My self driving car is my horse
A horse navigating San Francisco will, when uncertain, slow down. When genuinely confused, it stops or maybe redirects. What it won’t do is drag a pedestrian twenty feet down a city street.
On October 2, 2023, a Cruise robotaxi struck a woman who had already been hit by another vehicle, then pulled her approximately twenty feet while navigating toward the curb. California suspended Cruise’s operating permits. GM disclosed $3.48 billion in Cruise losses for 2023 alone, against more than $10 billion deployed total. The program is now folded.
Toyota
Toyota has been running a different experiment for seventy years, documented from the start. Embedded in their production system is jidoka — machines designed to detect abnormalities and stop, waiting not for another machine but for a person. The andon cord, which any worker on the line can pull to halt production entirely, is the physical implementation of this idea.
The reasoning: the moment a worker observes why something stopped and decides how to fix it is where improvement originates. Automate that observation step and you remove the failure, the observer, and the feedback mechanism that makes the system better over time. Toyota maintains deliberately manual segments of production — including a program of takumi, master craftsmen whose accumulated knowledge resists formalization — because the judgment they carry is where process knowledge lives. Akio Toyoda has said robots cannot feel whether something is right. That isn’t a poetic remark. It’s an engineering description of what feedback resolution requires.
The Rediscovery
In 2017, Tesla designed what Musk called an “alien dreadnought” — a Model 3 production line built to minimize human presence, including an automated conveyor system for battery module assembly. It couldn’t reach throughput or quality targets. The solution was a tent erected in the Fremont parking lot, where humans assembled battery packs by hand.
On April 13, 2018, Musk wrote publicly:
“Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.”
Toyota had formalized this in the 1950s.
Open and Closed
The distinction that matters isn’t automation versus manual. It’s whether the problem space is enumerable.
Warehouse robots, port cranes, mining trucks in geofenced pits — these work because the relevant world has been successfully bounded. The actors are defined, the geometry is mapped, the edge case distribution is actually finite. Waymo’s functional success is in Phoenix — exhaustively mapped, geofenced, controlled conditions. That’s a closed world, at a cost of roughly $30 billion across seventeen years. It operates well for the same reason an automated mine does: the domain has been closed, not because the technology solved an open-world problem, but because the geography was arranged to approximate a bounded one.
Public roads don’t close. Construction appears without notice, sensors occlude behind an adjacent semi, pedestrians assume positions that weren’t in the training data, weather degrades perception without flagging its own degradation. Systems trained on distributions fail outside them, and the tail of an unbounded distribution is where the incidents accumulate.
The Horse and the Drone
Police horses navigate San Francisco today. Tourist carriages navigate it. The capability exists operationally — not theoretically.
A hub-and-spoke model using horses for freight to fixed waypoints, with drones distributing from those points into dense urban corridors, joins two working capabilities at a logical seam. The technology investment falls on the drone leg, doing bounded point-to-point aerial delivery — the type of problem automation handles cleanly.
Working horses cost $5,000–15,000 to acquire and $4,000–8,000 per year to maintain. No sensor suite, software update, or remote operations center. No liability tail requiring a federal investigation. Against Cruise’s $10 billion in total losses — including the $8–12 million pedestrian settlement — and Waymo’s roughly $30 billion deployed before generating material revenue, a complete cost-benefit analysis of the horse-drone model would produce, at minimum, considerably more stable employment and substantially more horse manure than the sector currently generates.
The horse at Market and Castro, slowing at the intersection, reading the cyclist cutting left, the delivery truck double-parked, the pedestrian watching their phone — is not a limitation to engineer around. Though it has failure modes too, which so far don't require congressional testimony.