(Yicai) Nov. 2 -- Chinese video streaming platform Bilibili has hosted a forum on world models during its inaugural Super Science Night. There, entrepreneurs offered assessments on world models’ development stage, including admitting that they remain far from industrial-grade readiness and that their best-fit applications do not lie in gaming but in real-world scenarios.

“If we compare physics to a 100-kilometer marathon, we are only at the five-to-10-km mark,” said Chen Boyuan, co-founder and chief research and development officer of Chinese world models startup Inverse Matrix Technology. World models and physical artificial intelligence are still in the early stages, he noted, adding that the industry can stage polished demonstrations in controlled laboratory settings, but a large gap remains before they are ready for industrial deployment.

Even though the industry is at an early stage, Chen believes that the technology roadmap will soon converge, with applications landing in multiple real-world scenarios, much like the trajectory of large language models.

Founded earlier this year, Inverse Matrix aims to develop general-purpose world foundation models. It has completed two financing rounds, one in March and the other in June, raising over USD10 million and USD100 million, respectively, according to media reports.

World Models Are More Suitable for Real-World Scenarios

Many industry insiders previously suggested that gaming, film, and television were the most promising use cases for world models. However, based on the entrepreneurs’ assessment at Super Science Night, world models are better suited for domains that demand rigor and physical correctness, such as autonomous driving, robotics, and industrial simulation.

Although applying world models to gaming is theoretically feasible, the priority of games is visual polish rather than strict adherence to the laws of physics, said Jiang Zheyuan, founder and chairman of Chinese humanoid robot startup Songyan Dynamics.

Embodied AI training is one of the most promising entry points for world models, Jiang believes. When robots learn through trial and error in real environments, companies must bear costs for labor, equipment depreciation, and parts wear, he explained. If part of that training could be simulated in a sufficiently reliable world model, those costs could be replaced by far cheaper computing costs.

The most compelling application of world models is enabling machines to work in places humans cannot reach, such as high-temperature underground mines, Chen noted. Equipped on robots or software systems, world models could put their ability to understand the world into actual operation, extending the boundaries of human capability, he added.

Inverse Matrix is conducting deployment validation in high-value scenarios, such as data centers and chip manufacturing, where reliability requirements are extremely high, according to Chen. In cable plugging and unplugging in data centers, even a 90 percent success rate means the 10 percent failures could damage cables and create safety risks and operational losses.

Capital Is Flowing Into World Models

Even though commercialization is still in its early stages, capital markets have begun placing heavy bets on world models. Thirty-nine world model companies were founded and completed their first funding round this year, half of which closed their first fundraiser within two months, according to a report released by venture capital data provider ITjuzi in October. Last year, 30 such firms were established.

The accelerated inflow of capital is driven partly by industry leaders’ endorsement, and partly by the arrival of a technological inflection point, where LLMs can already solve many problems in the virtual world through AI agents, Chen told Yicai, adding that their applications in the physical world remain limited.

US computer scientist Fei-Fei Li and her French peer Yann LeCun are widely regarded as key figures behind the push for world models. In September, World Labs, the spatial intelligence company founded by Li, was acquired by US semiconductor giant Advanced Micro Devices for USD8.2 billion. Late last year, LeCun founded world model company Advance Machine Intelligence Labs, which has reached a valuation of USD3.5 billion.

World model entrepreneurs believe that, beyond all the hype, there is still a business question that must be answered. Once world models enter real industries, how much value will they actually create? Or will they drive up costs instead? Only real-world deployment can answer that.

Editor: Futura Costaglione