Self-driving cars as a target for philanthropy

by paulfchristiano

The cost of driving

Driving causes a lot of damage; in the US alone, each year there are about 2 million accidents, around 30,000 people die, the CDC estimates $100 billion of costs from injuries and lost productivity, and it looks like well over 50 billion hours are spent driving. (Interesting aside: these quantities would all be equal at $2 / hour and $3 million / life, which I believe are relatively close to the average american’s actual reservation prices.)

Another way to estimate the size of losses from auto collisions is to look at the auto insurance industry, which has revenues of around $180B, of which more than 2/3 goes to claims. Claims and CDC estimates seem likely to be (highly correlated) overestimates for damages, but I think the ballpark figure of $100B in damages, 30,000 lives, and 50 billion hours is probably about right. 

Tech advances might significantly reduce accident rates and significantly increase the value of hours spent driving, eventually by completely automating driving. Automating driving looks likely to have a wide array of positive impacts, via reducing congestion, facilitating driverless vehicles, increasing integration between vehicles and electronic systems, etc. I think that neglecting most of these sources of value and being fairly conservative, the value of self-driving cars to American consumers is around $100B. Including other sources of value, I would believe that the value is around an order of magnitude higher at $1T, though this is much more speculative and harder to justify (ETA: Carl points out that Brad Templeton has written at some length about the benefits of self-driving cars and the social & technological path to them.)

If we do the accounting at $15 / hour for drivers’ time (as some kibitzers have suggested), we would get an estimate an order of magnitude higher. I would be skeptical of that analysis—it seems clear that few Americans would pay $15 / hour to avoid driving, even if society were to adjust in the long run to use those extra hours more efficiently.

In fact driverless cars will probably not realize benefits at that scale for a few decades, and the number of deaths / year in the US is declining by about 5% a year. However, the value of auto insurance claims is rising faster than inflation, and the number of hours spent driving is rising. So I feel comfortable assessing the value of self-driving cars at $100B / year when they arrive, in the US alone.

The situation worldwide is a bit harder to assess (since I can’t find data quite as trivially, and conditions are changing more unpredictably over the coming decades). The US represents about 1/4 of the annual auto accidents, and 1/5 of the total world product. Both of those fractions will decrease substantially in the coming years as the rest of the world grows. I think it is quite conservative to adjust the value of self-driving cars up by a factor of 2 to account for international adoption, for a total of $200B / year.

The cost of not driving

To assess the value of self-driving cars as a technology, we need to also understand the extra cost of building the cars themselves. Some random startup founder claims that the sensors in modern driverless cars probably cost more than $250k—at that price point, even for $200B it wouldn’t be close to worth it to deploy driverless cars broadly. Most possible costs for self-driving cars fall into two regimes, at least within the US—either the price is low enough that they will be broadly adopted and the cost is negligible relative to $100B, or the cost is high enough that they won’t be broadly adopted and/or the costs exceed the $100B. Worldwide the situation is a bit more complicated, given large gaps in wealth.

I will be very surprised if the cost of self-driving cars stays high as the underlying technology improves, and would be willing to bet (literally, at 5 : 1 odds) that costs will eventually come down far enough for broad adoption. I don’t have data to support this claim, but I haven’t yet encountered anyone who argued the other side. It is often claimed that adoption will be uneven until costs come down, but this is a very different issue (see below) which affects timelines but not the value of mature technology for self-driving cars.

Is it worth considering intervening? 

It seems like self-driving cars will create a lot of value, but that those benefits will not be concentrated. It will be hard to pass much of the real value created by driverless cars onto customers as increased prices, primarily because of price competition amongst auto manufacturers. (I don’t think externalities are likely to be a huge deal here: I think that auto insurance price adjustments may suffice to handle the negative externalities of accidents, and perhaps overcompensate for them. There are other possible externalities, involving the organization of transportation systems, but these seem more speculative.)

This problem is particularly severe when we consider the legal and social challenges that may modestly (or significantly) slow down the adoption of driverless cars. Someone who works to resolve these challenges does not expect to get much competitive advantage thereby. For the most part, auto makers should view “self-driving” as another feature in a continuing struggle over market share (demand for cars is not very elastic to whether cars drive themselves). Making self-driving cars come faster across the board changes the importance of that feature, but doesn’t have any clearly positive effect for any particular manufacturer (unless one has a clear, large lead in that area). Overall I would expect very significant underinvestment (relative to socially efficient levels) in this area, at least without altruistic funding.

Markets also have steeper discount rates than I would like, and so are likely to base investment in a technology on its potential over the next decade. In the case of self-driving cars, estimating the impact over the next decade significantly understates the real social value.

Accelerating adoption of self-driving cars therefore strikes me as an interesting philanthropic project. This could be done by subsidizing development of pieces of the technology, lobbying, advertising and outreach, or something else. I doubt this is the best possible project even given unusually US-centric values, but it still seems worth looking into in order to better understand the space of high-impact interventions.

Looking into self-driving cars may also help sharpen our intuitions, or at least my intuitions, about interventions that sound good in principle. If many superficially attractive interventions look less attractive upon closer inspection, than I should be somewhat more suspicious of hard-to-inspect, apparently-good ideas. (Many schemes to reduce global catastrophic risk fall into this category.) Of course it would be more representative to actually pursue some interventions that looked like a good idea. But comparing “how things look at face value” to “how things look after a bit of thought” is a much cheaper proxy, and is reasonably representative for other domains where available evidence is scant and “face value” plays a larger role in final assessments.

Finally, I think that trying to support self-driving cars is similar to other, more complicated tech-boosting interventions. Some of those interventions are likely to be strong contenders for my intervention of choice. So understanding this intervention better makes sense from that perspective.

I’m keen to get info from or talk with people who better understand the state of affairs regarding self-driving cars, if any of you are interested in talking with me. I don’t really understand the situation, and didn’t put nearly enough effort into this blog post to remedy that.

Accelerating technological progress

For most of history, the developments necessary to automate driving were scattered across a wide range of fields, so that it would be hard to bring self-driving cars closer without boosting society quite broadly. In 1970 there was very little investment in self-driving cars per se; doubling that investment might have increased the apparent rate of progress towards self-driving cars,  but it would have contributed very little to actual progress, i.e. it would have had very little effect on the actual time until self-driving cars are widely adopted. Instead, the adoption of self-driving cars has been driven by progress in other areas (robotics, sensors, machine learning, semiconductor manufacturing, population growth, the industrialization of china, etc.). It’s not clear that you could have accelerated real progress on self-driving cars by a year, even if you had been willing to invest 10% of US GDP into the problem. I bet that you could, but I’m not too confident.

(When we look at technologies that are further out, the picture gets even murkier. If you are a philanthropist in 1600 and want to get humans to the stars as quickly as possible, your best bet ex ante would have been to speed up the industrial revolution. Even though few others were trying to get to the stars as fast as possible, accelerating our journey to the moon by a year would have been a very expensive project unless you happened to be blessed with unusually good insight about where society was going as a whole. Insight about the stars wouldn’t help.)

However, today it looks like the remaining obstacles for driverless cars are a combination of domain-specific technological problems and domain-specific social and legal problems. If this is true, it means that there is an unusually strong relationship between apparent progress in 2013 towards full adoption of driverless cars, and real progress towards the adoption of driverless cars.

(I’m a bit uncertain about the self-contained claim. The main way it could fail is if broader technological progress is needed to reduce the price of the sensors in self-driving cars to a reasonable price point, rather than design improvements in the cars themselves being enough to carry it. I expect the latter, but don’t know the details. I don’t think a similar state of affairs is plausible with respect to software, computing hardware, or actuators.)

This means that in order to evaluate the impact of apparent progress on driverless cars, we can compare the endpoint of their progression with the status quo. If you could purchase 1 year of progress towards driverless cars, it would have the effect of bringing about each level of future capability 1 year earlier. It would eliminate 1 year of driving using current automation, and create 1 year of driving using well-developed and deployed driverless cars. Above, I estimated the value of this change at $200B.

This holds however poorly understood the mechanisms underlying progress are, and however gradually progress on driverless cars translates into material improvements in quality of life. Of course, if the current mechanisms of progress are poorly understood, it may be hard to push on them, but as long as it is possible to accelerate progress locally then the global impact follows.

For example, suppose the ultimate bottleneck to realizing the gains from driverless cars is consumers gradually becoming more familiar with them over a period of decades. Then it might look like technological progress today is just a small piece of the puzzle, but if that is all that is happening today, then pushing technology a year ahead will get self-driving cars to market a year earlier, which means getting started on that process of familiarization a year earlier, which means that self-driving cars will be broadly adopted a year earlier.

If instead the ultimate driver is (e.g.) huge industry investment, it suffices to ask what processes will eventually cause that investment to occur. If that investment is underway now, then to speed up adoption you would need to make a comparably huge investment. However, if that investment is not underway now, then some other process must run its course before that investment begins. Maybe this is the existence of an early market for self-driving cars, maybe it is early tech developments, maybe it is organizational change within auto manufacturers, whatever. It must be some process that is happening today, and if nothing big and expensive is happening today, then it may not require big and expensive investment to speed up that process.

The point is: regardless of whatever processes will ultimately be important for bringing self-driving cars about, if you can understand what is changing today, then you can speed up the entire timeline. In reality we will be uncertain about what the important processes are today. To be sure we’ve had an impact, we need to speed up all of the processes which might be the relevant ones. If we can identify a process which constitutes 50% of our subjective expectation of progress, then all else equal we can get 50% of the total benefits by accelerating that process.

The unsettling possibility, for the would-be technology-pusher, is that the relevant processes are distributed broadly over society. As I remarked above, this doesn’t look very likely to me. It looks like we can identify a relatively small set of processes that bear a relatively large share of the responsibility for bringing self-driving cars closer. Of course, to the extent that we trust our analysis that self-driving cars are a big deal, we should be more skeptical of any story on which they are being bottlenecked by relatively overlooked processes.

Accelerating this particular technological progress

It’s not obvious that it’s possible to accelerate progress on driverless cars, even if one assigns a huge value to doing so.

For reference: I went into this analysis expecting that google spent O($50 million) on self-driving cars in total, and that Google was responsible for a significant fraction of progress in terms of both technology and social infrastructure. Based on those estimates, I thought that this was likely to be a good intervention. I think my estimates match the empirical situation reasonably well, but my error bars haven’t much decreased after looking around a bit. It’s hard for me to tell what’s going on.

Current work on technology:

Some relevant numbers:

  1. The R&D budget of google is $11B, of which it seems a very small fraction is invested in self-driving cars. Poorly sourced internet reports put the number of engineers working on the project as <15 as late as 2010, which is continuous with the academic predecessors but suggests to me that Google’s development budget is unlikely to exceed $50 million today. It seems like it may well be much lower than this, but it’s also plausible that I’m overlooking something important and it is much higher.
  2. The R&D budget of the auto industry seems to be about $70B +- $15B (an extrapolation from the info in this report) of which a significant fraction seems to be spent on electronics which are directly relevant to self-driving cars. This budget is divided about 8 ways, and there is much replication of effort.

There is a reasonably large body of academic work on self-driving cars, which feeds directly into Google’s project but which seems to be surprisingly unrelated to developments in the auto industry. Meanwhile most auto manufacturers have some offerings that could be branded a “self-driving car,” (see here for a very haphazard assortment, and you can assemble your own inventory pretty easily). Self-driving under normal conditions on a highway seems to be essentially resolved as a technical problem, and most development effort appears to go into driving on surface streets and robust responses to obstructions etc.

As far as I can tell, Google is the only player that currently has a feasible solution to the general problem. This (somewhat surprising, in light of big funding differences) asymmetry is supported by the observation that Google is the only player that appears to be regularly test-driving their cars on public roadways under normal conditions.

It seems very likely that marketable self-driving cars will make use of more robust and mature technology developed by automobile manufacturers, and also reasonably likely that the software being developed by Google will play an important role in making early models roadworthy. It is hard to know how to distribute responsibility for progress between the technological developments at Google and elsewhere. It seems like Google’s contribution is probably between 10 and 50% of the relevant technological development. (That is, I think that if I conducted a more extensive inquiry my subjective expectation of Google’s contribution would be between 10 and 50%.

Current work on social issues:

As best I can tell, google is responsible for most lobbying directly relevant to driverless cars, though it wouldn’t be too surprising if automakers contributed much more towards shaping indirectly relevant legislation. According to the WSJ, Google spent only $140k lobbying to get driverless cars approved in california; their total budget for federal lobbying is about $20M, though the vast majority concerns wrangling over anti-competitive practices. I would be quite surprised if their lobbying efforts for self-driving cars exceeded $2M (based on a very weak understanding of how lobbying works and standard prices, see e.g. here).

According to opensecrets.org, the automobile industry in general spends about $60M lobbying / year. Most of this lobbying appears to be unrelated to self-driving cars, though a significant fraction may be related to liability issues which have significant influence on the overall course of self-driving cars.

As best I can tell, there is very little investment in advertising or press management for self-driving cars. It’s not clear whether these are activities which could have any impact; it seems likely that even if these activities could have an impact, it would not be profitable for any of the potential vendors of self-driving cars.

It is hard to find much evidence for other efforts aimed at building social infrastructure for self-driving cars. This is weak evidence that such efforts don’t draw large amounts of money. Again, it’s not exactly clear what they would consist of (but see below).

What to do?

As I’ve probably made clear by now, I don’t have a deep understanding of this area. The first step of any effort to promote self-driving cars would be discussion with people who understand the issues better and can see where money could be productively applied. However, there are some natural interventions to consider, so I can at least list those.

Recall that my estimate for the value of self-driving cars was $200B / year, several times the entire R&D budget of the automobile industry. My baseline is purchasing acceleration at a rate of just over $500M / day, and any multiplier over that is gravy. I suspect it is possible to accelerate self-driving cars about 10-1000x that efficiently.

1. Fund existing research efforts

R&D departments need to decide whether to invest marginally more money in their projects. For projects that have large social value, we expect that spending is too low, and that marginal increases will speed up progress. We could try to estimate returns to this tech development to estimate how efficiently marginal dollars accelerate progress. But short of that, it is probably a sensible heuristic that the value of marginal investment in research is not too much lower than the value of average investment—perhaps 1 order of magnitude less efficient, but quite possibly not even that bad in the medium run, and probably not 2 orders of magnitude less efficient (I made this up based on a passing familiarity with economics and industrial organization).

If we think that Google makes up 1/4 of the relevant tech development and Google’s research budget is $50M / year, a factor of 10 efficiency loss on the margin would give $2 B / year of technological progress. I think this is a relatively conservative estimate, although further inquiry would quickly shed light on that question.

Funding R&D work elsewhere may be more expensive. However, if Google’s contributions are unusually significant primarily because of Google’s different priorities, then it may be easier to find programs at automobile manufacturers which are unusually cost-effective for accelerating long-run progress.

I have no idea how possible it is to set up the relevant partnerships between philanthropists and R&D departments. To the extent that there are strong interventions that would take the form of such partnerships, and such partnerships are not currently possible, understanding this issue and building relevant infrastructure may be a good use of resources.

2. Increase scope of testing

Right now there seem to be dozens of self-driving cars on the road, and they are legal in 3 states. These numbers could be scaled up significantly. More aggressive lobbying could lead to broader and more flexible legal arrangements for self-driving cars. Increased investment in test fleets could allow more testing of more different ideas in a wider variety of contexts. This intervention has the advantage of being more parallelizable, and therefore less likely to run into diminishing returns—Google will naturally start the highest-leverage testing soonest and leave less attractive opportunities for later, but at least separate efforts can run without stepping on each other’s toes.

3. Lobbying / creating legal infrastructure

In absolute terms, little money seems to go into lobbying or drafting acceptable legislation that can accommodate the deployment of self-driving cars (and improve the liability situation to make them more tractable, etc. etc.) It looks like the legal issues have some probability of holding things up at the moment, and liability concerns seem to be serious amongst automobile manufacturers. Diminishing marginal returns to lobbying seem potentially more severe than those to technological development; nevertheless, because automobiles are largely regulated at the state level it is possible to conduct many independent lobbying efforts in parallel, and successful policies in one state may be copied elsewhere. (Not even considering international politics.)

It seems that an investment of $10M would probably much more than the amount of relevance-adjusted money spent on lobbying. It is not clear what effect this would have, partly because it is unclear what the effect of more lobbying would be on the legal status of self-driving cars, and partly because it is not clear to what extent the legal status of self-driving cars influences their adoption.

4. Advertising and public relations

It seems that the adoption of self-driving cars may depend in large part on public perception of their usefulness, safety, etc. and in particular on minimizing the risk of significant negative public perception. $1B is enough to purchase very large advertising campaigns (this is the amount that Obama and Romney each spent in the 2012 election). It is not clear what effect such spending would have on the public perception of self-driving cars or what effect that perception has on adoption. But to the extent that positive changes in perception are responsible for eventual perception, it seems likely that such changes can be accelerated with large advertising budgets.

Conclusion

There are a handful of plausible approaches to hastening the arrival of self-driving cars. Most of these approaches would not be pursued by manufacturers even if they were efficient. There is uncertainty about which of these interventions would have a real effect on the timelines for self-driving cars, but even given complete uncertainty we could apparently have a large impact by investing in one of them at random (or all of them).

Based on the scale of current activities and the factors which plausibly contribute to adoption of self-driving cars, it seems to me that the efficiency of these interventions are roughly in the regime of $10 – $1000 of economic value created in a few decades for each $1 invested today. I am not convinced that creating economic value in the US is a high-priority use of funds, but for donors who disagree this seems worth considering.

I am interested to hear any thoughts and especially criticisms on this (admittedly very rough) analysis.

Advertisements