AI Safety Optimal Investment

It occurred to me today that I have never seen any attempt at calculating how much society should be willing to spend to prevent an AI Catastrophe. I am pretty good at this kind of thing, so here's a quick Fermi estimate:

Conceptually, this is a lot like buying a life insurance policy for the human race. There is some probability of a catastrophe, so the annual amount we should be willing to pay for insurance is the cost of the catastrophe times its annual probability.

First, we need the dollar value of the human race. This is relatively easy. We have revealed-preference estimates for the value of a flourishing human life in a rich society, so that should be about right for estimating the value people place on a good personal and genetic future, assuming a future where everyone's quality of life is about as good as the average citizen of the US in the early 21st century.

The Value of a Statistical Life in the US is about $10 million, or $1x10^7. Over the relevant time frame in which an AI catastrophe is likely, world population will likely be stabilizing at about 10 billion, or 1x10^10. So the value of the human race, in current US dollars, is about $1x10^17.

Now we find the probability of a catastrophe. For convenience, I use numbers presented in the article linked above, which is a pretty good summary of the current understanding of the field.

Artificial Superintelligence is unlikely to happen before 2025, but almost certain to arrive by 2125. So each year in that timeframe, there is a 1 in 100, or 1x10^-2, chance of ASI.

Now the numbers get very speculative, so in the grand tradition of Fermi estimates I will simply round everything to the nearest power of ten and say that there is a 10% chance that, when ASI develops, it will, without any safeguards, destroy the human race.

So, given a 1x10^-3 annual chance of losing something valued at $1x10^17, we should be willing to spend $1x10^14 to prevent that possibility.

$100 trillion a year is a lot of money. That is almost exactly equal to the total value of the entire world economy. It tells us that any policy short of destroying the entire world economy and/or killing millions of people would be worth doing if it prevented an ASI catastrophe with 100% certainty [Edit: assuming that the policy does not make other existential risks significantly more likely]. For example, a total ban on producing or maintaining any kind of computer, combined with a credible threat to nuke anyone who violates this ban, is a legitimate policy option with a positive expected payoff.

More realistically, we will continue on our current technological course and try to make things safer. Even if AI research only had a 1 in 100 chance of guaranteeing Friendly AI, we should be willing to pay a trillion dollars a year for it.

IBM Watson Personality Insights

IBM claims that it can provide insights to my personality by analyzing the text I write. Let's test this by separately feeding it the text of my last few blog posts and seeing what it says about each one:

Organ Donor Safety Exemptions:

You are shrewd, unconventional and can be perceived as indirect.

You are imaginative: you have a wild imagination. You are laid-back: you appreciate a relaxed pace in life. And you are intermittent: you have a hard time sticking with difficult tasks for a long period of time.

Your choices are driven by a desire for prestige.

You are relatively unconcerned with tradition: you care more about making your own path than following what others have done. You consider helping others to guide a large part of what you do: you think it is important to take care of the people around you.

Well-Being Analysis:

You are shrewd, inner-directed and can be perceived as indirect.

You are unconcerned with art: you are less concerned with artistic or creative activities than most people who participated in our surveys. You are intermittent: you have a hard time sticking with difficult tasks for a long period of time. And you are imaginative: you have a wild imagination.

Your choices are driven by a desire for prestige.

You are relatively unconcerned with both taking pleasure in life and tradition. You prefer activities with a purpose greater than just personal enjoyment. And you care more about making your own path than following what others have done.

So far, pretty consistent. But both of those blog posts were technical academic analysis. What happens when I give it a first-person account of a more emotional experience?

Chen Guangcheng:

You are social, boisterous and unconventional.

You are empathetic: you feel what others feel and are compassionate towards them. You are assertive: you tend to speak up and take charge of situations, and you are comfortable leading groups. And you are confident: you are hard to embarrass and are self-confident most of the time.

Your choices are driven by a desire for efficiency.

You are relatively unconcerned with tradition: you care more about making your own path than following what others have done. You consider helping others to guide a large part of what you do: you think it is important to take care of the people around you.

That is a big change in the 'you are' and 'driven by' lines, and the first paragraph is entirely different. The only commonality is the 'unconcerned with tradition' and 'helping others' parts of the last paragraph.

Now, what does it think about my account of going out and helping people who got their cars stuck in the snow?

Car Shoving:

You are heartfelt.

You are empathetic: you feel what others feel and are compassionate towards them. You are unconcerned with art: you are less concerned with artistic or creative activities than most people who participated in our surveys. And you are calm-seeking: you prefer activities that are quiet, calm, and safe.

Your choices are driven by a desire for well-being.

You consider helping others to guide a large part of what you do: you think it is important to take care of the people around you. You are relatively unconcerned with tradition: you care more about making your own path than following what others have done.

Again, it says something almost completely different. It is interesting to note that the personality analysis for the last two are backwards. I think that this paragraph describes the Me who attended the Chen Guangcheng talk, and the previous one describes the Me who went out to shove cars. It is particularly funny that it reacts to my description of shoving cars around in hilly slippery roads by saying 'you prefer activities that are quiet, calm, and safe'.

Now we go back to a more a more analytical post, but a different kind of analysis:

Media Musings:

You are shrewd, somewhat inconsiderate and can be perceived as indirect.

You are laid-back: you appreciate a relaxed pace in life. You are carefree: you do what you want, disregarding rules and obligations. And you are imaginative: you have a wild imagination.

Your choices are driven by a desire for efficiency.

You consider both independence and taking pleasure in life to guide a large part of what you do. You like to set your own goals to decide how to best achieve them. And you are highly motivated to enjoy life to its fullest.

The first line looks like its reaction to my other bits of analysis, and the 'desire for efficiency' is a repeat, but the rest is mostly things it has not said about me before. What might it say next?

Confusing Social Norms

You are a bit inconsiderate, somewhat critical and excitable.

You are melancholy: you think quite often about the things you are unhappy about. You are intermittent: you have a hard time sticking with difficult tasks for a long period of time. And you are unconcerned with art: you are less concerned with artistic or creative activities than most people who participated in our surveys.

Your choices are driven by a desire for connectedness.

You consider helping others to guide a large part of what you do: you think it is important to take care of the people around you. You are relatively unconcerned with taking pleasure in life: you prefer activities with a purpose greater than just personal enjoyment.

Some new and different stuff, some repeats. That post was more of an expression of confusion and questioning than an account or analysis. Let's chalk that one down to the small sample size, it was just 360 words after I chopped out the quotes and links.

The next one should be more informative, as it combines analysis and first-person accounts, and talks about something that is more connected to my identity:

Lego in Asia

You are inner-directed and skeptical.

You are calm-seeking: you prefer activities that are quiet, calm, and safe. You are empathetic: you feel what others feel and are compassionate towards them. And you are deliberate: you carefully think through decisions before making them.

Your choices are driven by a desire for prestige.

You are relatively unconcerned with both taking pleasure in life and tradition. You prefer activities with a purpose greater than just personal enjoyment. And you care more about making your own path than following what others have done.

The overall analysis sounds familiar and kind of accurate, but it jumps out at me that it says 'you are skeptical' to a blog post that I would characterize as being filled with the wonder of shared experience and progress and connectedness. It has not said that about any other post.

Cargo Cult Crafts:

You are excitable.

You are laid-back: you appreciate a relaxed pace in life. You are empathetic: you feel what others feel and are compassionate towards them. And you are calm-seeking: you prefer activities that are quiet, calm, and safe.

Your choices are driven by a desire for well-being.

You are relatively unconcerned with tradition: you care more about making your own path than following what others have done. You consider independence to guide a large part of what you do: you like to set your own goals to decide how to best achieve them.

Okay, that one just confused the algorithm. It says that I am both excitable and laid-back, that I 'prefer activities that are quiet, calm, and safe' when the entire focus of the blog post is about how I like to hack at real pumpkins with real knives and learn by taking risks, and that I am 'unconcerned with tradition' when I defend the traditions of my childhood against a shallow commercial substitute.

Its reaction to Important Information, Important Caveat is about the same as its reaction to most analysis posts, so no sense repeating it. But I want to feed it one last post, a rant, to see how it reacts:

Whack Rant

You are a bit compulsive, somewhat critical and skeptical.

You are intermittent: you have a hard time sticking with difficult tasks for a long period of time. You are unconcerned with art: you are less concerned with artistic or creative activities than most people who participated in our surveys. And you are melancholy: you think quite often about the things you are unhappy about.

Your choices are driven by a desire for efficiency.

You consider achieving success to guide a large part of what you do: you seek out opportunities to improve yourself and demonstrate that you are a capable person. You are relatively unconcerned with tradition: you care more about making your own path than following what others have done.

That is actually about how you would expect someone to react to an intellectual takedown of something I hated.

Overall, it is pretty clear that the system does not have any deep personality insights. It is reacting almost entirely to the rhetorical choices I make for each particular post. Anyone who has any writing skill or understanding of rhetoric knows that you should use a different voice, tone, and approach in different situations.

The only thing consistently output, in eight of ten writing samples, was that I am relatively unconcerned with tradition. Nothing else showed up in more than half of the results. This probably reflects my consistent use of scientific and analytical language.

Five of the results claimed that 'You are intermittent: you have a hard time sticking with difficult tasks for a long period of time.' and another five claimed 'You are unconcerned with art: you are less concerned with artistic or creative activities than most people who participated in our surveys.' I consider both of these claims to be dubious, and I am not really sure where they came from.

Four of the results claimed that I am shrewd, empathetic, driven by prestige, guided by helping others, and/or unconcerned with taking pleasure in life. Again, that seems kind of random and not really connected to who I am.

As a final bit of fun, let me plug the output of the system into its input, so we can see what it says about itself:

You are confident and generous.

You are assertive: you tend to speak up and take charge of situations, and you are comfortable leading groups. You are calm under pressure: you handle unexpected events calmly and effectively. And you are respectful of authority: you prefer following with tradition in order to maintain a sense of stability.

Experiences that give a sense of well-being hold some appeal to you.

You are relatively unconcerned with tradition: you care more about making your own path than following what others have done. You consider helping others to guide a large part of what you do: you think it is important to take care of the people around you.

Organ Donor Safety Exemptions

The following question was posted on the EA Policy Facebook page:

General idea: should exemptions be granted to certain safety laws if you sign up to be an organ donor?

Would such policies likely be a positive or negative for organ donation and safety compliance?

My response turned into a full blog post:

The biggest possible failure is that it discredits the idea of voluntary organ donation, resulting in less organ donation from other sources. I have no idea how big this effect might be.

Concrete examples help analysis, so I will consider this by looking at motorcycle helmet laws:

It is possible that telling people 'We will allow you to ride a motorcycle without a helmet if and only if you agree to become an organ donor, so we can harvest your organs after you crash and become brain-dead.' might shock them into wearing their helmet voluntarily, which seems like a simple good thing with no side effects aside from the possible organ donation discrediting.

Assuming that people are rational agents, some of whom experience large amounts of utility from riding without a helmet, we can expect the following:

1) People take the deal and ride without a helmet (good effect, social utility increases).
2) Motorcycle crashes decrease due to risk compensation (good effect).
3) The crashes that happen are more likely to cause death or brain injuries to motorcycle riders (bad effect).
4) More organs are available for use, both from the crash victims and people who take the deal and die from other causes (good effect).

I expect effects 3 and 4 to be the largest. More riders will die, or become disabled and require a lifetime of expensive medical care, and lives of people waiting for organ transplants will be saved.

Fatal motorcycle crashes tend to be an excellent source of donor organs, but I am fairly sure that most of them will cause a net decrease in QALYs, even the ones that result in many useful organs. Organ recipients tend to be old and unhealthy, so the organ does not give them that many QALYs. If the rider was young and healthy, they likely lose more QALYs than all the combined recipients of the organs would gain. If the rider was middle-aged and unhealthy, the organs are not likely to be as useful.

However, organ donation makes the QALY loss from those fatalities significantly less than most fatalities. (A nonfatal crash resulting in permanent disability will cause a larger QALY loss than a fatal crash.) If the rate of motorcycle crashes among people who take the deal is small enough, then the good effects of all of their organ donation will outweigh the bad effects of their crashes, and the policy is good.

I suspect that the crash rate is low enough that the policy of offering these exemptions is good. I think the breakeven would be something like ten new organ donors for every lethal or disabling crash, and I am almost sure that the lifetime risk of such a crash for motorcycle riders is significantly less than 10%.

Well-Being Analysis

This is a technical post, only of interest to people doing policy analysis. Much of it was originally written in response to a Facebook post related to Effective Altruism Policy Analytics, and I am reposting it here so it is in a more permanent place that I can link to.

One of the group members posted a link to the following paper:

Well-Being Analysis vs. Cost-Benefit Analysis

and asked me if we should consider using this method.

Briefly, well-being analysis is an attempt to measure the costs and benefits of a policy using a direct measure of human happiness or well-being, instead of using dollar values. My response was divided into two sections:

Well-being analysis in any form is not currently useful for people submitting policy comments:

1) Any comment that relied on a well-being analysis would likely be summarily dismissed. As the paper states, regulatory agencies are required by law to do cost-benefit analysis, and that is what they understand. They do not understand this new alien thing, have no way of judging its quality, and have no incentive to care about it.

2) The language of policy right now is monetary costs and benefits. We will influence policy by speaking the language currently used, and speaking it well. Lojban might be a better language than English, but it would be folly to submit comments in it.

3) There is a good chance that in 30 years, an overall framework like this will replace cost-benefit analysis and be required by law. This does not make it useful to us. 30 years ago, cost-benefit analysis was a bizarre fringe thing that nobody cared about and that would not affect policy in any way.

Well-bring analysis as presented in the paper has serious flaws:

4) Their example analysis dramatically undervalues monetary costs. They claim that "If that same individual's income decreased from $100,000 to $36,700, she would lose 0.11 WBUs" on a scale of 1 to 10. By comparison, "Unemployed individuals suffer a loss of 0.83 WBUs per year during the time that they remain unemployed." However, we know that people who have lost their jobs routinely choose to remain unemployed rather than settle for jobs that pay less than what they earned before.

4a) The data they use to determine well-being will systematically undervalue money, because they do a regression that separately counts the effects of money and health. Treating money and health as separate and exogenous is a grave error. Money and health are heavily correlated because money buys health. Most of what rich people do with their money is to rearrange their lives so they will be healthier and safer. When you take away money, health goes down. Their data on the 'well-being effect of income' is just the residual effect of money after 'correcting' for all the things that the money bought.

5) To a first approximation, the well-being analysis they present is just a way to ignore the compliance costs of any regulation that does not cause observable layoffs. Widespread use of this system in its current form would encourage a horde of very expensive regulations. Given how cheap it is for individuals to purchase lives with money, this is a huge problem. If 0.001% of the compliance costs of the EPA regulation they support would have otherwise gone to effective charities, the regulation has reduced well-being.

6) According to this analysis methodology in its current form, job losses dominate any consideration of economic efficiency, technological growth, or cheaper, better goods. Consistent use of it to analyze policies would result in support of policies that enforced stagnation on an economy, preventing almost any kind of innovation that had the potential to cause layoffs unless that innovation had an immediate and obvious health or safety benefit. Specifically, a "Ban computers" policy analyzed 40 years ago would probably score highly on a well-being analysis.

I do hope that, at some point in the future, all government actions will be evaluated on how they are expected to affect the well-being of all people, measured directly. Money is an imperfect measurement of how things affect well-being, because different people value money differently and a lot of important goods are hard to put money values on.

But right now, measuring things by their monetary impact, and the health and life impacts converted to their monetary values, is the best we can do. There is simply far more data available on money than there is on direct well-being measurements, and an imperfect system you can actually use is better than a theoretically ideal system that cannot be implemented.

Also, money is a much better measurement tool than most people realize. The money cost of doing a thing is a signal that includes a very good estimate of most of the resources consumed by that thing. Pollution and other externalities often make the money price of a thing an underestimate of its true cost, but in a competitive market it is rare for a money price to overestimate the cost of a thing. Whenever you think something should be cheap, but it costs a lot, that almost certainly means that you are missing something important. Regulators tend to miss important side effects and assume that things will be cheaper than they really are, because they only look sat narrow technical effects without considering larger impacts. Measuring the monetary impacts of a rule can correct for this.

However, as more research on well-being accumulates, and we get more data on how things affect quality of life, and we develop a better understanding of the complicated chains of causality that dominate all human interactions including the economy, we will probably move toward some kind of well-being analysis. The monetary costs and benefits will become a subset of the analysis. But it will likely take decades to get to that point.

Chen Guangcheng

Last evening, my girlfriend and I went to a book release event at the National Press Club. The author was Chen Guangcheng, the blind Chinese dissident who escaped house arrest three years ago. My girlfriend works with him at a policy think tank, and he had given her a signed book.

It was one of the most intense experiences I have had in a long time. I was confronted with a lot of thoughts and emotions.

Chen started by walking unsteadily to the podium, fumbling with braille notes, and giving a speech in broken English that included an excerpt from his book. As he described the effects of his imprisonment, he nearly broke down in tears and had to stop for almost a full minute.

On some level, he looked foolish or frail, but it was clear that we were seeing a great man afflicted by great trials. I found myself thinking how unfair it was that he lived in a world that, after all of his other problems, forced him to go through the hassle of learning a foreign language late in life just to continue doing his work and talking to people.

Later on, during the question and answer period, He spoke through a translator. I could tell from the way he spoke in Chinese that he was a good orator in his native language, speaking swiftly and confidently and with emotional resonance.

People tried to ask him questions about his real or percieved conflicts with people in the US. He deflected these in ways that were wise, but probably sounded a bit strange to most Americans.

While thinking about his life and the surrounding events, I realized several things about heroes. First, they will usually be difficult people. If they did not submit to armies of government thugs, there is no way they will submit to your assumptions and demands. They will do what they see is right, and your silly little social norms or strategic gamesmanship can be shoved where the sun does not shine.

Second, I was reminded of that I and many of my colleagues overvalue rational thought and analytical intelligence. A true hero is marked by insight and willpower, seeing what needs to be done and doing it. The media and academic world value a kind of cleverness and slickness that he did not have, but his presence reminded me that our culture often values and rewards the wrong virtues, and that we should remind ourselves about what the ancient thinkers like Aristote and Confucuious knew about virtue ethics.

Car Shoving

This morning, we had a very pretty snow; light powdery flakes that coat the trees and ground with a beautiful white dusting. It was the kind of snow that skiers love, the kind that anything will glide effortlessly over. About 9:00, I bundled up and went out for a walk to enjoy the scenery.

I quickly saw that what is good for skiing is terrible for driving. There were three pileups blocking the road outside my apartment. Nobody was getting anywhere. The rather steep hills had been turned into pure ice, and people were either sliding around or 'parked' at all kinds of angles on the side of the road. There were people attempting to drive around who should not have left their parking space, people who had no clue how to drive in snow and/or were driving crappy cars with bald tires. Some people had been stuck there since 6:30. It was a complete Charlie Foxtrot.

I joined a group of guys who were trying to clear up the mess near the entrance to the road. We told people coming in about the road conditions, trying to get them to turn around while they still could. For the cars on the road, we pushed them to get them started off the ice, and to help them steer around the other cars. Often this required telling people to put the car in neutral and coast backwards down the hill while we shoved the car away from the parked cars. Sometimes we were able to get a car out of the road and on its way, but for many of the ones near the bottom of the hill we had to just shove them off to the side to get them out of the road.

The de facto leader of the group was a charismatic young black guy who was clearly skilled, smart, aware of the situation, and knew how to handle winter conditions. Often he would offer to drive people's cars for them, using his skills to glide them down the hill and away from trouble. The thing that amazed me was the trouble he had getting people to listen to him. He would explain to people, repeatedly, that they had no chance of getting up the hill unless they had a 4-wheel drive and knew how to use it. He told them how he had personally seen 20 cars fail to make the hill. And people would repeatedly ignore him, try the hill, fail, get stuck or start sliding around, and have to rely on us to get them out of the road or back down the hill. Some people refused to cooperate, and we basically had to ignore them and route people around them.

We, and the other groups spontaneously forming at other parts of the road, did a decent job of clearing enough people out of the middle of the road to prevent it from getting completely stuck. But the situation was not really resolved until around 10:30, when the police came, parked their cars across the street to prevent anyone from entering it, and started ordering everyone to park legally on the curb or get off the road. I left then, after helping a couple last people on their way. Presumably the road has been cleared, plowed and salted by now.

At the start of my walk, I had been a bit chilly despite all the layers. But while moving cars, I was comfortable, even a bit warm toward then end. When I got back to my apartment around 11, I discovered that my t-shirt was completely soaked through with sweat. Right now, my thermostat claims that it is 46 degrees in my apartment, but I am sitting writing this in pajama bottoms and a (different) t-shirt, barely feeling cold at all.

So, the things I was reminded of this morning were:

1) People do not appreciate the usefulness of friction until it goes away.

2) Spontaneous social order can be a very helpful way of rapidly solving problems if enough people cooperate, and our society has amazing reserves of public-spiritedness.

3) In order to permanently fix things, you still need the men with uniforms and tasers on their hips, especially if there are people who will not cooperate with the spontaneous order.

4) Shoving cars around for an hour and a half is an excellent workout.

5) The human body is capable of amazing feats of thermoregulation if properly trained.