The problem is that "impact" is easy to establish for shiny new things, whereas for preventing problems – you know, maintaining things – the impact is how much

The problem is that "impact" is easy to establish for shiny new things, whereas for preventing problems – you know, maintaining things – the impact is how much better you have made the universe compared to what the universe would have been without your work. Of course, you have done your work, so we never get to observe that other, "counterfactual" universe that establishes your actual impact. To estimate the impact, we need to run controlled experiments. And we have lots of A/B testing infrastructure to help us run those experiments when it comes to launching shiny new things. But when it comes to maintaining things, not so much. So if there are two possible projects of equal difficulty and impact that you could take on, and one is to launch a shiny new thing and the other is to maintain an existing thing, you're always better off launching the shiny new thing because the cost of quantifying your impact will be low, whereas had you chosen the maintenance project, quantifying your impact would be prohibitively expensive. And that's why building shiny new things dominates maintaining existing things: you get full credit for your work, not a small fraction.