Cutting Middle Management Doesn't Make Organizations Faster

Think of a company as a control machine, where middle management acts as the state estimator. Meta's recent invitation for some non-managerial employees in its Applied AI division to return to managerial roles looks like an HR rollback, but from an information theory perspective, it exposes a harder problem: AI can compress information, but it cannot generate accountability out of thin air.

Middle Management is the Feedback Loop

I've been looking into world models and embodied intelligence evaluations for the past two months, and I increasingly feel that organizations are similar. The core contribution of Ha and Schmidhuber's work is decomposing agents into perception, memory, and controllers. In physical AI, this is also true; a controller without state estimation is blind.

Employees generate massive amounts of noise daily—changing requirements, dependency blocks, customer complaints, temporary priorities. What middle management used to do was compress high-entropy noise into low-entropy instructions, then map those instructions into schedules, resources, and escalation paths. Shannon talked about encoding, Wiener about feedback. Without state estimation and feedback loops, systems drift.

For a while, Meta pushed hard for flattening hierarchies, moving about 7,000 employees into the Applied AI engineering department, alongside multiple rounds of layoffs. The logic wasn't strange. If AI can summarize meetings, write code, and break down tasks, why keep people just to relay messages? But the reality is, if goals, acceptance criteria, permissions, and failure cases aren't explicit, AI can't handle coordination. It cannot cover up ambiguous requirements.

Two Paths: Cut People First or Build Systems First?

Path A is "AI replaces middle management." Flatten the hierarchy first, then let agents handle project management, code reviews, and meeting summaries. The short-term narrative is strong, and costs seem to drop quickly. This assumes task logs, permission boundaries, and acceptance standards are all structured.

Path B is "Human-AI hybrid management." Let AI handle low-entropy repetitive work first, such as generating status updates, tracking dependencies, and summarizing changes; managers handle high-entropy decisions, such as prioritizing goals, arbitrating conflicts, and taking responsibility for failures. This path is slower but stable.

Meta's shift looks like a retreat from Path A back to Path B. Beyond model capabilities, the organization lacks evaluation benchmarks. When I discussed graphing codebases recently, I mentioned that humans and AI viewing the same repository need impact scope, source, and change chains. Management is the same. Who approves, who is responsible, who changes priorities, who judges risk—if these aren't explicit, AI will only accelerate chaos.

Zuckerberg admitted in an internal memo that these adjustments were complex and that they made mistakes.

This sentence is more critical than layoff numbers. The error lay in forcing flatness when states were unobservable. Flatness requires every node to clearly see what it connects to and what it changes.

AI-Native Organizations Need Observability

From an embodied intelligence perspective, training world models requires rollouts, rewards, and failure cases. Organizational AI transformation needs these too. How a change impacts delivery, how a priority error causes delays—all must be recorded. Rewards need to measure a joint function of quality, risk, cost, and morale.

The truly hard part is turning company operations into a measurable system.

I worry many companies interpret "AI-first" primarily as workforce optimization. Being AI-first requires interfaces first, including task states, decision logs, permission boundaries, and evaluation sets. Without these, cutting middle management just severs the feedback loop.

If you only look at the news, Meta's move seems like appeasement. From a research perspective, it looks more like calibration after a cybernetic experiment failed. AI can replace much information transport, but it cannot replace responsibility allocation. Deleting nodes doesn't necessarily speed things up; the key is whether each node has observable states and feedback signals.