The Slot Machine With Better Graphics

I watched an engineer fail eleven times in under an hour. Eleven AI-generated attempts at the same function, each one discarded the moment it didn’t compile or didn’t pass the test, replaced instantly by the next generation. He called it moving fast, and by the metric he had in his head, he was right. When I asked what had actually been wrong with attempt three, he couldn’t tell me. He had already moved four attempts past it by the time I asked, and nothing about the process had asked him to remember.

"Fail fast" did not originate as permission to fail more. It came out of lean-startup practice, and the fast was never the point, it was a means to an end that everyone quotes and almost nobody finishes the sentence on: fail fast so you can learn fast, and stop spending real resources on an assumption that a cheap test could have killed for you first. The entire value of the philosophy lived in the second half. Speed bought you cheap access to the lesson. Nobody was supposed to want the failure itself.

What made that discipline durable for the two decades before AI showed up was not virtue. It was cost. Building the next attempt took real time, so the time it took was, whether anyone designed it this way or not, exactly the pause a person needed to ask what went wrong before spending more effort on a variation of the same mistake. Expensive iteration accidentally enforced reflection. Nobody had to be disciplined about it. The economics did the discipline for you.

AI dissolved that economics without replacing what it had been quietly providing. Generating the next attempt now costs nothing, so the pause that used to be built into the cost of trying again simply isn’t there anymore, and nothing else has stepped in to hold its place. The chain used to run: attempt, failure, forced pause, diagnosis, better attempt. Now it runs: attempt, failure, attempt, failure, attempt, failure, as many times as the second matters to you, and the diagnosis step is the one link nobody rebuilt, because it was never a step anyone had to design in the first place. It came free with expensive failure. Cheap failure doesn’t include it.

I want to be precise about what is actually missing, because it is not effort and it is not intelligence. It is a specific, almost boring habit: stating in one sentence, before generating anything else, exactly what was wrong with the attempt that just failed. That sentence is the entire mechanism by which a failure becomes information instead of noise. Skip it, and a hundred fast failures teach you nothing more than one slow one did, because volume was never the variable that mattered. Diagnosis was, and diagnosis takes a specific kind of attention that generating a new attempt does not require and, if you let it, will happily let you skip.

The test I now give any engineer leaning hard on AI is really two questions, asked out loud, every time, in order. First: what was actually wrong with the attempt that just failed, in one sentence. Second, and this is the one people skip even when they can answer the first: what specific change would fix it. AI will generate you the next attempt for free, as many times as you ask. It will not do either of those two things for you, and mistaking the free part for the whole job is exactly how a hundred fast failures end up teaching a person less than one slow one used to.

None of this is an argument for slowing down for its own sake, which would be its own kind of confusion. It’s an argument for relocating a discipline that used to live in the economics of the tool into the habits of the person using it, because the tool will not volunteer it back. Failing fast was never the hard part, and AI just made it trivial for everyone. What actually separates a good engineer now is the pair of skills sitting on the other side of that failure: naming precisely why something broke, and knowing precisely what fixes it. Those are the two things I watch for now, in myself and in anyone I hire. Everyone else is just getting really, really fast at pulling the same lever.