Hacker News Reader

Comment by tiffanyh | Hacker News Reader

Is DarkSky not an actual weather app?

https://news.ycombinator.com/item?id=35268026

> Meteorologists seemed to feel that Dark Sky was a graphics processing tool, not a weather app.

---

  “Any weather forecast beyond a couple of hours ... depend on supercomputer models that work according to the laws of physics ... But when we talk about Dark Sky, all it was doing was taking the visual input of the radar and extrapolating what was going to happen over the next couple of hours.”

https://slate.com/technology/2022/12/dark-sky-weather-app-ap...

View on HN

If that's all it took for Dark Sky on the backend to do what it did, then there's even less excuse for the lack of a good replacement in 2024.

Unlike the meteorologists, darksky actually worked and would tell you when it was about to rain

Every time I looked at the animated weather app, I would see the animation start in the past where clouds moved and grew and shrank naturally. As soon as it passed the present moment, the clouds would become fixed shapes and continue on whatever their current vector is. In the visualization there was no attempt to model clouds growing and shrinking. The clouds would suddenly start skidding across the screen.

I've read that is what the underlying precipitation "models" did as well, but obviously can't confirm.

Yes; they basically just extrapolated from these "rain blobs" on the visualization as the short-term forecast they provided to users. There are some long-since wiped blog posts that provide a bit more context on how they do a little bit of statistical processing of the general forecast model output to help with they "hyper-localization," but the reality is that it was terribly unsophisticated relatively to what is traditionally done in meteorology.

The rain nowcasting feature that Dark Sky popularized is now table stakes in any consumer weather app. There's little value in making these types of forecasts any more complex (e.g. using AI or other contemporary techniques) because they still have egregious and noticeable failure modes. And it's so trivial to make this type of forecast that there is open source software you can easily run to do it [1].

[1]: https://pysteps.github.io/

Look at windy.com, their cloud prediction works pretty much the same.

Except it did that and communicated about it better than any other "weather app"

Seems like an strawman comparison to me?

Dark Sky's hyper-local weather predictions weren't based on projecting hours into the future. They were based on very small-scale extrapolation and interpolation.

The meteorologists with their supercomputers aren't, to the best of my knowledge, even trying to do that. And Dark Sky wasn't trying to compete with their models, either. For that kind of long-term forecasting, it was simply passing along the forecasts that meteorological services used those models to produce.

How does that make it "not a weather app"? This is just dumb gatekeeping.

I imagine because it operated counter to basic meteorological theory.

The difference is that 15 minute rain forecasts don't need that to be usefully accurate.

A model of equivalent simplicity would be about as accurate as the farmer's almanac for a 10 day forecast.

It doesn't run counter to meteorological theory. It just doesn't apply sophisticated long term models.

Simpler models are fine for very short term forecasting which is what they were doing.

This is just weather modellers getting annoyed because someone else is showing them up - not with 10 day atmospheric pressure forecast accuracy, which only weather modellers care about - but with "is it going to rain in 2 hours" which normal people really care about.

Is predicting beyond a couple hours all that useful/important?

For the eclipse, I was using hyperlocal forecasts for the next day (and next two) for planning for which direction to go, where to stay, and where to head the next day.

The morning of the eclipse it was a question of "this town or that town or that other town at this time."

This may be an edge case, but there are certainly times where prediction for a certain time 24h or 48h in advance is useful.

What were you using for hyperlocal projective forecasts that far in advance?

Weather app, yr.no (e.g. https://www.yr.no/en/details/table/2-4235193/United%20States... ) , myradar, and pivotalweather (which gave access to a number of different models). In the case of things that gave maps, it was "look at the location on the map".

It wasn't necessarily "I want to know the weather for this block" but "I want to know the weather for this square mile."

One of the features that I liked in Dark Sky of very old (it was removed well before Apple decided to buy them) was the Lines feature which showed the models that it used. Knowing how to get at and read the models themselves (and asking the weather geek who I work with about the strengths and weaknesses of each model - "well X is good, but it over predicts precipitation and cloud cover within 50 miles of a lake that can produce lake effects, but if you're not near a lake its spot on for...").

Reply on news.ycombinator.com