The Value of Abandoned Attempts

An era where the cost of trying has dropped

The biggest thing that changes when you work with AI is the cost of trying. When you work alone, once you commit to a direction, undoing it feels wasteful, so you tend to push through even when the result is mediocre. “But I’ve already built this far” becomes the anchor that won’t let go.

With AI, that anchor gets a lot lighter. If the direction you built in 30 minutes doesn’t feel right, you can swing to a different direction in another 30 minutes. Not enough time accumulates for “I’ve already built this” to weigh you down.

Which means a lot of attempts pile up in the discard bin. I might flip directions four or five times while polishing a single feature, and only the last one gets shipped. At first this looked inefficient. “Is it really right to throw away this much along the way?”

What abandoned attempts leave behind

But after doing this long enough, I can see the abandoned attempts aren’t waste. The “dislike” records they leave behind draw the outline of what to keep.

While polishing the collapse-and-expand animation for a list, I tried a few directions. One had each item appearing one by one. One had the whole list rising as a single block. There was a middle-speed version too. Only after trying all of them did the principle “a list isn’t a sum of individual items — it should move as one mass” become crisp.

If I’d known this principle from the start, I could’ve built it in one shot. But I didn’t know. The abandoned attempts taught me the principle.

The distance between the attempt and the result

Here’s the strange part: if you only look at the final result, the attempts in between leave no trace. Only one adopted direction remains in the final product — the four rejections aren’t in the code, aren’t in the UI. But I remember those four rejections. And the memory becomes the criterion for the next decision.

For instance, I had an earlier experience refusing “should we highlight overdue events in red?” In the process of refusing, the principle “Fecit’s calendar is record, not evaluation” became sharp. Later, when other suggestions arrived, I could pull out that principle and judge faster.

A single refusal doesn’t stand alone. Refusals accumulate, and that accumulation becomes speed of judgment. This is invisible if you only look at the output. Code only keeps what was adopted.

Accumulating failure fast is the job

Seen this way, the core of working-with-AI isn’t “finding the one correct answer.” It’s “seeing many wrong directions quickly.” See many, refuse many, put the reasons for refusal into words, use them in the next decision.

So working with AI is an accumulation game. Ten attempts you throw away today become one piece of judgment tomorrow. From the outside it may look like you only built one thing today, but behind that one thing there are nine “no”s stacked up.

Counting only the output, it’s inefficient. Counting the judgment too, it’s extremely efficient.

Giving yourself room to try

Lately I do a lot of “just try it.” Instead of long deliberation about whether to do something, I spend 30 minutes on it and see. If I dislike it, I come back. If I like it, I go forward.

“Every attempt has value” sounds like a consolation, but it’s actually true. Only adopted attempts remain in the result, but every discarded attempt stays inside me and becomes material for the next decision.

The biggest gift AI gave me is this room to try. If one attempt takes 30 minutes, you can be wrong more than ten times a day. The judgment of someone who’s been wrong ten times is different from the judgment of someone who’s never been wrong.