News of a startup's funding round in Paris last month captured the attention of the global information technology industry. AMI Labs, a startup founded by Yann LeCun, the New York University professor known as the godfather of artificial intelligence, raised $1.03 billion (approximately 1.55 trillion won) in initial investment. Various venture capital firms as well as Nvidia and Samsung Electronics (005930.KS) participated in the funding. The list of individual investors included IT giants such as Amazon Chairman Jeff Bezos and former Google Chairman Eric Schmidt. With this massive capital, Professor LeCun's AMI Labs is now pursuing world model development.
World models are next-generation AI technology that global companies are eyeing. As the name suggests, they refer to AI models that understand and simulate how the real world operates. Nvidia defines world models as "neural networks that understand the dynamics of the real world, including physics and spatial properties." Simply put, they are models that understand the principles governing the real world and can replicate them inside a computer.
Physics Understanding Enables Simulation Predictions
Understanding dynamics is a crucial factor that differentiates world models from existing AI models. When you ask ChatGPT to draw a picture or use Nanobanana to create a video, AI produces images that resemble reality. It depicts scenes like an apple falling from the sky or a boat moving forward as oars paddle through water. However, generative AI does not actually know that gravity causes the apple to fall to the ground. It merely describes phenomena based on countless images and videos it has learned and user commands. World models, in contrast, understand the principles that cause phenomena to occur.
Even before world models, digital twin technology existed to convert real-world spaces and object placements into data and transfer them into three-dimensional computer environments. World models resemble digital twins in that they create a virtual replica that closely mirrors reality. However, world models include capabilities that transcend the inherent limitations of digital twins: prediction. While digital twins focus on replicating the real world in virtual space, world models go a step further by learning spatial configurations and physical dynamics of reality to predict future outcomes.
NC AI is a leading domestic company that has entered world model development. Last month, NC AI unveiled its self-developed "World Foundation Model." NC AI's world model also focuses on predictability. Jang Han-yong, head of NC AI's Physical AI Research Lab who leads world model development, emphasized, "When it comes to controlling robots in industrial settings, predicting a physically accurate future is far more important than flashy visuals." Jang explained, "NC AI's World Foundation Model is a latent world model that skips the video rendering stage to increase inference speed and concentrates computational resources on physical consistency."
World models are tools for prediction. Therefore, the quality of world models depends on how accurate their simulations are. The IT industry consensus is that simulation accuracy is determined by visual accuracy and physical consistency. Jang said, "The main challenge is creating virtual environments that look exactly like reality while simultaneously implementing physical laws such as gravity, collisions, and friction within the virtual world."
Hyundai Motor Also Researching; Interest from Manufacturing and Logistics
Predictability has immediately stimulated industrial demand. Interest is emerging first from manufacturing, logistics, and construction sectors that are far removed from IT. This is because using world models before attempting production equipment installation or work flow design allows companies to evaluate results in advance. From a corporate perspective, they now have a means to examine more accurate return on investment and accident probability before making changes to their workplace environments.
Cases of world model utilization are gradually increasing in Korea. Business is progressing in the form of IT companies integrating physical AI into manufacturing, logistics, and construction workplaces. LG CNS (064400.KS), a major system integrator, has already conducted world model construction projects for clients in manufacturing and logistics sectors. In manufacturing client cases, world models were used to optimize entire processes combining production lines and logistics flows within factories. Optimizing process recipes by predicting physical phenomena for each unit process is also possible. In logistics, when designing entire work procedures from receiving to shipping, the design structure was established by reflecting cargo flow and worker movement patterns. Additionally, LG CNS plans to introduce world models to its digital twin platform in the third quarter of this year to strengthen competitiveness in related businesses.
Ju Kyung-hee, senior consultant at LG CNS's Smart Factory Business Division, noted, "Demand for world models is high in industries where robots and automation equipment are rapidly spreading, such as manufacturing and logistics," adding, "There is great interest in predicting outcomes according to various scenarios beyond simple data analysis." Ju elaborated, "In these industries, overall productivity varies greatly depending on worker movement patterns and cargo volume changes, so companies want to use optimal world models to derive operational strategies."
Shtagen, an autonomous manufacturing platform startup, is collaborating with Hyundai Motor on world model-related projects. Since 2022, the two companies have been researching ways to introduce world models to Shtagen's autonomous manufacturing platform "Meta Robo" and replicate Hyundai Motor's (005380.KS) Ulsan plant in virtual space. The approach under review is to transform the Ulsan plant into a Software-Defined Factory (SDF). SDF refers to a next-generation smart factory platform that integrates and controls all elements of a manufacturing plant through AI-based software.
Kim Won-hyun, CEO of Shtagen, said, "The ultimate goal of Shtagen's world model-based SDF is to achieve process optimization," identifying "sim2real and real2sim-based robot control" as the most important technology. Sim2real refers to the concept of applying results tested in virtual environments to actual environments in reality, while real2sim means the opposite.
Three Years Until Commercialization
Experts estimate approximately three years until world model commercialization. While early adoption cases of world models are being discovered in industry, experts agree there are many hurdles to overcome before widespread adoption. Ju said, "The three prerequisites for world model commercialization are data standardization, high-performance GPU operating costs, and ROI verification," predicting, "Within two to three years, cases of world models being used for actual operational decision-making and robot operations will rapidly increase."
CEO Kim analyzed, "World model commercialization will start first in industries without people, because obtaining and predicting data related to human behavior is the most difficult." He observed, "Automobile manufacturing that pursues dark factories (unmanned factories) and semiconductor plants that have already achieved high levels of automation will be the first to adopt world models."