> ## Content Index
> Fetch the complete content index at: https://somethingsneverchange.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Speed Trap
- URL: https://somethingsneverchange.ghost.io/the-speed-trap/
- Published: 2026-08-04T15:14:00.000Z
- Updated: 2026-08-18T12:51:11.000Z
- Description: Building has phases, and each needs its own speed. Going fast and going somewhere are not the same thing.
- Author: Siddharth Joshi

*"Move fast and break things"* once sounded like rebellion. Now it sounds like the rule. Every founder deck, investor memo, and product all-hands treats velocity as proof of life. Among the people funding consumer AI, the line has hardened into doctrine: the edge is no longer the model, it is the speed of iteration around it. Ship before the incumbent's roadmap clears review. Be first to production and the market is yours.

In the short run this works often enough to look like law. Some are making real money by being first, and I won't pretend otherwise. But *"it makes money this year"* and *"it is the right way to build"* are different claims. 

### **What speed actually is**

Speed is a measure, not a direction. A speedometer tells you how fast you are going and nothing about where. For most of history, going fast was expensive. **You had to want a destination badly enough to pay for the motion.**

AI removes the cost. A prototype that took weeks takes a night; experiments and coding agents run while you sleep; the price of another try falls close to zero. This is not the first cycle to make speed cheap, but it is the most extreme version of a long pattern. Mobile compressed years into months, cloud compressed months into weeks, AI compresses the week into a night. 

What repeats is not the technology but the temptation: to mistake the measure for the direction, to feel that going fast is the same as going somewhere worth going. *It isn't, and the confusion has a cost. Left ungoverned, speed distorts judgment, compounds bias, and normalises risk. Each is worth seeing on its own.*

**Speed distorts judgment**. In February 2023, with its leadership reportedly in a *"code red"* over ChatGPT, Google published a short demo of its new chatbot, Bard. The clip was the company's own showcase, hand-picked to make the case for the product. In it, Bard claimed the James Webb Space Telescope had taken the first picture of a planet beyond our solar system. It had not; the European Southern Observatory's Very Large Telescope did, in 2004\. Reuters caught it within two days. [Alphabet's shares fell around 8%](https://www.reuters.com/technology/google-ai-chatbot-bard-offers-inaccurate-information-company-ad-2023-02-08/?ref=somethingsneverchange.ghost.io), erasing roughly $100 billion in market value. Nobody at Google had become less capable that week. The race had simply compressed the schedule, skipping the check that catches a wrong fact in your own advertisement.

**Speed compounds bias.** In 2021, Zillow's iBuying arm used an algorithmic pricing model to make instant cash offers on US homes. As the housing market shifted, the model began over-paying: it was tuning on its own recent acquisitions while broader market signals lagged. By the time the bias was caught, Zillow had bought thousands of homes it couldn't sell. [The write-down was $304 million](https://www.bloomberg.com/news/articles/2021-11-03/zillow-s-drop-brings-market-value-loss-to-30-billion-from-peak?ref=somethingsneverchange.ghost.io); Zillow Offers was shut and $30 billion wiped from market value. Two thousand people lost their jobs to a price the model set faster than anyone could check. Nobody at Zillow set out to build a model that over-paid. A small skew in the data, amplified one cycle at a time, went out as prices. The bias was always possible. Speed is what lets it propagate before correction arrives.

**Speed normalises risk.** Bias compounds inside one loop. Risk compounds across many. The clearest study of this is not from technology at all. In 1996 the sociologist [Diane Vaughan](https://press.uchicago.edu/ucp/books/book/chicago/C/bo22781921.html?ref=somethingsneverchange.ghost.io) examined why the space shuttle Challenger broke apart 73 seconds after launch on a cold January morning in 1986, killing all seven aboard. The seals on the solid rocket boosters, the O-rings, were known to be vulnerable in cold. Engineers had watched them erode on earlier flights. But each launch that survived the anomaly quietly reclassified it: what began as an alarming deviation became, flight by flight, an accepted feature of how things worked. Under constant schedule pressure, the baseline of "acceptable" drifted. The night before, engineers warned against launching in the forecast cold; managers overruled them. Vaughan called the pattern the normalization of deviance: an organization talking itself, one survived shortcut at a time, into treating a known danger as routine.

The mechanism is not exotic, and it is not confined to rockets. Every team has a check it skips because nothing broke the last ten times: the eval it postpones, the review it waves through, the edge case it files under "later." Speed feeds this directly. The faster you ship without consequence, the more the corner you keep cutting feels safe to cut, right up to the morning it isn't.

![](https://storage.ghost.io/c/97/1d/971d6e15-b552-48fb-92f4-cbc697e5bd5b/content/images/2026/08/st1.png)

### **The different speeds of building**

The strongest case for speed is not that it makes money. It is that speed is itself the check: shipping to real users beats internal review, because the market corrects faster than any process. And the market did: Bard's error was public within two days. But that correction works only where the error is correctable: what Jeff Bezos called two-way doors, decisions you can walk back. Zillow's balance sheet was not a two-way door. Neither was Challenger. The market corrects fast; it does not resurrect. 

None of this is an argument against speed. It is an argument against treating all speed as one thing. Building is not a single race run at a single pace. It has phases, and each phase rewards a different tempo. Every phase is faster than it used to be, but they are not the same. The trap has a name: borrowed tempo, running one phase at another phase's tempo.

**Exploration: play fast. *What can we build?***  
Early on, while you are still hunting for the problem worth solving, speed is an asset. This is where AI earns its reputation: generate, research, simulate, remix, throw away. The cost of a try has fallen far enough - and the skill it takes, low enough - that quantity becomes its own signal. Build a prototype in a day. Run a hundred rough versions and the shape of the answer starts to show, if the questions were good. Going slow here is the mistake. Curiosity and divergence are the point; caution only narrows the search before you've seen enough of it.

**Validation: slow down to see clearly. *What should we build?***  
Once a hypothesis forms, the tempo has to change. This is the shift most teams skip. The questions split into two. Does it work: is it accurate, where does it break, can it be trusted? And should it exist: does the customer actually want it, what value does it create, what are we displacing? Both questions need the slower tempo, and for the same reason: speed here doesn't save time; it spends it. It lets you confirm what you hoped, ahead of the data that would have corrected you. The work is still be faster than before, workflows, synthetic users and automated evaluation compress it. But faster than before does not mean running at exploration tempo.

At Amazon, with billions of products across thousands of categories, we moved fast inside each area - home, fashion, beauty, electronics - and then collided as we scaled, because no one had stepped back to ask how a customer who shops across all of them should be served. We shipped our org chart, not their experience. In 2019, with new 3D technology and COVID pushing customers online, we shipped a stack of ways to view a product - in your room, on your body, on your face, in a virtual space - before deciding where each belonged, or which to reach for when. Beauty and fashion try-on were even separate products. The speed was real. So was the cognitive load we handed the customer to sort out what we hadn't.

**Scaling: governed acceleration. *How should we build?***  
Once something is proven, you accelerate again, with steering built in. A useful distinction comes from my time at Amazon: some decisions are two-way doors, reversible, and should be made fast by whoever stands closest to them; others are one-way doors, irreversible, and deserve deliberation. A frontier model launch is today's clearest one-way door: Anthropic could restrict Fable after release, but there was no going back from launching it.

The same pull shows up at the door you open to grow. In my early years at Amazon, we opened the catalogue to a flood of new selling partners, and selection expanded overnight. Listing and data quality fell, established sellers were squeezed, and customers got a bigger store that was harder to trust. The selection metric went up. The thing the metric was standing in for went down. The aim is not to move slowly; it is to move wisely fast: fast enough to learn, slow enough to last. The skill is knowing which phase you are in, and refusing to borrow one phase's tempo for another. 

![](https://storage.ghost.io/c/97/1d/971d6e15-b552-48fb-92f4-cbc697e5bd5b/content/images/2026/08/st2.png)

The failure at scale is treating a one-way door like a two-way one: pushing an irreversible change at exploration speed. The edge case that was harmless in a demo becomes a systemic risk when it runs a million times an hour. Scaling well is not slowness. It is speed with review loops, fail-safes, and a clear sense of which doors you are walking through.

### Beyond the gospel of speed.

Speed is free now; nearly everyone has it. A thing that everyone has is not an advantage. What stays scarce is direction under speed: knowing when to sprint, when to slow down to see, and when to stop and think. That judgment is not faster than it used to be, and it never will be, because it is the part of the work that cannot be compressed without being lost. Speed measures the motion. It still falls to us to choose where we go.

![](https://storage.ghost.io/c/97/1d/971d6e15-b552-48fb-92f4-cbc697e5bd5b/content/images/2026/08/st3.png)

---

***Next:*** [***What Play Leaves Behind***](https://somethingsneverchange.ghost.io/what-play-leaves-behind/)***. The most serious capability of the AI era is trained by the least serious-looking work.***