Large-scale AI models—the vast repositories of linguistic, visual, and audio data that power generative artificial intelligence services—are becoming as important to the development of AI as operating systems are to the development of smartphones. , looks like a spatial platform (idea others are also eating noodles).What about Swiss startups? Jua has ambitions to use that paradigm to forge new frontiers in how AI can be used in the physical world. It cost him $16 million to build what is essentially a large-scale “physical” model of the natural world.
This company is still in a very early stage. Its first applications are in the modeling and forecasting of weather and climate patterns, initially with stakeholders in the energy industry. The company says it will be available in the coming weeks. Other industries that will be covered by the model include agriculture, insurance, transportation, and government.
468 Capital and Green Generation Fund are co-leading this seed round for the Zurich-based startup, with participation also from Promus Ventures, Kadmos Capital, Flix Mobilityfounders, Session.vc, Virtus Resources Partners, Notion.vc and InnoSuisse doing.
Andreas Brenner, Jua’s CEO and co-founder of the company with CTO Marvin Gabler, said increasing “instability” due to climate change and geopolitics is forcing the physical world to There is a growing need among organizations working in or anything else – for more accurate modeling and prediction.2023 was a high-water mark year for climate disasters, according to the US National Environmental Information CenterThe result is tens of billions of dollars in damages. This current state of affairs is driving organizations to plan for the right tools, not to mention better predictive tools for market analysts and others using their data.
In some ways, this is not a new problem, or even one that engineers have not yet tackled using AI.
Google’s DeepMind division built GraphCast.Nvidia is forecast net; Huawei has Pangu, a weather component announced last year; wave of interest. As highlighted in just last week’s article, projects are also underway to build AI models from weather data to focus on other natural phenomena. This report About a team working to bring new understanding to bird migration patterns.
Jua has two reactions to that. First, they believe their model is better than other models. The company claims that one of the reasons for this is because it captures more information and is 20 times the size of his GraphCast. Second, weather is only a starting point for considering broader physics questions and answers and challenges.
“Companies need to improve their ability to respond to all of this.” [climate] Volatility,” he said. “So, in the short term, that’s the problem we’re trying to solve. But as we look to the future, we’re building the first fundamental model of the natural world…we’re essentially We are building machine models that learn physics… and that is one of the key pillars of achieving artificial general intelligence, because understanding language is not enough. It is not enough.”
The company hasn’t launched its first product yet, but investors are acting on more than just the general AI hype.
Prior to Jua, Gabler led research at Q.met, which has been involved in weather forecasting for many years. He also worked on deep learning technology for the German government. Brenner works in the energy sector and previously founded a fleet management software startup. Together, these experiences lead to not only a technical awareness of the problem and potential solutions, but also a first-hand understanding of how the industry is experiencing this.
As we continue to develop our product, we also share our early work with prospective investors and customers to get their input on the data.
One of the objectives appears to be to take a new approach to the concept of what is included in a predictive model. For example, when building weather prediction models, “it’s pretty obvious to use weather stations,” Brenner said. But in addition, it incorporates what he describes as “much noisier data” to build the model, including recent satellite imagery, terrain, and other “more novel, recent data.” There is. “The main difference is that we are building this end-to-end system where all the data that was used at different steps in the value chain is all brought into the same pool,” he explained. The company said that training he has about 5 petabytes (5,000 terabytes) of data, compared to about 45 terabytes for GPT3 and (allegedly) 1 petabyte for GPT4. (However, understand that linguistic data may require less data than a physical world model.)
Another objective, and not a small one, is that the company is trying to build more efficiencies to reduce operating costs for itself and its customers. “Our system uses 10,000 times less computing than traditional systems,” Brenner says.
It’s worth noting that Jua is rising and getting funding, especially at this point.
Fundamental models are becoming a fundamental part of how the next generation of AI applications are developed, so companies that build and control them have a lot of value and potential power.
Today, the biggest drivers and change-makers in this space are companies like OpenAI, Google, Microsoft, Anthropic, Amazon, and Meta, all of which are American companies. As a result, other parts of the world, such as Europe, are actively searching for and funding their own champions to replace them. Notably, 468 Capital also supports Germany. Aleph AlphaThe company, like the basic model players in the US, also builds language models at scale, but appears to be working closely with potential customers. (One of the catchphrases is “Sovereignty in the age of AI”).
“Andreas, Marvin and the team are building the world’s first foundational AI for physics and the natural world. “We can provide powerful insights to disaster planning teams, agricultural organizations, airlines and charities,” 468 Capital general partner Ludwig Ensterer said in a statement.
AI companies that can help us better understand how climate change is affecting us, help us better plan for disasters, and perhaps one day even help us understand how to respond to disasters. has a distinctly “good guy” vibe. Reduce damage to the environment. And the big picture for startups looking to build AI that can understand the physical world is that it could potentially be applied to a much wider range of challenges in materials science, biomedicine, chemistry, and more. However, this outlook includes issues such as safety, reliability, etc., in addition to the feasibility of the model itself, even in rudimentary terms. It comes with a lot of questions about things you’re already thinking about. At this point.
“For a model to work and be accepted, you have to enforce consistency,” Gabler says. “To solve the problem correctly, we need to make sure that the model is actually learning physics from the ground up.”