Learning from Block's 40% Layoff—You Might Only Learn How to Fire People

Deep thoughts on AI and aspirations —— Bytedance Deep Thought Circle

Block laid off 40% of its workforce at the end of 2025, going from idea to execution in under three weeks. Outsiders see the percentage; insiders see the algorithm: the team ran a three-step minimization exercise—how many people to keep services running at 100%, how many to meet strict regulatory compliance, how many to deliver growth commitments plus agent rebuilds. Add error margin to those three numbers, and you have the final answer.

Many CEOs will be tempted by this news. But companies that directly copy this move will likely only learn how to fire people, not the conviction behind the firing. Where that conviction comes from is the truly valuable part of this story.

A Two-Thousand-Year-Old Hierarchy Problem

To understand what Jack Dorsey is dismantling, you first have to acknowledge why hierarchy exists. The organizational structure of modern companies has its prototype in the Roman legion: 8-person squads nested into centuries, cohorts, and legions, with a commander at each level responsible for aggregating and transmitting information. It was designed around a human limitation—a leader can effectively manage roughly 3 to 8 people, what we now call span of control. In the 1850s, railroad companies drew the first organizational charts; Taylor later refined them into functional pyramids. For the past two centuries, management theory has essentially been making improvements within this framework.

Hierarchy is fundamentally an information routing system, coordinating large-scale human effort under conditions of limited information transmission. It has an unavoidable cost: narrowing the span of control requires adding layers, and more layers means slower information flow. Tech companies have made various escape attempts—Spotify promoted cross-functional squads, Zappos tried holacracy, Valve operated flat—and after scaling, they all returned to hierarchical coordination. Not because they didn’t try hard enough, but because no alternative routing mechanism existed at the time.

Dorsey’s judgment: AI is the first real alternative. A system that can maintain a continuously updated model of the entire business and use it to coordinate work—something that previously required human managers to relay messages layer by layer.

Replacing Routing with Models Requires Two Foundations

Block’s approach sounds straightforward: instead of giving each person an AI assistant, let AI directly assume the coordination function that hierarchy performs. The company needs two world models—one about itself, one about customers.

The one about itself depends on a condition most companies don’t have. Block is a remote-first company where all decisions, discussions, code, designs, and plans exist as digital records. AI can continuously maintain a panoramic view: what’s being built, what’s blocked, where resources are allocated, what works and what doesn’t. In traditional companies, this is a manager’s job—passing contextual information up and down the chain of command. In a company where everything is text, machines read it themselves.

For the customer model, Block occupies a position few can replicate. Money is the most honest signal. People lie in surveys, ignore ads, abandon shopping carts, but every transaction, transfer, and payment is a fact that actually occurred. Through Cash App and Square, Block sees both sides of millions of daily transactions—the financial reality of each customer and merchant continuously layered into the system.

With these two models, you can talk about an intelligence layer on top. Payment, lending, card issuance—these capabilities are no longer products but building blocks. The intelligence layer combines blocks into solutions at the right moment. When a restaurant’s cash flow starts tightening before a seasonal dip, the system automatically assembles a short-term loan with an adjusted repayment plan, delivered before the merchant thinks to look for money. Where does the roadmap come from? When the intelligence layer tries to compose a solution but fails because a capability doesn’t exist, that failure signal becomes what to build next. Customer reality directly generates the backlog—no more guessing by product managers.

People retreat to the edge, but the edge is where action happens. The entire organization converges into three roles:

From Dorsey to any employee, it’s currently 5 layers deep maximum. His goal is to compress it to 2 or 3.

Why You Can’t Copy It Usually Comes Down to Foundations

My judgment: the replicability of Block’s approach is severely overestimated. It has two foundations, each filtering out most companies.

First, the work itself is machine-readable. If your company’s core knowledge lives in veteran employees’ heads, at dinner tables, in twenty-minute voice messages, AI has nothing to read—the world model is an empty shell. Copying Block’s middle-management cuts at that point is like pulling out routers when the new network isn’t installed yet. Second, customer signals are honest. Most companies understand customers through surveys and satisfaction scores—low-fidelity channels. An intelligence layer working with low-fidelity signals won’t compose the right solutions, at most pleasing solutions.

There’s also a narrative calibration needed. The most widely circulated version says AI makes most decisions. Even Block director Roelof Botha doesn’t see it that way—his version is much more moderate: AI helps transmit alignment, management sets frameworks, people at the edge handle corrections, information input, and directional adjustments. Radical claims are good for spreading; hybrid reality is good for operating. When learning from others’ organizational experiments, step one is distinguishing these two—don’t treat fundraising narratives as operating manuals.

One change is particularly concrete though. Block’s meetings went from flipping through slides to everyone bringing AI-built real-time prototypes within two months, using simulated or real data to slice out a cross-section of their work, editable on the spot. What changed isn’t really meeting format but evidence quality: discussions center on things actually being built, not things planned to be built, and the explorable breadth suddenly expands. Any size company can learn this tomorrow without buying a single layoff tool.

Two Things a 100-Person Company Can Do Tomorrow

Dorsey gave small companies a starting point. First, ask yourself: is this management layer necessary, and where does it block actually solving customer problems? A 100-person company has at most two or three layers—now is when you can change lightly. Wait until thousands of people and the cost is an entirely different magnitude. Second, put daily generated information—group chats, documents, meeting notes—into an intelligent system you can dialogue with. His exact words: this will improve your company understanding by two or three times, because before you could only rely on people telling you, and people don’t necessarily say things for various reasons.

Finally, there’s the litmus test he left behind, which I think is worth copying more than the entire four-layer architecture: What truly difficult-to-understand things does your company comprehend, and is that understanding deepening every day? If the answer is no, AI is just a cost optimization story for you—cut headcount, improve a few quarters of profit margins, then get absorbed by something smarter. If the answer is profound, AI is revealing what your company actually is. This question has no unified answer—a gas station and a law firm respond completely differently—but every company must answer it, and the sooner the better.