The model was trained on four decades of temperature, wind speed and air pressure data from 1979 to 2018 and can produce a 15-day forecast in just eight minutes—compared to the hours it currently takes.
A new artificial intelligence-based weather model can deliver 15-day forecasts with unrivaled accuracy and speed, a Google lab said, with potentially life-saving applications as climate change ramps up.
GenCast, invented by London-based AI research laboratory Google DeepMind, "showed better forecasting skill" than the current world-leading model, the company said Wednesday.
The European Center for Medium-Range Weather Forecasts (ECMWF) produces predictions for 35 countries and is considered the global benchmark for meteorological accuracy.
But DeepMind said GenCast surpassed the precision of the center's forecasts in more than 97 percent of the 1,320 real-world scenarios from 2019 which they were both tested on.
I’m honestly surprised at the 8-minute completion time. Given the performance of modern LLMs, trained on billions of parameters, I assumed that a weather model trained on numerical temperature,...
I’m honestly surprised at the 8-minute completion time. Given the performance of modern LLMs, trained on billions of parameters, I assumed that a weather model trained on numerical temperature, wind speed, and air pressure records would be considerably simpler than Gemini which is trained on a text corpus of the entire internet. Of course the hours-long ECMWF process is even more surprising.
Great news regardless! This seems like a really practical and low-hanging application of machine learning tech. I’m excited to see how these models continue to evolve over the next decade!
Weather data is exponentially more complicated than language, with much larger data sets. Here is one small sampling of available data from NOAA. The possible interactions are huge, and why even...
Weather data is exponentially more complicated than language, with much larger data sets. Here is one small sampling of available data from NOAA. The possible interactions are huge, and why even today most predictions fall apart after 4 days or so.
It's exacerbated by needing to calculate the weather for the entire planet, rather than just a small word prompt. It's more akin to having the model spit out a whole new model every day, which is then the input for the next model, instead of a small piece of output.
It's honestly miraculous if their claims about accuracy and completion time are true.
Weather simulations are brute-force calculations that scale poorly. It's one of the reasons supercomputers are built. Accuracy can be improved by increasing resolution (smaller cells and shorter...
Weather simulations are brute-force calculations that scale poorly. It's one of the reasons supercomputers are built. Accuracy can be improved by increasing resolution (smaller cells and shorter time steps) and running more simulations, so they can always use more computation.
It's a bit surprising how low-resolution these models are:
GenCast generates an ensemble of stochastic 15-day global forecasts, at 12-h steps and 0.25° latitude–longitude resolution, for more than 80 surface and atmospheric variables, in 8 min.
These are cells more than a mile wide and that's pretty good for a weather model.
I don't know about low hanging, but based on what little I think I know about weather forecasts and AI, it kind of seems like an obvious use case for AI. Weather will always be a best guess type...
This seems like a really practical and low-hanging application of machine learning tech.
I don't know about low hanging, but based on what little I think I know about weather forecasts and AI, it kind of seems like an obvious use case for AI. Weather will always be a best guess type of scenario and that feels like what AI is good at.
To me most AI news seems uninteresting at best or dystopian at worst buy I'm actually interested to see how this weather thing unfolds.
I wouldn’t really say it’s low hanging fruit. For one, it’s not the obvious solution, that is to have the network export posterior probabilities. Rather, it’s using a model merely as the state...
I wouldn’t really say it’s low hanging fruit. For one, it’s not the obvious solution, that is to have the network export posterior probabilities. Rather, it’s using a model merely as the state transition function from which probabilities are calculated afterwards. In that way it’s similar to earlier deepmind work (eg the monte-Carlo based alphago).
It also uses a diffusion model (yes, the diffusion in stable diffusion), which is a fairly recent development.
What does that mean then? Being reliant on government provided data makes sense, that's generally who operates weather satellites, being reliant on the models seems odd. It's neat they were able...
But he said these forecasting systems are reliant on the weather prediction models that are already running, such as that operated by ECMWF.
What does that mean then? Being reliant on government provided data makes sense, that's generally who operates weather satellites, being reliant on the models seems odd.
It's neat they were able to outperform models using old data. It seems a no brainer to be running it using current data to see if it actually can predict out 15 days and generate a track record of being right.
ECMWF are using a 30 petaflop, 1,000,000 CPU cluster for 6 hours daily to run the forecasts, and dedicating a further 12 hours on that cluster to internal research work, so it's expensive and...
ECMWF are using a 30 petaflop, 1,000,000 CPU cluster for 6 hours daily to run the forecasts, and dedicating a further 12 hours on that cluster to internal research work, so it's expensive and niche enough that only a few organisations want to make the investment.
So they're tracking all the butterflies in the world to see the flaps of their wings then? Weather, is unpredictable. It always has been. Unless we manage to actually control it, then we have, at...
So they're tracking all the butterflies in the world to see the flaps of their wings then?
Weather, is unpredictable. It always has been. Unless we manage to actually control it, then we have, at best, guesstimates of what it will do. In theory we can predict what will happen and sometimes we are right, but at any given moment a volcano can erupt and throw everything sideways. A whirlwind whips left and suddenly a tornado, but it chills down to gales really fast. No AI model or weather prediction software has been accurate to any degree past a few minutes. It's been a long standing joke that whatever the weather-person says is a lie.
If a model has managed to predict 15 days in a row, that's great. Now do a 14 day forecast accurately for the world for a month.
I’m honestly surprised at the 8-minute completion time. Given the performance of modern LLMs, trained on billions of parameters, I assumed that a weather model trained on numerical temperature, wind speed, and air pressure records would be considerably simpler than Gemini which is trained on a text corpus of the entire internet. Of course the hours-long ECMWF process is even more surprising.
Great news regardless! This seems like a really practical and low-hanging application of machine learning tech. I’m excited to see how these models continue to evolve over the next decade!
Weather data is exponentially more complicated than language, with much larger data sets. Here is one small sampling of available data from NOAA. The possible interactions are huge, and why even today most predictions fall apart after 4 days or so.
It's exacerbated by needing to calculate the weather for the entire planet, rather than just a small word prompt. It's more akin to having the model spit out a whole new model every day, which is then the input for the next model, instead of a small piece of output.
It's honestly miraculous if their claims about accuracy and completion time are true.
Weather simulations are brute-force calculations that scale poorly. It's one of the reasons supercomputers are built. Accuracy can be improved by increasing resolution (smaller cells and shorter time steps) and running more simulations, so they can always use more computation.
It's a bit surprising how low-resolution these models are:
These are cells more than a mile wide and that's pretty good for a weather model.
Here's an example of one such simplified calculation for fluid dynamics.
Here's a weather-specific one.
I don't know about low hanging, but based on what little I think I know about weather forecasts and AI, it kind of seems like an obvious use case for AI. Weather will always be a best guess type of scenario and that feels like what AI is good at.
To me most AI news seems uninteresting at best or dystopian at worst buy I'm actually interested to see how this weather thing unfolds.
I wouldn’t really say it’s low hanging fruit. For one, it’s not the obvious solution, that is to have the network export posterior probabilities. Rather, it’s using a model merely as the state transition function from which probabilities are calculated afterwards. In that way it’s similar to earlier deepmind work (eg the monte-Carlo based alphago).
It also uses a diffusion model (yes, the diffusion in stable diffusion), which is a fairly recent development.
What does that mean then? Being reliant on government provided data makes sense, that's generally who operates weather satellites, being reliant on the models seems odd.
It's neat they were able to outperform models using old data. It seems a no brainer to be running it using current data to see if it actually can predict out 15 days and generate a track record of being right.
ECMWF are using a 30 petaflop, 1,000,000 CPU cluster for 6 hours daily to run the forecasts, and dedicating a further 12 hours on that cluster to internal research work, so it's expensive and niche enough that only a few organisations want to make the investment.
So they're tracking all the butterflies in the world to see the flaps of their wings then?
Weather, is unpredictable. It always has been. Unless we manage to actually control it, then we have, at best, guesstimates of what it will do. In theory we can predict what will happen and sometimes we are right, but at any given moment a volcano can erupt and throw everything sideways. A whirlwind whips left and suddenly a tornado, but it chills down to gales really fast. No AI model or weather prediction software has been accurate to any degree past a few minutes. It's been a long standing joke that whatever the weather-person says is a lie.
If a model has managed to predict 15 days in a row, that's great. Now do a 14 day forecast accurately for the world for a month.