The current AI competition is nothing short of spectacular.
His name is Peng Mingyu. He is 24 years old, and his previous job was as a security guard.
Someone with seemingly no connection to AI whatsoever managed to fight his way through tens of thousands of contestants and secure a spot among the national Top 20.
△ Peng Mingyu, creator of Move Your Neck
This dramatic turnaround began during Peng Mingyu’s time as a security guard.
At the time, he was fascinated by the various Douyin effects he saw while scrolling through Douyin. After work, he began teaching himself how to create effects. As he gradually gained experience, Peng Mingyu went from a complete beginner to an advanced effects designer.
The most immediate payoff was that his side income from making effects exceeded his salary as a security guard. So he decided to quit and return home, making Douyin effects full-time.
However, creating effects required sitting in front of a computer for long periods, and Peng Mingyu often suffered from neck pain. In his own words, “Every time I finished making an effect, my neck would crack whenever I moved it.”
One day, while creating an effect for a game, he happened to discover that this type of motion-controlled game could actually relieve neck pain. Around the same time, he came across this AI competition, so he decided to enter and give it a shot.
Peng Mingyu’s product is called Move Your Neck. It is a browser-based, two-player motion-controlled health platform. There is no need to download anything or register—just open the webpage and start playing.
At the roadshow for the Top 20 finals, Peng Mingyu demonstrated how it worked:
As the demonstration shows, the mini-game lets you turn your nose into a table tennis paddle to play table tennis or use it as a fishing rod to go fishing—all while sitting in front of a computer and moving your neck in a fun way.
In addition, two other mini-games, Emoji Master and Imitation Master, get your entire upper body moving.
Describing how he managed to accomplish all this, Peng Mingyu said at the event:
I was basically illiterate—I couldn’t even speak a word of English. But even starting from zero and working alone, I still managed to build it in two months.
There’s no denying it: AI really deserves a lot of the credit.
This fascinating and inspiring story unfolded at the TRAE AI Creativity Conference, and the AI Peng Mingyu used was TRAE.
It is worth noting that this roadshow was only one part of the Creativity Conference. The venue was already packed with people. Take a look:
The reason was that, besides Peng Mingyu, there were many other fascinating contestants and projects. A total of 37,000 people registered for the competition, and the Top 20 were selected from 14,000 entries:
Some reached for the skies, some explored the depths, some built dreams, some worked with ancient texts, some created music, and some focused on parent-child experiences…
Here is a quick spoiler for the final results:
- First place: Token to Things, which won a RMB 300,000 prize;
- Second place: MotionFrame, which won a RMB 250,000 prize (RMB 200,000 for second place plus a RMB 50,000 social impact award);
- Third place: Open-Source Space: Satellite Watcher, which won a RMB 100,000 prize.
The lineup of award presenters was equally impressive. It included singer and Vibe Coder Hu Yanbin, Tim, founder of the filmmaking channel Yingshi Ju Feng, Wang Xingxing, founder, CEO, and CTO of Unitree Robotics (online), and “The Master” Lou Tiancheng, among others.
Hu Yanbin and Tim not only presented Peng Mingyu with his award in person, but also tried out Move Your Neck on site:
It was lively—extremely lively.
Behind all the excitement, however, this Creativity Competition not only vividly demonstrated that “anyone can create,” but, more importantly, gave people a sense of AI that is technically capable, powerful, and full of warmth.
A High School Student Takes Home the RMB 250,000 Grand Prize
Next, let’s take a closer look at the projects that won first, second, and third place, and explore what makes them so creative.
First Place: Bringing Tokens into the Physical World
The team that won first place at this year’s TRAE AI Creativity Conference targeted a very clear gap that has emerged as AI Coding has evolved:
If software can now be created simply by describing it, what about hardware?
△ The Token to Things team: Yang Zhenyu, Zheng Zhongpeng, and Lu Chenglong
The three members of Token to Things are Yang Zhenyu, Zheng Zhongpeng, and Lu Chenglong. All three are full-time entrepreneurs: one focuses on hardware, one handles industrial design, and one is responsible for software development and product implementation.
After meeting at Shenzhen Institute of Innovation and Entrepreneurship, they all hoped to enable people with no electronics or embedded-systems experience to build real hardware products simply by describing what they wanted—much like Vibe Coding.
So they created an AI hardware IDE and paired it with a desktop PCB milling machine.
For example, when a user enters, “Make me a musical keyboard,” the AI first confirms the requirements with the user, then generates a PRD, selects the components, and maps out the pin connections. It then automatically generates the code and completes compilation, flashing, and simulation. Once the user confirms that the design works, the system can continue by completing the PCB layout, routing the traces automatically, and generating the manufacturing files. Finally, it sends the files directly to the desktop machine, which mills the circuit board.
The entire workflow was already functional during the preliminary round, but it was more like a demo that worked only when every condition was perfectly met.
The real change came during the semifinals. Instead of staying behind closed doors and adding more features, the team handed the product directly to real users. They provided the software, board fabrication, and on-site support at youth challenge camps and Qihang Camp activities, interacting with approximately 500 users in total.
The problems surfaced immediately: some users did not know which development board to choose; some component documentation lacked pin information; some simulations ran successfully while the physical devices could not connect; and for others, switching computers caused the compilation environment to crash completely.
So during the semifinals, the team focused on one specific goal: making sure AI-generated results could actually be used for manufacturing.
When component information was incomplete, the system would pause and request the missing details instead of letting the AI guess. When compilation failed, it would determine whether the problem lay in the code, dependencies, or development-board configuration. When real users encountered bugs, Sentry recorded the circumstances, after which TRAE read the error context, identified the root cause, and fixed the issue directly—sometimes eliminating the need for anyone to describe the bug manually.
During their most intensive development cycle, they worked 14 hours a day for nearly five consecutive days, making 361 Git submissions. At last, real users were able to complete the entire workflow successfully.
The hardware itself was also repeatedly redesigned. By the finals, the new-generation PCB machine had been reduced by more than 50% in both size and weight. It also added automatic leveling and the ability to machine both sides of a PCB in a single setup. The machine was digitally integrated with the software as well: once the computer generated the circuit board, it could directly control the machine, which also provided real-time status updates.
In early August, they brought the equipment to a youth hackathon. A group of high school students completed their own hardware projects in eight days. During the on-site exhibition, a mother and daughter even wanted to build a piece of hardware themselves as a gift for someone else.
This is perhaps the most straightforward explanation of the name Token to Things:
To turn the tokens inside AI into something that can ultimately be held and touched.
At the awards ceremony, Tim also demonstrated Token to Things using a pair of roller skates, making a balloon figure perform the splits:
Second Place: The Project Reaches No. 22 on the Apple App Store Charts
The developer who won second place is Kyle, also known as “The Refactorer.” He is currently a high school sophomore, and his project is called MotionFrame.
△ Kyle, one of the creators of MotionFrame
Kyle and his teammate Xiaozhou are both from Zhongshan, Guangdong, and both enjoy playing badminton. Although they wanted to improve their skills, hiring a coach was simply too expensive. They had actually looked into it: a one-hour coaching session cost RMB 360, plus RMB 100 for court fees. At the time, Kyle had only RMB 200 in his pocket.
Moreover, even when a coach reminded them to “rotate your body” or that their “contact point is too low,” they often reverted to their old habits after leaving the court, because it is difficult for players to see their own movements. So, on the evening of July 11 this year, Kyle began building a project with the goal of creating an AI coach that could truly watch you play.
During the preliminary round, MotionFrame was only a demo capable of recognizing racket swings and offering suggestions. By the semifinals, however, Kyle had briefly considered adding multiple camera angles, 3D reconstruction, and a professional version for sports organizations. After writing code for several days, he realized the project was heading in the wrong direction and decided to shelve all of it.
For the next three days, he immersed himself in a badminton training facility, repeatedly filming and calibrating movements with coaches and players. When a mentor asked, “Why should users trust the AI?” he sought out professional coaches again and filmed both standard movements and common mistakes, giving the AI real examples against which to compare its judgments.
The semifinal version ultimately connected an iPhone, Apple Watch, and AirPods: the phone analyzed elbow angles, contact points, and swing speed frame by frame; the watch collected motion and heart-rate data; and the earbuds spoke directly to users during breaks between sets. The system also included 12 lessons, 100 training methods, and examples of 110 common mistakes.
During development, Kyle was responsible for setting the direction, making trade-offs, and testing on real devices, while TRAE converted clearly defined requirements into code. When the software needed hundreds of movement illustrations, TRAE generated them in batches within a few hours. After a real-world test uncovered nine issues, it fixed, compiled, and verified them one by one within two hours and 54 minutes.
One month later, MotionFrame passed Apple’s review and was officially released. At the finals, Kyle said that ten days after launch, the product had already climbed to No. 22 on the charts.
Third Place: A Father-and-Son Team
The contestants who won third place in this AI competition were a father-and-son team. They took their project all the way into space.
△ Tan Tan, creator of Open-Source Space: Satellite Watcher, and his son, a middle-school sophomore
The project is called Open-Source Space: Satellite Watcher. The father, Tan Tan, is 40 years old. He previously worked in solutions at a major internet company and entered the aerospace industry this year after being “graduated” from his former job. His teammate was particularly special: his son Bobo, who is still in middle school.
One father handled overall planning and development, while the other team member collected satellite data and models—and also served as his father’s product manager.
When Tan Tan first entered the aerospace industry, he discovered that the professional software used in the field had a very high barrier to entry. So when he registered for the competition on July 7, he set himself an ambitious goal: to use TRAE to build satellite situational-awareness and mission-simulation software.
The next day, he described the system’s objectives, data sources, and technical approach to TRAE in natural language. He asked TRAE to first generate a complete development task plan, then continue with the implementation.
The first complete execution of the development tasks took just over four hours, and the code in the project workspace had already reached 1.63 GB. After that, father and son tested the system while continuously reporting issues, resolving problems involving 3D textures, model compatibility, and server deployment. Three days later, they had a preliminary-round version.
At that point, however, it was still primarily in the “watching satellites” stage: it placed real orbital data on a 3D Earth, displaying each satellite’s current position and its upcoming flight path.
During the semifinals, Tan Tan began making it actually manage satellites.
The new version not only added domestic AOE orbital data and more 3D satellite models, but also created a virtual “TREA-01” satellite mission center.
Users could assign remote-sensing tasks to satellites, and the AI would automatically calculate suitable mission windows. When space debris approached, the system would trigger an alert, generate an avoidance plan, and replay the entire mission process. It even added a first-person cockpit view: the camera moved directly alongside the satellite, with Earth’s surface alternating between light and shadow below, creating the sensation that “you are the satellite.”
Development continued through repeated iterations during the semifinals. After qualifying on July 24, Tan Tan first fixed bugs and added database synchronization. Over the following days, he had TRAE generate new development documentation for telemetry, tracking, and command functions, then continued adding remote-sensing mission scheduling, AI mission planning, report export, and mission replay.
Some solutions that had looked promising on paper did not work well in practice. For example, he abandoned the original presentation of satellite obstacle avoidance and redesigned it from scratch.
At the finals, the father and son described their division of labor even more directly: Dad was the “architect and coder,” while his son gathered information, pointed out problems, and occasionally contributed new ideas.
But Tan Tan ultimately hopes to do more than build this one piece of software. After the competition, he plans to gradually open-source these capabilities, allowing more people to access aerospace software, understand satellites, and even help improve the broader “Open-Source Space” ecosystem.
Some Contestants Even Turned Dreams into Reality
Another team went a step further and turned dreaming into an app.
Li Yixuan, captain of the DreamTrace project, is still a university student. She came up with the product after dreaming that she attended her idol’s concert, received a handwritten autograph, and took a photo together. When she woke up, the excitement remained, but the details of the dream gradually faded. She had long wanted to create a tool that could preserve dreams.
For this competition, she and her 21-year-old teammate formed the “Anything Goes in Dreams” team. At the event, Li Yixuan jokingly summed up their roles: “I do the dreaming, and he turns my dreams into reality.”
During the preliminary round, DreamTrace was still just a demo focused mainly on displaying pages. By the semifinals, with TRAE’s assistance, the team had completed the backend and data capabilities and connected features including AI dream interpretation, AI dream illustration, real-world connections, and dream stories.
When users first wake up, they can record their dreams directly by voice. The AI can then turn the text into images, look for connections between dreams from a period of time and real-life experiences, and even continue the story as a novel or picture book.
Finally, users can publish their dreams to the community. Instead of a standard Like button, DreamTrace designed something more interesting: “I’ve dreamed this too.”
A dream that might otherwise disappear just minutes after waking can thus be recorded—and its dreamer may even find someone else who had a similar experience.
There’s no denying that these projects are both highly creative and genuinely fun.
How Can TRAE Make Programming Accessible to Everyone?
After looking at the contestants and their projects, one common thread quickly becomes clear:
There were geology students, laid-off UI designers, high school students, and even a former security guard who described himself as “illiterate.” Many of them had never belonged to the traditional programmer community at all.
Yet what they ultimately delivered was more than a few attractive demos. With TRAE’s help, some built geological modeling tools that ran on real devices, some launched products on the App Store, and others had already begun seeking real users for testing.
The fact that this is possible today is, to some extent, inseparable from the changes TRAE has undergone over the past two years.
When it first launched in early 2025, it was closer to an AI IDE. The AI primarily stayed within the development environment, helping users write and modify code. Six months later, SOLO emerged, and things began to change significantly: users only needed to provide a goal, and the AI could break down the tasks, read the project, call tools, and complete the work step by step.
At that point, TRAE was no longer solving only “How do I write this piece of code?” It was also addressing “How do I build this thing?”
As capabilities such as Skills, Rules, MCP, custom agents, Plan, and Spec were gradually added, many development tasks have now become a matter of clearly explaining the idea and constraints. The AI can then plan independently, call tools, modify files, run commands, and return the result for review.
Looking back at this evolution, what matters is not how many times the name changed or how many new entry points were added. What matters is that development steps that once had to be completed manually are being handed over to AI, one by one.
Of course, this does not mean that building products has suddenly become easy.
The first obstacle used to be, “I don’t know how to code.” Now, the real challenges are increasingly becoming, “What exactly should I build?”, “What did I get wrong?”, and “How should I change it next?”
In other words, the barrier of code is moving further down the process, while understanding, judgment, and trade-offs are moving to the forefront.
AI Competitions Can No Longer Be Won on Technical Skills Alone
This was one of the clearest impressions after watching the entire competition.
At today’s stage of AI Coding, simply “helping you write a piece of code” is increasingly becoming a baseline capability. The factors that truly set projects apart are beginning to move earlier in the process—before the code is written.
The Token to Things team first had to understand what hardware was missing in the age of AI. Kyle had to spend real money consulting badminton coaches to learn how expensive professional instruction was for a high school student. Li Yixuan wanted to continue the stories from her dreams…
AI cannot discover these problems out of thin air.
Only after identifying the problem does the next stage begin: breaking down the requirements, determining whether the AI has done the work correctly, testing when bugs arise, and deciding whether to tear down and rebuild what has already been written when the project goes off course.
So, looking back at this year’s competition, technology still determines whether a product can run, but creativity is increasingly determining why it is worth building in the first place.
In the future, not everyone necessarily needs to become a programmer.
A teacher can create tools for their students. A parent can develop an app around their child’s needs. An ordinary person can even “build” a piece of software on the spot to address a small problem that has repeatedly bothered them in daily life.
As the ability to make software becomes more widely accessible, the truly scarce resource is returning to people themselves: whether you live attentively, whether you can identify a problem worth solving, and whether you are willing to take action and actually solve it.
That may be the most significant aspect of the 37,000 people who entered this competition.
So, do you have an idea you’ve always wanted to build but never pursued because you thought, “I don’t know how to code”?
Why not give it a try with TRAE?