Peak Chaos (45% to 55%)

At the midpoint of AV road share, the system no longer risks mere strain, but failure. When the roads contain a complex mix of fully human-driven vehicles, ADAS-supervised vehicles, conditionally automated vehicles, and geo-fenced robotaxis—none dominating—no single behavioural logic will govern traffic.

I think this is not quite right on two counts:

1. The peak is likely to be fairly flat, since the randomness of specific interactions will make 40% saturation feel a lot like 60% saturation. Elsewhere, Andrew says that with fewer than 30% AV’s, it will be “the calm before” the storm. It can’t both be true that 50% is a disaster and 25% is basically fine. All the same types of situations will emerge at both saturation levels, just at different frequencies.
2. I suspect that peak chaos is going to occur at a much lower level of saturation.

Mental modeling

Full disclosure: this post began as an email to Andrew (whom I like and respect!) that I decided to formalize instead.

If I were to start modeling road chaos with the introduction of a new type of driver, I would want to include three key ingredients:

1. How frequent are the opportunities for confusion
2. How quickly do human drivers adjust to machine drivers
3. Whether machine drivers are ‘smoother’ than human drivers in the final equilibrium

Machine drivers have to adjust to human drivers very early in the process or they won’t be allowed to stay on public roads; we can take that as given.

Analogy on two wheels

From long experience as a cyclist, my guess is that the crossover point – where further saturation starts reducing chaos – happens at a pretty low rate of saturation, more like 5% than 50%. When I cycle in suburban areas where cyclists are rare (especially back in the 90s), drivers can be genuinely confused with how to deal with me. They might proceed very slowly behind me, or pass dangerously, or honk.

But at a modest saturation rate (maybe 1%), most drivers adjust and learn how to pass a cyclist, or handle a cyclist passing them in the red light queue. But at the same time, the number of opportunities for error is 50 times higher at 5% cyclists than 0.1% cyclists. Do drivers get 50x better with regular exposure to cyclists? No. No they do not.

In particular, there are a small number of drivers who have no clue how to drive around bikes. As bike saturation rises, that small number is forced by circumstance to acclimate. At the extreme you have the Netherlands, where it’s inconceivable that a driver not know how to operate around bikes. In DC or Boston, the 5% or less bike share causes more disruption than does the 50% bike share (or whatever) in the Netherlands.

Back to robots

To my mind, disruptive human-AV interactions will occur when (a) there’s a bunch of vehicles involved and (b) at least one of them hasn’t adjusted. Given that it’s humans, not machines, who still have to learn to deal with the other, conditions (a) and (b) are going to be satisfied most commonly when there is still a large human majority on the road, but some of those humans haven’t yet learned how to interact with a machine driver. 

Long before we reach 50% saturation, humans will internalize lessons like “if I double-park, AVs will just sit there, cause a huge backup and people will be angry at me” or “AVs are going to follow yield-to-the-right rules strictly”. Once the frequency of un-accultured human drivers starts to shrink to a small minority, then the worst is over. Peak chaos might occur when 5% of vehicles are AVs and 20% of humans are still clueless. These are just vibes numbers, of course, and cluelessness is not a binary variable. But I’ll bet Andrew a beer that the number of human-AV collisions peaks (in a given geography) before AV saturation reaches 25%.

Deeper differences

So far, I’ve written this agnostically, as if humans and machines have random differences and we could swap their places in the model. But it’s not random. Machines are far more predictable and uniform in behavior than humans. So it will be pretty easy to learn their tendencies. 

The part that I don’t know is whether machine drivers will ultimately be smoother than humans. On foot, humans are very good at negotiating each others’ space with instinctual balance and nuanced body language. We’re less fluent in vehicles, but still pretty darn impressive given that we didn’t have any evolutionary hardware for solving spatial problems at 100 kph. Will AVs be able to copy all of that, or develop their own smoothness? Or will they be held back by the necessarily-high degree of conservatism in programming? If AVs are just plain clunky and slow, then city driving will become slower and clunkier as AVs take market share even without any misunderstanding between modes.

The Build Now Act uses an existing stream of federal funding for larger local governments, Community Development Block Grants, to reward localities that improve their net housing growth rate over the course of a decade, and penalize those that worsen their growth rate. The amounts of money in play are modest: the median grant is less than $1 million per year, and the penalty is 10%.

A big part of the legislative process for the Build Now Act was to exempt most places from eligibility, which means they can neither gain nor lose funding.

• Any place too small to receive direct CDBG funding (e.g. cities below 50,000 people) is implicitly ineligible and doesn’t show up on the map below.
• Any place with low housing costs is ineligible.
• Any place with a high rental housing vacancy rate is ineligible. Unfortunately, most of the variation in rental vacancy rates seems to be driven by mis-categorization of short-term rentals. This makes San Francisco ineligible.
• Any place where a federally-declared emergency or disaster, such as a major winter storm, occurred in the past 3 years.
• Any county that “lacks the legal authority to enact or update zoning and permitting ordinances.” It remains wide open to see how HUD will interpret this. In the version mapped here, I made conservative estimates.

Throughout the legislative debate over Build Now, I maintained a dataset that helped me loosely estimate how the law might work in practice. My data are not perfect and have some known errors. And, of course, they are backward-looking. Build Now will not take effect until FY 2029, I believe, at which point the evaluation decade will have flipped and the map of disasters will have shifted.

In the map embedded below (and hosted by Datawrapper here), I show which Entitlement Communities would be ineligible and which would gain or lose funding based on the data I have available. NOTE: The amount of data seems to be straining the Datawrapper server. Reload this page several times and eventually the data will show up.

The tooltips show each community’s key stats, including reasons for ineligibility or funding gain/loss. Here’s the Bay Area, where high rental vacancy rates (i.e. lots of short-term rentals for business travelers) make most municipalities ineligible:

And here’s detail from Utah, where South Jordan would lose funding despite having the 25th-highest housing growth in the country from 2020-2024, because its housing growth rate from 2015-2020 was even higher, so it has worsened its performance.

Don’t take this map so much as an indication of which places will actually gain or lose funding when the law comes into effect. Rather, it’s an indication of what a typical year might look like. The biggest takeaway is that most places are ineligible. The second takeaway should be that the law ought to be amended to better match what proponents say it’s meant to do: provide more federal funding to support places that are growing, even if those are in cheaper parts of the country or experienced a nasty snowstorm 2 years ago.

UPDATE: What would I do instead?

If you think the Build Now Act is unnecessarily complicated and poorly targeted to support or incentivize housing production, then you’re right. How can it be fixed? One way, which I & others proposed several years ago, is to build a complex but capable formula.

The other way, which I’ve come around to, is brutally simple:

• Define eligibility in a way that includes at least 2/3 of jurisdictions. Disaster exemptions should be only for disasters that destroy homes, like the Altadena fires.
• Take 10% of CDBG funding away from every eligible community
• Rebate that funding pot back to every eligible community based on how many net homes they built in the previous year.

More homes, more funding. Any politician can explain that to his constituents.

Another type of restriction that deserves attention is deed restrictions. Although El Paso, Texas, has fairly permissive zoning and new state preemptions that legalize residential in commercial zones and small-lot subdivisions of 5-acre+ parcels, those permissions cannot be used on the huge swath of commercial and residential land restricted by deed.

Alex Hoffman, an urban planner at the City of El Paso, mapped out deed restrictions in the city – yielding a map of rigidities that looks suspiciously like a zoning map. Alex even found 5.5 acres restricted to quadplexes only: no triplexes or five-plexes, thank you very much!

Even in neighborhoods where single-family, multifamily, commercial, and public uses exist alongside each other, most of those uses are locked in by deed.

There’s a range of reform options for deed restrictions, including some combination of the following:

• Cancel parts entirely
• Terminate deed restrictions after some period of time
• Allow owners to “buy out” their deed restrictions

Personally, I favor putting time limits on deed restrictions. This comports with the longstanding jurisprudential “rule against perpetuities”: no generation should be allowed to lock up the future. And unlike entirely canceling or banning deed restrictions, it allows most of their positive features – especially predictability in the development process – to work while eliminating most of their negative features (locking in land uses that have become sub-optimal).

This argument made me wonder: what is actually being built out there? The Census has datasets! If you go here and click the “Housing Units Completed By Purpose and Design” link you will find data going back to the 1970s. Single-family houses peaked in 2006, at 1.6 million. After nosediving during the Great Recession, house completion recovered partially- but even in 2025 just over 1 million were completed, almost a 40 percent drop. About 18 percent of these 1 million are attached, a slight increase from 2006 (when just over 14 percent were).

By contrast, the conventional apartment complex (by which I mean units in buildings with 20 or more units) has recovered from the recession. 173,000 such units were completed in 2006, and 420,000 units were built today- even more than the 1974 level of just over 350,000. (But before you scream “Aha! We are building enough apartments!” remember that this is only a 20 percent increase, in a country where population has increased by about 60 percent).

What about “missing middle” housing? These categories of housing have nosedived over the past 20 years, to an even greater extent than single-family houses: from 41,000 to 16,000 in the 2-4 unit category, from 42,000 to 11,000 in the 5-9 unit category, and from 57,000 to 37,000 in the 10-19 unit category. For all three categories, the 2006 construction level was less than half the 1974 level, so on balance less multifamily housing was built in 2025 than in 1974.*

In particular, condo construction has simply collapsed. Americans completed 192,000 multifamily for-sale units in 1974, 127,000 in 2006, and 23,000 in 2025. By contrast, multifamily rental completions are only about 25 percent below their 1974 level (461,000 in 2025, 596,000 in 1974).

So to sum up: single-family construction, condo construction, and “missing middle” construction have not recovered from the recession. Larger-scale apartment construction has recovered from the recession but is still not at 1970s levels. This suggests to me that there isn’t really enough housing built to satisfy either family demand OR nonfamily demand.

*Some pro-natalists might think that not building apartments is a fine thing, because without it renters will be forced into supposedly more family-friendly single-family or missing middle housing. So let’s look at the five states with over 60 live births per 1000 women age 15-44: Alaska, Kentucky, Nebraska and the Dakotas. If fertility was related to apartment buildings, these would be states where the overwhelming majority of renters lived in houses and duplexes. This is certainly true in Alaska and Kentucky (where only 18 and 22 percent of renter-occupied units are in apartment buildings with over 10 units). But in South Dakota, 40 percent of renter-occupied housing units are in such apartment buildings- more than in Maine, Massachusetts, New Hampshire, Oregon or Vermont (the lowest-fertility states). The comparable numbers for Nebraska and North Dakota are 40 percent and 57 percent. So there isn’t really a strong correlation here.

But a recent paper by some engineering scholars suggests that this argument is oversimplified. They suggest that emissions from “operational impacts” (for example, day-to-day electricity and natural gas use) are more significant than construction impacts. As a result, the article concludes that in Toronto, annual greenhouse gas emissions for new high-rise multifamily buildings are on average 24 percent less than low-rise multifamily buildings.

I note, however, that the study is inconclusive in some ways- most notably because it includes a fairly small number of Toronto buildings, and (perhaps more importantly) because its high-rise category includes all buildings in the sample that are over five stories. Thus, it does not address whether there is a difference between 5-10 story midrises and much larger buildings.*

*Readers might be wondering why the authors did not find enough buildings to make such distinctions. First, the authors chose to focus on post-2009 buildings, thus limiting sample sizes. Second, the authors’ data was based on an earlier study with a limited building sample.

I discovered pretty quickly that the four most unequal large metros (measured by GIni coefficient) are Miami, New York, Los Angeles, and San Francisco: all high-cost metros. So at first glance, it thus appears that there is a strong correlation between inequality and housing prices.

But then I looked at the fifth through tenth most unequal metros, and a different picture emerged. These were, in order of Gini coefficient: Memphis, Houston, Boston, Tampa, Cleveland, and Chicago. All of these except Boston are low and medium cost metros; for example, the median home price in Memphis is under $300,000. And the differences between these metros and the high-cost metros are not enormous: San Francisco’s coefficient is .489 and Memphis’s is .486.

The least unequal metros with over a million people were mostly low and medium-cost metros,* but with a few exceptions: in Washington, Riverside and Denver the median home price was just over $600,000.

So what distinguishes the high-inequality/high-cost metros from the other high-inequality metros? As one might guess from conventional economic theory, income. The highest-cost high-inequality metros (Miami, New York, Los Angeles and San Francisco) have median household incomes over $90,000. So do Washington and Denver (although Riverside barely misses the cut at $87,000). By contrast, Memphis’s median income is just under $65,000, Cleveland’s is just under $68,000 and Houston’s is just under $80,000. In other words, high inequality plus high income = high prices. High inequality + low income = low prices.

*The bottom 10 included: Salt Lake City, Washington, Virginia Beach, Riverside, Grand Rapids, Kansas City, Minneapolis, Indianapolis, Denver, and Raleigh.

But state law is so stacked against housing that even when a locality is eager to allow new homes, it cannot necessarily do so. Residents of Sag Harbor whose anti-growth sentiments had failed to win in the normal political process instead sued over technicalities in the state environmental review, a review process written so broadly that basically any proposal can be subject to litigation if someone wants it badly enough.

Using Sag Harbor as my point of departure, I discussed new research on how New York can re-balance the checks and balances in its permitting process in a New York Daily News op-ed yesterday. Read the whole thing.

• Modern Housing for America, Gail Radford, 1996, (Housing Act of 1934 and Housing Act of 1937)
• The Wagner Housing Act: A Case Study of the Legislative Process, Timothy McDonnell, 1957 (Housing Act of 1937)\
• Housing Reform During the Truman Administration Richard Davies, 1966 (Housing Act of 1949)
• Government and Slum Housing: A Century of Frustration, 1969, Lawrence Friedman (Housing Act of 1954)
• Robert Clifton Weaver and the American City: The Life and Times of An Urban Reformer, Wendell Pritchett, 2008, (Fair Housing Act and HUD Act)
• Building Foundations: Housing and Federal Policy, edited by Denise DiPasquale & Langley C. Keyes, 1990 (Cranston-Gonzalez)
• The Housing Policy Revolution: Networks and Neighborhoods, David Erickson, 2009 (Cranston-Gonzalez)
• Act of Congress: How America’s Essential Institution Works, and How It Doesn’t, Robert Kaiser, 2013 (Dodd-Frank)
• Federal Government & Urban Housing (3rd ed.), R. Allen Hays, 2012 (all)

What’s right with the DSB argument? It does make sense to me that if rents go down far enough, new construction might not be profitable. For example, if construction costs are $200,000 per house and a new house will sell for $150,000, there will not be much new construction. (Even then, there will be some construction because developers might think prices will eventually go back up).

What’s wrong with the DSB argument? First of all, it assumes that rents will go down to some extent due to new construction- which, if true, would be a huge win compared to the status quo of ever-escalating rents. Let us suppose that in neighborhood X, rents have escalated to $2000 per month, but that developers will stop building if rents go below $1500 per month. First, this means that rents will go down 25 percent until developers stop building- which would be a very good thing.

Second, if developers will stop building when rents go below $1500, this means that they will start building when rents go above $1500. So this means that, in the absence of other factors limiting construction or raising construction costs,* rents will never go back up again. As soon as rents go up, developers will start building again to curb the rent increase.

By contrast, under the “business as usual” scenario (in which government limits construction to cater to the desires of local homeowners) rents keep going up and up forever. So even if the “developers will stop building” argument is correct, zoning deregulation will cause a decline in rents that will be far better than business as usual policies. The only unknown fact is how significant the rent decrease will be: because zoning exists everywhere (at least in the United States) we do not have any real-world examples.

To draw an analogy to climate policy: let us suppose that some policy would magically freeze carbon emissions at slightly below current levels. Would this be ideal? No. Would this be better than business as usual? Of course!

*Obviously I am oversimplifying here because often such factors exist: for example, Kevin Erdmann has written extensively about the impact of tight lending policies on housing markets.

I would have guessed that the most mixed-use American metros were old ones, with many neighborhoods built before zoning and widespread car ownership: New York, Philadelphia, Boston, Chicago.

But no: Austin is the US champ, followed by San Jose and Miami. At the opposite end, Chicagoland is off the charts in its degree of use segregation. There doesn’t seem to be a rhyme or reason: the four Texas cities are scattered across the chart. The two halves of the Bay Area (SF and San Jose) are on opposite ends, and not in the way I’d have expected. The Midwest is represented all over the chart.

Of course, having commercial and residential space in the same block group doesn’t mean that the two are integrated in a traditional neighborhood form.

Do Ho et al find this result because of patterns in how block groups are constructed? Or because some cities have much bigger downtown peaks? The latter appears to be part of the explanation for Chicago and Denver. Looking under the hook, the coefficient of variation of commercial floorspace per block group is 16 in Chicago, about 5x more than the median city, despite those two cities having the highest commercial space share. In the index, that makes a block group with 1000 residences and 10 shops look much more segregated if it’s in Chicago than if it’s in, say, Milwaukee.

The results don’t appear to come from some cities having bigger block groups than others, or larger commercial shares overall (except Chicago and Denver).

I would welcome complimentary approaches to understanding this question. Are my intuitive guesses ultimately more reliable? Or is the data-speaks approach telling us something important about mixed uses in the US?

(As for the rest of the paper, it’s very creative in squeezing a lot of results from limited data. But I’m skeptical of the core findings, which are based on the assumption that the crowding of workers per unit of commercial floorspace and residents per unit of residential floorspace are accurate proxies for prices.)