Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 17 September 2026.
The bozo explosion always sounded too much like a personnel problem.
A players hire A players. B players hire C players. Keep going and eventually everybody sucks.
Sure. Versions of that happen.
What scares me more is a company full of individually competent people making locally sensible decisions until the aggregate starts eating itself.
They know the local game. They know which project gets budget, what phrasing survives review, which metrics carry a promotion, how much bad news a VP wants before the meeting gets weird, which executive's favorite initiative needs "more data" before anybody can call it dead.
Then they act accordingly.
Congratulations. The company has started teaching people how to kill it.
Yahoo had one beautifully nasty loop
I went looking through Paul Graham's essays after thinking about organizational death spirals, and pieces of the same machine are sitting all over them.
What Happened to Yahoo has the cleanest loop. Graham argues that Yahoo let programmer quality fall early. Strong programmers want to work with strong programmers, so lower quality made the company less attractive to the people who might have raised it again. Hiring got weaker, which made future hiring harder, which weakened the place further.
He calls it a death spiral.
The important part for me is the recursion.
One bad hire is a person.
A hiring system that makes the next bad hire more likely is a process.
A company where the changing population alters who applies, who stays, who gets promoted, and who gets to hire the next generation has started evolving in a direction.
The bozo explosion fits inside that picture, but "bozo" makes the whole thing sound like the problem is stupidity.
A much scarier company can keep hiring bright people.
It can teach those people the wrong lesson.
Every company has a hidden curriculum
The lesson arrives through consequences.
Suppose an engineer raises a serious problem with a launch. The launch slips. A senior manager gets embarrassed. The engineer spends the next month defending why they escalated.
A different engineer sees a similar problem six months later and phrases it more gently.
"We may want to keep an eye on this after launch."
The launch ships and survives the week. Their manager says they handled the situation pragmatically.
Everybody nearby just learned something.
Policy stayed the same. The organization taught the lesson anyway.
Companies do this constantly. Which projects get rescued? Who gets promoted after a miss? Who gets described as "strategic"? Who gets called difficult? Which person can say "this is stupid" and keep their career? Which person gets celebrated for landing the quarter after quietly borrowing from the next one?
The employee handbook carries less information than the observed gradient.
People are astonishingly good at finding it.
The winner gets to write the next exam
Graham's The Lesson to Unlearn is about bad tests. Give people a consequential proxy for the thing you care about and they'll learn to optimize the proxy. Students want to learn, grades become the consequential test, and ambitious students become extremely good at getting grades.
The corporate version gets recursive because the people who score well on the internal test often gain authority over the test.
A manager learns that projects with clean dashboards survive budget season.
So they build clean dashboards.
They get promoted.
Now they evaluate other managers.
Which managers look competent to them?
Probably the ones who produce the kind of legible success they learned to produce.
A local survival tactic has become a selection criterion.
Give it another generation and it becomes culture.
A company can drift through accumulated local choices. The people who thrive under the existing incentives become the people who allocate money, conduct interviews, write performance reviews, define leadership qualities, and decide what counts as good judgment.
The test starts breeding people who are good at the test.
Then the winners inherit the answer key.
Smart people can make the spiral run better
Intelligence sounds like the obvious corrective until you ask what the intelligent person is optimizing.
A smart employee can notice that the official goal and the career goal have diverged earlier than everyone else.
Maybe customer retention is crucial to the company while the promotion committee rewards launches.
The smart employee launches.
Maybe reliability is crucial to the product while leadership rewards visible feature work.
The smart employee learns how much reliability work earns social credit and stops there.
Maybe everybody says they value candor while the people bringing ugly news spend twice as long defending themselves.
The smart employee figures out how to package ugly news so nobody important feels implicated.
Excellent local judgment.
Terrible aggregate result.
The Power of the Marginal gets at a harsher version of this. Graham argues that some hierarchies become tests for the wrong qualities. Rising through them can demand so much attention to politics that the selection process favors people unusually willing or able to play that game.
Once the test becomes an anti-test, competence at passing it becomes part of the problem.
You can end up with executives who are genuinely excellent at executive selection pressure.
They can read a room. They know where the budget lives. They know when to attach themselves to a project, when to distance themselves, which failure needs a task force, which success deserves a victory lap, how to turn an ambiguous outcome into an achievement, how to make a reorg look inevitable two weeks after fighting against it.
Those are real skills.
The customer may receive zero benefit from them.
Information starts arriving pre-defended
Founder Mode comes at the problem from information routing. Graham dislikes the idea that a CEO should treat each branch of the org chart as a black box and interact mostly through direct reports. His proposed alternative includes much more skip-level contact and direct engagement with details.
Forget the founder-versus-manager argument for a second. The sensor problem is useful by itself.
Suppose every layer of management has to summarize reality for the layer above.
Each summary is written by somebody whose team, budget, status, or judgment may be evaluated using the summary.
The bug becomes a risk.
The risk becomes a challenge.
The challenge becomes an investment area.
By the time it reaches the executive meeting, the company has a strategic opportunity to improve resilience.
Maybe everybody told the truth.
They also translated it through four careers.
Senior leadership can become sincerely detached from what the company feels like at the bottom because every information channel passes through somebody with a reason to make the information survivable.
Then leaders make worse decisions using cleaner decks.
Worse decisions create more pressure.
More pressure makes clean decks more valuable.
Oh, great. Another loop.
Conformity has population dynamics too
Graham gets even closer to the chain-reaction idea in How to Think for Yourself. He argues that successful startups begin with unusually independent-minded people, then growth can dilute that population because conventional-minded people are more common. He floats the idea of organizational policies acting like control rods that slow a chain reaction of conformism.
Control rods are exactly the right image here.
A culture can move because the composition of the room changes.
Then the changed room affects what can comfortably be said.
What can be said affects what people think through openly.
The narrowed conversational range affects who enjoys working there.
Those people leave or stay.
Now the composition changes again.
The same recursive pattern shows up from another angle.
A founder hires fifty people and thinks the company got fifty people larger.
The company also became a different social environment for the fifty people who were already there.
Every hire changes the audience.
Every promotion changes the expected behavior of the people watching.
Every departure tells a story too.
If the person who kept saying "the numbers don't support this" leaves, the remaining people inherit a quieter room.
A death spiral can begin with success
Failure gives us obvious examples because the numbers eventually look ugly.
Success can start the same process with nicer furniture.
A product works. Revenue grows. The company develops a repeatable sales motion. Somebody discovers the exact customer segment that buys quickly. Somebody else finds the pricing trick. Quarterly planning becomes more predictable. The company gets extremely good at operating the thing it already knows how to operate.
Now experimentation competes against a machine that already makes money.
People attached to the machine have budgets, headcount, historical wins, executive sponsors, and metrics everybody understands.
The strange new idea has a prototype and a person talking too quickly.
Guess who wins the planning meeting.
Graham's September 2026 essay Making Startups Powerful argues that hired executives often work on shorter horizons than founders and that mature companies can drift from inventing new things toward extracting more from existing ones. You can debate how much of that comes specifically from founders. The time-horizon mechanism works either way.
Change the horizon and different decisions become rational.
A ten-year investment can look brilliant from 2036 and irresponsible from this quarter.
A leader whose compensation, reputation, or expected tenure lives inside the quarter has a coherent reason to prefer the second view.
Again, personal rationality and organizational survival can separate.
The company starts protecting the people who can navigate the company
The Corrective Mechanism Is the Threat was about a related failure at the level of a person or regime: the thing chosen to protect the goal slowly becomes the thing protected as the goal.
Organizations add inheritance.
One person's defensive move can become another person's incentive.
One manager's incentive can become a promotion rule.
The promotion rule changes who has authority.
The people with authority choose the next managers.
Eventually the company can protect the kind of person who is especially good at navigating the company.
Customers may barely enter the loop anymore.
This is where "culture" stops sounding soft to me.
Culture is partly the accumulated answer to a hard practical question:
What happens to people who do this here?
Tell the boss they're wrong.
Kill your own project.
Expose a metric everybody likes.
Hire somebody who scares you a little.
Admit the quarter came from pulling demand forward.
Spend six months on work whose value will become obvious after your next performance review.
The organization answers each question through consequences.
People remember.
Reality needs more than one way into the room
The obvious response is to hire virtuous people who love truth and resist incentives.
A plan that requires everybody to be a saint is fragile.
A healthier company makes correction cheaper and gives reality multiple routes into decisions.
Customers can reach people with authority. Engineers can surface failures outside the chain that owns the project. Metrics have definitions that survive a bad quarter. Leaders talk to people several levels down. Projects can die while their owners keep their dignity. Postmortems increase the status of the person who found the problem early. A respected senior person can say "yeah, I fucked that up" and remain respected.
Politics, pride, career incentives, conformity, and bad judgment stay.
The checks change which behavior is locally survivable.
Graham's control rods metaphor works because the chain reaction is the point. A merely corrigible organization is enough. The chain needs enough interruption that one local distortion struggles to reproduce itself into the next generation.
The best check may be brutally simple: keep some important tests close to the outside world.
Does the customer stay?
Does the software work?
Did the launch actually improve anything?
Can the person doing the work tell the decision-maker what happened directly, with little ego management?
Reality is wonderfully rude when it retains veto power.
The local game eats the real game
A company can die while everybody inside it gets better at their jobs.
They get better at forecasting what leadership wants.
They get better at selecting projects that survive planning.
They get better at presenting ambiguity as progress.
They get better at hiring people who feel like strong hires inside the current culture.
They get better at avoiding career-limiting surprises.
Every skill can be real.
The tragedy is that the local game has drifted away from the real one.
At the beginning, people bend themselves a little to survive the organization.
Later, the organization depends on people who are unusually good at bending themselves that way.
Then those people choose successors.
Eventually you can have a building full of people who are excellent at succeeding inside the company and a company that's terrible at succeeding anywhere else.
Ruin emerged from ordinary local wins.
The company taught them how.