GPT-6 Has Arrived: Is the Moat of Embodied Intelligence Still There? – An In-depth Analysis of Anxiety, Opportunities, and the Underlying Challenges
Hello everyone, I'm your financial journalist. Recently, the tech world has been in an uproar as OpenAI released the Astra version of GPT-6 and explicitly stated their plans to develop humanoid robots. It's like a giant with a “dragon-slaying sword” suddenly saying, “I'm going to farm.”
For the embodied intelligence startups that are already working hard on robot hardware and algorithms, this is not just a source of pressure but also an existential crisis: If AI becomes intelligent enough, will they no longer need specialized “robot experts”? Will my job be taken away?
Today, we will break down this issue in simple terms, using the perspectives of several industry leaders from the Bund Conference, such as Zhu Zheng from Jijia Vision, Zhu Xing from Ant Lingbo, and Wang Qian from Zibianliang, to clarify what GPT-6 really means, where the moat of embodied intelligence lies, and whether we should panic or stay calm.
---
1. The Wolf Has Really Arrived: Why Is Everyone Suddenly Anxious?
First, we must admit that there are real reasons for this anxiety; it's not unfounded.
1. The Sense of a “Dimensional Reduction Attack”
Previously, people thought that developing language models (like ChatGPT) and robots (embodied intelligence) were two different things. But now, GPT-6 Astra can not only chat but also directly control robotic arms. In the demonstration videos, it can pick up chopsticks and insert them into a cup with great skill.
Zhu Zheng from Jijia Vision put it bluntly: “The wolf has arrived; stop dreaming.” His argument is that language models have already “devoured” multi-modal capabilities (seeing images, listening to sounds) and are now encroaching on embodied intelligence. Moreover, giants like OpenAI have a significant advantage in terms of talent, computing power, and funding. If they decide to get involved directly, startups will find it very difficult to compete.
2. The Urgency of the Time Window
Zhu Zheng believes that the next one to two years are a critical period. Current embodied intelligence companies have not yet established a true, irreplicable “moat.” If the giants enter now, their strong general cognitive abilities could quickly close the technological gap, potentially eliminating startups that are still in the early stages of product development.
**3. The Essence of Anxiety: Fear of “Premature Realization”
The core of this anxiety is that the value of embodied intelligence originally required years of development by startups. But now, large model companies could potentially “realize” that value by releasing a more powerful model and claim all the benefits for themselves. It’s like you’ve painstakingly developed a new drug, only for a big company to release a generic formula that replaces it.
---
2. Don’t Panic Just Yet: It’s Not That Easy for Large Models to Enter the Physical World
Although the anxiety is real, there were also calm voices at the conference. Zhu Xing from Ant Lingbo and Wang Qian from Zibianliang raised a sharp question: If OpenAI is so powerful, why doesn’t it focus on autonomous driving and take over Tesla?
**1. “Understanding the Principles” Doesn’t Mean “Being Able to Act”
Language models are good at “semantics” and “logic”—for example, knowing what a “cup” is and understanding the instruction to “put the cup on the table.” However, the physical world is complex, dynamic, and full of friction.
- Semantic Layer (Brain): Knowing where things are and where to place them—GPT-6 is strong at this.
- Physical Layer (Limbs): Knowing how to apply force, adjust angles, and handle object slippage—this is where GPT-6 currently falls short.
2. The Gap Between Data and Verification
Wang Qian pointed out that the progress of language models is backed by massive programming data and a comprehensive verification system. While video generation has improved, it hasn’t directly solved the problem of robot motion control.
Large model companies still need to address several challenges when entering the physical world:
- Real Data: They lack data on robots falling or failing to grasp objects in the real world.
- Hardware Adaptation: Different robotic arms and sensors require different control strategies.
- Verification Systems: They need a complex evaluation system to prove that robots are performing correctly.
3. The Gap in Industry Capability
Zhu Xing used autonomous driving as an analogy: Even if a model is leading in capability, it can’t skip the need to understand the context or immediately have a mature data pipeline. Determining what constitutes “good data” and accumulating the “know-how” for data collection is a more substantial barrier.
In short, **OpenAI has the strongest “brain,” but it doesn’t have the most flexible “body” or the experience needed to navigate the complexities of the physical world.”
---
3. Unraveling the Truth: What Can GPT-6 Really Do? What Can’t It Do?
To understand GPT-6’s real role in embodied intelligence, Xu Huazhe from Breakout Robot and the RoboDojo research team conducted some specific tests. The results were interesting: GPT-6 is neither a “god” nor a “useless tool”; it’s more of a “powerful advisor.”
1. Strengths: Semantic and Spatial Understanding
In the tests, Astra performed exceptionally well:
- It could identify objects on a table.
- It could pick up thin chopsticks and even insert them into a cup.
- It could plan non-grasping actions like “pushing” and “shifting.”
This shows that GPT-6 is very strong at “knowing what to do” and “understanding spatial relationships.”
2. Weaknesses: Complex Interactions and Fine-Motion Control
However, when tasks became more complex (e.g., using chopsticks to pick up something, stacking clothes, or passing objects with both hands), Astra’s success rate was low.
Reason: Physical interactions involve many invisible variables, such as friction, object elasticity, and slight changes in center of gravity, which cannot be determined by appearance alone and require real-time tactile feedback and adjustments.
3. The Key to Success: A Hybrid Architecture
RoboDojo’s research revealed a crucial point:
- Pure GPT-6 Astra control: 26% success rate.
- GPT-6 Astra + specialized embodied model (π₀.₅) hybrid control: 48% success rate.
- Details: In the hybrid architecture, GPT-6 only corrected 14.4% of the actions; the remaining 85.6% were done by the underlying embodied model.
This demonstrates that a combination of general reasoning and local operation experience is the best approach. GPT-6 handles high-level planning, while the underlying model handles the specifics.
Conclusion: GPT-6 doesn’t “solve” embodied intelligence, but it significantly enhances the system’s potential. It’s like an experienced commander, but the soldiers (the underlying control models) are still essential.
---
4. Where Is the Moat? Redefining the Barrier from “Data” to a “Closed Loop”
Since GPT-6 is so powerful, where does the moat of embodied intelligence companies lie? The saying “I have data” no longer holds true.
1. The “Data Barrier” is Being Diluted
Li Tianyu from Yuance Future shared a trend: They are already using GPT-6 to generate simulation scenarios.
- Previously: Building a 3D simulation required extensive manual modeling, which was costly and time-consuming.
- Now: GPT-6 can quickly generate logical and complete 3D scenarios through programming.
This means that “having data” is no longer a competitive advantage, as the cost of data production is decreasing. The advantage once gained through data accumulation is being eroded by the tool capabilities of general models.
2. The Real Barrier: The Ability to Continuously Identify Problems
If data is easily accessible, what remains challenging?
- Understanding when robots are likely to fail.
- Entering the relevant scenarios to collect “failure” data.
- Identifying the missing information from failures.
- Improving through training and evaluation.
This is the real barrier. Large model companies may be good at “mass-producing” data, but embodied intelligence companies excel at “learning from real-world failures.”
The moat is no longer about “how much data you have” but about “whether you have a closed-loop system that can continuously learn from failures and translate it into more reliable products.”
3. Physical Native vs. Semantic Fine-Tuning
Shen Yujun from Yingshen Intelligence and Min Wei from Yingshen Intelligence emphasized the concept of “physical native” models:
- Old Approach: Fine-tuning on language models (VLM) or video models to create embodied models.
- Risk: If the base model (like GPT-6) evolves, your fine-tuned models will quickly become obsolete.
- New Approach: Developing models based on physical laws and spatial changes (e.g., 4D world models).
The old approach is like building a house on sand; the sea level rises and it’s destroyed. The new approach is like building on rock—more stable.
---
5. The Future Pattern: Not “Who Eats Who,” but “Symbiosis and Division of Labor”
Finally, we need to move beyond a zero-sum mindset. The emergence of GPT-6 doesn’t mean the end of the embodied intelligence industry; rather, it represents a reorganization of industry roles.
1. Infrastructure vs. Application Delivery
Min Wei from Yingshen Intelligence believes that OpenAI might become an infrastructure provider, similar to water, electricity, and internet services.
- OpenAI/Anthropic: Providing powerful general cognitive abilities as an API or base model.
- Embodied Intelligence Companies: Responsible for hardware integration, scenario adaptation, physical interaction optimization, and after-sales service.
The commercial logic is that in a practical robot product, the semantic capabilities of large models may account for less than half of the value; the rest comes from understanding physical laws, mechanical engineering, and complex interaction. These are complex tasks that giants may not want to do and cannot be fully outsourced.
2. Reverse Empowerment: Robots Making AI Smarter
Shen Yujun proposed the idea that embodied intelligence could influence large models. Robots continuously collect real-world feedback through sensors, which is valuable learning material for digital models.
- Data on robot failures can teach AI about gravity.
- Data on failed grasps can teach AI about friction.
This shows that the physical world offers new paths for AI evolution. Embodied intelligence companies are not only users of AI but also providers of essential “nutrients” for its development.
3. Changing Competition Metrics
Future competition will not be about who has the largest model parameters but about:
- Who can achieve higher success rates with less interaction experience?
- Who has lower deployment costs?
- Who can handle on-site exceptions and control maintenance costs better?
- Who can create a fast feedback and optimization cycle?
---
Summary: Recommendations for Ordinary People and Professionals
1. For Investors: Don’t blindly invest in “pure software” embodied intelligence companies, nor completely dismiss hardware companies due to GPT-6. Focus on those with real-world data, the ability to develop physical native models, and control over deployment costs.
2. For Entrepreneurs:
- Don’t rely on fine-tuning large models for success; it’s a dead end.
- Dive into the physical layer: Focus on sensor integration, mechanical structure optimization, real-world data collection, and failure analysis.
- Embrace Tools: Use tools like GPT-6 to reduce simulation and data cleaning costs and invest the saved resources in core physical interaction research.
- Build a Closed Loop: Your moat is not the data itself but your ability to derive “physical intuition” from it.
3. For Ordinary People: GPT-6 won’t instantly make robots widespread, but it will make them smarter and more affordable. In the next few years, you’ll see more “semi-automatic” robots in factories and homes. They may still make mistakes, but they will become more reliable.
In One Sentence:
GPT-6 is an “accelerator” for embodied intelligence, not an “exterminator.” It speeds up the time it takes for the industry to prove its value but also raises the bar. The boundaries are still moving, but those who can transform the complexity of the physical world into unique product capabilities will maintain their position.