I'm just one person, and my programming needs are somewhat unusual (building various kinds of statistical forecasting models). But I'm just not seeing the consistent productivity gains from LLMs that I would have expected if you'd asked me 6 months ago.

For low-complexity tasks, they're almost always helpful. For medium-complexity tasks, probably split about evenly between 1) works like magic on the first try; 2) wrong but directionally helpful or works after refinements, and 3) proves to be a net waste of time.

They're usually worth "gambling" on. If it would take me 30 minutes to figure something out on my own, and 1) saves me 25 minutes, 2) saves me 15 minutes and 3) costs me 30 minutes, the expected value is positive. But there's still quite a bit of 3).