Is the Terminator Coming to Corvallis? - Daily Kos

No, Oregon State University isn’t building Skynet. But something considerably more interesting is happening in Corvallis. The artificial-intelligence world may be approaching a major transition—from machines that primarily learn from language to machines that learn models of the physical world itself . And OSU, with a $50 million gift from NVIDIA founder and Oregon State alumnus Jensen Huang and his wife, Lori Mills Huang, is building an unusual collection of computing, robotics, simulation and real-world research infrastructure just as that transition is gathering momentum. The science-fiction version is the Terminator. The considerably less dramatic scientific term is embodied AI built around world models . And understanding the difference between today’s large language models and these emerging systems helps explain why what is happening at Oregon State could become important. There is also a personal reason this caught my attention. I spent 11 years as the engineering librarian at Oregon State, working with the faculty and students whose research culture is now expanding into this new generation of AI and robotics. That sent me down this particular rabbit hole. Most of us have come to know artificial intelligence through large language models—ChatGPT, Claude, Gemini and their relatives. They are astonishing systems, trained on enormous quantities of human-produced information to predict and generate language. But some leading AI researchers believe that this approach runs into a fundamental wall. Yann LeCun, Meta’s chief AI scientist and one of the pioneers of modern deep learning, has argued for years that language models lack something humans and animals acquire almost effortlessly: a working understanding of the physical world. A child learns that unsupported objects fall, that something hidden behind a chair probably still exists, that pushing an object can move it, that actions have consequences, and eventually that sequences of actions can accomplish goals. Much of that knowledge isn’t learned from sentences. It comes from watching, touching, moving, predicting—and discovering when predictions are wrong. That is the idea behind what researchers increasingly call world models . Instead of merely predicting the next word, a world model attempts to predict the next state of the world . Given what is happening now, what is likely to happen next? If I take an action, what will its consequences be? Meta’s V-JEPA 2, for example, is explicitly described as a world model trained from video for understanding and predicting events in the physical world, including robot planning and control in unfamiliar environments. The larger goal is AI that can learn, plan and operate in the physical world rather than remaining principally an intelligence of symbols and screens. And that creates an enormous new problem: Where does all the experience needed to train such an intelligence come from? That question made me look differently at something being built here in Oregon. Jensen Huang Comes Home to Oregon State NVIDIA founder and CEO Jensen Huang and his wife, Lori Mills Huang, met while engineering students at Oregon State University. Both are OSU graduates. In 2022 they gave $50 million toward what became the Jen-Hsun and Lori Mills Huang Collaborative Innovation Complex in Corvallis. The complex, now expected to open in late 2026, is a roughly 143,000-square-foot interdisciplinary research facility centered around one of the nation’s most powerful university supercomputers. Oregon State says its research will include artificial intelligence, robotics and intelligent systems, materials, climate, water and other computationally demanding fields. The Huangs themselves described the supercomputer as a kind of “time machine” for accelerating research, specifically naming robotics along with climate science, oceanography and materials science. But the interesting part isn’t just the computer. The building is being equipped with an extended-reality theater using high-resolution motion capture , an Invention Studio and Cyber Physical Playground where next-generation robots can be built and tested for real-world integration , flexible laboratories connecting research with field deployment, and facilities spanning physical science and engineering. In other words, there will be computation, simulation, sensing, physical machines and places for those machines to interact with environments. That starts looking suspiciously like an infrastructure for embodied AI. There is even something wonderfully Oregonian about the building itself. I first heard about the project from a retired Oregon State forestry professor, who also happened to be the person who clued me in that the CEO of NVIDIA was an OSU alumnus. He pointed out the unusual wooden structural system going into the building. It turns out that isn’t architectural window dressing. The Huang Complex is itself an experiment in advanced mass-timber construction , developed in collaboration with OSU’s College of Forestry and TallWood Design Institute. Mass-timber columns, beams and a composite deck are being engineered to provide the stability required by sensitive scientific equipment while substantially reducing embodied carbon compared with conventional concrete construction. Some of the timber was even selected from OSU’s McDonald-Dunn Research Forest. So the building that may help researchers teach machines about the physical world is itself partly an engineering experiment in that physical world. Oregon State Isn’t Alone What makes this more interesting is that similar pieces are appearing at several major American research universities. Carnegie Mellon , arguably America’s premier robotics institution, actually lists World Modeling as a Robotics Institute research field. This year CMU and Fujitsu also created a Physical AI Research Center devoted to AI systems capable of operating safely in real environments. Its new Robotics Innovation Center includes high-bay robotics laboratories, motion capture, a drone cage, a water tank and outdoor testing facilities. MIT’s CSAIL has organized an Embodied Intelligence research community around essentially the same problem: combining perception, sensing, language, learning and planning to create intelligent physical agents. MIT researchers explicitly identify scarcity of real-world interaction data as one of the major obstacles. One response is simulation—creating rich virtual worlds in which robots can accumulate experience before being deployed physically. Stanford has its Robotics and Embodied Artificial Intelligence Lab, which describes its mission simply: creating intelligent systems that learn through interaction with the physical world. And this year Northeastern University went even further, launching a Physical AI Research initiative and NU-WORLD , an open world-model platform. Its researchers describe the goal as enabling AI to learn from rich multimodal information, construct physics-consistent models, reason under uncertainty and act in the real world. These institutions aren’t necessarily executing some centrally planned national program. They are pursuing different pieces of a scientific problem that is beginning to converge. And that may be where Oregon State fits. From Language AI to Physical AI Imagine the emerging research loop: physical world → cameras and sensors → data → world model → simulation → prediction → robot action → physical world again The robot acts. Reality answers. The AI discovers whether its prediction was correct. Then it learns. This is fundamentally different training material from another trillion words scraped from the Internet. And Oregon State has some unusual advantages here. It isn’t only a computer-science university. OSU has robotics, engineering, forestry, agriculture, oceanography, climate science, materials research and field stations dealing with extremely complicated physical environments. Forests are worlds. Farms are worlds. Ocean environments are worlds. Factories and laboratories are worlds. Robots operating within them could eventually generate precisely the kind of multimodal, cause-and-effect experience that world-model researchers need. None of this means Oregon State has announced that the Huang Complex will become a “world model center.” It hasn’t. But the pieces are striking. A university with existing robotics expertise is receiving enormous AI computing capacity from one of the central figures in the AI hardware revolution, while simultaneously constructing facilities for simulation, motion capture, robotics, sensing and real-world experimentation. Meanwhile, some of America’s leading AI institutions are converging on the proposition that the next major step after language models may be machines that don’t merely describe the world but learn how the world works by predicting and interacting with it . Oregon State may turn out to have a very interesting seat at that table The Darker Side to Teaching Machines About the World There is another side to this technological transition that deserves attention now, before the infrastructure becomes ubiquitous. World models need experience. Experience means data. And an AI trying to understand the physical world potentially wants vastly more than the books, websites and conversations consumed by today’s language models. It wants video, sound, spatial relationships, movement, physical interactions and the consequences of actions over time. That could produce extraordinary scientific tools and extraordinarily capable robots. It could also produce the most comprehensive surveillance infrastructure human beings have ever constructed. We can already see a primitive version of this problem. AI-enabled camera networks can identify vehicles and reconstruct movements across communities. Today’s systems may principally recognize license plates or objects. Tomorrow’s physical-AI systems will increasingly be designed to understand scenes, relationships, activities and changes occurring through time. The same general technological progression that allows a robot to ask, “What is happening around me, and what is likely to happen next?” can allow a surveillance network to ask exactly the same question about us. That doesn’t make world-model research bad. It means its governance matters enormously. And there is a second issue that is particularly relevant here in Oregon: the physical cost of all this computation. Data centers require land, electricity, transmission capacity, cooling infrastructure and sometimes enormous public investment. Yet the number of permanent local jobs produced by a hyperscale data center can be surprisingly small compared with the resources devoted to it. That distinction matters. A university research complex filled with scientists, engineers, graduate students, laboratories, robots and collaborative research is not economically equivalent to a giant warehouse filled primarily with servers. The first can become an intellectual ecosystem. It trains people. It generates discoveries. It spins off companies. It attracts researchers. It creates relationships with local schools and industries. And, ideally, knowledge produced there becomes part of a public scientific commons. A conventional data center may provide construction work, taxes and some permanent employment, but communities are entitled to ask whether those benefits justify the electricity, water, land and infrastructure it consumes. That conversation is already happening around Corvallis. I recently had exactly this discussion at a local beer festival with a friend involved in city government. His objection wasn’t to artificial intelligence or advanced computing. It was to data centers in and around Corvallis that consume substantial local resources without producing enough lasting local employment to justify the bargain. I think that is precisely the distinction we should be making. The question isn’t whether we build AI infrastructure. It is what kind of AI infrastructure we build, who benefits from it, and what the community receives in exchange. The Huang Complex potentially represents the more attractive version: enormous computational resources physically connected to researchers, students, robotics laboratories, scientific disciplines and the surrounding world. But even that model raises questions that universities should confront early. Who owns the sensor data collected to train embodied AI? Can people inadvertently captured in that data be identified? Can research datasets created for robotics later be repurposed for surveillance? Who controls world models trained partly from public environments? What happens when university research migrates into commercial systems? And how much electricity and other physical infrastructure should communities devote to AI computation? Those aren’t arguments for stopping the research. They are arguments for developing the governance alongside the technology. Because the transition from language models to world models may ultimately mean something much larger than better ChatGPTs. We may be beginning to build machines that continuously observe the world, construct internal representations of it, predict what will happen next, and act upon those predictions. Oregon State could become one of the places where that transition is explored. If so, Oregon has an opportunity to ask an equally important question while the technology is still young: Can we learn how to build physical AI without accidentally building the infrastructure for an automatically observed society? Jasciu writes develops the governing logic of Constitutional Civic Realism here on the DailyKos at My Profile and for academia at: Zenodo