Dear Aventine Readers,

Since the 1990s, meteorologists have become better at predicting the chance of seasonal hurricanes; the error rate in one-to-three day forecasts has been reduced by 75 percent in the last two decades. Now, new AI-based prediction models promise even greater accuracy, potentially allowing meteorologists to deliver far more precise forecasts about all types of weather activity: rainfall, wind speeds, temperature. So instead of entire cities or regions buckling up for the sky to drop a foot of water, forecasters will be able to aim warnings at specific regions or neighborhoods. 

This week we look at the development of GAIA, a new AI prediction model based on enormous amounts of high-resolution satellite images and weather data. In the testing phase, GAIA already outperformed more traditional weather prediction models. The hope is that GAIA can become almost a simulation of the earth’s atmosphere, allowing meteorologists more insight into complex systems like hurricanes and, ultimately, giving them the opportunity to save more lives.

Thanks for reading and all best, 

Danielle Mattoon 
Executive Director, Aventine

For meteorologists, the 2025 hurricane season, which starts in June, came in like a lamb. As of the final weeks of the summer, there had been only two named hurricanes and five tropical storms in the Atlantic Basin, fewer than in a typical year. 

That calm was broken by some extraordinary events in late September. Two storm systems, Humberto and Imelda, came within 500 miles of each other — the closest two storms have come in more than 50 years. Then on October 27, Tropical Storm Melissa intensified into a Category 5 hurricane, and tied with the Labor Day Hurricane of 1935 as the third-most powerful Atlantic hurricane on record, killing an estimated 75 people in Haiti and Jamaica and causing billions of dollars in damage. 

Wallace Hogsett, a meteorologist, monitored all this activity from the relative safety of the National Hurricane Center in Miami, a good 30-minute drive from the ocean. There, he scans through mountains of data — pressure drops, increasing wind speeds, ocean swells — looking for early signs of a hurricane. 

“Sometimes we struggle to predict a storm’s intensity,” said Hogsett. “Other times, it’s the path. Where is it going to go? What are the hazards? Where is the heaviest rainfall going to be?” Lives and property hang in the balance; since 2000, hurricanes in the United States have killed more than 2,000 people and cost more than two trillion dollars in damages. Faster and more accurate forecasts could help blunt some of these storms’ devastating effects. 

Across the country in Mountain View, California, an engineer and artificial intelligence expert named Dr. David Bell is working on a possible solution to Hogsett’s problem. For the past year and a half, Dr. Bell has directed a team of data scientists at the Universities Space Research Association (USRA), working with NASA and Boston Consulting Group, to build an artificial intelligence model that can forecast extreme weather with unprecedented accuracy. 

The system is called Geospatial Artificial Intelligence for Atmospheres or GAIA, named for the Greek deity who represents the earth. Built on 25 years of high-resolution weather data from satellites, GAIA has been trained to spot the subtle atmospheric shifts that precede catastrophic events like hurricanes and atmospheric rivers, those long trails of water vapor that carry precipitation from the tropics.

GAIA is what researchers call a foundation AI model: a system trained on enormous amounts of data that can be adapted to many different tasks. One of the best known uses of a foundation model is Open AI’s ChatGPT. Trained on text, the model on which ChatGPT is based can answer questions or draft essays because it has absorbed the patterns of the English language. GAIA applies a similar architecture to Earth’s atmosphere. Instead of words, it was trained on petabytes of satellite images and atmospheric measurements to learn complex patterns that help predict extreme weather events. Once a model is trained, it can perform individual tasks more quickly and less expensively than current forecasting systems, which run on supercomputers. 

“It’s still early days, but it could be transformative,” Dr. Bell said.

For the past 50 years, the gold standard in weather prediction has been the numerical weather model, which takes current atmospheric conditions and projects them hours, days and weeks into the future using sophisticated formulas. “If you take hurricanes as an example, we’ve reduced the error rate in our one-to-three day forecast by 75 percent in 20 years,” said John Ten Hoave of the National Oceanic and Atmospheric Administration (NOAA). 

The improvement is impressive, but the approach has some significant downsides. To gauge a storm’s path, for example, meteorologists often need to run multiple simulations at once, but because numerical models can only explore one scenario at a time scientists are limited in how many simulations they can run. 

“With our current models, we’re often limited to run five simulations per storm event,” Ten Hoave said. “But with a foundation model, we could run hundreds of simulations, which would give us a much better sense of the uncertainty around each forecast.”

Bell knows the limitations of numerical weather models: Before leading the team behind GAIA, he built task-specific AI models using satellite data with NASA and the US Geological Survey. It was encounters with models like ChatGPT that got him excited about applying a similar architecture to weather. “It was a real wow moment,” he recalled. 

Bell and his team began developing GAIA in April 2024. They first downloaded 25 years of historic data from a trio of satellite systems — Europe’s Meteostat, Japan’s Himawari satellite and the National Oceanic Atmospheric Administration’s GOES East and West satellites — that together offered a near global view of Earth. The data consisted of high-resolution images of Earth and atmospheric measurements like cloud cover, temperature and moisture. Despite the sheer volume of data — a total of 77 petabytes or roughly four times the volume of information held in the Library of Congress — it took only about a week to download and convert the files into a format the model could accept. 

Using a network of graphics processing units (GPUs) as well as a novel algorithm that includes masked autoencoders, GAIA learned the language of the Earth’s atmosphere. Masked autoencoders are learning models that work by quite literally masking portions of input data so the system learns to fill in the blanks using only the available data. In the case of GAIA, the algorithm randomly hid large chunks of satellite data, forcing the model to use historic data to restore the missing information, a little like restoring a damaged photograph. The team repeated this process tens of thousands of times before they were satisfied with the results. 

To test the model’s accuracy, the team compared GAIA’s predictions to real-world data. To their delight, GAIA reconstructed the missing data with uncanny accuracy compared to the actual records. 

After the model mastered gap-filling, the team moved on to trickier tasks, like asking GAIA to predict heavy rainfall based on historical data – the kind of rainfall that floods cities, washes out roads and, in the worst cases, feeds hurricanes, which are among the most costly and deadliest of disasters. GAIA’s results were then compared to the real-life rainfall. Again, Dr. Bell was impressed. “The quality is starting to exceed what the numerical weather predictions do.” 

The first version of GAIA designed for research groups was released in May, and the latest iteration has been trained intensively on hurricane data. When it becomes fully operational, GAIA could be fed new data about every 30 minutes, with each satellite pass, and then adjust its forecasts accordingly. This will allow for far more targeted forecasting than numerical models, which can only be updated every four to six hours. 

Bell can’t say when the model might be available to forecasters like Hogsett. But GAIA isn’t the only product on the market. In the past couple of years, Google, Microsoft and Nvidia launched AI models of their own focused on dramatically improving weather prediction. 

“If a tornado is coming, sometimes the whole county gets an alert,” said Brian LaMarre, a former meteorologist with the National Weather Service. “ But I think in the very, very near future, probably the next few years, we're going to see a greater precision alert with artificial intelligence.” 

As the model matures, there are plans to release updated versions combining additional data — just like OpenAI did with ChatGPT. Future versions will likely incorporate data from low Earth orbit satellites, for instance, to better cover the poles — regions that existing satellites cannot see. Dr. Bell and his team envision other applications for GAIA, supporting agriculture, aviation and a multitude of other industries affected by weather.

“I think we’re moving toward a future where just like ChatGPT, there is going to be a single service that answers questions about oceans, land, weather and space,” said Dr. Bell. The ultimate goal is a dynamic, real-time model of Earth’s atmosphere. It’s not just about better forecasts, said Dr. Bell, but about giving people the tools to adapt in a world that’s changing fast.

How AGI became the most consequential conspiracy theory of our time, from MIT Technology Review
6,350 words, or about 25 minutes

Maybe you’re terrified of what a superintelligent AI might do to humanity. Maybe you’re bullish that it will cure cancer, climate change and poverty. Or maybe you just find the whole AGI thing a little… cultish. Whatever your view, this essay is worth your time. “If you're building a conspiracy theory, you need a few things in the mix: a scheme that’s flexible enough to sustain belief even when things don’t work out as planned; the promise of a better future that can be realized only if believers uncover hidden truths; and a hope for salvation from the horrors of this world,” writes Will Douglas Heaven. “AGI just about checks all those boxes. The more you poke at the idea, the more it starts to look like a conspiracy.” Along the way, the piece traces the origins of the term AGI, examines how it has shaped the tech industry and the broader economy, and asks what happens when Silicon Valley organizes itself around a quasi-religious vision of digital omnipotence. It is nuanced, thoughtful and provocative, and will inform your understanding of the debate around AGI regardless of what you believe going in.

The fallout from the AI-fuelled dash for gas, from The Financial Times
2,500 words, or about 10 minutes 

A gas renaissance is underway, driven by the soaring electricity demands of AI data centers. But the dynamics of that revival may be turning ugly. At the heart of the issue is a bottleneck: The supply of giant turbines for gas-powered plants — made mainly by Mitsubishi, Siemens and GE — can’t keep up with demand. This is understandable, given that the plants producing these turbines are experiencing a fair amount of whiplash: Only a few years ago, production was slowing down in response to the global focus on renewable energy when climate change, not data centers, was a primary driver of decisions around energy production. But the upshot is more complex than you might expect. Wealthier regions like the US, Europe and the Middle East can afford to pay steep premiums to secure turbines. Emerging markets cannot, and the consequences could ripple across the global economy. Vietnam, a beneficiary of Western manufacturing shifts away from China, may be unable to build the energy infrastructure needed to sustain its export growth, in turn rattling supply chains worldwide. Some developing countries that had planned to expand gas power may turn back to coal, using Chinese-made hardware and reviving the dirtiest fossil fuel on the planet. Even US projects like Alaska LNG, which depends on Asian demand for gas, could be undermined. 

Editing nature to fix our failures, from Noema
3,800 words, or about 15 minutes

Genetic engineering could, in theory, help the natural world cope with — or even overcome — some of the damage humans have caused to the environment. It might produce new forms of coral that can withstand ocean acidification, give plants like the American chestnut resistance to blight introduced by global travel, or even bring extinct species back to life. But just because we can rewrite nature, should we? This essay dives into the technical, ethical and ecological dilemmas that surround conservation powered by genetic engineering. What are the ripple effects of saving certain species rather than others? Does the ability to edit nature give us moral cover to keep consuming and polluting? And what happens if we go beyond restoration, reshaping ecosystems to serve human ends? There are no obviously correct answers, and humankind is on the clock to fix its mistakes before it's too late. But gene editing is now a tool that lies at our disposal, and the high-stakes choices we make about how it is used could determine what this planet looks like for the rest of time.