Everything in AI seems to be focused on LLMs lately. But in my opinion, powerful problem-specific models like this are even more interesting. The SOTA AI models used in weather forecasting are already outperforming the classic NWP models while being orders of magnitude more efficient (inference). Most are based on multi scale (hierarchical) Graph Neural Networks, an architecture which is not often talked about. The original Graphcast paper is worth a read if you think this is interesting: https://arxiv.org/abs/2212.12794
You say this as if you don't need he MWP models to train the AI models? The accuracy of the AI Prediction depends entirely on the quality of the training dataset...
Pretty much all of the AI weather prediction models are trained on ECMWF ERA5, which is kinda like a numerical weather prediction model run to forecast at t=0. ERA5 is historical weather data, but it’s a “reanalysis” of it.
Indeed, you can imagine this as some sort of advanced physics-based interpolation of various measurements (land stations, satellite data, ...) to fill in every cell in a latitude-longitude grid. This is not only used for ERA5 (training data for the models), but also to determine the initial conditions for every grid cell which are used to roll out the forecast. So AI weather models depend greatly on the NWP/physics used in for reanalysis and initial conditions. That being said, there is also research being conducted in training models straight from the raw data (weather stations, satellite, ...), thus bypassing the "interpolation" step.
Yeah. I can’t remember names off the top of my head, but there are a few companies, and I think many researchers, working on AI “data assimilation” for this.
ECMWF has an experimental AIFS direct observational prediction model (AIFS-DOP) that has become competitive with their physics based IFS model on certain metrics just in the past year.
Historical weather data is discrete. You need continuous state for weather modelling which is currently achieved through conventional reforecasts using those historical observations.
A more interesting question is...does differential equations based models like mamba/state space models perform better on this sort of physics problem than pure transformer LLMs?
I can assure you that anyone who touches numerical simulations of any kind (physicists, engineers, chemists, biophysicts...etc) has tried their hand at ML based surrogate models in the last 5 years, so it's not like they aren't being tested. From my experience, they aren't very robust. Weather modeling is actually one of the very few areas where it seems to work half decently.
Why does it work for weather at all? Is there something that the mathematical models are over-simulating? Is weather easier to predict than we thought? Just curious what the intuition is to regarding the success of ML weather modelling...
There is just A LOT of data available- usually an order of magnitude more than in any other related problem.
And general weather forecasts are not that hard - we have semi useful forecasts for more than 50 years. It’s when you want to do something special: long range, nowcasting of convective storm, other extreme weather etc. that is hard. And even then it’s as much a problem of input data accuracy than the models themselves.
Traditional physics based weather models also rely heavily on physical parameterization for sub grid scale processes (think clouds, microphysics of rain sleet snow, etc) so even the deterministic physics models are learned approximations from data.
everything in AI is not focused on LLM, if you think so then that's because you are in LLM bubble. The big idea with LLM is that it's generative AI, the generative could be anything! Not just large languages, we have seen break through in image generation, video, audio, but guess what. Anything that you have enough data and given data you can predict what comes next can have gen AI applied, so we are seeing it with physical actions so robots get trained to generate the next move, and I think the same thing applies to weather forecast. It's predictable too given enough data
One of my professors is referenced in the Wikipedia page of graph neural networks. It's funny that he explained them in the worse way possible and I eventually understood them better with another professor
Insightful paper, thanks for sharing. Two things stand out to me.
First, it reinforces that you want methods that get better with more data. It emphasizes that the current approach cannot improve based on historic data - that’s the opportunity that ML based approaches exploit.
Second, it highlights that mature legacy solutions are tough competitors. They benefit from extensive tuning and real world feedback. Even when you have a genuinely better approach, it will take meaningful time & effort to achieve the current standard.
Patience and solid long term strategy are needed to make progress in these situations. You need confidence that your approach will win long term, backed by enough money & time to prove yourself correct.
Everything in the western world isn't focused on LLM. The top western players are heavily focused on AGI.
Meanwhile the Chinese are using LLMs and other non-AGI AI tech at the edge wherever they think to put it for task-specific productivity or optimization. They don't really care about AGI, or more accurately: they're working on getting their society more efficient and decarbonized, and then they'll be free to work on AGI with far fewer resources.
OpenAI, Anthropic, et al are working toward someday having AGI, and if they ever do, when they do, the Chinese will be hopelessly far ahead of us on energy, manufacturing, logistics (especially low/zero carbon transport of goods and people) and so on.
Once the Chinese figure out how to train an AI for ULEV lithography, especially once they figure out how to train it for semiconductor design or validation - it's game over for the semiconductor industry, and the big AI players will follow, because they won't possibly be able to compete against a Chinese version of NVIDIA with TSMC-like capabilities, or Chinese AI companies running on those much cheaper chips, with cheap, zero carbon power.
How does predicting a typhoon prevent billions in damage?
It could save thousands of lives because people can be evacuated if you can predict a few hours or a day further ahead, or the path more accurately. You can save some damage by moving ships and vehicles.
But you can't evacuate buildings or infrastructure.
You can board up a helluva lot more stuff in a week than in two days. Crucially, you can move more of the most expensive stuff out of the storm surge zone, which is where the biggest damage happens and try to flood proof more of the things which can’t be moved.
> You can board up a helluva lot more stuff in a week than in two days
TFA says they might be able to predict one extra day ahead (three days instead of two). No prediction system will ever give you a week's notice on a typhoon.
Actually it doesn't even say that - they claim to have the same "accuracy" at 3 days that older methods have at 2 days. What does that actually get you? Were the older models so much less accurate at 3 days (compared to 2) that it prevented evacuation of key areas? Looking at the paper, it doesn't really seem like this can be answered yet because there's not enough data over a long enough time.
Keep in mind, DeepMind has a very well documented history of releasing enormously hyped up PR pieces with grandiose claims that are never backed up in real world usage, or are simply lies.
I know this is uncharitable and I am wrong but I am having trouble coming up with concrete scenarios where you die with 2 days notice but survive with 3. I am nonethless a believer that more accurate forecasting has value.
You can walk 50km in those extra 24 hours which will save your life if you’re in a storm surge area without other means of transportation, which is easily the case when everyone else is evacuating alongside you.
Could you imagine a scenario where from warning to complete evacuation takes more than two days? Evacuating a whole area is a hard task, particularly once you start looking at more complex problems (elderly, prisons, hospitals).
I feel like the details of this are highly dependent on the confidence of the warning; moving large numbers of people (particularly elderly) will result in some deaths regardless. I guess more time to do it should help though.
It's not uncharitable, because the system doesn't actually claim to give you an extra day of notice. It would be more accurate to say: "the model can reach a given level of forecast accuracy roughly a day farther in advance". So your question becomes: are there scenarios where I am 80% sure this is a Cat 5 hurricane 3 days before, where I would die if I was only 65% sure it was Cat 5 on that same day? The answer is - probably not, because even in the example used in the PR article, the NHC was already issuing strong early guidance 5 days before landfall.
I don't think the amount of notice is as important as "this totally is for sure going to wreck you" is. We generally know something is going to hit somewhere at sometime. Which isn't specific enough for everyone to act on
I've always assumed that insurance and/or government departments that would spend money due to storms would be the ones funneling money to these sorts of efforts. It's not exactly something you can easily sell directly to individuals who would benefit. It would be pretty dystopian for them to sell subscriptions for an extra 24 hrs notice on the next typhoon :P
I just discovered typhoon/cyclone predictions and they're insane. I get mine via https://zoom.earth (whose iPhone app is terrific).
Here's a selection from Typhoon Dolphin, currently sitting off the east coast of China.
Dolphin continues its slow, trochoidal Z motion, generally heading westward deeper into the East China Sea. Over the past 12 hours, the system completed another cyclonic loop and has decelerated, exhibiting continued meandering prior to establishing a sustained westward track.
The erratic motion witnessed over the past two days is attributable to a weak steering environment produced by a break in the subtropical ridge 2 over Korea, combined with the dynamics where the inner core is cocooned within a much larger parent circulation.
While the general steering pattern is weak, a mesoscale deep-layer ridge is seen building over southern Japan.
Here's Chan-hom, which threatens to make my birthday a windy day here in northern Japan.
Intensity guidance is in good agreement overall. However, the JTWC forecast is placed lower than all the guidance save for Google DeepMind over the next 36 hours, before joining the consensus envelope (which peaks at 95 km/h (50 knots) at 60 hours) through the remainder of the forecast.
If you’re just getting into this tropical tidbits[0] is my go to for more raw data. Less pretty than zoom earth but also an interesting place to see what the models are predicting on each of their runs which is then interesting to compare to actual forecast guidance.
> One key limitation of our approach is in how uncertainty is handled. We focused on deterministic forecasts and compared against HRES, but the other pillar of ECMWF’s IFS, the ensemble forecasting system, ENS, is especially important for 10+ day forecasts. The non-linearity of weather dynamics
means there is increasing uncertainty at longer lead times, which is not well-captured by a single deterministic forecast. ENS addresses this by generating multiple, stochastic forecasts, which model the empirical distribution of future weather, however generating multiple forecasts is expensive. By contrast, GraphCast’s MSE training objective encourages it to express its uncertainty by spatially blurring its predictions, which may limit its value for some applications.
As someone who has learned bayesian statistics in social sciences, isn't this a big deal? There is a reason why risk estimates need to be well understood and *explainable* for certain fields like this. Are you willing to bet a government response should issue an evacuation order 30 miles from the center of a hurricane at location X if the model can't tell you why it produced an uncertainty estimate there — or worst the model changed its mind later?
From the tagline in the article: "WeatherNext enables accurate cyclone forecasts that can give an extra day of warning. Now we are open sourcing the model."
Predicting big weather events is not that hard even with 50 year old technology.
What's hard is predicting details, like exactly where it will rain, what the slope of the beach is today (many people don't even know this changes drastically daily and why it is important), wave height, ocean depth today where people swim, water temperature, shorebreak, and knowing with certainty when rain becomes ice/sleet/snow and what routes will be affected, accurate wind speed, accurate temperature throughout different parts of the region, and what the weather next week will be.
We can't do any of those things with conventional equipment, but we can with training data and algorithms. So I'm very excited about the role of algorithmic prediction in weather, but not for the kind we already know how to forecast (without AI) but being able to glean useful insights that matter to people who live, work and play in the weather.
It would be interesting to research more into parameter analysis to learn more about what parameters a trained network has "extracted" that has the strongest predictive power; assuming the model itself has such preductive power. I guess I'm imagining some kind of alternative to embeddings where you get answers like:
Predicted percipitation P(accumulated humidity over n days, net wind vector over n days, humity today, temperature today, ...)
> We can now generate a single 15-day forecast in less than a minute on a TPU, empowering forecasters to quickly evaluate the probability distribution of potentially devastating tail-risks.
"This has surprised scientists, and it remains an open research question to fully understand how our models produce such accurate predictions at this resolution."
Also crazy.
Seems important to understand why something does what it does, in the very least to know when it might not?
on one side, deepmind makes a lot of advancement in science related application. but, on the commercial side, they struggle to compete with other major LLM providers.
Accurate weather forecasting has been one of the major achievements of the 20th and 21st century. Computing power is a central piece of this story, but it's also important to remember that the government infrastructure in place to collect ground-truth current weather data is utterly critical to these model's successes. From launching weather balloons to running global weather-monitoring satellites, the scientists and systems at NOAA/NWS (and in this case, the UK counterparts) provide critical expertise and data.
I say this because it seems that earlier announcements where industrial deep neural nets "outperformed NOAA" likely encouraged the slash-and-burn Trump administration in its gutting of critical activities and centers of expertise at NOAA. The impression that industry can predict weather better than the government agencies totally misses that the industrial models utterly rely on government data for inputs. In fact, almost all weather reports you see---weather.com, TV, etc.---are just lightly repackaged products that NOAA provides for free on weather.gov (which you can access for free without ads).
Google has the early warning system that gives people maybe 20-30s to e.g. turn off gas, stop vehicles, get under something solid. There was a lot of news recently about how this saved many thousands of lives in Venezuela I think it was.
But hey let's all keep shitting on Google because their coding agent is slightly worse than SOTA.
This needs to be tied to a whole house shutoff system because if I get an alert I’m not thinking about shutting off my gas or water. Having a system that shut those off immediately would be great
Forecasting earthquakes via ML should be possible but is very strongly limited by data. We have ~50 years of reasonably good seismological catalogs for most of the world. The seismic cycle (the sequence of major earthquake, reloading, major earthquake on a single section of fault) is generally thousands of years except at the fastest-slipping faults. There are very few sections of faults where we have seismological observations of multiple events, and for >90% of faults, we don't even know when the last earthquake was. There are geologic methods to help with this, but they are labor intensive and often yield error bars of hundreds to thousands of years, because the earthquakes don't produce radiocarbon signatures directly; the geologists use e.g. charcoal older and younger layers as available to bracket the timing, and many faults do not have suitable geologic sites to preserve the earthquake deformation and bracket the timing.
I do think it's possible that thorough exploration of the data that do exist can yield broader patterns that apply to many regions, but earthquake behavior has a lot of complexities and different fault systems may behave differently.
A lot of the hope is for coupling physical simulators to ML and the existing datasets to better understand the physics and then work from there, but this is typically cutting-edge HPC work, which limits the pace of research and the number of researchers.
You seem to know much more about this topic than I do, and your assessment agrees with mine. Predicting rare and sudden events is a very difficult modeling problem generally, and the data limitations is real. I agree that it will probably need to come from coupling of ML with physical simulators plus more extensive 3D map data on force vectors, material properties, etc.
I don't mean to be disrespectful... but, why would you consider tech intended to save lives and resources less of a deal, than slightly faster flights over the Atlantic?
Cool how they integrated both huge machine-scale data and smaller human-curated data for this project.
> The model was co-trained on two distinct data modalities: global weather dynamics and expert-curated historical cyclone observations. By training end-to-end on nearly 20 terabytes of global atmospheric data and the historical IBTrACS database spanning nearly 5,000 historical storms, the model learns complex atmospheric patterns and how to model extreme weather.
And this is why GDM has to go. It's crazy that when Google is struggling so badly that efforts like this that have no path to revenue at all were funded.
GDM management really thought that they were some kind of charity. UNREAL.
Are you being sarcastic? Even if “Google is struggling so badly” (which it really is not - the narrative will flip again at some point), these efforts will have a lasting impact on the world. Not everything good is about bringing in revenue.
shareholders are pretty unhappy about demis. think about alphafold. huge investments from the company, tens of billions. at a critical time. and absolutely 0 revenue. it got demis a nobel though. as a shareholder you'd be unhappy too.
Everything in AI seems to be focused on LLMs lately. But in my opinion, powerful problem-specific models like this are even more interesting. The SOTA AI models used in weather forecasting are already outperforming the classic NWP models while being orders of magnitude more efficient (inference). Most are based on multi scale (hierarchical) Graph Neural Networks, an architecture which is not often talked about. The original Graphcast paper is worth a read if you think this is interesting: https://arxiv.org/abs/2212.12794
You say this as if you don't need he MWP models to train the AI models? The accuracy of the AI Prediction depends entirely on the quality of the training dataset...
I would imagine this would be trained on actual historical weather data instead?
Pretty much all of the AI weather prediction models are trained on ECMWF ERA5, which is kinda like a numerical weather prediction model run to forecast at t=0. ERA5 is historical weather data, but it’s a “reanalysis” of it.
Indeed, you can imagine this as some sort of advanced physics-based interpolation of various measurements (land stations, satellite data, ...) to fill in every cell in a latitude-longitude grid. This is not only used for ERA5 (training data for the models), but also to determine the initial conditions for every grid cell which are used to roll out the forecast. So AI weather models depend greatly on the NWP/physics used in for reanalysis and initial conditions. That being said, there is also research being conducted in training models straight from the raw data (weather stations, satellite, ...), thus bypassing the "interpolation" step.
Yeah. I can’t remember names off the top of my head, but there are a few companies, and I think many researchers, working on AI “data assimilation” for this.
ECMWF has an experimental AIFS direct observational prediction model (AIFS-DOP) that has become competitive with their physics based IFS model on certain metrics just in the past year.
https://arxiv.org/html/2606.19093v1
Ah, yes, that’s one of them! Not to be confused with AIFS and AIFS Ensemble that are competitive with IFS, but start with the same DA as IFS.
Historical weather data is discrete. You need continuous state for weather modelling which is currently achieved through conventional reforecasts using those historical observations.
From a quick read: ECMWF and IBTrACS data - the former is model based (with measurement data crunched), the latter purely observational.
A more interesting question is...does differential equations based models like mamba/state space models perform better on this sort of physics problem than pure transformer LLMs?
Rumor is that part of the disruption at GDM these past few months also involved people not wanting to be bound to strictly LLM research.
Is there any website publishing these forecasts? I imagine NWS/NOAA isn’t doing anything different yet on their public websites.
Yes. AIFS directly by ECMWF and AIGEFS by NOAA. Every vibecoded weather app these days has them. Google those terms you’ll find them.
Anything more daily human friendly/consumable?
https://sites.gsl.noaa.gov/desi/
I can assure you that anyone who touches numerical simulations of any kind (physicists, engineers, chemists, biophysicts...etc) has tried their hand at ML based surrogate models in the last 5 years, so it's not like they aren't being tested. From my experience, they aren't very robust. Weather modeling is actually one of the very few areas where it seems to work half decently.
Why does it work for weather at all? Is there something that the mathematical models are over-simulating? Is weather easier to predict than we thought? Just curious what the intuition is to regarding the success of ML weather modelling...
There is just A LOT of data available- usually an order of magnitude more than in any other related problem.
And general weather forecasts are not that hard - we have semi useful forecasts for more than 50 years. It’s when you want to do something special: long range, nowcasting of convective storm, other extreme weather etc. that is hard. And even then it’s as much a problem of input data accuracy than the models themselves.
Traditional physics based weather models also rely heavily on physical parameterization for sub grid scale processes (think clouds, microphysics of rain sleet snow, etc) so even the deterministic physics models are learned approximations from data.
everything in AI is not focused on LLM, if you think so then that's because you are in LLM bubble. The big idea with LLM is that it's generative AI, the generative could be anything! Not just large languages, we have seen break through in image generation, video, audio, but guess what. Anything that you have enough data and given data you can predict what comes next can have gen AI applied, so we are seeing it with physical actions so robots get trained to generate the next move, and I think the same thing applies to weather forecast. It's predictable too given enough data
Please don't generate cyclones!
One of my professors is referenced in the Wikipedia page of graph neural networks. It's funny that he explained them in the worse way possible and I eventually understood them better with another professor
Insightful paper, thanks for sharing. Two things stand out to me.
First, it reinforces that you want methods that get better with more data. It emphasizes that the current approach cannot improve based on historic data - that’s the opportunity that ML based approaches exploit.
Second, it highlights that mature legacy solutions are tough competitors. They benefit from extensive tuning and real world feedback. Even when you have a genuinely better approach, it will take meaningful time & effort to achieve the current standard.
Patience and solid long term strategy are needed to make progress in these situations. You need confidence that your approach will win long term, backed by enough money & time to prove yourself correct.
Everything in the western world isn't focused on LLM. The top western players are heavily focused on AGI.
Meanwhile the Chinese are using LLMs and other non-AGI AI tech at the edge wherever they think to put it for task-specific productivity or optimization. They don't really care about AGI, or more accurately: they're working on getting their society more efficient and decarbonized, and then they'll be free to work on AGI with far fewer resources.
OpenAI, Anthropic, et al are working toward someday having AGI, and if they ever do, when they do, the Chinese will be hopelessly far ahead of us on energy, manufacturing, logistics (especially low/zero carbon transport of goods and people) and so on.
Once the Chinese figure out how to train an AI for ULEV lithography, especially once they figure out how to train it for semiconductor design or validation - it's game over for the semiconductor industry, and the big AI players will follow, because they won't possibly be able to compete against a Chinese version of NVIDIA with TSMC-like capabilities, or Chinese AI companies running on those much cheaper chips, with cheap, zero carbon power.
Maybe was this that was the last drop for Sundar.
Demis: "I have a new amazing breakthrough"
Sundar: "Great! We really need a answer to Sol and Fable"
Demis: "They are completely owned in typhoon forecasting"
Ironically typhoon forecasting, at this moment, is more valuable. These predictions are matters of life, death, and billions of dollars in damage.
How does predicting a typhoon prevent billions in damage?
It could save thousands of lives because people can be evacuated if you can predict a few hours or a day further ahead, or the path more accurately. You can save some damage by moving ships and vehicles.
But you can't evacuate buildings or infrastructure.
You can board up a helluva lot more stuff in a week than in two days. Crucially, you can move more of the most expensive stuff out of the storm surge zone, which is where the biggest damage happens and try to flood proof more of the things which can’t be moved.
> You can board up a helluva lot more stuff in a week than in two days
TFA says they might be able to predict one extra day ahead (three days instead of two). No prediction system will ever give you a week's notice on a typhoon.
Actually it doesn't even say that - they claim to have the same "accuracy" at 3 days that older methods have at 2 days. What does that actually get you? Were the older models so much less accurate at 3 days (compared to 2) that it prevented evacuation of key areas? Looking at the paper, it doesn't really seem like this can be answered yet because there's not enough data over a long enough time.
Keep in mind, DeepMind has a very well documented history of releasing enormously hyped up PR pieces with grandiose claims that are never backed up in real world usage, or are simply lies.
30% more time to know exactly where to board up etc. seems very significant
Wouldn't it be 50% more time?
Why couldn’t we get a week?
I know this is uncharitable and I am wrong but I am having trouble coming up with concrete scenarios where you die with 2 days notice but survive with 3. I am nonethless a believer that more accurate forecasting has value.
You can walk 50km in those extra 24 hours which will save your life if you’re in a storm surge area without other means of transportation, which is easily the case when everyone else is evacuating alongside you.
Could you imagine a scenario where from warning to complete evacuation takes more than two days? Evacuating a whole area is a hard task, particularly once you start looking at more complex problems (elderly, prisons, hospitals).
I feel like the details of this are highly dependent on the confidence of the warning; moving large numbers of people (particularly elderly) will result in some deaths regardless. I guess more time to do it should help though.
The 2 vs 3 days makes less of an impact on personal decision making but has massive benefits for decision making at the country wide response level.
Hurricane Maria went from Cat 2 to Cat 5 in less than 24 hours, and turned making a direct hit to Dominica in 2017.
It's not uncharitable, because the system doesn't actually claim to give you an extra day of notice. It would be more accurate to say: "the model can reach a given level of forecast accuracy roughly a day farther in advance". So your question becomes: are there scenarios where I am 80% sure this is a Cat 5 hurricane 3 days before, where I would die if I was only 65% sure it was Cat 5 on that same day? The answer is - probably not, because even in the example used in the PR article, the NHC was already issuing strong early guidance 5 days before landfall.
I don't think the amount of notice is as important as "this totally is for sure going to wreck you" is. We generally know something is going to hit somewhere at sometime. Which isn't specific enough for everyone to act on
You live on a chain of small islands and travel by boat.
Hurricane blast radius is very small
Imagine you've got to evacuate a hundred thousand people. That extra day is incredibly valuable.
Valuable, agreed. But lucrative?
Are Sol and Fable lucrative? I suspect they also are valuable (to clients) but not lucrative (yet).
I think both have value, but in opposite ways. While WeatherNext prevents costs, models like fable or sol "create profit".
I can think of 10 examples how one could make money with fable. With WeatherNext? Only 10 examples of preventing costs.
Taking this, maybe naive, thought further, profits have no upper limit (except resources) while costs can only save so much?
Yes of course. My comment was facetious.
Trade agricultural futures based on it maybe?
I've always assumed that insurance and/or government departments that would spend money due to storms would be the ones funneling money to these sorts of efforts. It's not exactly something you can easily sell directly to individuals who would benefit. It would be pretty dystopian for them to sell subscriptions for an extra 24 hrs notice on the next typhoon :P
by giving it away they capture none of the value except for some PR.
I agree, but the shareholder mentality undervalues the heck out of that.
I just discovered typhoon/cyclone predictions and they're insane. I get mine via https://zoom.earth (whose iPhone app is terrific).
Here's a selection from Typhoon Dolphin, currently sitting off the east coast of China.
https://zoom.earth/storms/dolphin-2026/Here's Chan-hom, which threatens to make my birthday a windy day here in northern Japan.
https://zoom.earth/storms/chan-hom-2026/If you’re just getting into this tropical tidbits[0] is my go to for more raw data. Less pretty than zoom earth but also an interesting place to see what the models are predicting on each of their runs which is then interesting to compare to actual forecast guidance.
1: https://www.tropicaltidbits.com/
I am getting into cyclone predictions (for maritime scheduling)
Is there a basic/freemium resource for past events? Mostly just very coarse spatial/temporal maps of past events
For atlantic basin hurricanes (and the occasional one that could impact Hawaii he also does fantastic youtube videos
This is really cool, please more of this from the AI folks! That’s way more impactful and interesting than another coding agent
> One key limitation of our approach is in how uncertainty is handled. We focused on deterministic forecasts and compared against HRES, but the other pillar of ECMWF’s IFS, the ensemble forecasting system, ENS, is especially important for 10+ day forecasts. The non-linearity of weather dynamics means there is increasing uncertainty at longer lead times, which is not well-captured by a single deterministic forecast. ENS addresses this by generating multiple, stochastic forecasts, which model the empirical distribution of future weather, however generating multiple forecasts is expensive. By contrast, GraphCast’s MSE training objective encourages it to express its uncertainty by spatially blurring its predictions, which may limit its value for some applications.
As someone who has learned bayesian statistics in social sciences, isn't this a big deal? There is a reason why risk estimates need to be well understood and *explainable* for certain fields like this. Are you willing to bet a government response should issue an evacuation order 30 miles from the center of a hurricane at location X if the model can't tell you why it produced an uncertainty estimate there — or worst the model changed its mind later?
From the tagline in the article: "WeatherNext enables accurate cyclone forecasts that can give an extra day of warning. Now we are open sourcing the model."
Seems to be the same methodology as the ECMWF AI ENS operational since mid last year:
https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs...
https://www.nature.com/articles/s44387-026-00073-7
Predicting big weather events is not that hard even with 50 year old technology.
What's hard is predicting details, like exactly where it will rain, what the slope of the beach is today (many people don't even know this changes drastically daily and why it is important), wave height, ocean depth today where people swim, water temperature, shorebreak, and knowing with certainty when rain becomes ice/sleet/snow and what routes will be affected, accurate wind speed, accurate temperature throughout different parts of the region, and what the weather next week will be.
We can't do any of those things with conventional equipment, but we can with training data and algorithms. So I'm very excited about the role of algorithmic prediction in weather, but not for the kind we already know how to forecast (without AI) but being able to glean useful insights that matter to people who live, work and play in the weather.
> what the slope of the beach is today (many people don't even know this changes drastically daily and why it is important
So why is it important? As far as I know the slope changes AFTER the weather not before as a prediction mechanism but happy to learn
Check out my pytorch reproduction of the paper here for those interested: https://github.com/NVIDIA/physicsnemo/pull/1660
It would be interesting to research more into parameter analysis to learn more about what parameters a trained network has "extracted" that has the strongest predictive power; assuming the model itself has such preductive power. I guess I'm imagining some kind of alternative to embeddings where you get answers like: Predicted percipitation P(accumulated humidity over n days, net wind vector over n days, humity today, temperature today, ...)
Sounds a like a model Predict Wind needs.
> We can now generate a single 15-day forecast in less than a minute on a TPU, empowering forecasters to quickly evaluate the probability distribution of potentially devastating tail-risks.
Crazy
"This has surprised scientists, and it remains an open research question to fully understand how our models produce such accurate predictions at this resolution."
Also crazy.
Seems important to understand why something does what it does, in the very least to know when it might not?
You can't, and it's one of the biggest problems when trying to use AI for anything.
What a beautiful outcome of this age of AI!!! And they are open sourcing it too!!
#HappyNews
on one side, deepmind makes a lot of advancement in science related application. but, on the commercial side, they struggle to compete with other major LLM providers.
which for google shareholders is pretty bad. we don't get anything from them releasing this.
SOTA to be abandoned for something that makes money.
Accurate weather forecasting has been one of the major achievements of the 20th and 21st century. Computing power is a central piece of this story, but it's also important to remember that the government infrastructure in place to collect ground-truth current weather data is utterly critical to these model's successes. From launching weather balloons to running global weather-monitoring satellites, the scientists and systems at NOAA/NWS (and in this case, the UK counterparts) provide critical expertise and data.
I say this because it seems that earlier announcements where industrial deep neural nets "outperformed NOAA" likely encouraged the slash-and-burn Trump administration in its gutting of critical activities and centers of expertise at NOAA. The impression that industry can predict weather better than the government agencies totally misses that the industrial models utterly rely on government data for inputs. In fact, almost all weather reports you see---weather.com, TV, etc.---are just lightly repackaged products that NOAA provides for free on weather.gov (which you can access for free without ads).
They should try to forecast earthquakes, that would really be a breakthrough If anything better than random comes out
Google has the early warning system that gives people maybe 20-30s to e.g. turn off gas, stop vehicles, get under something solid. There was a lot of news recently about how this saved many thousands of lives in Venezuela I think it was.
But hey let's all keep shitting on Google because their coding agent is slightly worse than SOTA.
> But hey let's all keep shitting on Google because their coding agent is slightly worse than SOTA.
Reminder, we can do two or even more things. In fact, we can even simultaneously hold contradictory opinions.
And it's been built into every Android phone for years, for free. While Apple is still completely Missing In Action.
This needs to be tied to a whole house shutoff system because if I get an alert I’m not thinking about shutting off my gas or water. Having a system that shut those off immediately would be great
I don't think the parent was shitting on Google
Forecasting earthquakes via ML should be possible but is very strongly limited by data. We have ~50 years of reasonably good seismological catalogs for most of the world. The seismic cycle (the sequence of major earthquake, reloading, major earthquake on a single section of fault) is generally thousands of years except at the fastest-slipping faults. There are very few sections of faults where we have seismological observations of multiple events, and for >90% of faults, we don't even know when the last earthquake was. There are geologic methods to help with this, but they are labor intensive and often yield error bars of hundreds to thousands of years, because the earthquakes don't produce radiocarbon signatures directly; the geologists use e.g. charcoal older and younger layers as available to bracket the timing, and many faults do not have suitable geologic sites to preserve the earthquake deformation and bracket the timing.
I do think it's possible that thorough exploration of the data that do exist can yield broader patterns that apply to many regions, but earthquake behavior has a lot of complexities and different fault systems may behave differently.
A lot of the hope is for coupling physical simulators to ML and the existing datasets to better understand the physics and then work from there, but this is typically cutting-edge HPC work, which limits the pace of research and the number of researchers.
You seem to know much more about this topic than I do, and your assessment agrees with mine. Predicting rare and sudden events is a very difficult modeling problem generally, and the data limitations is real. I agree that it will probably need to come from coupling of ML with physical simulators plus more extensive 3D map data on force vectors, material properties, etc.
This was my immediate hope too, as fault line resident
Is this even feasible with our current sensor data?
It seems especially useful for cargo ships, with better predictions they could save some fuel and be safer.
Predict Wind has a sailing ship course auto router.
Wake me up once commercial airplanes can take advantage of this and take us across the Atlantic in less than 5 hours.
I don't mean to be disrespectful... but, why would you consider tech intended to save lives and resources less of a deal, than slightly faster flights over the Atlantic?
No disrespect taken :) It was (obviously) a joke, I'm not seriously waiting for us to hurl airplanes through cyclones to make air travel faster.
planes fly above the weather, so kinda irrelevant. you can cross the atlantic fast with something like the Concorde
Well, that explains why even cyclones don't make us faster!
Obviously the technology I'm talking about would involve the planes going into the cyclone so plane can go faster.
Cool how they integrated both huge machine-scale data and smaller human-curated data for this project.
> The model was co-trained on two distinct data modalities: global weather dynamics and expert-curated historical cyclone observations. By training end-to-end on nearly 20 terabytes of global atmospheric data and the historical IBTrACS database spanning nearly 5,000 historical storms, the model learns complex atmospheric patterns and how to model extreme weather.
Next step: steering them. (As in Permutation City's "Operation Butterfly".)
And this is why GDM has to go. It's crazy that when Google is struggling so badly that efforts like this that have no path to revenue at all were funded.
GDM management really thought that they were some kind of charity. UNREAL.
Are you being sarcastic? Even if “Google is struggling so badly” (which it really is not - the narrative will flip again at some point), these efforts will have a lasting impact on the world. Not everything good is about bringing in revenue.
shareholders are pretty unhappy about demis. think about alphafold. huge investments from the company, tens of billions. at a critical time. and absolutely 0 revenue. it got demis a nobel though. as a shareholder you'd be unhappy too.
i feel you. as a google shareholder i am disappointed too.
Ask Gemini why google maps doesn't have a weather layer. Its justifications are defensive rubbish, even for Gemini.