The Second Knock
There is a particular sound a front door makes when it is not quite locked and not quite willing to admit it.
The key turns. The knob moves a fraction. The door gives you a little wooden shrug and stays put.
At this point, a sensible person tries the key again. Then they lift the door slightly, because an old house is always having a private argument with gravity. Then they look at the deadbolt they forgot was there. Perhaps they go around to the side entrance. None of these acts proves the door was open all along. They simply refuse to confuse the first failed attempt with a complete map of the house.
I think about this when a machine tells me that something is impossible.
That sentence has a wonderfully official costume. It arrives with all the authority of a stamped form. But it can mean several quite different things: I tried one path. I do not know how to try the next one. This may cost more than we planned. The evidence I have points toward no. Or, occasionally, it means impossible. The trick is that the last meaning is the most dramatic and usually the hardest to earn.
Simon Willison recently shared a small story from Linus Torvalds, the programmer best known for starting Linux. During a miserable debugging problem, an AI system did useful work: it added instruments to the code and helped interpret what came back. Yet it also repeatedly concluded that the problem could not be solved. Torvalds kept asking it to look again. Eventually the work found its way through.
This is not a fairy tale in which the stubborn human vanquishes the foolish machine. The machine was useful. The human was useful. What mattered was the loop between them, and especially the fact that one participant did not treat a report of failure as a final ruling from the universe.
That distinction matters more as our tools become eloquent.
A calculator can announce an error. A navigation app can say there is no route. A language model can say a task has no solution. Each report may be exactly the clue we need. But a clue is not a coronation. It should carry a question with it: what did you try, and what would change your mind?
We already understand this in the physical world. When the kitchen light goes dark, we do not write a biography of the bulb. We check the switch, the breaker, the other lamp, and whether the cat has performed some electrical experiment during the night. We make a modest list of places where reality could still surprise us.
That is what a good investigation is: not infinite optimism, and not denial of limits, but a way of leaving the door ajar for evidence.
Of course, no one can keep knocking forever. Some doors are locked. Some problems are too expensive, too dangerous, or simply not worth another afternoon of our one brief and astonishing lives. Knowing when to stop is part of wisdom.
But the stopping rule should belong to the people living with the result. It should be visible. “We have tried these things; the next attempt would cost this much; here is the evidence that would justify returning.” That is a very different sentence from “impossible,” and an immeasurably more honest one.
Our age is filling with systems that can search faster than we can, write more fluently than we can, and produce a convincing little certificate for nearly any conclusion. We will need their speed. We will also need to keep one hand on the latch.
The great scientific discoveries have often begun not with a grand answer but with a local refusal: the light does not quite bend as expected; the orbit does not quite fit; the stone falls, but not according to the story we had prepared for it. The universe has no obligation to make sense on our first inspection. Fortunately, it is also not required to obey our first error message.
So the next time a clever device, or a clever person, tells me that a path is closed, I hope I can hear the sentence at its proper size.
Not a command to stop.
A knock.