Latest news with #BCGX
.jpg&w=3840&q=100)
WIRED
2 days ago
- Science
- WIRED
Orbital Intelligence: When Satellites Meet Machine Learning
How BCG X, research institutions, and space agencies are using generative AI to supercharge weather forecasting with the GAIA Foundation Model. The 20,000-Foot (or Mile?) View Here's a fact that almost everyone on the planet is becoming increasingly familiar with: As the Earth's climate warms and its weather systems become less reliable, so do the weather prediction capabilities underpinning the business practices of countless agriculture, insurance, public safety, and scientific research organizations around the world. Here's a less obvious fact: As those prediction capabilities deteriorate, so do many of the public and private services we take for granted every day. 'Having better, more reliable, more detailed intelligence about what's going on in the weather system has a lot to do with who's going to win and lose in financial industries like insurance and lending, in infrastructure sectors like energy, and places like state and local government,' says David Potere, geospatial tech leader and BCG X managing director and partner. 'As an example, the way we characterize risk affects the homes we buy, the businesses we invest in, the cities that grow or don't. And right now, there is a known gap in the insurance industry being able to cover a rapidly changing game board.' The impact that this weather intelligence gap—and countless other gaps like it—has on organizational margins can trickle down to consumers in harsh ways. New volatility in the climate system can manifest as extended droughts and high winds that fuel record-breaking wildfires, or back-to-back 100-year storms that cause property damage at massive scales. A societal inability to forecast those kinds of events drives up our insurance rates, undercuts public safety measures, and strains governmental relief efforts. The world needed a novel solution to a rapidly growing problem. The BCG X AI Science Institute may have found it alongside a growing new class of gen AI-powered weather models. Turning to Eyes in the Sky Enter GAIA (Geospatial Artificial Intelligence for Atmospheres) Foundation Model, an open source foundation model built in partnership between the BCG X AI Science Institute and several of the world's leading aerospace organizations to help researchers all across the world better understand and anticipate weather's next move. Similar to large language models (LLMs) trained on text, the GAIA Foundation Model is a gen AI vision model trained on 25 years of satellite imagery that allows researchers to study climate and weather patterns at a greater speed and accessibility than ever before. Specifically, GAIA works with images from a constellation of school bus-sized satellites that 'stare' at the planet from a stationary position more than 22,000 miles above the surface, capturing high- resolution images of the entire 'disk' of the Earth every 30 minutes. This provides a continuous, real-time stream of images and atmospheric data. Taken in concert with a global array of thousands of hyper-detailed weather ground stations, meteorologists can essentially visually map weather developments in near real time across the entire globe. 'There are naturally gaps in this record, including a 'soft spot' when it comes to tracking weather in polar regions,' says Potere.. 'What we're talking about is investments on the ground through generative AI capabilities like GAIA that have the potential to unlock a synthetic fourth satellite constellation.' That kind of visualization capability, produced via open source gen AI technology, is groundbreaking on its own. But the setup behind that tooling is equally innovative. Consider compute power: Depending on the bands and mosaicking process, global satellite imagery can clock in at 3298 x 9896 pixels (and more), and a 15-year span of data measured every 30 minutes yields 263,000 images—more than the total frames in a typical Hollywood feature length film. That's 17 TB of data to be crunched per training session for the GAIA model. The team is also working with live weather data, tapping into the same operational satellites that weather forecasters use on the news at night. These foundation model approaches require a lot of GPUs—a common reason why visual-based gen AI tools have traditionally been a lesser-explored space. 'Up until now, the sheer compute and the algorithms and the know-how you need to be able to translate pixels into answers has been very rare,' Potere says. Tackling the Compute Problem BCG X made two conscious decisions when scoping the endeavor that not only proved to be novel but allowed them to bring the project online in just one year rather than the 18 to 24 months typical for other projects. The first was to commit to creating an environment that could be deployed in the cloud, rather than being tethered to a purpose-built supercomputer. According to Tom Berg, BCG X lead engineer for the project, 'There was something really daunting here; it's almost become an accepted truth that to roll up your sleeves and build your own foundational model is too expensive, if you look at the immense resources the hyperscalers are using. One of the things we wanted to show is that if it works, you don't have to have a dedicated supercomputer to do these kinds of builds.' To that end, the GAIA team turned to a national network of university computing resources distributed across the United States. This constellation of off-the-shelf GPUs (ranging from state-of-the-art to 10-year-old GPUs) is precisely what BCG X's development team had in mind. 'That profile, rather than matching a supercomputer, gave us a lot of parameters to work with,' Berg says. 'It's a very, very adaptable system, and at one point we were using 15 percent of the NRP's entire cloud.' Still, such a setup provided some interesting challenges. Where a dedicated supercomputer has all of its processing power in one building with one uniform power configuration, Berg and Potere's team would instead be connecting GPUs on opposite sides of the Earth. There were also acute issues like power outages, or a university unexpectedly cycling their data centers. Crucially, GAIA was sharing compute space with hundreds of other research applications running at the same time. 'You're basically on a busy public road rather than a dedicated racetrack,' Berg says. The team's second operational decision was to initially narrow their focus to precipitation and top-of-cloud temperature data—as opposed to modeling all aspects of every layer of the atmosphere. Because that selected data closely corresponds to a range of weather phenomena, it provided researchers with the flexibility needed to prove out the foundation model and run experiments at a still-manageable level of initial effort. Critically, by focusing on this 'lower complexity' problem statement, the GAIA team was able to immediately scale their modeling to global atmospheric conditions, putting a dent in the problem of weather predictability. That surgically targeted start allowed the team to reach an equally targeted—yet incredibly meaningful—outcome. Building a Global Toolset and Modeling Solutions A key reason why the team was able to build so rapidly: open source tools and resources, oftentimes combined with earlier research from equally pioneering research teams. 'We have the benefit of standing on the shoulders of some of the earliest groups working on this,' Potere says. 'Even now, the literature has gotten three times denser since we started, but there was something of a literature, so we certainly benefited from the second mover advantage.' That open source, iterative mindset will now define the project's next phase, as well: To give back to the research community and contribute to its ever-evolving toolset, BCG X and their collaborators released the GAIA Foundation Model to the global open source community. In other words, they modeled the Earth, for the Earth. And their work couldn't have come at a better time. As governments, businesses, and research institutions increasingly grapple with the new normal of disruptive weather volatility, gen AI weather and environmental models like GAIA can fuel faster and better decision making—something experts and organizations across the world need more every day. Climate change may very well be the defining issue of our time—and GAIA may very well be part of how the world as a whole is able to meet it. Learn more about Boston Consulting Group here.
Business Insider
11-05-2025
- Business
- Business Insider
The AI mistake companies are making — and how they can fix it, according to a BCG tech leader
Like a first date that's gotten awkward, some companies struggling to win at artificial intelligence might be trying too hard. They might take on too many projects or fail to understand that AI windfalls often come from rewiring how people work, not from "super-cool" AI engines or large-language models, said Sylvain Duranton, global leader of BCG X, the tech build and design division of Boston Consulting Group. Those types of missteps can balloon into big-time frustrations for business leaders, he told Business Insider. Duranton said that if CEOs' big question around AI in 2024 was which model to use, their ask in 2025 is "Where's my money?" Indeed, he said, there are often challenges around implementing broad use of AI. "Scaling this thing from a tech standpoint — it is hard," Duranton said. To help companies salvage their AI efforts, he said, his "golden rule" is that organizations allocate about 10% of their effort and money to the algorithms — to build AI engines or train LLMs. Another 20% should be reserved for data and technology. Essentially, that's to make the AI work in a company's tech environment, Duranton said. The bulk of the effort — the remaining 70% — should go to changing the way people work, he said. "Assuming you have a technology that can scale, you need to bring that into the hands of the people. It's a massive change effort," said Duranton, who's based in the company's Paris office and oversees BCG X's global army of nearly 3,000 technologists, scientists, programmers, engineers, and others. Some companies are struggling Companies' frustration is real. In the final months of 2024, BCG surveyed some 1,800 C-suite execs from big companies in nearly 20 countries and found that while 75% of respondents ranked AI among their top-three priorities, only 25% reported seeing "significant value" from the technology. To find more value, Duranton recommends that companies not try to do everything all at once. He said the scope of change that companies need likely can't be achieved with tens or hundreds of use cases. "That's not the plan. The plan is to focus on a very few things, and the things that matter," he said. Duranton said companies sometimes also look to "incremental initiatives." He said he thinks that's often a mistake. Instead, he said, companies should home in on a few "quintessential" things. For a retailer, Duranton said, this might be using AI to ensure a brick-and-mortar store has the perfect product mix for that location to better withstand competitors nearby and online. Retail CEOs understand the stakes, Duranton said. They'll often tell him something like, "I know that if I don't do that better than others, I'm cooked," he said. Duranton said another imperative for a retailer might be to develop an AI agent that can shop for customers — one that's so good that users won't want to switch to a competitor. "With those two things, you have both a strategic agenda and an AI agenda," he said, referring to making sure inventory is dialed and the shopping bot. The trick, then, is to keep the focus on those efforts, Duranton said. "Those quintessentials, that's where you put all your money, all your energy," he said. That's necessary, Duranton said, if companies want to take the 10%, 20%, 70% approach he recommends. AI's Bermuda Triangle He said scaling AI is also often difficult because companies can feel pressure to compromise on expenses, quality of results, or the speed at which they're produced. "You have a sort of Bermuda Triangle, where it is either costly and relevant with a good latency, or you have to compromise on one of the three, and you have to optimize," Duranton said, referring to AI results. He said it's often easy to demonstrate some tech wizardry in a demo. What's difficult, Duranton said, is handling millions of requests every day and producing timely and relevant results. "It's a different ballgame," he said. Ultimately, to succeed with AI, Duranton said, companies will need to bring along people, not just bots. "Invest in change-management, not just technology, and have your fearless and strongest leaders be in charge," he said.
