2026 World Model Turf War: Poach Talents First, Define The Rules Later, The Number Of Real Competitors Can Be Counted On One Hand | Full Panorama

"Workplace Bonus" (ID: ZhiChangHongLi)

Written by / Jiang Yutian     Cover image / Still from The Thirteenth Floor

Art Editor / Jiang Yutian, Text Editor / Chen Tong

Technology has not been implemented in real scenarios, which does not prevent capital from entering the market. ·01

Talent anxiety under the narrative bubble · 02

How do academics view the current industrial status? ·03

List of leading figures: The creators of world models ·04

In the first seven months of 2026, 666 billion yuan of capital flowed into the world model track in China's primary market, with 58 companies securing financing. Four months ago, there were roughly 100 companies in China claiming to develop world models — if screened strictly by pre-training standards, no more than 5 companies are actually working on real world model products.

"There is a consensus in the academic circle on the conceptual interpretation of world models, but the business community does not share this view," said one interviewee.

The same is true in the public view. Apart from industry insiders, most consumers have a vague or even unfamiliar understanding of the meaning and value positioning of "world models". A researcher from a large tech company who worked at the exhibition booth for three days at WAIC told Workplace Bonus that the most frequent question raised by visitors during those days was: What on earth is a world model?

Last month, Li Feifei launched Atlas on the official website of World Labs, which once again boosted the popularity of the "world model" topic in the talent market. This multimodal world model marks a clear camera pose for each image at the input stage, based on which the model understands the 3D structure of the scene, so that creators can precisely control camera movement, rather than only generating a visually appealing video.

Even though there is a huge gap between technological development and external public cognition, world models have still become one of the tracks with the most severe talent shortage in the market. The narrative bubble has risen, how to distinguish which people and teams are making real products?

Unimplemented technology does not stop capital from entering the market

"The number of domestic companies that have truly reached the world model standard can be counted on one hand at present." Talking about the technical standards of "world models", YT, a PhD intern from a large tech company, told Workplace Bonus.

By the beginning of August, 23 new world model companies had been established this year, 18 of which completed the first or first two rounds of financing within several months of establishment, and 2 were labeled as unicorns at the seed round or Pre-A round.

The reason why world models can attract so much capital is that they are widely regarded as the "foundation of next-generation AI", the bridge connecting digital AI and the physical world, and the infrastructure for physical AI applications such as embodied intelligence, autonomous driving and robotics.

At present, this track with an unclear boundary is gathering capital and talent at an unprecedented speed. Overseas leading AI laboratories and robotics companies are developing world models as an independent research and talent direction. Google DeepMind has publicly recruited research scientists in this direction, focusing on multimodal generative models, real-time video generation and interactive world models; robotics company 1X has set up positions with an annual salary of 200,000 to 300,000 US dollars plus equity. [1] Runway has set up world model engineer positions with an annual salary of 270,000 to 370,000 US dollars plus options [2] . xAI offers an annual salary range of 180,000 to 440,000 US dollars [3] .

Li Feifei's World-Labs recruits talents based on professional competence regardless of academic background, and Yann LeCun's newly founded AMI Labs and other major technology companies and laboratories have begun to set up dedicated world model R&D teams, publicly recruiting talents in physical simulation and video prediction.

In this high-threshold field, talent is a scarce resource, and enterprises have fully advanced their talent recruitment plans. They no longer wait for the graduation season, but start to contact master and doctoral students in advance, and complete early screening through top research internships.

However, for entering this industry, "the window period for technical positions basically no longer exists, but there is still a window period for product positions." YT said.

[1] The 1X World Model Lab | Est. 2026

[2] Member of Technical Staff, Research Engineer-Runway-Madrona

[3] https://yubhub.co/jobs/job_28b01ce3-8a3/

Talent anxiety under the narrative bubble

This year, AI positions are expanding as a whole, high-end and compound talents are generally in short supply, and the demand for world model positions has increased significantly. A headhunter believes that the current situation is that "there is no one to train, so we can only poach talents from other teams". World models have high requirements for talents, and recruitment is quite difficult.

From a general perspective, ordinary large model algorithm positions are already relatively crowded, and what is lacking are high-matching compound talents who are proficient in both algorithms and engineering implementation; from a specific perspective, the intersection of VLA, world model/World-Action Model, RL and motion control, which is the brain/operation foundation of embodied intelligence, is one of the directions with the largest talent gap and salary flexibility.

Practitioners in the world model field basically come from the same technical pedigree. The most direct source is computer vision, 3D graphics and neural rendering, and talents who master the new technology stack are relatively scarce. At present, practitioners come from diverse backgrounds including metaverse, video generation, LLM, mathematical and physical background, as well as robotics, mechanics and electronics. There is a saying that former "metaverse" practitioners are the largest source of talents for current world models; the most core base model/WAM is still mainly composed of doctors in mathematics and physics and LLM/video generation researchers.

According to YT's experience, positions in laboratories are mostly filled through personal referrals and academic supervisor networks, while career changers are more distributed on the engineering side or peripheral positions. Positions in enterprises are more open.

If we summarize the minimum access threshold for these positions, practitioners need to have a master's or doctoral degree from universities in relevant directions, solid foundations in machine learning/deep learning, and at least one relevant scientific research or project experience. "We believe that a university diploma is a stepping stone, because it can prove the candidate's value density and the quality of being AI Native." said Cao Zhenhua, a researcher at Tsinghua University engaged in AI medical direction.

World model positions span multiple fields including generative AI, simulation and causal reasoning, reinforcement learning and real machine deployment, large-scale distributed training, etc., and the number of candidates who meet the requirements and have both academic and scientific research endorsements in each subdivision direction is very small. Therefore, the current realistic talent sources are mainly talents overflowing from large companies and master/doctoral students from universities.

For companies or research institutes with complete infrastructure, after stripping off repetitive labor such as data processing, the core base model team of a world model can operate with roughly 10 members.

Due to the large talent gap, enterprises have advanced the talent grabbing stage to the first or second year of doctoral students, and some large factories grab talents under the policy of "enjoy full-time treatment during internship". The highest internship salary for top university researchers at Momenta can reach 3600 yuan per day [4] .

The salary increase for internship positions is just the tip of the iceberg.

The industry consensus is that the talent window period for world models may only last two to three years.

The annual total compensation package for the chief world model/VLA algorithm position at a well-known domestic listed company is about 9 million to 10.2 million yuan. [5] The situation overseas is even more exaggerated. Waabi offers a salary range of 155,000 to 269,000 US dollars per year for world model research engineers, plus equity and bonuses. [6] To build the Super Intelligence Laboratory (MSL), Meta offered top researchers from OpenAI and Google DeepMind a 4-year compensation package of 300 million US dollars [7] . (MSL is positioned for AGI/ASI, and world model is one of its research directions.) The annual salary of domestic new doctoral graduates is generally in the range of 600,000 to 1.5 million yuan [8] .

For people who want to enter this industry, they need to carefully evaluate the long-term value of the company when making choices. If you join a company that lacks a clear profit model, has vague product planning or ignores talent training, you may face the risk of business stagnation or even bankruptcy, which will affect your personal career development. Some startups blindly chase the trend without establishing a sustainable business logic, and eventually cannot sustain their operations. Therefore, job seekers should pay attention to the core competitiveness, strategic direction and team stability of the company, and avoid ignoring potential risks just because of short-term popularity.

As a front-line researcher, YT admits that "the industry cannot be planned in the long term, because hotspots are constantly changing". So what methods can be used to judge whether a hotspot is a real demand?

One method is to test the sustainability of demand with the time scale, and observe whether the demand still exists after extending the cycle. For example, when world models are combined with humanoid robots, we need to treat them separately: world models solve the essential problem of interactivity, and humanoid robots are the narrative carrier. By observing whether the industry has a long-term consensus of 5-10 years, conduct a comprehensive evaluation from aspects such as academic research, continuous R&D investment of giants, long-term capital layout, and independence of underlying demands.

Another method is to evaluate the gap between publicity and actual capabilities through the observation of technical routes. First, judge whether the industrial technical route is feasible, and then analyze the gap between the two. The further the distance between publicity and actual capabilities, the greater the possible bubble. For example, when evaluating the authenticity and implementation feasibility of technology, it should be noted that the definition of world models is broad, and some projects only replace the labels of their original simulation, 3D or video generation businesses. "Pre-training" is one of the key criteria to distinguish authenticity from falsity — whether the company trains the model from scratch with massive data to understand physical laws, rather than only performing fine-tuning or application.

Finally, we can start from the essential needs of human society, assume that we replace the technical carrier, and see whether the technology still has demand and is still valid. For example, YT entered the industry from the metaverse, and although the technical carrier has switched several times, the technical direction has always focused on virtual interaction and content generation, without fundamental changes. "Narrative bubbles are inevitable. The most important thing is to polish the technology well. Even if commercialization is not realized now, the efforts you put in are valuable."

"There are indeed many technical hotspots, but there is no need to worry too much." Most practitioners believe that technological breakthroughs are happening every day, but the technology is not as complicated as it seems. After practitioners master the technology thoroughly, they can soon reach the cutting-edge. Building a solid foundation in machine learning/deep learning from top universities is sometimes more useful than reading more papers and doing more experiments.

[4] https://www.zhipin.com/job_detail/890d316f1df5aa850nR63d28E1tU.html

[5] https://www.liepin.com/a/79057849.shtml

[6] Research Engineer, World Models-Waabi-Venture Partners

[7] https://dxpress.gelonghui.com/live/1959365

[8] https://www.shobserver.com/staticsg/res/html/web/newsDetail.html?id=1100951

https://www.nowcoder.com/jobs/detail/464360

How do academics view the current industrial status?

When we talk about world models, what on earth are we talking about? Is it the realistic video generated by Sora, or the autonomous decision-making of robots? From the academic point of view, if this concept is not clarified, the industry cannot move from Demo to large-scale implementation.

The perspective of Li Feifei, the founder of world models, is to observe the function of technology in the closed loop of world models. The whole process is: the agent takes an action, the action changes the state of the world, the agent receives the observation, and takes the next action based on the observation. This also constitutes the underlying logic of current industrial division of labor:

The first category is renderers that output pixels: such as text-to-video and image generation. They are good at pleasing human eyes, have the most mature commercialization, but do not understand physics and geometry. You cannot expect to use them to design buildings or train robots, because they are good at pleasing human eyes, but cannot guarantee the correctness of physics and geometry.

The second category is planners that output actions: such as VLA models and embodied intelligence brains. They are the current gathering place of hot capital, responsible for telling the agent what to do next. But at present, no planner has been fully verified in the complexity of the real world and long-sequence tasks.

The third category is simulators that output states: this is the most difficult part of R&D. It outputs geometric data and physical parameters, highly simulating the real world. The biggest problem is that virtual objects generated by AI often have hidden defects, which will cause model penetration or collapse once connected to the physics engine. It is the bridge connecting the digital and the real world, and also the biggest bottleneck at present.

At present, world models are most widely used in the following five major industries.

● Autonomous driving

Autonomous driving is the battlefield closest to commercial returns at present, and it is also the field where world model applications are the most systematic and have the clearest commercialization path.

On the technical path and commercialization rhythm, different enterprises have different strategies. Huawei adopts a heavy investment strategy, carrying out deployment by end through the architecture system. The cloud-side world engine is responsible for generating massive training scenes, and the vehicle-side world behavior model directly processes sensor data and outputs driving instructions. In 2026, Huawei plans to invest more than 18 billion yuan in this field. In January this year, NIO pushed the new NWM version to more than 460,000 vehicles, which is the earliest released and earliest mass-produced autonomous driving world model in China.

Relevant practitioners mainly come from the intelligent driving teams of vehicle enterprises, autonomous driving companies and university laboratories. Huawei, XPeng, NIO, Li Auto, Waymo as well as Tsinghua University, Zhejiang University, Westlake University, etc. are the main talent sources, and the AI departments of large Internet factories are also transporting talents. In terms of background, a master's degree or above in computer science, automation, mathematics, vehicle engineering, electronic engineering, etc. is the basic threshold.

The industry generally believes that autonomous driving is the vertical scenario with the most mature commercial implementation conditions in the world model track, but the Sim2Real (simulation to reality) gap is still the top problem in the industry. No matter how realistic the simulation world is, it cannot reproduce all the physical characteristics of the real world, and real-world data must be used to fill the gap. The recall incident of Waymo was caused by the decision-making error of autonomous driving on waterlogged sections, and the simulation could not reproduce the real hydrological physical characteristics.

● Humanoid robots and physical AI

The overall scale of humanoid robots and physical AI has exploded in the past two years, with extremely