The software industry has gone through a lot of changes in my lifetime. When I entered college, most universities still didn’t have real computer science departments – they were just tucked inside math or electrical engineering as a sub-specialty. The web exploded a few years later, which led in turn to an explosion in non-technical designers and web devs, then offshore CS programs, onshore MA programs, offshore contractors, etc. The zero interest rate period (and all the ridiculous startups that got funded), Covid, massive layoffs, and AI are all just points on the timeline heading into the future.

Last week, I think we glimpsed another important inflection point, and I think everyone missed it because they were focused on the wrong thing. Meta tried an experiment, it blew up in their faces, and we all got to enjoy a little Schadenfreude at their expense.

To recap, Meta moved about a third of their engineering workforce into a special unit focused on data classification, with the goal of improving their coding LLM with high quality gold data. Meta has tens of thousands of engineers, they decided that this was a priority, so they allocated 6500 resources to the effort.

Let’s see how it turned out.

“It’s literally the gulag,” one of the employees claims. “You have zero purpose in life all of a sudden, you barely interact with anyone, you just have these tasks every week.”

Another employee describes some of the tasks—generating puzzles to test how reliably AI models from Meta and other companies can solve them—as easy compared to the software development work they had been doing previously. But the new projects feel menial, and “almost all” employees seem unhappy, they say. “Most people find the work soul-crushing,” the third employee says.

There are so many layers to this, I honestly find it fascinating. The outer skin of the onion is that Meta leaders made a management choice, then executed it in such an inept, ham-fisted way that it caused a revolt. Almost as if they had no insight into the fact that their engineers had interior lives, and might care about levels of Maslow’s hierarchy above basic food and shelter.

Then, the engineers started complaining that their employer was treating them as though they were just another set of resources to consume or inputs to optimize (have they not been paying attention?). One – apparently without irony – called his well-paid, cushy job “literally the gulag.” Mr. Solzhenitsyn would beg to differ.

Anyway, the press loves a bad guy, Zuck is always up to the task, and voilà, the narrative. Meta leadership apologizes, pinky promises to do better, and everything goes back to normal, right?

Lol, no, I don’t think so. I think we’re seeing the future. There’s been a lot of talk about model collapse, when models start ouroborosing their own outputs. Meta leadership came up with a plan to address the problem.

They made a mistake, but it wasn’t the mistake the press is reporting. Their real mistake was in feeding the pigs off the fine china, in digging the latrine with the nice silverware. They used the most expensive, most pampered, most entitled workers to perform work that wasn’t hard, wasn’t distasteful, was just… boring. Meta leadership “learned their lesson” in the same way A/B testers learn their lesson when one of their experiments fails in an interesting way. I’m sure they’re already looking at the data, figuring out what worked and what didn’t, and planning the next evolution of the program.

A quick search will find plenty of articles on Meta’s Africa-based content moderators (and the trauma they suffer in order to protect people in rich countries from the horrifying cesspool of what people post online). Meta has plenty of experience scaling up offshore third-party classification services, and with one massive tech layoff after another over the past five years, there’s a glut of engineers on the market. Why not use those engineers to generate gold data? Why not use them to classify? There are plenty of people who will sign up to spend 10 hours a day gold farming doing leetcode problems to pay the rent.

I think we’re witnessing the creation of a new job category, almost exclusively located in low-cost countries. When workers can be anywhere – when they “barely interact with anyone” – when there’s no coordination required because each task is completely self-contained, with no history, surrounding code, or broader context – when the tasks are “easy compared to the software development work they had been doing previously” – then there’s no need for expensive Silicon Valley salaries, and certainly no need for expensive Silicon Valley attitudes. Gig workers in foreign countries are so much easier to control. Training the model will be like driving an Uber – you get a prompt, provide a service, get your fee, and hope for a five star rating.

While workers in this new category won’t directly compete with the current software engineering category, it will still apply downward pressure on the job market. Once Meta launches and scales their offshore operations of easily replaceable, inexpensive parts who don’t need expensive benefits and (for instance) won’t complain about their employer tracking their key strokes, we’ll see another round of onshore cost-cutting, with other mega corporations watching in interest.