it’s not impossible; you can make a profit!

Consider a 50/50 mix of Assets A and B. So $500 initially goes into Asset A and $500 to Asset B. At the end of year 1, A has doubled in price (netting you $500) and B has halved (losing you $250). So at the end of year 1 you have gained 25%.

To keep the weights the same at the start of year 2 as they were at the start of year 1, you redistribute your new total of $1250 across A and B according to the same 50/50 mix ($625 in each). What happens now? Now Asset A halves (losing you $312.50) and Asset B doubles (gaining you $625). And by the end of year 2, you end up with $312.50 in A and $1250 in B, for a total of $1562.50! You’ve gained $562.50 by investing in two assets whose prices ultimately are the same as where they started!

This is the magic of diversification. By investing in multiple independent assets instead of just one, you can up your rate of return and decrease your risk, sometimes dramatically.

Another example: In front of you are two fair coins, A and B. You have in your hand $100 that you can distribute between the two coins any way you like, so long as all $100 is on the table. Now, each of the coins will be tossed. If a coin lands heads, the amount of money beside it will be doubled and returned to you. But if the coin lands tails, then all the money beside it will be destroyed.

In front of you are two coins, A and B. You have $100 dollars to distribute between these two coins. You get to choose how to distribute the $100, but at the end every dollar bill must be beside one coin or the other.

Coin A has a 60% chance of landing H. If it lands H, the amount of money placed beside it will be multiplied by 2.1 and returned to you. But if it lands T, then the money placed beside it will be lost.

Coin B has only a 40% chance of landing H. But the H outcome also has a higher reward! If the coin lands H, then the amount of money beside it will be multiplied by 2.5 and returned to you. And just like Coin A, if it lands T then the money beside it will be destroyed.

Coin A: .6 chance of getting 2.1x return
Coin B: .4 chance of getting 2.5x return

The coins are totally independent. How should you distribute your money in order to maximize your return, given a specific level of risk?

(Think about it for a moment before reading on.)

If you looked at the numbers for a few moments, you might have noticed that Coin A has a higher expected return than Coin B (126% vs 100%) and is also the safer of the two. So perhaps your initial guess was that putting everything into Coin A would minimize risk and maximize return. Well, that’s incorrect! Let’s do the math.

We’ll start by giving the relevant quantities some names.

X = amount of money that is put by Coin A
Y = amount of money that is put by Coin B
(All your money is put down, so Y = 100 – X)

We can easily compute the expected amount of money you end up with, as a function of X:

Alright, so clearly your expected return is maximized by making X as large as possible (by putting all of your money by Coin A). This makes sense, since Coin A’s expected return is higher than Coin B’s. But we’re not just interested in return, we’re also interested in risk. Could we possibly find a better combination of risk and reward by mixing our investments? It might initially seem like the answer is no; after all Coin A is the safer of the two. How could we possibly decrease risk by mixing in a riskier asset to our investments?

The key insight is that even though Coin A is the safer of the two, the risk of Coin B is uncorrelated with the risk of Coin A. If you invest everything in Coin A, then you have a 40% chance of losing it all. But if you split your investments between Coin A and Coin B, then you only lose everything if both coins come up heads (which happens with probability .6*.4 = 24%, much lower than 40%!)

Let’s go through the numbers. We’ll measure risk by the standard deviation of the possible outcomes.

This function is just a parabola (to be precise, Risk^{2} is a parabola, which means that Risk(x) is a hyperbola). Here’s a plot of Risk(X) vs X (amount placed beside A):

Looking at this plot, you can see that risk is actually minimized at a roughly even mix of A and B, with slightly more in B. You can also see this minimum risk on a plot of return vs risk:

Notice that somebody that puts most of their money on coin B (these mixes are in the bottom half of the curve) is making a strategic choice is strictly dominated. That is, they could choose a different mix that has a higher rate of return for the same risk!

Little did you know, but you’ve actually just gotten an introduction to modern portfolio theory! Instead of putting money beside coins, portfolio managers consider investing in assets with various risks and rates of return. The curve of reward vs risk is famous in finance as the Markowitz Bullet. The upper half of the curve is the set of portfolios that are not strictly dominated. This section of the curve is known as the efficient frontier, the basic idea being that no rational investor would put themselves on the lower half.

Let’s reframe the problem in terms that would be familiar to somebody in finance.

We have two assets, A and B. We’ll model our knowledge of the rate of return of each asset as a normal distribution with some known mean and standard deviation. The mean of the distribution represents the expected rate of return on a purchase of the asset, and the standard deviation represents the risk of purchasing the asset. Asset A has an expected rate of return of 1.2, which means that for every dollar you put in you expect (on average) to get back $0.20 a year from now. Asset B’s expected rate of return is 1.3, so it has a higher average payout. But Asset B is riskier; the standard deviation for A is 0.5, while B’s standard deviation is 0.8. There’s also a risk-free asset that you can invest in, which we’ll call Asset F. This asset has an expected rate of return of 1.1.

Asset F: R = 1.1, σ = 0
Asset A: R = 1.2, σ = 0.5
Asset B: R = 1.3, σ = 0.8

Suppose that you have $1000 that you want to invest in some combination of these three assets in such a way as to maximize your expected rate of return and minimize your risk. Since rate of return and risk will in general be positively correlated, you have to decide the highest risk that you’re comfortable with. Let’s say that you decide that the highest risk you’ll accept is 0.6. Now, how much of each asset should you purchase?

First of all, let’s disregard the risk-free asset and just consider combinations of A and B. A portfolio of A and B is represented by the weighted sum of A and B. w_{A} is the percentage of your investment in the portfolio that goes to just A, and w_{B} is the percentage that goes towards B. Since we’re just considering a combination of A and B for now, w_{A} + w_{B} = 1. The mean and standard deviation of the new distribution for this portfolio will in general depend on the correlation between A and B, which we’ll call ρ. Perfectly correlated assets have ρ = 1, uncorrelated assets have ρ = 0, and perfectly anti-correlated assets have ρ = -1. Correlation between assets is bad for investors, because it destroys the benefit of diversification. If two assets are perfectly correlated, then you don’t get any lower risk by combining them. On the other hand, if they are perfectly anti-correlated, you can entirely cancel the risk from one with the risk from the other, and get fantastically low risks for great rates of return.

Let’s suppose that the correlation between A and B is ρ = .2. Since both R_{P} and σ_{P} are functions of w_{A} and w_{B}, we can visualize the set of all possible portfolios as a curve on a plot of return-vs-risk:

The Markowitz Bullet again! Each point on the curve represents the rate of return and risk of a particular portfolio obtained by mixing A and B. Just like before, some portfolios dominate others and thus should never be used, regardless of your desired level of risk. In particular, for any portfolio that weights asset A (the less risky one) too highly, there are other portfolios that give a higher rate of return with the exact same risk.

In other words, you should pretty much never purchase only a low-risk low-return item. If your portfolio consists entirely of Asset A, then by mixing in a little bit of the higher-risk item, you can actually end up massively decreasing your risk and upping your rate of return. Of course, this drop in risk is only because the two assets are not perfectly correlated. And it would be even more extreme if we had negatively correlated assets; indeed with perfect negative correlation (as we saw in the puzzle I started this post with), your risk can drop to zero!

Now, we can get our desired portfolio with a risk of 0.6 by just looking at the point on this curve that has σ_{P} = 0.6 and calculating which values of w_{A} and w_{B} give us this value. But notice that we haven’t yet used our riskless asset! Can we do better by adding in a little of Asset F to the mix? It turns out that yes, we can. In fact, every portfolio is weakly dominated by some mix of that portfolio and a riskless asset!

We can easily calculate what we get by combining a riskless asset with some other asset X (which can in general be a portfolio consisting of multiple assets):

So σ_{P} = w_{X}σ_{X}, from which we get that R_{P} = (σ_{P}/σ_{X})R_{X} + (1 – σ_{P}/σ_{X})R_{F} = R_{F} + σ_{P} (R_{X} – R_{F})/σ_{X}

What we find is that R_{P}(σ_{P}) is just a line whose slope depends on the rate of return and risk of asset X. So essentially, for any risky asset or portfolio you choose, you can easily visualize all the possible ways it can be combined with a risk-free asset by stretching a line from (0, R_{F}) – the risk and return of the risk-free asset – to (s_{X}, R_{X}) – the risk and return of the risky asset. We can even stretch the line beyond this second point by borrowing some of the risk-free asset in order to buy more of the risky asset, which corresponds to a negative weighting w_{F}.

So, we have a quadratic curve representing the possible portfolios obtained from two risky assets, and a line representing the possible portfolios obtained from a risky asset and a risk-free asset. What we can do now is consider the line that starts at (0, R_{F}) and just barely brushes against the quadratic curve – the tangent line to the curve that passes through (0, R_{F}). This point where the curves meet is known as the tangency portfolio.

Every point on this line is a possible combination of Assets A, B, and F. Why? Well, because the points on the line can be thought of as portfolios consisting of Asset F and the tangency portfolio. And here’s the crucial point: this line is above the curve everywhere except at that single point! What this means is that the combination of A, B, and F dominates combinations of just A and B. For virtually any level of desired risk, you do better by choosing a portfolio on the line than by choosing the portfolio on the quadratic curve! (The only exception to this is the point at which the two curves meet, and in that case you do equally well.)

And that is how you optimize your rate of return for a desired level of risk! First generate the hyperbola for portfolios made from your risky assets, then find the tangent to that curve that passes through the point representing the risk-free asset, and then use that line to calculate the optimal portfolio at your chosen level of risk!

If in the future we develop the ability to make accurate simulations of other humans, a lot of things will change for the weirder. In many situations where agents with access to simulations of each other interact, a strange type of apparent backward causality will arise. For instance…

Imagine that you’re competing in a prisoner’s dilemma against an agent that you know has access to a scarily accurate simulation of you. Prior to your decision to either defect or cooperate, you’re mulling over arguments for one course of action over the other. In such a situation, you have to continuously face the fact that for each argument you come up with, your opponent has already anticipated and taken measures to respond to it. As soon as you think up a new line of reasoning, you must immediately update as if your opponent has just heard your thoughts and adjusted their strategy accordingly. Even if your opponent’s decision has already been made and is set in stone, you have no ability to do anything that your opponent hasn’t already incorporated into their decision. Though you are temporally ahead of them, they are in some sense causally ahead of you. Their decision is a response to your (yet-to-be-made) decision, and your decision is not a response to theirs. (I don’t actually believe and am not claiming that this apparent backwards causality is real backwards causality. The arrow of time still points in only one direction and causality follows suit. But it’s worth it in these situations to act as if your opponent has backwards causation abilities.)

When you have two agents that both have access to simulations of the other, things get weird. In such situations, there’s no clear notion of whose decision is a response to the other (as both are responding to each other’s future decision), and so there’s no clear notion of whose decision is causally first. But the question of “who comes first” (in this strange non-temporal sense) turns out to be very important to what strategy the various agents should take!

Let’s consider some examples.

Chicken

Two agents are driving head-on towards each other. Each has a choice to swerve or to stay driving straight ahead. If they both stay, then they crash and die, the worst outcome for all. If one stays and the other swerves, then the one that swerves pays a reputational cost and the one that stays gains some reputation. And if both swerve, then neither gains or loses any reputation. To throw some numerical values on these outcomes, here’s a payoff matrix:

This is the game of chicken. It is an anti-cooperation game, in that if one side knows what the other is going to do, then they want to do the opposite. The (swerve, swerve) outcome is unstable, as both players are incentivized to stay if they know that their opponent will swerve. But so is the (stay, stay) outcome, as this is the worst possible outcome for both players and they both stand to gain by switching to swerve. There are two pure strategy Nash equilibria (swerve, stay) and (stay, swerve), and one mixed strategy equilibria (with the payoff matrix above, it corresponds to swerving with probability 90% and staying with probability 10%).

That’s all standard game theory, in a setting where you don’t have access to your opponent’s algorithm. But now let’s take this thought experiment to the future, where each player is able to simulate the other. Imagine that you’re one of the agents. What should you do?

The first thought might be the following: you have access to a simulation of your opponent. So you can just observe what the simulation of your opponent does, and do the opposite. If you observe the simulation swerving you stay, and if you observe the simulation staying you swerve. This has the benefit of avoiding the really bad (stay, stay) outcomes, while also exploiting opponents that decide to swerve.

The issue is that this strategy is exploitable. While you’re making use of your ability to simulate your opponent, you are neglecting the fact that your opponent is also simulating you. Your opponent can see that this is your strategy, so they know that whatever they decide to play, you’ll play the opposite. So if they decide to tear off their steering wheel to ensure that they will not swerve no matter what, they know that you’ll fall in line and swerve, thus winning them +1 utility and losing you -1 utility. This is a precommitment: a strategy that an agent uses that restricts the number of future choices available to them. It’s quite unintuitive and cool that this sort of tying-your-hands ends up being an enormously powerful weapon for those that have access to it.

In other words, if Agent 1 sees that Agent 2 is treating their decision as a fixed fact and responding to it accordingly, then Agent 1 gets an advantage, as they can precommit to staying and force Agent 2 to yield to them. But if Agent 2 now sees Agent 1 as responding to their algorithm rather than the other way around, then Agent 2 benefits by precommitting to stay. If there’s a fact about which agent precommits “first”, then we can conclusively say that this agent does better, as they can force the outcome they want. But again, this is not a temporal first. Suppose that Agent 2 is asleep at the wheel, about to wake up, and Agent 1 is trying to decide what to do. Agent 1 simulates them and sees that once they wake up they will tear out their steering wheel without even considering what Agent 2 does. Now Agent 1’s hand is forced; he will swerve in response to Agent 2’s precommitment, even though it hasn’t yet been made. It appears that for two agents in a chicken-like scenario, with access to simulations of one another, the best action is to precommit as quickly and firmly as possible, with as little regard for their opponents’ precommitments as they can manage (the best-performing agent is the one that tears off their steering wheel without even simulating their opponent and seeing their precommitments, as this agent puts themselves fully causally behind anybody that simulates them). But this obviously just leads straight to the (stay, stay) outcome!

This pattern of precommitting, then precommitting to not respond to precommitments, then precommiting to not respond to precommitments to not respond to precommitments, and so on, shows up all over the place. Let’s have another example, from the realm of economics.

Company Coordination and Boycotts

In my last post, I talked about the Cournot model of firms competing to produce a homogenous good. We saw that competing firms face a coordination problem with respect to the prices they set: every firm sees it in their rational self-interest to undercut other firms to take their customers, but then other firms follow suit, ending up with the price dropping for everybody. That’s good for consumers, but bad for producers! The process of undercutting and then re-equilibrating continues until the price is at the bare minimum that it takes for a producer to be willing to make the good – essentially just minutely above the cost of production. At this point, producers are making virtually no profit and consumer surplus is maximized.

This coordination problem, like all coordination problems, could be solved if only the firms had the ability to precommit. Imagine that the heads of all the companies meet up at some point. They all see the problem that they’re facing, and recognize that if they can stop the undercutting, they’ll all be much richer. So they sign on to a vow to never undercut each other. Of course, signing a piece of paper doesn’t actually restrict your future options. Every company is still just as incentivized as before to break the agreement and undercut their competitors. It helps if they have plausible deniability; the ability to say that their price drop was actually not intended to undercut, but a response to some unrelated change in the market. All that the meeting does is introduce some social cost to undercutting and breaking the vow that wasn’t there before.

To actually permanently fix the coordination problem, the companies need to be able to sign on to something that truly and irrevocably ties their hands, giving them no ability to back out later on (equivalent to the tearing-off-the-steering-wheel as a credible precommitment). Maybe they all decide to put some money towards the creation of a final-check mechanism that looks over all price changes and intervenes to stop any changes that it detects to be intended to undercut opponents. This is precommitment in the purer sense of literally removing an option that the firms previously had. And if this type of tying-of-hands was actually possible, then each company would be rationally incentivized to sign on! (Of course, they’d all be looking for ways to cheat the system and break the mechanism at every step, which would make its actual creation a tad bit difficult.)

So, if you give all companies the ability to jointly sign on to a credible precommitment to not undercut their opponents, then they will take that opportunity. This will keep prices high and keep profits flowing in to the companies. Producer surplus will be maximized, and consumers will get the short end of the stick. Is there any way for the consumers to fight back?

Sure there is! All they need is the ability to precommit as well. Suppose that all consumers are now given the opportunity to come together and boycott any and all companies that precommit to not undercutting each other. If every consumer signs on, and if producers know this, then it’s no longer worth it for them to put in place the price-monitoring mechanism, as they’d just lose all their customers! Of course, the consumers now face their own coordination problem; many of them will still value the product at a price higher than that which is being offered by the companies, even if they’re colluding. And each individual reasons that as long as everybody else is still boycotting the companies, it makes little difference if just one mutually beneficial trade is made with them. So the consumers will themselves face the problem of how to enforce the boycott. But let’s assume that the consumers work this out so that they credibly precommit to never buying from a company that credibly precommits to not undercutting its competitors. Now the market price swings back in their favor, dropping to the cost of production! The consumers win! Whoohoo!

But we’re not done yet. It was only worth it for the consumers to sign on to this precommitment because they predicted that the companies would respond to their precommitment. But what if the companies, seeing the boycott-tactic coming, credibly precommit to never yielding to boycotters? Then the consumers, responding to this precommitment, will realize that boycotting will have no effect on prices, and will just cause them all to lose out on mutually beneficial trades! So they won’t boycott, and therefore the producers get the surplus once more. And just like before, this swings back and forth, with the outcome at each stage depending on which agent treats the other agent’s precommitment as being more primal. But if they each run their apparently-best strategy (that is, making their precommitments with no regard to the precommitments of the other party so as to force their hand and place their own precommitments at the beginning of the causal chain), then we end up with the worst possible outcome for all: producers don’t produce anything and consumers don’t consume, and everybody loses out.

This question of how agents that can simulate one another AND precommit to courses of action should ultimately behave is something that I find quite puzzling and am not sure how to resolve.

The Cournot model is a simple economic model used to describe what happens when multiple companies compete with one another to produce some homogenous product. I’ve been playing with it a bit and ended up solving the general linear case. I assume that this solution is already known by somebody, but couldn’t find it anywhere. So I will post it here! It gives some interesting insight into the way that less-than-perfectly-competitive markets operate. First let’s talk about the general structure of the Cournot model.

Suppose we have n firms. Each produces some quantity of the product, which we’ll label as . The total amount of product on the market will be given the label . Since the firms are all selling identical products, it makes sense to assume that the consumer demand function will just be a function of the total quantity of the product that is on the market: . (This means that we’re also disregarding effects like customer loyalty to a particular company or geographic closeness to one company location over another. Essentially, the only factor in a consumer’s choice of which company to go to is the price at which that company is selling the product.)

For each firm, there is some cost to producing the good. We capture this by giving each firm a cost function . Now we can figure out the profit of each firm for a given set of output values . We’ll label the profit of the kth firm as . This profit is just the amount of money they get by selling the product minus the cost of producing the product: .

If we now assume that all firms are maximizing profit, we can find the outputs of each firm by taking the derivative of the profit and setting it to zero. . This is a set of n equations with n unknown, so solving this will fully specify the behavior of all firms!

Of course, without any more assumptions about the functions and , we can’t go too much further with solving this equation in general. To get some interesting general results, we’ll consider a very simple set of assumptions. Our assumptions will be that both consumer demand and producer costs are linear. This is the linear Cournot model, as opposed to the more general Cournot model.

In the linear Cournot model, we write that (for some a and b) and . As an example, we might have that P(Q) = $100 – $2 × Q, which would mean that at a price of $40, 30 units of the good will be bought total.

The constants represent the marginal cost of production for each firm, and the linearity of the cost function means that the cost of producing the next unit is always the same, regardless of how many have been produced before. (This is unrealistic, as generally it’s cheaper per unit to produce large quantities of a good than to produce small quantities.)

Now we can write out the profit-maximization equations for the linear Cournot model. . Rewriting, we get . We can’t immediately solve this for , because remember that Q is the sum of all the quantities produced. All n of the quantities we’re trying to solve are in each equation, so to solve the system of equations we have to do some linear algebra!

Translating this to a matrix equation…

Now if we could only find the inverse of the first matrix, we’d have our solution!

I found the inverse of this matrix by using the symmetry in the matrix to decompose it into two matrices that were each easier to work with:

As a hypothesis, suppose that the inverse matrix has a similar form (one value for the diagonal elements, and another value for all off-diagonal elements). This allows us to write an equation for the inverse matrix:

To solve this, we’ll use the following easily proven identities.

Alright awesome! Our hypothesis turned out to be true! (And it would have even if the entries in our matrix hadn’t been 1s and 2s. This is a really cool general method to find inverses of this family of matrices.) Now we just use this inverse matrix to solve for the output from each firm!

Define:

And there we have it, the full solution to the general linear Cournot model! Let’s discuss some implications of these results. First of all, let’s look at the two extreme cases: monopoly and perfect competition.

Monopoly: n = 1

Perfect Competition: n → ∞

The first observation is that the behavior of the market under monopoly looks very different from the case of perfect competition. For one thing, notice that the price under perfect competition is always going to be lower than the price under monopoly. This is a nice demonstration of the so-called monopoly markup. The quantity intuitively corresponds to the highest possible price you could get for the product (the most that the highest bidder would pay). And the quantity , the production cost, is the lowest possible price at which the product would be sold. So the monopoly price is the average of the highest price you could get for the good and the lowest price at which it could be sold.

The flip side of the monopoly markup is that less of the good is produced and sold under a monopoly than under perfect competition. There are trades that could be happening (trades which would be mutually beneficial!) which do not occur. Think about it: the monopoly price is halfway between the cost of production and the highest bidder’s price. This means that there are a bunch of people that would buy the product at above the cost of production but below the monopoly price. And since the price they would buy it for is above the cost of production, this would be a profitable exchange for both sides! But alas, the monopoly doesn’t allow these trades to occur, as it would involve lowering the price for everybody, including those who are willing to pay a higher price, and thus decreasing net profit.

Things change as soon as another firm joins the market. This firm can profitably sell the good at a lower price than the monopoly price and snatch up all of their business. This introduces a downward pressure on the price. Here’s the exact solution for the case of duopoly.

Duopoly: n = 2

Interestingly, in the duopoly case the market price still rests at a value above the marginal cost of production for either firm. As more and more firms enter the market, competition pushes the price down further and further until, in the limit of perfect competition, it converges to the cost of production.

The implication of this is that in the limit of perfect competition, firms do not make any profit! This may sound a little unintuitive, but it’s the inevitable consequence of the line of argument above. If a bunch of companies were all making some profit, then their price is somewhere above the cost of production. But this means that one company could slightly lower its price, thus snatching up all the customers and making massively more money than its competitors. So its competitors will all follow suit, pushing down their prices to get back their customers. And in the end, all the firms will have just decreased their prices and their profits, even though every step in the sequence appeared to be the rational and profitable action by each firm! This is just an example of a coordination problem. If the companies could all just agree to hold their price fixed at, say, the monopoly price, then they’d all be better off. But each individual has a strong monetary incentive to lower their price and gather all the customers. So the price will drop and drop until it can drop no more (that is, until it has reached the cost of production, at which point it is no longer profitable for a company to lower their price).

This implies that in some sense, the limit of perfect competition is the best possible outcome for consumers and the worst outcome for producers. Every consumer that values the product above the cost of its production will get it, and they will all get it at the lowest possible price. So the consumer surplus will be enormous. And companies producing the product make no net profit; any attempt to do so immediately loses them their entire customer base. (In which case, what is the motivation for the companies to produce the product in the first place? This is known as the Bertrand paradox.)

We can also get the easier-to-solve special case where all firms have the same cost of production.

Equal Production Costs

It’s curious that in the Cournot model, prices don’t immediately drop to production levels as soon you go from a monopoly to a duopoly. After all, the intuitive argument I presented before works for two firms: if both firms are pricing the goods at any value above zero, then each stands to gain by lowering the price a slight bit and getting all the customers. And this continues until the price settles at the cost of production. We didn’t build in any ability of the firms to collude to the model, so what gives? What the Cournot model tells us is certainly more realistic (we don’t expect a duopoly to behave like a perfectly competitive market), but where does this realism come from?

The answer is that in a certain sense we did build in collusion between firms from the start, in the form of agreement on what price to sell at. Notice that our model did not allow different firms to set different prices. In this model, firms compete only on quantity of goods sold, not prices. The price is set automatically by the consumer demand function, and no single individual can unilaterally change their price. This constraint is what gives us the more realistic-in-character results that we see, and also what invalidates the intuitive argument I’ve made here.

One final observation. Consider the following procedure. You line up a representative from each of the n firms, as well as the highest bidder for the product (representing the highest price at which the product could be sold). Each of the firms states their cost of production (the lowest they could profitably bring the price to), and the highest bidder states the amount that he values the product (the highest price at which he would still buy it). Now all of the stated costs are averaged, and the result is set as the market price of the good. Turns out that this procedure gives exactly the market price that the linear Cournot model predicts! This might be meaningful or just a curious coincidence. But it’s quite surprising to me that the slope of the demand curve () doesn’t show up at all in the ultimate market price, only the value that the highest bidder puts on the product!

Op-amps are magical little devices that allow us to employ the laws of electromagnetism for our purposes, giving us the power to do ultimately any computation using physics. To explain this, we need to take three steps: first transistors, then differential amplifiers, and then operational amplifiers. The focus of this post will be on purely analog calculations, as that turns out to be the most natural type of computation you get out of op amps.

1. Transistor

Here’s a transistor:

It has three ports, named (as labelled) the base, the collector and the emitter. Intuitively, you can think about the base as being like a dial that sensitively controls the flow of large quantities of current from the collector to the emitter. This is done with only very small trickles of current flowing into the base.

More precisely, a transistor obeys the following rules:

The transistor is ON if the voltage at the collector is higher than the voltage at the emitter. It is OFF otherwise.

When the transistor is in the ON state, which is all we’ll be interested in, the following regularities are observed:

Current flows along only two paths: base-to-emitter and collector-to-emitter. A transistor does not allow current to flow from the emitter to either the collector or the base.

The current into the collector is proportional to the current into the base, with a constant of proportionality labelled β whose value is determined by the make of the transistor. β is typically quite large, so that the current into the collector is much larger than the current into the base.

The voltage at the emitter is 0.6V less than the voltage at the base.

Combined, these rules allow us to solve basic transistor circuits. Here’s a basic circuit that amplifies input voltages using a transistor:

Applying the above rules, we derive the following relationship between the input and output voltages:

The +15V port at the top is important for the functioning of the amplifier because of the first rule of transistors, regarding when they are ON and OFF. If we just had it grounded, then the voltage at the collector would be negative (after the voltage drop from the resistor), so the transistor would switch off. Having the voltage start at +15V allows V_{C} to be positive even after the voltage drop at the resistor (although it does set an upper bound on the input currents for which the circuit will still operate).

And the end result is that any change in the input voltage will result in a corresponding change in the output voltage, amplified by a factor of approximately -R_{C}/R_{E}. Why do we call this amplification? Well, because we can choose whatever values of resistance for R_{C} and R_{E} that we want! So we can make R_{C} a thousand times larger than R_{E}, and we have created an amplifier that takes millivolt signals to volt signals. (The amplification tops out at +15V, because a signal larger than that would result in the collector voltage going negative and the transistor turning off.)

Ok, great, now you’re a master of transistor circuits! We move on to Step 2: the differential amplifier.

2. Differential Amplifier

A differential amplifier is a circuit that amplifies the difference between two voltages and suppresses the commonalities. Here’s how to construct a differential amplifier from two transistors:

To make a differential amplifier, we require that A_{DM} be large and A_{CM} be small. We can achieve this if R_{f} >> R_{C }>> R_{C}. Notice that since the amplification is a function of the ratio of our resistors, we can easily make the amplification absolutely enormous.

Here’s one way this might be useful: say that Alice wants to send a signal to Bob over some distance, but there is a source of macroscopic noise along the wire connecting the two. Perhaps the wire happens to pass through a region in which large magnetic disturbances sometimes occur. If the signal is encoded in a time-varying voltage on Alice’s end, then what Bob gets may end up being a very warped version of what Alice sent.

But suppose that instead Alice sends Bob two signals, one with the original message and the other just a static fixed voltage. Now the difference between the two signals represents Alice’s message. And crucially, if these two signals are sent on wires that are right side-by-side, then they will pick up the same noise!

This means that while the original message will be warped, so will the static signal, and by roughly the same amount! Which means that the difference between the two signals will still carry the information of the original message. This allows Alice to communicate with Bob through the noise, so long as Bob takes the two wires on his end and plugs them into a differential amplifier to suppress the common noise factor.

3. Operational Amplifier

To make an operational amplifier, we just need to make some very slight modifications to our differential amplifier. The first is that we’ll make our amplification as large as possible. We can get this by putting multiple stages of differential amplifiers side by side (so if your differential amplifier amplifies by 100, two of them will amplify by 10,000, and three by 1,000,000).

What’ll happen now if we send in two voltages, with V_{in} very slightly larger than V_{in}‘? Well, suppose that the difference is on the order of a single millivolt (mV). Then the output voltage will be on the order of 1 mV multiplied by our amplication factor of 1,000,000. This will be around 1000V. So… do we expect to get out an output signal of 1000V? No, clearly not, because our maximum possible voltage at the output is +15V! We can’t create energy from nowhere. Remember that the output voltage V_{out} is equal to 15V – R_{C}i_{C}, and that the transistor only accepts current traveling from the collector to the emitter, not the other way around. This means that i_{C}cannot be negative, so V_{out} cannot be larger that +15V. In addition, we know that V_{out} cannot be smaller than 0V (as this would turn off the transistor via Rule 1 above).

What this all means is that if V_{in} is even the slightest bit smaller than V_{in}‘, V_{out} will “max out”, jumping to the peak voltage of +15V (remember that A_{DM} is negative). And similarly, if V_{in} is even the slightest bit larger than V_{in}‘, then V_{out} will bottom out, dropping to 0V. The incredible sensitivity of the instrument is given to us by the massive amplification factor, so that it will act as a binary measurement of which of the voltages is larger, even if that difference is just barely detectable. Often, the bottom voltage will be set to -15V rather than ground (0V), so that the signal we get out is either +15V (if V_{in} is smaller than V_{in}‘) or -15V (if V_{in} is greater than V_{in}‘). That way, perfect equality of the inputs will be represented as 0V. That’s the convention we’ll use for the rest of the post. Also, instead of drawing out the whole diagram for this modified differential amplifier, we will use the following simpler image:

Ok, we’re almost to an operational amplifier. The final step is to apply negative feedback!

What we’re doing here is just connecting the output voltage to the V_{in}‘ input. Let’s think about what this does. Suppose that V_{in} is larger than V_{in}‘. Then V_{out} will quickly become negative, approaching -15V. But as V_{out} decreases, so does V_{in}‘! So V_{in}‘ will decrease, getting closer to V_{in}. Once it passes V_{in}, the quantity V_{in }– V_{in}‘ suddenly becomes negative, so V_{out} will change direction, becoming more positive and approaching +15V. So V_{in}‘ will begin to increase! This will continue until it passes V_{in} again, at which point V_{out} will change signs again and V_{in}‘ will start decreasing. The result of this process will be that no matter what V_{in}‘ starts out as, it will quickly adjust to match V_{in} to a degree dictated by the resolution of our amplifier. And we’ve already established that the amplifier can be made to have an extremely high resolution. So this device serves as an extremely precise voltage-matcher!

This device, a super-sensitive differential amplifier with negative feedback, is an example of an operational amplifier. It might not be immediately obvious to you what’s so interesting or powerful about this device. Sure it’s very precise, but all it does is match voltages. How can we leverage this to get actually interesting computational behavior? Well, that’s the most fun part!

4. Let’s Compute!

Let’s start out by seeing how an op-amp can be used to do calculus. Though this might seem like an unusually complicated starting place, doing calculus with op amps is significantly easier and simpler than doing something like multiplication.

As we saw in the last section, if we simply place a wire between V_{out} and V_{in}‘ (the input labelled with a negative sign on the diagram), then we get a voltage matcher. We get more interesting behavior if we place other circuit components along this wire. The negative feedback still exists, which means that the circuit will ultimately stabilize to a state in which the two inputs are identical, and where no current is flowing into the op-amp. But now the output voltage will not just be zero.

Let’s take a look at the following op-amp circuit:

Notice that we still have negative feedback, because the V_{–} input is connected to the output, albeit not directly. This means that the two inputs to the op amp must be equal, and since V_{+} is grounded, the other must be at 0V as well. It also means that no current is flowing into the op amp, as current only flows in while the system is stabilizing.

Those two pieces of information – that the input voltages are equal and that no current flows through the device – are enough to allow us to solve the circuit.

And there we have it, a circuit that takes in a voltage function that changes with time and outputs the integral of this function! A natural question might be “integral from what to what”? The answer is just: the integral from the moment the circuit is connected to the present! As soon as the circuit is connected it begins integrating.

Alright, now let’s just switch the capacitor and resistor in the circuit:

See if you can figure out for yourself what this circuit does!

…

…

Alright, here’s the solution.

We have a differentiator! Feed in some input voltage, and you will get out a precise measurement of the rate of change of this voltage!

I find it pretty amazing that doing integration and differentiation is so simple and natural. Addition is another easy one:

We can use diodes to produce exponential and logarithm calculators. We utilize the ideal diode equation:

The logarithm can now be used to produce circuits for calculating exponentials and logarithms:

Again, these are all fairly simple-looking circuits. But now let’s look at the simplest way of computing multiplication using an op amp. Schematically, it looks like this:

And here’s the circuit in full detail:

There’s another method besides this one that you can use to do multiplication, but it’s similarly complicated. It’s quite intriguing that multiplication turns out to be such an unnatural thing to get nature to do to voltages, while addition, exponentiation, integration, and the rest are so simple.

What about boolean logic? Well, it turns out that we already have a lot of it. Our addition corresponds to an OR gate if the input voltages are binary signals. We also already have an AND gate, because we have multiplication! And depending on our choice of logic levels (which voltage corresponds to True and which corresponds to False), we can really easily sketch a circuit that does negation:

And of course if we have NOT and we have AND, then we have NAND, and if we have NAND, then we have all the rest of binary logic. We can begin to build flip-flop gates out of the NAND gates to get simple memory cells, and we’re on our way to Turing-completeness!