AI models are getting smarter by that day. Whether it's inference or solving equations, LLMS (short for large-scale language models) has proven to be the smartest child in today's class. But that's the coding that makes things really interesting. Large language models (LLMs) such as GPT-4, Claude 3.5 Sonnet, and Gemini 2.5 Pro have already written and debugged code like professional coders. And what began as an AI assistant for programmers now raises concerns about a complete replacement of human coders. Senior AI researchers at Alibaba think this future is not too far away.
Binyuan Hui, a staff research scientist on the Qwen team at Alibaba, recently shared his thoughts on X (formerly Twitter), claiming that “LLM inevitably outweighs humans in coding.” Not only does AI reflect the way humans learn, it's become so sophisticated that it even outcodes humans.
In his post, Hui explains that human coders usually go through two stages. He says that the beginning is memory and imitation. The coder learns syntax, studies examples, and reproduces great projects. The second stage is trial and error. At this stage, the programmer writes, runs the code, fixes bugs, and improves it through feedback.
Now, as LLM is becoming smarter, Hui argues that AI is also following the path of human learning. Pretraining allows the model to absorb and compress a huge amount of code, much beyond what humans might remember. “Rehnecortion Learning (RL) runs the model through feedback, but at speed and scale, humans couldn't match, and millions of deployments were completed in a short time,” Hui wrote.
He further allows individual programmers to debug one project at a time, but LLMS suggests that millions of iterations can be performed in parallel, allowing them to accelerate the growth curve. “This advantage of parameter capacity and iteration means it's only a matter of time before the model moves in front of a human,” he adds.
Hui's predictions are not isolated. A recent Stanford University study already reveals early signs of this change in the workforce. A study that analyzes pay data from millions of US workers found that entry-level software engineers have been hit hardest since the launch of CHATGPT in late 2022. The survey showed a 16% decline in employment for workers aged 22 to 25 in AI-exposed industries such as coding and customer service.
Research shows that instead of hiring new students, companies are leaning towards AI for repetitive coding tasks, while maintaining and expanding the role of senior engineers. In reality, AI acts as an assistant to experienced developers, but researchers at Stanford University warned that these same forces could ultimately extend to more advanced levels of work.
But the ability to code is just the beginning. Hui points out that with constant improvement and training, AI models may begin to prove themselves. He suggests that once the model reaches the artificial general information (AGI) stage, it is at the level where companies like Meta, Open and Google are racing. At that point, he insists that AI optimizes not only its own code. “Code is not only the foundation of human productivity, but also the starting point for recursive improvements. When models can write their own code and optimize, what we see is no longer AGI, but the first sign of ASI.
– end
