> WeatherNext 3 will power weather features in Google Search, the Gemini app and Google Maps, as well as the Google Maps Weather API and Google Earth Engine.
- Units are imperial rather than metric (i.e. not basing it on users locale)
- The concept of init time is not initially clear at all, so it's confusing when you click the calendar icon to see the weather forecast for a future date only to see it isn't an option, you have to use the slider.
- When changing the sidebar to the "detailed" view, most of the hovering element is cut off by the container so you can't actually read it.
And it took longer to write this comment than it did to find these issues. It's fine since it's clearly marked as experimental, but I disagree that it's "really easy to use" :p
Init time makes sense to me, it's when the forecast/model is initialized, letting you look back in the past to watch the change to the present and future.
The sidebar thing is weird, but it also can expand out.
> Would that be practical for weather forecasting or not really?
What are the conditions at 5000 feet, 10000, etc? What is the location of the jet stream and its strength? The reduced number of (e.g.) weather balloons is hindering forecasts (per the links).
A barometer will give you surface pressure, but that's a field that tends to vary relatively slowly over the surface of the Earth. The calibrated weather stations that exist at every airstrip do a reasonable job of providing these conditions over land, and the residual of "pressure from phones" probably won't help all that much.
The data that would be most valuable to initial conditions is upper-atmosphere winds -- this is the kind of data given by weather balloons. In clear air there's no great way to measure this from either the ground or from space.
One important supplemental data source here are aviation reports, from planes flying at altitude and particularly trans-oceanic routes. When air traffic was largely curtailed during the early phase of the Covid pandemic, weather forecasting suffered a bit for the lack of data (see eg https://www.ecmwf.int/en/about/media-centre/news/2020/drop-a...).
The online viewer could really, really use Wind Direction as a compass bearing. Its super important considering wildfire/bushfire, air quality, ocean-going conditions, and a myriad of other things.
It's great that they're working on this, but I am puzzled at how absolutely awful the forecasts are in the Google weather app for my area. The forecast will show no rain, the radar view shows nothing, meanwhile it's pouring outside and every other app I check shows it. I know I can't expect it to be perfect, but being terribly wrong even one in ten times is enough to tarnish its reputation permanently.
Weather 2 doesn't seem to have been an ensemble model. Weather 3 is, so theoretically it can get more accurate outcomes by taking the probabilistic analysis of several models concurrently to determine the most likely weather conditions.
I'm building a tool right now that uses an ensemble to determine wind gust likelihood, which is useful for safety critical work on construction sites and the like.
WeatherNext 2 was based on the FGN architecture described in [1]. It was explicitly designed and trained to produce ensemble forecasts (it was trained in such a way that the output ensemble optimized a CRPS metrics). In fact, it was a set of 4 different model weights, each of which was seeded with a random noise vector to produce an array of 16 forecasts for a total of 64 ensemble members. WeatherNext 3 trimmed that down from 4 to 2 separate model weights to use.
The Google daily and hourly forecasts aren't great, but the weather map forecast is excellent. Where does that come from? What's generating the predicted movement of the storm clouds?
This isn't a super serious comment (and an opportunity for someone to speak up on this) but...
I've heard some commentary in the past that 5G interferes with data measurements heavily used for weather forecasts (something about satellite measurements?)
I remember being _very_ impressed like 15 years ago about how I would have _hourly_ weather forecasts (in particular around the rain) that seemed like magic! And then things... seemed to slowly get worse (at least in Tokyo)
a couple years ago I was chatting with a friend in Kyoto. They used to live in Tokyo and had made the _exactly_ aligned comment like: "I was used to the rain forecasts being not so accurate anymore. After moving to Kyoto they seemed to be quite good! And now they're also bad here. Am I cursed?"
I looked at some 5G rollout maps and you could see Kyoto rollout happened a bit around the time frame they were complaining about....
Anyways I think for most people (at least for myself) weather forecasting seems like this odd dark magic that can't work at all, but there was a window in which it _felt_ like it was super accurate. At least in my personal experience
> I've heard some commentary in the past that 5G interferes with data measurements heavily used for weather forecasts (something about satellite measurements?)
The short version of the "spectrum" issue is that 5G is being allocated in bands very close to the microwave spectra where atmospheric water vapor emits. A bevy of public and commercial satellites in low-Earth orbit passively monitor these microwave spectra and produce extremely important information that is assimilated into numerical weather models.
The federal government sets limits on "out-of-band emissions" for operators emitting in the allocated 5G spectra. These emissions can bleed over into the microwave water vapor bands, creating noise that masks the natural presence of water vapor. The limits for this out-of-band emission is on the order of ~10-20 dB, and there's some work in the atmospheric science literature suggesting that this is enough to confound certain water vapor retrievals. That basically means we lose these observations that help constrain the forecast.
There isn't much indication that this is a serious issue in day-to-day meteorology at the moment. But it's an issue which will be significantly more difficult to unroll and claw back than to simply protect key spectra in the first place.
A couple of years in Tokyo gave me great trust in certain weather apps for eerily precise minute-by-minute rainfall predictions. After a few years back in Sydney, my trust in them has dropped back down to the usual uncertain baseline, and my reliance on them has been mostly replaced with my own rough assessment of the air and sky.
On recent visits to Tokyo, with my current habits, the general crowd (who presumably check some app) has been the more reliable forecast. Either everyone has an umbrella in their hand or they don't. And I discover whether I should have brought my umbrella or if I'll be lugging it around pointlessly, just far enough from my hotel to be stuck with my choice
does google actually use their own models in consumer products? i'm not an android user, but the weather widget in the google search results for my area always seems to be attributed to weather.com.
Same here, if its raining Windy.com radar will always show it, but often Google Weather does not show it, despite saying its updated "now". Sometimes it seems to be a tiling issue, where there will be a blob of rain on the map but cut off at some boundary over the forecast area.
At one point they did release the most accurate weather model by far, but it actually decreased usage compared to showing a weather forecast more skewed towards "happy" temperatures and weather. I'm certain this is still the case and likely explains why you are seeing a more positive (no rain) version of the weather.
It's because the current administration cut NOAAs funding, which means fewer weather balloons, which means less data to make a prediction. And it means less real time updates.
So yeah, when it rains, it might take a few hours for that to flow into your weather app.
There is little to no evidence that the current degradation of the US upper air backbone is regularly contributing to degraded forecast skill. That might change as we head into the more active northern hemisphere winter.
They are, as are plenty of folks in the community. I expect to see some good talks on this at the AMS Annual Meeting in January, based solely on my own peer network and what colleagues have mentioned they're working on.
I've stopped using google weather here in the UK as well because of how inaccurate its been over the past few years and have just fallen back on the BBC. Maybe its related, maybe not.
Yeah, the Met Office and BBC (which I believe use data from Meteo France) are both much more accurate for "later today" weather than Google for me in London.
Google does update its forecast more frequently, and usually is as accurate as the others for the hour ahead. And the longer range 3+ day forecasts aren't noticeably worse either. It's just the intermediate 4 - 48 hour stuff that they're surprisingly terrible with, but that's exactly the time period I most care about in everyday use!
I'm interested on how you drew the cause/effect of this, is there non-biased sources? Has google said as much, and how does this explain why some models are more accurate?
The only forecasted rainfall for local areas within a very short timeframe. This made it possible to utilize simple factors to predict the near future.
It is built into Apple Weather after Apple purchased them.
While that might be true but apple weather still fails compared to dark sky. I still don't use it for radar like I did with dark sky - they effed up the UI.
> It is built into Apple Weather after Apple purchased them.
Did Apple lobotomize the tech when they integrated it, or lose access to whatever upstream data Dark Sky had access to? Apple Weather is comically bad (though not nearly as bad as Google's weather searches).
Apple Weather (followed by Xweather (formerly Aeris Weather)) seems the most accurate forecast for Prague when it comes to TP/FP for rain in Prague by my long term experience when I compared multiple data sources
Something changed because Apple Weather has never been as accurate as Dark Sky for me, especially for rain forecast.
I used to be able to go on walks without an umbrella because of how reliable it was at predicting when rain was going to restart. If I do that with Apple Weather, I get home wet.
Dark Sky is still one of my all time favorite app. The accuracy of the forecast within the coming hour is to this day unmatched. It constantly reminded me of Back to the Future 2, where Doc says the rain will stop in “in exactly 5 seconds”
well, that's interesting. the design it's a bit of a letdown - very very very low information density, lots of horizontal scrolling, a tab bar item dedicated to notifications (which one never changes) and another one dedicated to community (which is empty of content). Hopefully good data and good notifications make up for the poor design.
Dark sky was awful and all of its acclaim comes from confirmation bias. In seattle I'd see its forecast change by the minute. The whole point of a forecast is to know what will happen and if that keeps changing, then we do not know.
I never used that, but I like Rain Parrot in Australia. Not sure if it's similar. It will tell me that it's likely to rain in x minutes, then stop y minutes later, and it's almost always fairly accurate.
"We create a 3-channel image where the red channel represents velocity in the x-direction, the blue channel represents velocity in the y-direction, and the green channel represents change in storm intensity" and then use computer vision techniques to predict what happens next
> It’s crazy we’ll never have forecasts as good as dark sky again.
'Nowcasting' is an area of active research, both with machine learning and with physics-informed or visual flow approaches.
Part of the problem from the machine learning side is that these are _huge_ problems. NVidia's StormCast (https://research.nvidia.com/publication/2024-08_kilometer-sc...) works globally at kilometer scales, and you can imagine how big those grids are. Even with patch training, you're dealing with very large datasets.
At the same time, this is not exactly a high-profile area of research. National weather centres focus on actionable medium-range weather predictions, and meteorologists can look at radar themselves and perform mark-one-eyeball predictions for very short-range watches and warnings. Some private-sector actors will pay for short range predictions, but they're often looking for something hyperlocalized (e.g. weather at this particular construction site, for crane safety) or specialized (near-real-time cloud and wind predictions for renewable energy).
Most public-accessible weather predictions are downstream of either a public-sector effort (which doesn't internalize benefits, leading to under-resourcing) or a byproduct of another private-sector offering.
There is a great article from the German Meteorological Service on WeatherNext 3 and how AI-based weather models will probably co-exist with physics-based ones.
There's another tool which seemed to coordinate with the launch of WeatherNext 3 last week, the "Operational WeatehrBench" from Brightband -> https://owb.brightband.com/
It features a couple of AI and NWP model forecasts for comparison.
This seems to be a bit of "we're throwing a _bunch_ of inputs into this machine learning set and then pulling out outputs". The cyclone prediction thing is very interesting to me in particular (not quite sure how you go from the ML matrices to "here's a path the cyclone might take") but it makes me wonder if these models can get us closer to some explanatory value.
I imagine a lot of predictive sciences are ultimately about mixing together a bunch of inputs to attempt to decipher some output. Do we end up being able to take stuff from here and figure out some new ideas about modelling the climate as a whole?
Sure we do, you can follow the weights so to say, to see which data turns out to be more important vs less important for predictions of a higher quality
> The cyclone prediction thing is very interesting to me in particular (not quite sure how you go from the ML matrices to "here's a path the cyclone might take")
In a high-level view, it's the result of specialized decoding heads.
Traditionally one would take gridded forecast outputs, then process those with comprehensible actions like "find all local pressure minima in the ocean, then filter to ones which correspond to warm cores, etc." to infer (diagnose) the presence of a cyclone.
One problem with this is that gridded forecasts suffer from known biases and tradeoffs. For example, a forecast on a ~25km grid is just on the edge of being able to represent the eye of a hurricane (50km scales), and it certainly can't accurately represent the sharp transition of wind in the eyewall. That means that the forecast winds are almost certainly a smoothed (and therefore less intense) version of what observers would see.
It's kind of like a post-processing or bias correction (see for example https://www.ecmwf.int/en/about/media-centre/aifs-blog/2026/a..., which applies in physical space), but by having access to the model latent space and by being included in model training it is (probably!) higher-quality than a pure, after-the-fact approach.
85 comments
[ 0.24 ms ] story [ 7.5 ms ] threadLink?
> WeatherNext 3 will power weather features in Google Search, the Gemini app and Google Maps, as well as the Google Maps Weather API and Google Earth Engine.
https://dataconomy.com/2026/09/04/weathernext-3-ai-forecasts...
So I assume the main way would be googling "weather Los Angeles" and it will be powered by the WeatherNext 3 models
They also open sourced the last one and are doing B2B/enterprise arrangements so maybe other weather apps are experimenting with it.
- Time is UTC rather than local by default.
- Units are imperial rather than metric (i.e. not basing it on users locale)
- The concept of init time is not initially clear at all, so it's confusing when you click the calendar icon to see the weather forecast for a future date only to see it isn't an option, you have to use the slider.
- When changing the sidebar to the "detailed" view, most of the hovering element is cut off by the container so you can't actually read it.
And it took longer to write this comment than it did to find these issues. It's fine since it's clearly marked as experimental, but I disagree that it's "really easy to use" :p
Everything seems dark and desaturated, as if the the only clue we have (color!) for matching things up has been deliberately reduced.
And then: It sure does feel like the legend has even more of whatever-that-is going on than the map does.
I find it difficult to look at the map and understand the information it relays at the same time.
Init time makes sense to me, it's when the forecast/model is initialized, letting you look back in the past to watch the change to the present and future.
The sidebar thing is weird, but it also can expand out.
* https://apnews.com/article/weather-forecasts-worsen-doge-tru...
* https://www.independent.co.uk/news/world/americas/us-politic...
A good book on the history of forecasting, The Weather Machine: A Journey Inside the Forecast:
* https://www.andrewblum.net/the-weather-machine-2
Would that be practical for weather forecasting or not really?
What are the conditions at 5000 feet, 10000, etc? What is the location of the jet stream and its strength? The reduced number of (e.g.) weather balloons is hindering forecasts (per the links).
The data that would be most valuable to initial conditions is upper-atmosphere winds -- this is the kind of data given by weather balloons. In clear air there's no great way to measure this from either the ground or from space.
One important supplemental data source here are aviation reports, from planes flying at altitude and particularly trans-oceanic routes. When air traffic was largely curtailed during the early phase of the Covid pandemic, weather forecasting suffered a bit for the lack of data (see eg https://www.ecmwf.int/en/about/media-centre/news/2020/drop-a...).
I guesstimate that it has less than 50% accuracy for my area
As a former Googler, I wouldn't at all be surprised if this is an integration that is "planned" -- but just not done yet.
And some good handful of people are planning to wring a promo out of work. "Implemented weather UI in Android that is 63% more accurate." ;)
I'm building a tool right now that uses an ensemble to determine wind gust likelihood, which is useful for safety critical work on construction sites and the like.
[1]: https://www.nature.com/articles/d41586-026-02643-w
I've heard some commentary in the past that 5G interferes with data measurements heavily used for weather forecasts (something about satellite measurements?)
I remember being _very_ impressed like 15 years ago about how I would have _hourly_ weather forecasts (in particular around the rain) that seemed like magic! And then things... seemed to slowly get worse (at least in Tokyo)
a couple years ago I was chatting with a friend in Kyoto. They used to live in Tokyo and had made the _exactly_ aligned comment like: "I was used to the rain forecasts being not so accurate anymore. After moving to Kyoto they seemed to be quite good! And now they're also bad here. Am I cursed?"
I looked at some 5G rollout maps and you could see Kyoto rollout happened a bit around the time frame they were complaining about....
Anyways I think for most people (at least for myself) weather forecasting seems like this odd dark magic that can't work at all, but there was a window in which it _felt_ like it was super accurate. At least in my personal experience
The short version of the "spectrum" issue is that 5G is being allocated in bands very close to the microwave spectra where atmospheric water vapor emits. A bevy of public and commercial satellites in low-Earth orbit passively monitor these microwave spectra and produce extremely important information that is assimilated into numerical weather models.
The federal government sets limits on "out-of-band emissions" for operators emitting in the allocated 5G spectra. These emissions can bleed over into the microwave water vapor bands, creating noise that masks the natural presence of water vapor. The limits for this out-of-band emission is on the order of ~10-20 dB, and there's some work in the atmospheric science literature suggesting that this is enough to confound certain water vapor retrievals. That basically means we lose these observations that help constrain the forecast.
There isn't much indication that this is a serious issue in day-to-day meteorology at the moment. But it's an issue which will be significantly more difficult to unroll and claw back than to simply protect key spectra in the first place.
On recent visits to Tokyo, with my current habits, the general crowd (who presumably check some app) has been the more reliable forecast. Either everyone has an umbrella in their hand or they don't. And I discover whether I should have brought my umbrella or if I'll be lugging it around pointlessly, just far enough from my hotel to be stuck with my choice
So yeah, when it rains, it might take a few hours for that to flow into your weather app.
It's a direct result of DOGE.
Google does update its forecast more frequently, and usually is as accurate as the others for the hour ahead. And the longer range 3+ day forecasts aren't noticeably worse either. It's just the intermediate 4 - 48 hour stuff that they're surprisingly terrible with, but that's exactly the time period I most care about in everyday use!
It really does come down to data source quality and access…that’s the crux of it.
It is built into Apple Weather after Apple purchased them.
Did Apple lobotomize the tech when they integrated it, or lose access to whatever upstream data Dark Sky had access to? Apple Weather is comically bad (though not nearly as bad as Google's weather searches).
I used to be able to go on walks without an umbrella because of how reliable it was at predicting when rain was going to restart. If I do that with Apple Weather, I get home wet.
https://acmeweather.com/
https://www.theverge.com/tech/883089/acme-weather-forecast-a...
"We create a 3-channel image where the red channel represents velocity in the x-direction, the blue channel represents velocity in the y-direction, and the green channel represents change in storm intensity" and then use computer vision techniques to predict what happens next
post was written after a discussion here on hn: https://news.ycombinator.com/item?id=3187326
'Nowcasting' is an area of active research, both with machine learning and with physics-informed or visual flow approaches.
Part of the problem from the machine learning side is that these are _huge_ problems. NVidia's StormCast (https://research.nvidia.com/publication/2024-08_kilometer-sc...) works globally at kilometer scales, and you can imagine how big those grids are. Even with patch training, you're dealing with very large datasets.
At the same time, this is not exactly a high-profile area of research. National weather centres focus on actionable medium-range weather predictions, and meteorologists can look at radar themselves and perform mark-one-eyeball predictions for very short-range watches and warnings. Some private-sector actors will pay for short range predictions, but they're often looking for something hyperlocalized (e.g. weather at this particular construction site, for crane safety) or specialized (near-real-time cloud and wind predictions for renewable energy).
Most public-accessible weather predictions are downstream of either a public-sector effort (which doesn't internalize benefits, leading to under-resourcing) or a byproduct of another private-sector offering.
https://www.dwd.de/DE/wetter/thema_des_tages/2026/9/6.html (German only)
There's also https://nickleenders.github.io/verisky-scoreboard/history.ht... which tracks the trends over time.
It features a couple of AI and NWP model forecasts for comparison.
This seems to be a bit of "we're throwing a _bunch_ of inputs into this machine learning set and then pulling out outputs". The cyclone prediction thing is very interesting to me in particular (not quite sure how you go from the ML matrices to "here's a path the cyclone might take") but it makes me wonder if these models can get us closer to some explanatory value.
I imagine a lot of predictive sciences are ultimately about mixing together a bunch of inputs to attempt to decipher some output. Do we end up being able to take stuff from here and figure out some new ideas about modelling the climate as a whole?
In a high-level view, it's the result of specialized decoding heads.
Traditionally one would take gridded forecast outputs, then process those with comprehensible actions like "find all local pressure minima in the ocean, then filter to ones which correspond to warm cores, etc." to infer (diagnose) the presence of a cyclone.
One problem with this is that gridded forecasts suffer from known biases and tradeoffs. For example, a forecast on a ~25km grid is just on the edge of being able to represent the eye of a hurricane (50km scales), and it certainly can't accurately represent the sharp transition of wind in the eyewall. That means that the forecast winds are almost certainly a smoothed (and therefore less intense) version of what observers would see.
The WN2 approach (paper: https://www.nature.com/articles/s41586-026-10953-2) adds a direct readout head to the model: given latent-space access to the full forecast, it tries to predict the bona-fide cyclone observations (https://www.ncei.noaa.gov/products/international-best-track-...).
It's kind of like a post-processing or bias correction (see for example https://www.ecmwf.int/en/about/media-centre/aifs-blog/2026/a..., which applies in physical space), but by having access to the model latent space and by being included in model training it is (probably!) higher-quality than a pure, after-the-fact approach.
Never let a good crisis go to waste for marketing, especially if the product you're promoting is contributing to exacerbating it.