Horse and Buggy Software
Part 2 of Retcon Reckoning (The Great AI Replacement)
Technology does not eliminate jobs, the received wisdom goes (and what is received wisdom but a retcon), it relocates them; the received wisdom is not wrong, exactly, in the way that a map is not wrong when it omits the elevation, in describing the territory with enough accuracy to be useful and enough omission to be dangerous, depending on where you are going and how much the climb matters.
The horse didn’t survive the automobile as an industry, but the people responsible for the care and feeding of horses became saddled with the job of maintaining the machines needed for the automobile; they became mechanics, and machinists, and assembly line workers, and the vast ecosystem of human labor required to extract oil from the ground and refine it and move it through a distribution network and deliver it to the stations where the cars that needed it could find it, which over the following century produced employment at a scale that the horse economy could not have imagined, and which is the version of the story that gets told, and which despite also being a retcon, is true, as far as it goes.
What the story tends to gloss over is what happened to the mechanic when the car got computerised. The computer didn’t eliminate the mechanic. It shifted the mechanic’s work from the physical diagnosis of mechanical failure—a task that rewarded accumulated craft knowledge, that got better with decades of practice, that was legible to the person doing it in ways that could be taught and refined and passed on—to the act of connecting the car to a device and reading what the device said, and then ordering the part the device specified and installing it, which is a different category of work, requiring less judgment, commanding less pay, and carrying less of what the previous version of the job was made of. Rather than disappearing, mechanic got cheaper (on the supply side that is, on the demand side the price of car servicing has increased substantially, as anyone who has received, following a four-minute consultation between a mechanic and a screen, a bill of $1,753.41 before parts and tax, well knows) and the craft that made the mechanic irreplaceable got thinner, and the knowledge that accumulated over a career got shallower, and none of this showed up as unemployment because the person was still employed, doing something still called by the same name, though his coveralls were noticeably cleaner.
What also got thinner was the mechanic’s understanding of what the device was reading. The device diagnosed. The mechanic knew how to read the device. These are not the same kind of knowing, and the difference between them becomes visible when the device is wrong; when the fault code points to a sensor and the sensor is fine and the actual problem involves something the device doesn’t surface, even though the sensor could detect it, something that would have been obvious to the mechanic who could hear it and smell it and feel it through thirty years of accumulated attention to how things fail. That mechanic exists. That mechanic is even more expensive and that mechanic is not the mechanic the diagnostic device made economically rational to employ.
the condition under which “moving up the stack” is a description of progress rather than a description of a person standing on a ladder whose lower rungs are being removed.
But someone still had to design the diagnostic device, and the software that ran on the diagnostic device, and the chips the software ran on, and the compilers that translated the code into instructions the chips could execute, and the fabs where the chips were made, and the lithography systems inside the fabs, and the ultra-pure water systems the lithography required, and the supply chains that delivered everything to everything else, and the grid that powered the supply chains, layer beneath layer in a dependency stack that expanded horizontally as it deepened, generating employment at every level, most of it further from the physical world than the level below it, most of it more abstract, most of it more dependent on the layers beneath remaining stable and legible and available to the people working above them, which they were, for a long time, which is the condition under which “moving up the stack” is a description of progress rather than a description of a person standing on a ladder whose lower rungs are being removed.
Meanwhile the software got buggier. As a direct consequence of the same optimization that made the mechanic cheaper: a shift from trying to prevent failures to trying to recover from them faster, which makes complete sense when the systems have grown too complex to keep stable and which produces, as a direct and measurable consequence, more failures, more frequently, in systems that a decreasing number of people understand well enough to articulate. As Mitchell Hashimoto recently observed, the AI infrastructure build-out is repeating the DevOps debate about MTBF (mean time between failures) versus MTTR (mean time to recovery), and the industry appears to be making the same choice DevOps made, which was to optimize for recovery, and not reduced failures.
Defensible as a choice, but also a choice that accepts failure as a chronic condition and defines success as not staying down too long, which describes a different relationship to reliability than the one the previous generation of engineers was hired to maintain. Anthropic’s service status page offers a public record of this relationship in practice. The brownouts and degradations that have become a feature of operating at the frontier are not accidents of scale (though Anthropic seems to struggle with them much more than the other labs.) They are the output of a system optimized for recovery, running exactly as designed, at a reliability level the optimization target was designed to accept. The software is buggy on purpose.
