Musk announced today the acquisition of cursor, and OpenAI released a new version of codex that leans towards developers. In addition, with the previous anthropoic focus on enterprise user technology, GLM's disguised price increase restricts ordinary users from leaning towards developers... These various behaviors indicate that achieving model leadership and even agi through continuous improvement of code generation capabilities has become a consensus among a wider range of market players This is the current optimal training path selection. Compared to natural language, code has three key advantages: One is strong verifiability, which provides high-quality feedback signals on whether it can run and pass testing; The second is the natural structure, which combines grammar, logic, and modularity, essentially compressing the expression of human thinking processes; The third is the high complexity of the task, which naturally includes multi-step reasoning, state management, and debugging iterations. This allows the model to continuously self reinforce in a closed loop of "write code run error fix", which is extremely close to a true intelligent system. Therefore, the industry is not betting on the application of "writing code" itself, but on using the code environment to train models' reasoning ability, problem decomposition ability, and iteration ability. The code is just a carrier, the goal is to give the model the ability to solve problems at the engineer level. This is also why, in the context of traditional scaling laws with diminishing margins, code has become a new growth engine: it provides denser, more reliable, and more automated training signals, while also having a clear path to commercialization. But code is not the end. Its limitations are equally evident: complete regularization, lack of physical constraints, and biased symbolic systems. This means that relying solely on code paths is difficult to solve uncertainty, noise, and complex environmental problems in the real world. So Yang Likun's second path also gained a lot of market attention: putting the model into a "feedback based reality system". The core is not the robot itself, but building a closed loop similar to a code environment - decision-making, execution, feedback, and correction. This path also has closed-loop capability, but its scalability is limited by the physical world and cannot expand as rapidly as code. But its value lies in being closer to the real world and closer to the "action capability" required by AGI. If code path training is about 'how to think', then real-world closed-loop path training is about 'how to act'. The ultimate form is likely to be the fusion of two paths: code capability provides "thinking", and reality closed-loop provides "action", together forming a true agent system. This is also the core of the current transition from "model" to "system". The path of AGI is likely to be: first learn to think in code, and then learn to act in reality.