Google LLC’s artificial intelligence research organization DeepMind today announced the release of two very different but exceedingly powerful large language models.

One of them, GenCast, is focused on weather forecasts, and the other, called Genie 2, is designed to create elaborate, visually striking virtual worlds for video games.

GenCast appears to have the most practical implications. Announced in a blog post by DeepMind researchers Ilan Price and Matthew Wilson, it’s not only better than any previous weather forecasting algorithm they have created, but even outperforms what is considered to be the most accurate system in use right now, maintained by the European Center for Medium-Range Weather Forecasts.

DeepMind published its paper on GenCast in the journal Nature, where it talks about how using machine learning for weather prediction has helped it to achieve fewer errors in forecasting than traditional systems, which utilize “physics-based simulations of the atmosphere.”

GenCast, the researchers explained, was trained on decades worth of “reanalysis data,” and is capable of generating 15-day weather forecasts in just eight minutes, compared with hours for the supercomputers running ECMWF.

Naturally, DeepMind wanted to see how GenCast stacks up against ECMWF, which is the system that’s used by 35 countries for their weather forecasting needs. In a series of tests that compared the 15-day weather forecasts of GenCast and ECMWF’s system, DeepMind’s model proved to be more accurate 97.2% of the time. With lead times greater than 36 hours, GenCast was even better, showing more accuracy 99.8% of the time.

The researchers explained that GenCast is a diffusion model, based on the same kind of technology that powers its Gemini family of generative AI models. DeepMind trained the system on almost 40 years’ worth of high-quality weather data curated by the European Center for Medium-Range Weather Forecasts, and the predictions it generates are said to be “probabilistic,” so they account for various possibilities, expressed as percentages.

AI researchers consider probabilistic models to be more nuanced than deterministic ones, which can only offer a guess at what the weather might be like on any given day.

What’s most impressive about GenCast is that it can get by with far less computing power than the system used by ECMWF. DeepMind said it can generate 15-day weather forecasts in just eight minutes using a single TPU v5 tensor processing unit. In comparison, the ECMWF system requires an enormous supercomputer with tens of thousands of processors to generate the same, physics-based forecast.

“I’m a little bit reluctant to say it, but it’s like we’ve made decades worth of improvements in one year,” Rémi Lam, lead scientist on DeepMind’s previous AI weather program, told The New York Times. “We’re seeing really, really rapid progress.”

DeepMind said it’s making GenCast open-source on GitHub, so anyone can download it along with sample code. In addition, it also plans to integrate GenCast with Google Earth, where anyone will be able to use it to generate weather predictions.

Genie 2 conjures up 3D worlds

As for Genie 2, this is another far-reaching development that could pave the way for developers to create expansive and playable 3D game worlds using a simple text prompt.

It’s said to be the next evolution of DeepMind’s Generative Interactive Environments tool, which uses AI to build interactive virtual environments. The original model, Genie 1, could only create 2D worlds, but Genie 2 ventures into the 3D realm, making it far more impressive.

According to DeepMind, Genie 2 is a “world model” that’s able to simulate a virtual world along with animations, physics and even support interactions between all of those elements. Users can create their worlds using a prompt image to depict it, or alternatively they can first generate an image with a text prompt, then use that as the basis of the world they want to create.

The possibilities are endless. Ask for a sailing simulation and Genie 2 will immediately create it. Or if you want a cyberpunk Western, it can create that too. All that’s required is a reference image to start, and if there isn’t one, it can even create that first.

The worlds support interactions between the player’s character, which can be controlled by a human or an AI agent, although at present they suffer from problems with stability. Quite simply, they’re not that stable, and the model begins to lose coherency after around 20 seconds or so.

DeepMind said that’s partly the result of Genie 2’s ability to support “counterfactuals,” which are the different paths a player can take from a fixed starting point. For instance, the player can turn right or left to show a different part of the world, but the model must still take into account what’s happening outside of the player’s view, in case they turn back toward that scene.

Still, the potential is clear, and DeepMind said it can also support various perspectives, such as an isometric, third-person or first-person view. It also depicts complex physics, such as what happens if a player or an object hits the water, how it ripples. In addition, it showed an ability to model smoke, gravity and reflections.

The environments could be useful for training AI models. In one demonstration, DeepMind’s team showed how an AI-controlled character could be told to go through a specific door with a text prompt. The AI can recognize the command and apply it to the rendered environment, then immediately proceed.

DeepMind didn’t say much about its plans for Genie 2, such as if and when it might be released. It also refused to say what kind of computing power it requires. The lack of stability suggests there’s still much work to be done, but the fact it’s being worked on suggests that AI-generated gaming worlds at some point in time are a distinct possibility.

Images: DeepMind