“ It's an unrivaled mindset and a zero-sum game.
It's been a while since we've been able to make a robot think like this, and everyone in the house comes up and says two blessings. Recently, someone has opened a table.
On 29 October, Chiang's Innovation Day. Co-founder Liu Qing spoke on the main forum: high dynamic power control linkage, big model success rate "basically completely zero".
Six days later, at the World Internet Conference, Wu-Tang's founder, Wang Xing, took up the phrase. He acknowledged that this part of the AI model's exposure to the physical world was not particularly good, and that the alignment between input, output and the real world was poor.
Two phrases, saying the same shortcut, the moment the machine touches something, how hard is it to get it wrong?
The first full-size humanoid robot was available in 1973, but the brain was still lacking. The magic and test results of OpenAI's GPT-6 Astra are no stranger, 120 simulation missions, with an average success rate of 66.3%.
How can this be done? Even the two men above who dare to answer are the same, and the two speeches together are an invisible sword about Mount Sinai.
01.
Accounting as a Growth Account
Two people, two positions.
Astra runs slowly in the simulation, landing on the real machine, and rethinking for more than 18 seconds; the underlying control is 100 to 1,000 Hz, and the response must be within 10 milliseconds. 18 seconds to 10 milliseconds, 1,800 times.
He also calculated that it would take 30 seconds and a minute to live, give the big model a little bit of a try, maybe a couple of hours to try and get it done.
It's the level of certainty that guards the mind.
The Linux OS 1.0, which was released on Innovation Day, made a full-fledged determination time delay optimization, with a high-frequency load of 1000Hz, extending the momentum from end to end by less than 4 microseconds. The momentum is the magnitude of the fluctuation, not the delay itself.
Note: Chiang Kai-shek's clever hand, power control and tactile execution, is also the most unusual feature of a large model. Image source: Chiang Kai-shek
In the meantime, we're going to have to look at the trends.
He defines the GPT moment as having physical intelligence, where robots complete 80% of tasks with a single voice command in 80% of unfamiliar scenarios. By 2027, it is likely the critical point.
In his words, the world's top models are close to this target, only a few millimeters away. The model is stronger than the year before, and this little thing is missing, and the model itself will grow.
One sees an incredible 18 seconds on the production line, and one sees a few millimeters shrinking each year.
The answer is different because the business of the two people is not the same. The robots are already on top of the workforce, selling the labor that is available right now, losing hands once, and losing the real gold and silver on the customer product line. Wang Xing is selling the body and the platform, and the buyer's picture is that it can be upgraded backwards, these millimeters cannot be filled, and the legs of the eyes stand still.
One is zero, and the other is three years, and no one is talking nonsense.
02.
Pricing chain meters
The sword is not just about two people; its two answers are already pricing the entire industry.
According to data from the China Telecommunications Institute, a significant proportion of humanoid robots operate for less than 10 days a month, accounting for more than 60 percent of the student population, and only 13 percent of those who actually work in smart manufacturing.
So the guests were divided into two dials. The dials that dared to put the general big model in the front row, most of which had not yet entered the real scene, wrestled only once and again; the dials that had entered the real scene, the key moves did not dare to hand over to the big model. Every time lost on the production line, the losses were real gold and silver.
The cost falls first on the three types of companies. There are those who do not have brains, either themselves replenish the model, or return to the place of substitution; there are those who do not have brains, or are compiled by the body as a tool, or are overlooked; the hardest hit are the two unreliable middle ones, who do not borrow the brains of others, and themselves do not grow up to be real, only to regret the result.
The cost is also in the price of the data. Wang said that deploying thousands or tens of thousands of robots to harvest data is cheap and easy; he also admitted that the utilization rate of the data is still low, including with U-tree, and a large amount of data is wasted.
Scaling, the scale of which is the number of real machines, is difficult to replicate by pure model companies. One considers data as inventory, one considers data as a barrier.
The industry has not yet figured out how much money the data should be worth, and has begun to bid on each other for this uncertain price.
03.
Practicing with a Knife
I don't want to put the general model in a layer of physical AI, and I don't want to put it around the neck, it has to be put in a bottom-up scenario, it has to be done, somebody has to be responsible.
In a Shanghai convenience store, the G2, which is equipped with the Intel operating framework for cooking, packing, picking up and filling, has a success rate of 8.6% from the first day and climbed to 85.7% in five days.
In March 2026, the company's flat-line plant, with a production rate of 99.99%, ran steadily for seven months, with a six-day full-line live broadcast, over 64 hours of cumulative operation, completed 64824 production line missions, produced 17625 plates, and an average of 22.99 seconds of shooting.
Chart: The site of the event. Its robots have been running in real-life locations such as convenience stores, electronics factories.
In the meantime, Yuri dares to embrace the general model, because the basic task is his own.
From the structure of the human body to the system of perception, since last year, with pre-trained reinforcement training, the self-developed whole-body movement control model. The human-shaped robot G1 got a hint word, can color itself, wash pencil, pencil ink. The general model's brain can't stand for a while, the robot's legs are still standing, which is the bottom it dares let go.
Note: U-tree G1 humanoid robot, which can receive natural language prompts to complete tasks such as painting.
Interestingly, the two groups are walking on the same terrain. Zhiyuo has announced that Linux OS 1.0 will be fully open by the end of the year.
At the end of his speech, he laughed. Many people asked him what AI would do with his body. His answer was that it would become the side of AI, not the side that was hit.
A more realistic one is that the human-power ratio went from 1 to 1 to 3 to 1, with the goal of 10 to 1 at the end of the year to be 20 to 1 next year. 20 to 1 is 5 percent. Customers only need to keep 5 percent of the human resources budget.
It's more of a sword than a sword. It's more of a sword than a sword. It's more of a sword than a sword. It's more of a sword than a sword. It's more of a sword than a sword. It's more of a sword than a sword.
At this point, everyone can ask a question, does it get up?
