Meta Invites Some Individual Contributors in AI Department to Reassume Management Roles
Meta is inviting some individual contributors in the Applied Artificial Intelligence (AAI) department to re-assume management roles. Four insiders confirmed to Business Insider that this arrangement is voluntary, and the company declined to comment.
The AAI was newly established this year, focusing on model training, led by Engineering VP Maher Saba and reporting to CTO Andrew Bosworth. About 7,000 people are included in this department, some of whom were former managers who transitioned to individual contributor roles. Earlier internal communications had also allocated some individuals to groups like "Agent Transformation Accelerator". Around 8,000 people, or about 10% of the workforce, received layoff notices during the same period, with public layoff analysis showing that managers made up the highest proportion. In identifiable samples from California and Washington, over 1,400 managers were included, with nearly half being software engineering managers.
This reappointment of management positions is a partial adjustment to the previous year's strategy of "cutting middle management, small teams, and fewer levels." Mark Zuckerberg has opposed "manager managing manager" since 2023's "Year of Efficiency" and advocates for a startup-like flat organization, believing that AI can take on more execution work. The AAI once had a subordinate-to-manager ratio of about 50 to 1. Employees internally described the reassignment as being "drafted," with some writing on Discord, "I got drafted"; some were allowed to leave the department after complaints. Zuckerberg acknowledged in an internal memo that organizational changes caused confusion and that mistakes would happen again, promising not to conduct large-scale layoffs at the company level within the year while also setting limits on the expanded management span.
Background: At the end of Q1, the employee count was approximately 77,986; layoffs and role changes impacted about 20% of the workforce, in addition to closing around 6,000 vacancies. Public reports indicate that the company's annual spending target on AI chips and infrastructure exceeds $130 billion. Comments from the CTO's side described previous personnel adjustments as "terrible." AAI work has been characterized as "hard work" supporting the Super Intelligence Lab with data labeling, human feedback reinforcement learning, etc., with engineers estimating that 30% to 50% of the core product and infrastructure teams were pulled for labeling.
In market mechanisms, this is not a secondary market transaction but a repricing of organizational capacity. Flattening saves on middle management salaries but pushes the coordination costs of the 7,000-person training cluster to the front line; when the span reaches 50 to 1, morale and output decline simultaneously, and management roles become scarce patches. Beneficiaries are those needing to lead model training rhythms and former managers wishing to return to reporting lines; those under pressure are employees recently changed to individual contributors who must decide whether to return under the guise of "volunteering," as well as the efficiency narrative still valued under startup storytelling. Funds continue to flow into computing power, while the workforce oscillates between "cutting managers" and "bringing managers back."
Supplementary structure: In recent years, Meta has simultaneously established the Super Intelligence Lab, changed some titles to "AI builder," and operated with smaller pods. The AAI has undergone three adjustments within the same year: from mandatory reassignment to allowing departures, and now inviting some to return as managers.
Source: Public information
ABAB AI Insight
Zuckerberg's organizational experiment began in 2023's Year of Efficiency by cutting "manager managing manager" and tied the 10% layoffs with the reassignment of 7,000 people to AAI in the same week: on one hand, focusing on removing engineering managers from the public list, while on the other hand, converting former managers into individual contributors for labeling and training. Internal memos described AAI as a transit station, not an endpoint; the result was a span of 50 people per manager, leading to anger directed at management during all-hands meetings. Recalling some to management is not a philosophical shift but a patch after coordination failure—training clusters need scheduling, and a flat structure cannot provide that.
The capital path is clear: the hundreds of billions in chip and cabinet spending must be diluted across fewer heads; middle management is the first to be used to offset CapEx. However, model training is a highly collaborative assembly line, requiring people to schedule data labeling, RLHF, evaluation, and release windows. Money continues to flow into Nvidia's supply chain and self-built data centers, while the workforce acts as a reversible valve between the "individual contributor pool" and "team-leading managers." Voluntary return reduces the political cost of another forced draft and keeps high-salary ICs who do not want to manage in training positions.
This is analogous to Amazon cutting and then supplementing its management layer during AWS's expansion phase, and Google reverting some directors back to ICs before restoring reporting lines in large model teams. The industry phase reflects a control reversal during an expansion period: the computing budget is locked in, but organizational experiments are still trial and error. Meta simultaneously maintains the Super Intelligence Lab and the AAI labor camp, with the former needing star researchers and the latter requiring a schedulable assembly line of 10,000 people; these two structures cannot share the same flat myth.
Structural judgment belongs to the management segment of the industrial chain reconstruction. The mechanism is that AI has not eliminated coordination; it has merely shifted coordination from product managers to training managers. When one person has to oversee 50 labeling and fine-tuning tasks, the salary saved from reduced levels will be recouped in the form of delayed releases, morale loss, and forced reassignment. Thus, the company treats management positions as switchable resources—turned off during layoff seasons and turned back on when clusters go out of control.
ABAB News · Cognitive Law
The levels saved by flattening will return with interest in coordination.
The larger the computing bill, the more managers need to be cut.
Voluntary return is the second version interface of forced drafting.