The Quiet Failure: Block's World Model Manifesto and the Line AI Can't Cross

I run four or five AI agents across my own projects, and I've learned to fear the failures that don't announce themselves. The loud ones I catch in minutes: a crash, a red test, a diff that's obviously wrong. The dangerous ones are the agents that keep producing confident, plausible output long after the ground truth has quietly drifted out from under them. You don't notice until a decision built on that output goes bad, and by then the trail is cold.

Organizations have a version of this, and it's about to get a lot more common.

When Zappos adopted Holacracy and eliminated management titles in 2015, roughly 29% of the workforce left. When Valve tried its flat structure, employees compared it to "Lord of the Flies": informal power hierarchies replacing visible ones, with zero accountability. Those failures were loud. Everyone saw them break.

The world model failure is the quiet kind. The system ties a churn spike to a feature launch when the real cause was a billing change, the team kills the wrong feature, and the postmortem blames "market conditions." As technologist Nate Jones puts it: "The most dangerous version of a world model is the one that works well enough that nobody questions it until the decision quality degrades."

On March 31, 2026, Block published From Hierarchy to Intelligence, the most ambitious manifesto anyone has written for replacing middle management with AI. Within weeks, a technical lead named Richard Hesse threatened to quit to get his infrastructure team rehired after Block admitted some layoff cuts were "clerical errors." The gap between the manifesto and the correction is the story.

What Dorsey Gets Right

Credit where it's due, and it's more than most critics are giving.

The historical arc in From Hierarchy to Intelligence is genuinely excellent. Dorsey and co-author Roelof Botha trace organizational hierarchy from the Roman contubernium through the Prussian General Staff's invention of middle management "to support incompetent generals," Taylor's scientific management pyramid, and McKinsey's matrix organization in 1959. The argument is elegant: hierarchy was never a natural law. It was an information-routing technology, the best one available for two thousand years. As they write: "The question was never whether you needed layers. The question was whether humans were the only option for what those layers do. They aren't anymore."

It's the most compelling reframe of organizational structure I've seen. Previous flat-org advocates (Zappos, Valve, the agile movement) struggled because they removed the coordination mechanism without replacing it. Dorsey's claim is that AI is the replacement: coordination shifts from humans to a continuously updated model of the company. The four-layer architecture is coherent, and Block's signal advantage is real. As Dorsey puts it: "Money is the most honest signal in the world." A fintech sitting on transaction data from both the buyer side (Cash App) and the seller side (Square) has a richer foundation for a world model than almost anyone. As a technical architecture document, I'd give it high marks.

But architecture documents don't lay people off. Organizational decisions do, and those depend on a number the manifesto never confronts.

The 60/40 Split

Managers spend 60–70% of their time on administrative logistics: status syncs, information routing, report generation, scheduling, approvals. This is the work the manifesto targets, and Dorsey is right that most of it can be automated. It probably should be.

The remaining 30–40% is judgment work. Mentoring a struggling engineer through a career crisis. Negotiating priorities between two teams that both think they should go first. Knowing that one name on a layoff list means losing the only person who understands the infrastructure that keeps three products running. These are interpretation tasks, and they require context that no model currently captures. The word "judgment" barely appears in the manifesto. "Mentoring" doesn't appear at all.

Neil Thompson, a researcher at MIT, framed the distinction precisely: "If part of your job gets automated, and it's something that really didn't use the expertise that you needed, that's great. You get to spend more of your time on the part of your job that is really valuable." The implication cuts both ways: you have to preserve the expertise parts. Block's manifesto automates the logistics and assumes the judgment will find new footing at "the edge." It doesn't design for that transition.

And the judgment layer is already strained. The average American manager now oversees 12 direct reports, nearly double the count since Gallup started tracking in 2013; Meta runs a 50:1 ratio in its applied AI engineering division. Only 27% of managers globally feel enthusiastic about their work, and most have never had formal training. The system is already broken; the question is how to change it.

Josh Bersin, who studied 70+ companies implementing AI productivity tools, found a pattern: "If you look at the 10-100X transformations from AI, none of them start with job displacement. They all start with business re-engineering."

Block started with displacement.

The Rehiring Problem

What does it look like when the judgment layer actually fails? Block gave us a case study within weeks.

In February 2026, Block cut over 4,000 employees, roughly half its workforce, and Dorsey explicitly tied the layoffs to AI productivity gains. The stock rose nearly 18% the day of the announcement as investors bet on efficiency.

Then the corrections started. Richard Hesse, a Block technical lead, threatened to quit to get his infrastructure team rehired, and Block admitted some of the cuts were "clerical errors." Let that land. The company that just published the most sophisticated argument for replacing human judgment with AI described the failure of its own decision-making process as a clerical error. If the world model carried an accurate picture of which knowledge is irreplaceable, a technical lead wouldn't have needed to threaten to quit. The system got the logistics right (here are the headcount numbers) and the judgment wrong (here's who you can't afford to lose).

The employees filled in the rest. Seven anonymous Block workers gave on-record accounts to The Guardian disputing the AI capability claims. One, whose actual job was helping others use AI tools, reported that roughly 95% of AI-generated code at Block still needed human fixes. Another said: "Everyone that I know that's still there has a ton of dread because they just realized their workload has quadrupled or 10xed and AI is not going to fix it."

The financial context complicates the narrative further. Block had grown from 4,000 to 13,000 employees during the pandemic hiring spree, and the stock had fallen 40% before the layoff announcement. Bloomberg called it AI-washing; Mizuho analyst Dan Dolev said bluntly: "The vast majority of these cuts were probably not due to AI."

The narrative presented the layoff plan with structured confidence, senior leaders acted on it, and the failure stayed invisible until someone with enough leverage made it visible. Most people don't have Richard Hesse's leverage. McKinsey partners warned in Harvard Business Review that cutting middle management hastily "can be a costly mistake" because middle managers are "the glue that holds teams and enterprises together," and Forrester predicts that 50% of AI-attributed layoffs industry-wide will be quietly rehired.

The Interpretive Boundary

There's a name for what Block's manifesto is missing: the interpretive boundary.

Every AI system in an organization implicitly sorts its output into two bins. "Act on this" covers factual, verified, low-risk conclusions: status rollups, threshold alerts, routine approvals. "Interpret this first" covers judgment calls the system isn't equipped to make: trends requiring context, correlations that might be causation, personnel decisions with human consequences. Your system is already making that distinction. The only question is whether you've designed it explicitly.

Nate Jones's analysis of world model architectures identifies three, each with its own blind spot:

The Vector Database approach is fast to deploy and adequate for information logistics, but semantic retrieval has no mechanism to distinguish surfacing from interpreting. Rankings become editorial judgments by default. At scale, "the ranking becomes a reality no one intended."

The Structured Ontology approach (think Palantir) defines objects and relationships explicitly, prevents hallucinations, enforces precision. But it can only represent what you've already categorized. It's "accurate about what it knows. Silent about what it doesn't know."

The Signal Fidelity approach (Block's model) is built on high-quality transaction data. The signal is clean, and that cleanliness is precisely the vulnerability. As Jones puts it: "High signal fidelity at the input layer creates an illusion of high judgment quality at the output layer."

That last line is the key insight. Transaction data is clean, factual, "the most honest signal in the world," as Dorsey says. But the decisions built on it — who to lay off, which products to kill, which markets to enter — are judgment calls wearing the costume of data. The cleaner the input, the more trustworthy the output looks, regardless of whether the interpretive leap was warranted.

Richard Teachout's analysis of decision boundaries in AI systems puts it plainly: boundaries "reflect where humans must remain in the loop, not because machines are incapable, but because consequences extend beyond the model's view." Good architecture does not grow that boundary on its own; you draw it deliberately, then enforce it. Skip that, and you get multi-agent systems that fail through "silent failure states", agents optimizing for subgoals that don't align with the mission. At organizational scale: the world model optimizes for efficiency, the humans who would have caught the error are gone, and the system looks like it's working. That's the quiet failure.

The Counter-Model

DBS Bank in Singapore shows what getting the boundary right looks like. By 2025 it had generated over S$1 billion in economic value from AI and data analytics, deployed more than 300 AI models, and reskilled all 40,000 employees. It didn't replace management; it automated the 60% logistics, retrained the humans for the 40% judgment, and created new oversight roles whose explicit job is to sit at the interpretive boundary. Augmentation, producing measurable results at scale. Block started from the wrong end.

If you're flattening your org right now, the boundary has to be drawn by hand. Classify every AI output in your management stack as "act on this" or "interpret this first." Make the system communicate confidence levels, not just conclusions. And capture the feedback loop, because as Jones argues: "A knowledge base might record what happened. A world model is supposed to record what happened, what was done about it, and what happened next." Without that, the model never learns from its own judgment errors.

The Unresolved Question

Dorsey, to his credit, acknowledges the risk: "Block is in the early stages of this transition. It will be a difficult one, and parts of it will likely break before they work."

That's intellectually honest. But "break before they work" hits differently when the things that break are people's careers. Richard Hesse's infrastructure team wasn't an abstraction.

And here's the question that keeps me up at night. When the quiet failure happens, when the world model confidently surfaces the wrong signal and the team acts on it and the damage is done, who is responsible? The system? The engineer who built it? The executive who replaced the manager who would have caught it? We don't have an answer. The accountability frameworks haven't caught up to the architecture.

Dorsey is right that humans were never the only option for routing information through an organization. He's just wrong about the timeline for the part that's left. Block didn't get the architecture wrong; it got the sequence wrong, removing the humans at the interpretive boundary before it had anything that could stand there in their place. And that 40% is exactly where a confident wrong answer costs someone their job before anyone with the leverage to object is still in the room to object.