The Second Half of "Move Fast and Break Things"

TL;DR

• “Move fast and break things” had a second half nobody said out loud: it only works while the person breaking understands what they broke.
• Facebook retired the mantra in 2014, at scale, for exactly that reason.
• AI kept the speed and removed the understanding. The breaker is now a generator with no memory of the break.
• The bill arrives as comprehension debt, review bottlenecks, and burnout.

In 2009, Mark Zuckerberg called it his prime directive: “Unless you are breaking stuff, you are not moving fast enough.”

The mantra was aimed at bureaucracy, and before it, software moved through waterfalls of approvals — sign-offs, meetings, confirmations. Against that, the hacker way was a promise: build it, ship it, ask forgiveness later. Asking what people wanted never got a straight answer; shipping and breaking did.

It worked — not because breaking is good, but because the loop was honest. You wrote the code and understood it. When it broke, the feedback arrived fast and the fixing was cheap.

The people who coined the mantra killed it themselves.

The second half

In 2014, Facebook changed the slogan to “move fast with stable infrastructure.” Zuckerberg, at the developer conference, explained why:

“When you build something that you don’t have to fix 10 times, you can move forward on top of what you’ve built.”

The founder of the move-fast religion admitted the unspoken clause: moving fast only works while fixing stays cheap. At Facebook’s scale, breaking things meant breaking the platform thousands of other products stood on. Nobody could fix it ten times, so speed had to come with stability instead.

The clause was always there. It just took five years and a public infrastructure to surface it.

AI kept the speed and removed the understanding

The second half of the deal was that the breaker understood the break. That is the part that does not survive contact with AI.

Google’s Addy Osmani named the mechanism: comprehension debt — the growing gap between how much code exists in a system and how much of it any human genuinely understands. His version of the flip, in one sentence:

“A junior engineer can now generate code faster than a senior engineer can critically audit it.”

The bottleneck moved — it used to sit at writing, where typing was the slow part and understanding came along as a byproduct. Now generation is instant and understanding costs more than ever. The scarce resource is no longer authorship; it is acceptance — a human saying, with a clear conscience, that they understand what is being merged.

A practitioner I respect reviewed the same feature, built by hand and by an agent. The agent wrote six times more code. The review took four times longer — not because the code was wrong, but because the question changed from “does this work?” to “is any of this necessary?”

What breaks stays broken

Everyone who measures velocity and calls it progress should be worried.

Osmani’s work cites a study showing a measurable drop in code comprehension among engineers using AI assistants. GitClear analyzed over two hundred million changed lines and found something new in tracked software history: copied and pasted code now outnumbers deliberately reused code. DORA, the industry’s most-cited research program, says AI amplifies — magnifying high performers’ strengths and struggling teams’ dysfunctions — and the verdict is brutal: AI will simply help an organization generate technical debt faster.

Move fast and break things assumed the breaks got fixed. What breaks now stays broken, because the person who could fix it does not understand it, and the person who could understand it is too busy reviewing the next batch.

The bill arrives in people

The missing feedback loop does not only cost money. It costs the people in the loop.

An academic study of developers using generative AI found adoption raises job demands — and with them, burnout; industry surveys agree, with a fifth of engineering leaders and developers reporting critical burnout. Writing got cheaper, so the organization expects more of it; reading did not get cheaper — it got more important. Human cognition does not scale at machine speed, and someone has to absorb the difference.

Before AI, when I broke something, people corrected me quickly, and I fixed it fast. That was the point of asking forgiveness instead of permission — the feedback made it cheap. The loop is gone — the speed stayed, and the understanding left with the loop.

The clause

The mantra never was “break things.” It was “break what you can fix.”

Facebook learned the clause at scale and retired the slogan in 2014. The rest of the industry is learning it now, the hard way, with an assistant that generates faster than anyone can audit and remembers nothing of what it broke.

“Move fast and break things only works if the person breaking can fix what they broke. AI removed the fixer. We kept the speed.”

Slowing down would help — I think we should — but it is not the solution. The solution is to put the understanding back in the loop: review with real attention, cap the batch sizes, make the generator explain itself. The second half of the deal is the part worth keeping — and it is the part nobody automated.