A short video transcript (Two Minute Papers) where Dr. Zsolnai summarizes a new DeepMind paper that introduces an efficient graph‑neural‑network–based weather forecasting model. The model reportedly predicts up to 10 days ahead by iterating accurate 6‑hour advances, matches or outperforms high‑resolution billion‑dollar forecasting systems on most test cases (90–99%), uses under 40 million parameters, is open source, is already deployed in practice, and promises much faster, cheaper warnings for extreme weather.
Chapters
Scientists at DeepMind just created a revolution in weather forecasting with their new AI. This paper is history in the making. Why? Well, you see, weather forecasting is typically done not by people and not even government agencies, but often intergovernmental organizations. Why is that? Well, because it's just too hard to be able to measure. All the information needed for a good weather prediction costs billions and billions of dollars. Almost no one can do it. You have to take into consideration temperature, precipitation, wind, humidity, pressure, vertical velocities and so much more and then simulate their physics into the future. This needs physicists, meteorologists, computer scientists and, and of course, billions and billions of dollars. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. So if it costs so much, we get the results instantly, right? Well, not quite. This is very expensive and it still takes a long time. But sometimes you don't have time and need a quick prediction. You can't wait for 24 hours for the algorithm to tell you what's going to happen in the next 10. During that time, storms, hurricanes, floods may happen and we have to know as much in advance as possible, so we need something faster. Now we are experienced fellow scholars over here, so we already know that AI techniques exist that can perform physics simulations. We start running a fluid or smoke simulation and we say okay, we stop the simulation here. Now little AI, tell me how it continues and magically so it does. However, wait. The main goal of these simulations is to be pleasing to the eye and not be that accurate. So they don't have the accuracy for weather predictions until now. Yes, here it is an absolutely amazing new paper that promises. Wow, that is crazy. They promise a prediction for the next 10 days in just one minute. One minute. Wow. I will immediately say that. OK, I get that. But then some trade offs surely have to be made. The quality can't be so good, right? I mean just matching the quality of the billion dollar system is in a minute would be absolutely impossible. But in any case, let's have a look. What? Are you seeing what I am seeing? Holy mother of papers. We are measuring the root mean squared error here. So lower is better. And the new super quick technique not only matches the industry standard high resolution forecast, but it even outperforms it. And it is better than the current techniques on about 90 to 99% of the test cases. That is insanity. It takes the current state of the quantities we can measure and then it advances the time by six hours accurately. Then we can take this predicted model as the current state and advance that with another six hours and so on. Thereby we can make predictions for even 10 days each into the future. And the list of good news just keeps going. I have three more incredible things for you about this paper. 1. What I am really shocked by is this. Now hold on to your papers fellow scholars, because it uses less than 40 million parameters. Hmm, that sure sounds like a lot, right? Well remember GPT4, the system that powers ChatGPT has more than 1.5 trillion parameters as far as I know. Yes, you are hearing correctly. This system is 50,000 times smaller than GPT4 and yet can pull off this absolute miracle. I will note that the paper did not claim that this would run on your phone, but the size and efficiency of this system could potentially allow us to do that. My goodness, this kind of superpower right in your pocket. I am out of words. And it gets even better. 2. It is already being used out there in the wild. Not so long ago, going from a paper to a product out there in the real world took several years, sometimes even decades. And this, this one is already out there. This paper is simply so good, no one can ignore it. Loving it. 3. The models are open source. Yes, with a little expertise you can download and play with them right now or even use them in an online notebook. I also loved how they used a graph neural network to accomplish this. These are typically used for things that have a graph like structure like like people and their relationships in a social network or molecular dynamics. However, here instead the temperature, pressure and other quantities are measured in these little points and the graph neural network ensures that there can be communication between the neighbors in this mesh. They can pass messages to each other. And the surprise is that such a tiny network can be so powerful at weather prediction that it rivals and often beats billion dollar systems. Bravo DeepMind. So an AI based technique that you can run in your pocket soon and it will give some of the best warnings for extreme weather events, flood risk and extreme temperatures. This is a paper that is going to save lives. This is history in the making. What a time to be alive. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the for GPU cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers because with the Lambda GPU cloud you can now get on demand H100 instances and they are one of the first cloud providers to offer publicly available on demand H100 access. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, mit and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today.
