In 2036 the State of California granted an incentive to drivers of electric trucks. This was just one act in a decades-long attempt to remove low-efficiency, high-emission vehicles from the roads, but because the situation was increasingly dire, the grant was unprecedented in its generosity. A confluence of political promises and lobbying efforts, the Clean Miles Rebate Project (CMRP) came to include two key terms:
- Rebates would be issued for the cost of charging zero-emission commercial trucks, adjusted for Provable Miles Driven (PMD).
- Light-duty trucks would be eligible for these rebates.
The requirement that mileage be provable was intended to prevent obvious abuses of the system. Hypothetically, a truck owner could use its vast battery as an electrical reservoir, let the vehicle sit in the driveway undriven, run their home from its charge, and be reimbursed by the state. This behavior was portrayed as a form of fraud, even though Pacific Gas and Electric was in open support of it as a method of reducing load on the state’s power grid. Still, a measure targeting vehicles was not meant to pay for energy infrastructure, so when auto manufacturers championed PMD-tracking technology they were commended for voluntarily cooperating with regulators.
In reality, proving mileage in commercial vehicles was a long-solved problem. Telematic devices sent robust engine diagnostics with sub-second update frequency. Odometer readings collated with GPS tracking, cell tower pings, and local network scanning provided an irrefutable log of every inch of vehicle motion. For commercial fleets, PMD reporting was trivial. It was mandatory. That market was saturated.
This technology had little demand in the consumer market though. Drivers never had a reason to match every mile traveled to the exact amp-hour listed on their energy bills. CMRP gave them a reason. This was why Daimler Truck AG advocated for the measure. Having acquired several telematics companies in the 2020’s, Daimler had been selling packaged PMD systems in its heavy-duty lines for years. In 2035 the company extended this feature to the series of light-duty trucks they were releasing under their Mercedes-Benz brand. The X-Class Marathoner Edition provided the feeling of piloting a Freightliner to drivers on their trips to wholesale big-box stores. The gimmick was insufficient to compete with American pickup brands whose electric offerings were inferior but popular. Daimler had already seen some success in promoting PMD to european regulators, so they estimated a push in California would provide a return on investment.
According to a report published in 2040, the returns on light-duty PMD would have been modest for Daimler (and scarcely break-even for efficiency-minded drivers) had they not integrated their Remote Pilot System into those same models. Unlike telematic devices, semi-autonomous driving technology had languished for over a decade. Consumer excitement had subsided after many promises and failures. Regulators were even less excited. Systems like Daimler’s were only approved for use in controlled lots where they aided in parking and reshuffling trucks for loading. Consumer vehicles faced less targeted restrictions, but criminal liability for negligent use of driving assists had the same chilling effect. This was the status quo until 2036.
Daisy Lane began as a research project within Amazon Logistics whose mission statement was to “optimize contractor welfare through deployed application of vehicle guidance enhancements.” Reporting from late 2032 suggests the company was spun out because it proved more useful as a sponge for public infrastructure spending, and allegedly to divert attention from the intrastate high-speed rail project that was gaining publicity at the time. Daisy Lane made its first PR splash when it committed to covering the entirety of El Camino Real with its devices. By the time of the company’s liquidation in 2040 the 600-mile long California highway, which stretches from San Francisco to San Diego, would see only 84 miles covered. They did manage to cross 600 miles by striking a deal with municipalities in the South Bay: the only region to be fully certified for remote driving by any government was a network of roads connecting San Jose’s suburbs to either side of the Dumbarton Bridge.
The inclusion of the bridge in the core network would prove to be a fateful decision. Ironically, Daisy Lane considered the bridge to be an ideal environment for their Autonomy Enhancing Sensors (AES). The AES devices, nicknamed “stems” by customers, were deployed at quarter-mile intervals – a relatively low density for the system, but ample coverage for a straight road with no intersections. The stems provided constant readings of the road conditions and live video feeds for remote analysis. Autonomous vehicles that subscribed to Daisy Lane’s service could supplement their proximity scans with a stream of data from the road lying ahead, beyond the vehicle’s range of vision. Daimler’s RPS did exactly this. For all of 2035, though, this investment in autonomous assists failed to respark consumer interest. The only real novelty was the multi-city legal approval Daisy Lane had managed to wrangle, but even this was underwhelming for autonomy enthusiasts. The need for supplemental devices was further evidence the industry was in a slump.
Still, a small community managed to form around the technology. Owners of AES-compatible vehicles shared stories, memes, and financial speculation in a Discord server named “Stems” after the devices. The last of these topics proved to be the most enduring. Entire channels were dedicated to optimizing gig economy jobs using semi-autonomous vehicles. Expanding the capacity of a delivery contract by adding a follow-on vehicle was seen as an essential upgrade for those who could afford the investment. As users became more skilled they were able to increase their margins by adding more and more simultaneous routes. “Remote piloting” required an A1 commercial license, but those were easy to come by since an online certification program had launched the same year Daisy Lane incorporated.
One of the channels in the Stems server was #mileage. This was mainly a place to post landmark updates – “crossed 50k fully remote” – but it was also full of idle chatter about vehicle efficiency. On January 30th, 2036, a user posted a bragging explanation of how their truck, the Mercedes-Benz Marathoner, was about to be the most economic vehicle on the market. Their theory was that CMRP, which was set to go into effect at the end of February, would offset the cost of charging enough to compensate for the Marathoner’s excessive bulk. They also proposed that the rebates would be processed quickly because of Daimler’s built-in PMD system and commitments by both FasTrak (the region’s toll-taker) and PG&E to integrate with the high-tech initiative.
The community was skeptical. The Marathoner was often mocked for being a pickup truck that was only driven by people who never picked anything up. Still, there was enough interest in cost-reduction strategies that many encouraged the original user to give it a try and post the results. They said they would and the channel forgot about them for the next three months.
On May 21st, 2036, the driver of the Marathoner announced they had purchased a second brand-new Marathoner and a third lightly-used model. This announcement was made in the #new-wheels channel and was immediately followed by a post in #mileage with the results of their experiment: the rebates were profitable. They included extra emphasis on “profitable” and explained in detail how they had earned back more money from the program than they had spent charging their vehicles. They supported their claims with screenshots of energy bills, their FasTrak account, and Daimler’s PMD dashboard which appeared to show accurate mileage for the roads traveled. They were able to reproduce this effect with the second vehicle and expected to do the same with the third. They stated the difference was enough to justify the vehicle’s premium price, which they estimated would pay for itself in under five years.
These posts were also met with doubts, but it was enough to inspire other Marathon owners to connect their accounts and test out the rebate program. By the middle of June there were enough testimonials that #mileage no longer included conversations about anything else. The only remaining doubts were about whether the profits were the result of a glitch (most agreed it was) and whether it was legal to take advantage of the apparent mistake. Legal or not, it seemed likely the state, county, or energy company would demand the money back once they identified the problem. Boosters of the scheme expressed certainty that the only remediation would be to update CMRP itself, which was politically fraught due to its popularity with companies operating heavy-duty fleets. There was also the suggestion that maybe no one would notice because of how many internet services had to communicate to make the program function. This sounded ridiculous to some, but to others it sent a different message: the sooner you start driving, the more you’ll earn.
By late summer, Stems was a server with only one topic: how to make money from CMRP. Threads where users discussed delivery optimization fell dormant as everyone discovered it was easier and (potentially) more profitable to run a fleet of Marathoners through Daisy Lane routes all day, every day, without stopping for pickup or dropoff. In #new-wheels users were tracking Mercedes-Benz imports to states with the lowest sales tax on light-duty trucks. Groups began to pool their money to purchase shares of vehicles. Those with less capital – often users who were previously relying on gig work – pledged their time instead of their money. A single pilot could only monitor about ten trucks during rush hour, but at 3 a.m. they could easily run fifty.
Every new truck brought more data that could be used to find the optimal route for “mileage farming” – driving empty vehicles in circles to earn CMRP credits. During the initial wave of excitement users shared their data openly, forming a server-wide collaboration. There were debates about how many times you could loop the same neighborhood before the stream of identical trucks would draw suspicion. Some operators farmed as far from home as possible while others thought keeping the vehicles local would bolster a hypothetical legal defense. The result was that there were some roads where flocks of Marathoners were common, but their distribution throughout the South Bay was broad enough that they did not draw much attention.
The community changed again in the fall when Stems faced three simultaneous crises:
- Schools returned to session, expanding the hours roads were congested by human-transport vehicles.
- Purchases of AES-compatible vehicles reached an all-time high, including models whose mileage-based earning potential had not been validated.
- A double-refund bug was discovered in the Dumbarton Bridge toll system.
The last of these should have been a boon for mileage farmers, but the bridge was already a contentious route. Under a measure that predated CMRP, tolls for electric vehicles were refunded during off-peak hours. Farmers avoided commuter traffic anyway, so the refund turned the bridge into a long, steady road with total AES coverage. It was a popular route for pilots of large fleets, but some Stems members were concerned that overuse of the toll rebate system would draw attention from regulators. If an audit revealed how many Marathoners were looping the South Bay, the whole scheme might be shut down. It was a user from the anti-bridge faction who leaked the details of the bug to the whole server.
In a private channel, a small group of pro-bridge fleet operators had been discussing the optimal speed at which to pass through the radio toll collectors. The consensus was that fifteen miles-per-hour above the posted limit could get a truck through the slow section quickly without it being cited for a speeding violation. Into this conversation stepped a newer fleet owner who was trying to understand why they seemed to be earning more money when their vehicles had been forced to drive slower. The group dug into these low-mileage journeys looking for the user’s error, but instead they found a new exploit: driving through the toll booth at under five miles per hour could trigger the radio transponder twice for a single vehicle. This should double-charge for a single journey, but the software must have accounted for this error, as the vehicles were only billed once. The system that tracked rebates, however, did not share this double-counting prevention system. The end result was that a slow-moving Marathoner could spend $8 to get refunded $16, a profit that was orders of magnitude greater than looping any other route.
The anti-bridge faction argued that this exploit was so blatant it was guaranteed to trigger an audit. They urged the Stems community to move all of their vehicles away from Dumbarton routes. This was not the message that was received by AES speculators who had already been looking for quicker means to pay down their debts. The leak was posted in #mileage on October 7th, 2036, at 1:03 p.m. Pacific Time. Within two hours every active user was arguing on every channel about whether to ban the exploit or whether it was simply the “new meta” that every fleet operator had to incorporate into their business. Some questioned whether the exploit was even real, but their doubts were drowned out by those who boasted about getting in their laps of the bridge before peak toll hours went into effect at three.
Anxious hours passed while Stems waited for the bridge to switch out of human-transport mode. Conspicuously absent from any public channels was chatter about the traffic conditions near the bridge; no one wanted to admit they had cars waiting on the east side, ready to swarm the tollbooth if everyone else did. Dash-cam footage made public during news coverage and court cases after the incident confirms reports from human drivers who approached Dumbarton from Thornton Avenue that afternoon: waves of identical Mercedes-Benz light-duty trucks were mounting in the marshlands. Fortunately, the elementary school on Thornton was out of session before the tide turned.
Instead of discussing the situation on the ground, Stems was embroiled in the backlash against those who were against the exploit. The popular opinion was that anti-bridgers wanted the route for themselves so they could farm loops without competition. This claim inspired more pilots to send their vehicles to catch anti-bridgers in the act, which led other pilots to send in the rest of their fleets to get in line before it was too late. In interviews after the fact, even the original leaker would admit to dispatching some of their own vehicles out of fear the Dumbarton bug would lead to the end of CMRP farming and this would be the last chance to limit their losses.
The number of vehicles using the Dumbarton Bridge had been increasing since its repairs in the late 2020’s. In 2035 – after Daisy Lane had installed its remote driving assists but before CMRP went into effect – the median number of westbound crossings on weekdays was 48,000. According to toll data, on that Tuesday in question in 2036 crossings had been slightly below average during rush hour, likely due to automated congestion of the eastward onramps. In the nine hours after peak tolls ended Tuesday evening and resumed Wednesday morning the bridge had roughly 77,000 westbound crossings, with an equal number of return trips made by vehicles that turned right around on the other side to attempt another chance at the rebate.
Estimates suggest over a thousand semi-autonomous trucks filled all six lanes at all times while maintaining speeds over sixty miles-per-hour. While there were constant fender-benders around the toll booth slow-zone, there were no serious collisions until 12:15 a.m. when a pilot crushed a smaller vehicle between two of their trucks. Stems was briefly aflame with accusations the pilot had targeted the non-Marathoner intentionally, but evidence suggests the pilot was simply exhausted. The community was also aware the Dumbarton situation was unprecedented and, in their excitement, were much more focused on reopening the lane for traffic. Using the coordination of several fleets (and the fact none of the vehicles held passengers) they were able to push the inoperable car to the end of the bridge and off into the marshy shoulder. This procedure would be repeated eight times before sunrise, though municipal workers reported over a hundred vehicles in various states of disrepair had to be disposed of after the incident.
Residents of the towns nearest the bridge’s ends had been calling the police constantly since the initial wave on Tuesday. Most departments asked for assistance from the California Highway Patrol, but the CHP was busy splitting its attention between blocking lanes to stem the tide and trying to figure out which traffic violation the empty vehicles were committing. Patrol cars that blocked off routes through small neighborhoods kept pushing the problem around; the Marathoners would reroute in an orderly manner, flood another neighborhood, and return when the patrol car left. Some departments have claimed the CHP made the problem worse by diverting trucks away from the interstate and onto local roads.
Those who lived off of Thornton Avenue had seen the worst the earliest. Citizens volunteered their cars to blockade the thoroughfare, but they realized quickly that shifting the traffic to the side streets was much more dangerous. Instead they ceded the avenue to the trucks and dedicated their resources to cutting off every shortcut the automated route-finders would be tempted to take. A night of trial and error led to the discovery that the trucks would not go down smaller lanes that were not covered by Daisy Lane devices. The stems that had previously faded into the background – matte green poles branching off of other street poles – became weeds to cut down. For the early hours of Wednesday morning, Thornton was a tidy line of automated traffic.
In the Stems server no one had been watching Thornton Avenue. The mission of clearing busted vehicles from the bridge used up all of their attention as commuter hours approached: the number of accidents increased as pilots became more eager to complete laps before the toll changed. Farmers were happy to let their software handle the job of looping back and getting in line to approach the bridge again. The quickest turnarounds were all clogged with vehicles doing the same thing, so the software had grown more adventurous. By destroying the right stems, the residents had managed to tame the coiling beast. This worked because the software preferred routes that had live data feeds. Street-based cameras could shave milliseconds off of pedestrian detection time, which could be difference between life and death for the victim of a six-ton Marathoner. Thornton had carved out safe zones, but few other neighborhoods were prepared when the tide shifted again.
At exactly 4:59 a.m. hundreds of pilots sent the same instruction to thousands of trucks: leave the bridge. Some were sent back to their charging stations, others were put into loops that would keep them nearby for when commuter hours ended. This was when the traffic problems became truly dangerous. During the night, alternatives to the bridge were easily accessible to the few in-vehicle human drivers traversing the South Bay. On Wednesday morning, however, congestion throughout the region surpassed its regular peak two hours ahead of schedule and remained at critical levels not seen since the earthquake of 2027. Semi-autonomous pilots (already fatigued from a night of farming) made numerous errors, but these conditions were even more perilous for regular people: insurers reported that, when medical expenses are taken into account, for every dollar of damage claimed by pilots $1,120.00 was issued to human passengers.
Digging into the Stems message logs reveals the unified effort for managing accidents on the bridge did not carry over to commuter hours. The consensus was that it was impossible to account for the complexity of in-vehicle driver behavior, so only the most micromanaging (and least tired) pilots attempted to update their strategies in real time. By 7 a.m. a majority of fleets had switched into fully autonomous operating mode with algorithms that were opportunistic – “attempt to avoid congestion” – while relying on AES whenever possible. Pilots would later claim in court that they were unaware of the gaps in AES coverage that local residents had created with their “sabotage campaign.” This argument was often used to reach settlements in civil suits, but was never successful against charges of criminal negligence. The precedent had been established that vehicle manufacturers were not accountable for improper algorithm usage, and Daisy Lane devices were only supplements to those systems. Still, relatively few criminal cases were pursued, and all of those were for pedestrian fatalities.
With the algorithmic shift, the residents of Thornton Avenue found their blockades were suddenly flanked by vehicles attempting to cut through their neighborhood in the opposite direction. They rallied quickly, adjusting their parked cars and disabling more stems until the (almost fully autonomous) Marathoners were forced back into containment. Most of the coordination of this effort was carried out on the ground by word of mouth, but a handful of social media posts had been boosted by media outlets covering the traffic debacle. The image that achieved the most reach featured a husband and wife in matching pajamas gripping either handle of a huge pair of bolt cutters. When residents of other neighborhoods woke up to find trucks flocking through their streets, this image was probably the first solution they saw in their search results.
One of the assurances Daisy Lane had given to municipalities was that special accommodation would be provided for school zones. Devices in these areas emitted signals that put hard caps on speed for vehicles in their proximity. To provide better visibility, stems were deployed at double the density used on other roads. This provided a marginal improvement to safety statistics, but it improved perception of the devices significantly. Until Wednesday the 8th, this gave some parents the confidence to drop their kids off using remote pilot mode. But on that day, even those who had used stems saw them as a threat. Schools were the first places people thought of when they decided the devices needed to be removed.
From the perspective of the software, a road with dense AES coverage is the safest choice. Removing a stem from a school dropoff lane would reduce coverage by a few percent, but that route would still have higher coverage than an adjacent block which had lost its only sensor. This meant that, to avoid being stranded, the vehicles started to seek out school zones. Few schools were able to remove every last device before children began to arrive – again, because of the increased density.
At 7:40 a.m. four unsupervised Marathoners were following behind a school bus traveling north on Clark Avenue in East Palo Alto. They had broken away from a larger pack of trucks that were looping the congested freeways waiting for the bridge to reopen. Readings show the lead vehicle was maintaining a safe distance from the bus, but the other three trucks were operating with a small gap optimized for fleets. When the bus arrived at the school’s horse-shoe driveway, the lead truck identified a gap in AES coverage: the bus was blocking a sensor. The street continued straight. The driveway turned right, wrapping around an island with a few parking spots. The lead Marathoner followed the bus, which came to a stop a few car lengths later. The software correctly determined that it should wait for the school bus, but that meant it was stopped halfway into the driveway and half sticking into the street.
At 7:44 the second truck identified the driveway as an unnecessary turn; the bus was no longer blocking the required sensor, so the street was available. However, it had already begun to turn right to follow its leader, so it was hugging the right shoulder as it passed the truck stopped behind the bus. The unblocked stem had gained an angle on the street, but it could not see past the stopped truck. On the other side was a compact car parked in a designated dropoff zone. Seconds after the new lead truck initiated its pass, a young boy opened the rear driver’s side door of that car and stepped into the street.
The first pedestrian struck by a Marathoner that day was a fifth grader named Luis Forenza. Luis’s parents have made their archive of the data surrounding his killing available to the public, which is how we know the moment-by-moment details of the tragedy. Eight other pedestrian fatalities were privately confirmed to be caused by Marathoners dispatched for the bridge toll exploit, five of whom were children. That single day in 2036 was responsible for 2% of pedestrians killed by light-duty trucks in California that year.