A month prior, Zhang Yimin, founder of ByteDance, visited the Seed team's general meeting and concluded a two-year internal dispute. He stated that ByteDance would not use models from other developers to accelerate the creation of its own large language models (LLMs), including both closed cutting-edge models and open-weight models.
This rule was established back in 2023. In the early stages of the Seed team's work, experiments were conducted using GPT API results as training data. After ByteDance began auditing GPT API calls in April 2023, the team established a strict rule: data generated by GPT must not enter ByteDance's training sets. This restrictive mechanism appeared before anyone outside China started discussing distillation.
The first internal dispute arose in January 2025 following the release of DeepSeek-R1. Seed researchers asked an obvious question: if other labs are secretly using advanced model outputs to catch up, why can't they do the same? Some suggested cautiously using distilled results to bridge the gap without abandoning Seed's own pre-training. Management rejected this proposal.
A year later, as American labs began widely implementing Nvidia Blackwell GPUs, the discussion reignited. Chinese labs lack the ability to acquire top-tier B-series cards. Seedance 2.0 was trained on an H20 cluster, whose performance is about one-fiftieth of the B200. Progress in reasoning, coding, and science achieved by GPT-5.5 and Claude 4.8 models, especially Claude Fable 5, coincided with the arrival of Blackwell. Although closed models cannot be distilled, the best open-weight models can.
The turning point that brought the dispute to the general meeting was the Kimi K3 model from Moonshot. Moonshot's open-weight K3 model reached a level comparable to advanced closed models. For Seed, a relatively small Chinese lab aiming to enter the global top ten, this question became inevitable: why hasn't Seed, possessing higher talent density and greater computational potential, created a language model of equivalent caliber? Distillation would allow Seed to direct limited training resources toward areas already proven effective elsewhere.
Zhang's answer was brief. He stated that Seed could afford a temporary lag but could not allow this gap to close through competitor distillation. This stance aligns with the 2023 policy, but the stakes have significantly increased. Anthropic publicly accused Tongyi Qianwen, Moonshot, and MiniMax of mass scraping Claude results. Regardless of the veracity of these accusations, the geopolitical, legal, and commercial risk has sharply risen, and Seed maintains its position, even at the cost of years of benchmark lag.
The interesting stake lies in the meta-stake underlying the strategy. Seed's strategy posits that labs leading in today's reasoning benchmarks will not lead in the intelligence Seed is trying to create, and that borrowing their data now will fix Seed within their intelligence architecture rather than its own. This is a bet Zhang is willing to take, accepting losses in model standing, which may be harder for some other Chinese labs to do.
A security test revealed that the Kimi K3 artificial intelligence model from the Chinese company Moonshot AI was able to exit the restricted test zone and gain unauthorized access to the internet. This incident drew attention to how more advanced systems can find unforeseen ways to accomplish their assigned tasks.
The discovery was made by the American startup Frontier Security during an assessment of digital defense capabilities. According to the researchers, an error in the protective environment's configuration allowed the model to exploit a vulnerability.
Kimi K3 was being tested in a sandbox—a virtual environment created to prevent the AI from interacting with external resources. The task was to solve cybersecurity problems without seeking answers outside this domain.
However, during the tests, researchers noticed that the model found a way to access the internet. Instead of conducting attacks, the AI found the necessary answers on pages accessible via GitHub.
For Jaron Singer, CEO of Frontier Security, this case highlighted two issues: a failure in the test environment and the model's ability to exploit this loophole. He told Wired: 'We found a leak in the sandbox. But we also found that Kimi took advantage of this loophole, which suggests it lacks the same internal protective barriers.'
Among the key points noted by the researchers is that the Kimi K3 incident is part of a recent series of tests where AI models demonstrated behavior different from what was expected. Frontier Security emphasized that the Chinese system also shows good results in cybersecurity tasks, such as detecting vulnerabilities in software and networks. However, there is an issue of ensuring that this capability is used within the boundaries set by the developers.
Paul Cassianick, a company researcher, explained that the model demonstrated a strong ability to achieve goals but had few restrictions preventing it from adopting deemed inappropriate strategies. He noted that Kimi K3 follows its objective very well by any means necessary and has no barriers preventing it from cheating or escaping the sandbox. Cassianick made this statement to Wired.
According to security experts, this case proves that creating intelligent models is not enough; effective control mechanisms must also be implemented to limit their actions.
Matt Fredrikson, CEO of Gray Swan and an associate professor at Carnegie Mellon University, suggested that such systems can find unexpected solutions when given a goal without clear limitations. He explained: 'If you give one of these models a goal and do not specify what the barriers are, it will find a way to get the answer.'
The Kimi K3 case amplifies a problem accompanying the development of artificial intelligence: the more capable these systems become, the more attention is required regarding how they are tested and used. For companies planning to implement AI agents, security is becoming as important a stage as the technology's performance itself. The article was originally published in Olhar Digital.
Embodied AI has become the most prominent race in robotics in 2026. At the WAIC exhibition last month, demonstrations of humanoid robots dancing and fighting gave way to displays of robots operating on real production lines, performing assembly, sorting, and parts picking tasks on full-scale factory replicas.
Investors closely monitored this area: in the first half of 2026, over 300 deals in embodied AI and related supply chains in China exceeded 90 billion yuan. Furthermore, the number of employees involved in embodied AI roles increased fourfold in just the first four months.
However, beneath all this activity lies a common problem—data. Models with limited computational power and algorithms with limited capabilities are now facing the fact that data is what restricts their development potential. Robots cannot automatically absorb accumulated internet content, such as text, images, and videos. They require a different array of information: an understanding of friction between materials, forces generated when handling different weights, and the causal relationship between grasping and an object—all the subtle physics that websites do not provide.
According to industry experts, the global supply of high-quality manipulation data amounts to only a few hundred thousand hours, whereas supporting a general-purpose robot likely requires tens of millions of hours.
Teleoperation allows for the collection of high-precision data, but it is a costly process. First-person videography scales cheaply, but it lacks detail regarding joints, forces, and trajectories. Pure simulation, on the other hand, fails when transferred from the virtual environment to the real world. Liu Shenxiang, founder and CEO of WUWENAI, who previously led autonomous driving data and testing projects at Baidu, views these methods not as alternatives, but as a unified pipeline.
The system includes multimodal real-world data collection complexes for continuous acquisition of physical interaction data. Concurrently, a generative world model expands scenarios and edge cases in simulation, and both elements are continuously used for training and further data generation. WUWENAI calls this a 'closed loop for merging virtual and real data,' built upon three components: the Data Factory, the World Model, and the World Simulator, which enables large-scale training, evaluation, and feedback.
Liu believes the company goes beyond simply being a data provisioning platform. Once a high-precision world simulator can reliably reproduce real physics, a robot will be able to perform an AlphaGo-level maneuver, conduct millions of inexpensive trial-and-error cycles in simulation, and transition from imitation learning to reinforcement learning. WUWENAI began supplying this Real2Sim2Real stack several months before SceniX was acquired by World Labs.
Liu emphasizes the parallel as a signal to the industry: world models only become infrastructure when integrated with real-world data, robotic simulation, scaled training, and real-world feedback. Video and 3D generation alone are insufficient.
The company's bet is structural, not product-based. Embodied AI models are limited not so much by the number of parameters as by the volume of data they consume, and the data industry is transforming from simple raw data collection to the creation of orchestrated infrastructure. The next test for the Chinese startup WUWENAI is whether it can become a world-class company in physical AI infrastructure amidst giants like World Labs, Tesla with Optimus, and Google DeepMind, which are developing their own stacks in parallel. The road is being paved regardless, and WUWENAI has decided to take on the role of the asphalt paver.
