Training the Machine That Replaces You
Meta, layoffs, and the backward feedback loop of AI labor
Meta is installing software on its employees’ computers that captures their keystrokes, mouse movements, and screenshots and using that data to train AI agents.
Wow!
The programme is called the Model Capability Initiative.
Employees cannot opt out on work-provided laptops.
Employees protested on internal forums, but who has that ever stopped?
Meta also announced it will cut approximately 8,000 jobs, roughly 10% of its global workforce, starting May 20. The company is also cancelling 6,000 open roles it had planned to fill, an effective reduction of 14,000 positions.
You see what’s happening?
The workers who survive the layoffs are being required to generate the training data that will teach AI systems to replicate their computer-use patterns.
The workers who are let go, are those the ones whose patterns the AI has already learned?
The people who remain are working to make themselves unnecessary, in exchange for keeping their jobs a little longer.
This is not some dystopian science fiction but actually happening.
The Numbers That Tell the Story
At the same moment Meta announced the layoffs, it revealed a new executive stock option program tied to reaching a $9 trillion market capitalization by 2031, roughly six times Meta’s current valuation.
The potential payout: up to $921 million each for CTO Andrew Bosworth, CPO Chris Cox, and COO Javier Olivan.
The stock appreciation that would generate those executive payouts will be funded, in part, by the labor cost reductions that the layoffs produce. So the people who decided to cut 8,000 jobs stand to make nearly a billion dollars each if the cuts help the stock achieve the target they set.
There is one line from Zuckerberg’s Q1 earnings
If you take the quantitative claim seriously as a measurement rather than as marketing, it represents one of the largest productivity shifts in the history of software engineering. Not sure how to measure productivity gains of high level languages over assembly in 60s or of frameworks and libraties in 90’s but surely not closing a factor of 50x!
But even if the productivity multiplier is less what Zuckerberg claimed, say, 10x rather than 50x, the organizational implication is the same.
Meta said, the company is reorganizing teams into AI-focused “pods” and transferring engineers from across the organization into the Applied AI group. The goal, per an internal memo, is “a step change in engineering productivity and product quality” through “fundamentally rewiring how we operate.”
This is not cost-cutting language. It is restructuring language.
The Industrial Analogy That Does and Doesn’t Hold
Every major labor displacement in the industrial era has followed a rough pattern. A new technology , the steam engine, the assembly line, the computer reduces the human labor required for a given task.
Employment in the affected sector contracts, often painfully. But total employment eventually recovers and often grows, because the technology makes the goods or services cheaper, expanding demand, and because the productivity gains generate income that gets spent on new things, creating new categories of work that didn’t previously exist.
The AI transition is different because it is displacing human minds, not human muscles. The question of what humans can do that AI cannot is genuinely open in a way that was not open when tractors replaced farmhands.
The historical pattern of technological displacement and reabsorption has always assumed that the displaced workers would eventually find their way into roles that required cognitive work. That assumption is now the variable, not the constant.
This is not a counsel of despair. There are plausible mechanisms by which human labor remains valuable even as AI becomes more capable.
Managing AI systems, providing judgment in high-stakes decisions, doing the work of understanding and communicating with other humans.
However, the speed of the current transition is compressing the timeline in which adaptation would normally occur.
Picture in Full
Pull all of this together and the picture is not comfortable. Meta cut 8,000 jobs while forcing its remaining employees to train the AI that will eventually replace them, awarded its executives $921 million in potential stock options tied to a valuation target that requires eliminating the kind of human labor they are currently eliminating, reported $56 billion in quarterly revenue and got punished by the market for not spending enough.
History offers some comfort here. Industrial transitions are always uncomfortable in progress and easier to understand in retrospect. The people living through electrification of manufacturing in the 1920s did not know the outcome.
The people living through the computerization of back-office work in the 1980s and 1990s did not know that it would ultimately generate more jobs than it displaced, in categories none of them could have named in advance.
What is different this time is the feedback loop that runs backward.
The system in which the workers who remain are employed specifically to train the system that will make them unnecessary.
That is genuinely new. And “genuinely new” in the history of technology has meant, roughly in equal measure, outcomes we hoped for and outcomes we didn’t see coming.