For all my 25 years in the industry, large companies have needed junior engineers to write code. Senior engineers were too scarce and expensive to write it all, so we roughly divided the labor: juniors hacked, seniors guided. Juniors’ on-the-job training was mostly a side effect: it was practically inevitable that they’d gain the expertise and wisdom to be senior engineers, just by writing code that met seniors’ standards. If a junior engineer left after a couple of years, the company lost some of its investment in training her, but that was acceptable. The code she wrote was worth it.

During my time as an engineer at MongoDB, before I switched to research, I mentored dozens of interns, new grads, and junior engineers as they began their careers at MongoDB. For me, this was enjoyable and beneficent work. For MongoDB, it was a way to produce code and cultivate the senior engineers we would need in the future. There was a period in the 2010s when it was the only way to hire large numbers of engineers: they were so in demand, and salaries had inflated to such ludicrous proportions. This was the ZIRP era, when the big tech companies hoarded labor. Amazon, Google, etc. deliberately hired more people than they needed, as a strategic reserve of talent, and to starve their competitors. MongoDB’s only hope was to catch engineers before they even graduated college. Many left after a few years, but their contributions as juniors were worth our investment. Some of these hires stayed; they’re now our most senior experts and leaders.

The Declining Investment #

The market for software engineering labor changed post-Covid. Interest rates rose and companies did mass layoffs. Elon Musk demonstrated at Twitter that savage, arbitrary cuts didn’t bring the site down. At MongoDB we’ve kept hiring interns and new grads, but across the industry their prospects diminished.

Covid hurt junior engineers in another way: companies permitted hybrid or remote work, senior engineers stopped coming to the office as much, and in-person mentorship suffered. This was the first tragedy of the commons. For seniors like me, working from home is pleasant and productive. But for the juniors I no longer sit with, it’s a loss. They need to build an intuitive theory of programming—a theory of the company’s software specifically, and software engineering in general. This is mostly transmitted implicitly, not through explanation but by working side-by-side and solving problems together. Doing this work over Zoom is only a fraction as effective.

The Impending Collapse #

Now comes the second tragedy of the commons: companies are, or may soon, stop hiring junior engineers and just replace them with AI. Senior engineers will be retained to oversee the coding agents. (For example, Amazon just announced that all AI code must be reviewed by senior engineers.) As two Microsoft executives wrote last week:

The result is a new incentive structure: Hire seniors, automate juniors. But without EiC [early-in-career] hiring, the profession’s talent pipeline collapses, and organizations face a future without the next generation of experienced engineers.

The authors propose “preceptorship at scale,” where senior engineers pair with juniors and help them guide AI agents. The juniors will absorb the seniors’ wisdom, judgment, and taste, born of long experience manually coding. This sounds great! But why would any company invest in juniors this way, instead of free-riding on other companies and poaching their engineers once they’re ripe? The authors make a normative argument and fail to address the economic incentives: any company that invests in juniors in the age of AI will likely lose out to a company that only hires seniors. If junior engineers’ output while they learn on the job is no longer economically valuable, the free market won’t replenish the stock of senior engineers.

Some Uncomfortable Solutions #

I assume that college grads will always need on-the-job training, at great expense to their employers, rather than graduating fully formed. I don’t know if any academic or professional degree truly prepares its graduates for work, but a Bachelor’s in Computer Science certainly doesn’t. Given the new incentives of the AI era, colleges or other institutions might try to accomplish this, but I doubt it’s possible. An educational program can’t simulate a workplace realistically.

We talk about “apprenticeship” today, but perhaps indentured servitude better describes a sustainable model in the future. Companies will hire new grads and train them to be expert AI-overseers, in exchange for a multi-year lock-in. The employer ensures it gets its money’s worth from the engineer. If the practice is widespread, would-be free-riders will be starved of talent and forced to train their own. This would ensure the supply of senior engineers, but it would create a huge deadweight loss compared to the free market we’ve enjoyed so far: lower profits, lower salaries, less production. We’ve already seen how inefficient such a system is: Silicon Valley beat the Boston area largely because California engineers could switch jobs easily. Maybe AI will make up for that loss so dramatically we won’t notice.

Or maybe we need an apprenticeship levy like in the UK: a tax on all employers, which is forgiven if they train junior employees. The UK program is a mere 0.5% of payroll—we’d need a bigger and more targeted fee for software firms, but it could work. Government intervention is the classic answer to market failures like this one.

My Conclusion To This Foolish Speculation #

All articles about AI are titled with question marks this year. I think we’ll need another generation of senior engineers to guide AI agents? But it seems equally likely that AI will soon enter a recursive self-improvement loop. It will get so smart so fast, “human software engineer” will be a forgotten archaism. Today we think it’s funny that a “computer” was once a human with a slide rule. In a few years we’ll say, as we lounge together in self-piloting electric flying cars, “Did you know that ’engineer’ used to be a human occupation?” The singularity seems more plausible to me every day.

Images: The Vault Of The Atomic Space Age.