A YouTube talk by Dr. Károly Zsolnai-Fehér (Two Minute Papers) presenting three recent AI research advances with potential to save lives: (1) a flood-prediction method that skips explicit rainfall forecasting and transfers models trained in data-rich countries to data-poor countries, claimed to outperform existing global flood awareness systems and run on low-cost devices or phones; (2) a diffusion-model-based weather-prediction technique that generates large volumes of plausible historical-like weather data (avoiding expensive physics simulations), improving extreme-event prediction accuracy at lower computational cost; (3) an AI system to detect and avoid long-lived aircraft contrails, including a trial with American Airlines (70 flights) showing ~0.3% extra fuel use and a 54% reduction in persistent contrails, yielding an approximately 20× net climate benefit. The video also includes speaker framing (50-hour round trip to visit labs) and a closing sponsor mention of Lambda GPU cloud (on-demand H100 instances, persistent storage).
Questions this video answers
50 hours round trip.
About 5%.
70 flights.
About 0.3% extra fuel usage.
54% reduction.
Chapters
I flew to San Francisco to talk to scientists at the OpenAI lab, Google in Mountain View, Nvidia and more. My travels took 50 hours round trip. I think I might need some advanced AI cloning technology to get there next time. Now here are some of the incredible things that I've seen at Google. They said they don't let in anyone external in that building for any reason and I could get in there and talk to some legendary scientists. Scientists. So that was an incredible honor. Thank you. And now I'll show you three absolutely incredible research papers that do the greatest thing any paper can do and that is saving human lives. You will see what's new in flood detection, weather prediction and something about this too that really surprised me. I'll note that this video is not sponsored by them and they will see this video for the first time together with you fellow scholars when it appears. Dear fellow scholars, this is two minute papers with Dr. Zsolnai one. They are Google so they see that people around the world search about floods several hundred million times each year. So this is a super important problem. However, current systems are not up to the task. Why is that? Well, one of the hardest variables to predict in weather is rainfall. It is the main source of uncertainty and you need those to predict floods. However, here is an insane idea. Just skip it entirely. Don't even think about the rain. Ok, but what now? Well now try to predict floods from everything else that we know instead. But there is a problem, a huge problem. The countries that need these predictions the most are the places that don't have the training data for the learning algorithm as they don't have the means to collect it. And this new technique helps them reuse data from the USA and other countries and apply it to less fortunate countries. Is that even possible? Well get this. This new AI system is more accurate than the global flood awareness systems used around the world. That is absolutely amazing. A great help for non advanced nations and not only for their governments, but for citizens too. You see, some of these are getting cheaper and cheaper to run. So much so that some of these even run on your phone in your pocket. 2. They also have a crazy new AI technique for weather prediction. However, wait, there are already real good ones out there. For instance Nvidia's ForecastNet and DeepMind's GraphCast. So what is going on here? Well, there is a problem. Most techniques need lots of physics computation to generate simulated weather data to be able to teach the AI. So here is once again a crazy idea. Don't do that and instead take A bunch of historical weather data and use diffusion based AI models to predict from historical data. Yes, these are the AI models that create images from your text where they start out from noise and over time they reorganize this noise to resemble your text input more and more over time. It can generate videos too. Except that here it generates plausible weather data to provide the AI with a textbook to learn the intricacies of extreme weather events. And as a result it has so much more data to work with and is more accurate than previous methods at a fraction of the computational cost of previous techniques. Loving this one. This will once again be able to predict extreme weather events and save many, many human lives overtime. Now three, I was really surprised by this. This one is about using AI to make flights a little more sustainable. Did you know that aircraft exhaust lines called contrails, are very good at spreading out overtime and trapping heat? Yes. So much so that they have a measurable impact on the temperature of the planet. Not huge, but measurable and meaningful. This I found very surprising because as we have tons of flights around the world, there are lots of contrails, but only few of them stay for long. Probably only 5% of them. And that 5% has a meaningful effect on our weather. OK, so now what? Well, it would be great to have some sort of simulation that would be able to predict which planes on which routes would are going to create these trails. However, there is a problem here too. To have an AI learn the evolution of these trails and how they behave over time, we would need tons and tons of training data. So, ok, just pull up lots of satellite imagery and off you go, right? Well, no, not at all. And therein lies the problem. Contrails are very easily confused by these famously wispy and serious clouds. But get this. Scientists found a way to tell them apart reliably. So this problem is now learnable. Fantastic. Now this is nice, but this is numbers on paper. Does this prediction technique actually work? Does it actually help us? Well, get this. They really put their money where their papers are because they had a trial with American Airlines for 70 flights where they proposed tiny, tiny changes to their routes by using the AI's results. And the additional fuel usage was very little, about 0.3% difference. And that led to a 54% reduction of these trails, cutting the heat trapping effect by half. Yes, if planes intentionally avoid flying through these regions, then they can avoid creating warming contrails with minimal impact on fuel usage. Now wait, 0.3% of extra fuel usage is still quite meaningful. Yes, that is true. However, the net effect of this technique is still 20x better for us and the planet than what we have now. And I hope that with this video more of you fellow scholars hear about these incredible research works that will save lives. What a time to be alive. There are many, many more out there and unfortunately I found that almost no one is talking or reporting on them. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold onto 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.
