_robertharding_Alamy.jpg?#)
Vibe Coders and the Last 20%
In the field of AI, achieving 80% accuracy is relatively easy, but the final 20% is exceptionally difficult. As a developer, I deeply understand the impact of this "80/20 rule" in real-world projects. Whether it's autonomous driving or AI programming, improving precision often requires far more effort than anticipated. Take autonomous driving as an example: despite significant technological advancements over the past few decades, fully reliable self-driving cars remain out of reach. Even the most advanced systems cannot completely avoid errors in extreme scenarios. This "last mile" problem is not only a technical challenge but also a test of developers' patience and attention to detail. In 2023, Roose, in his coverage of chatbots, underestimated the severity of the "hallucination" issue. Recent studies show that even advanced models like GPT-4.5 still produce outputs with "hallucinations" in over one-third of standardized tests. As a developer, I am well aware of the complexity behind this phenomenon. AI still has fundamental limitations in understanding and generating content, especially when dealing with ambiguous or open-ended problems, where error rates rise significantly. While AI-generated code can quickly accomplish basic tasks, its error rate in complex scenarios is equally concerning. These errors may persist for decades, particularly when building innovative applications. The precision issues in AI programming not only affect development efficiency but also pose potential security risks. Writing code is just the first step; debugging and maintenance are the real challenges. For non-programmers, fixing AI-generated code is nearly impossible. For instance, if the "Lunchbox Buddy" image recognition system malfunctions, users could face significant frustration. In scenarios like automated email writing, code errors could lead to serious consequences. Non-programmers struggle to ensure the stable operation of these systems, let alone resolve complex issues like backend network calls. Roose's article overlooks the difficulties of debugging and maintenance, making his perspective overly optimistic and even resembling commercial promotion. As a developer, my experience with tools like Cursor AI is mixed. While they can improve development efficiency, the actual workload for debugging and maintenance often far exceeds expectations. Roose's examples are too simplistic to reflect the real challenges of AI programming. The effort required to debug and maintain complex systems is much greater than he suggests. Many parents and educators may discourage children from learning to code, believing that AI can already replace human developers. However, the reality is that large-scale projects still require experienced software architects, and this demand will not change for decades. In 2016, Geoff Hinton predicted that deep learning would surpass radiologists within five years. Yet, nine years later, radiologists have not been replaced; instead, many regions face a shortage of them. This case reminds us that the practical application of AI technology is far more complex than theoretical predictions. As tech writer Mayo Olshin pointed out, "If we blindly trust AI's output due to a lack of expertise, the long-term harm will far outweigh the short-term productivity gains." This insight profoundly reveals the risks of over-reliance on AI. As a developer, I believe the advancements in AI are undoubtedly exciting, but we cannot ignore the challenges of improving precision and the complexities of debugging and maintenance. Blind optimism and hype not only mislead the public but could also hinder the healthy development of the technology. In practice, developers—especially "Vibe Coders"—value flexibility and creativity in their work. AI-generated code often falls short in meeting these needs. Vibe Coders prefer manual debugging and optimization to ensure code quality and performance, rather than relying on automated tools. While this approach is time-consuming, it better addresses complex scenarios and personalized requirements. For Vibe Coders, the human touch in coding remains irreplaceable, and AI is seen more as a supplement than a replacement.
In the field of AI, achieving 80% accuracy is relatively easy, but the final 20% is exceptionally difficult. As a developer, I deeply understand the impact of this "80/20 rule" in real-world projects. Whether it's autonomous driving or AI programming, improving precision often requires far more effort than anticipated. Take autonomous driving as an example: despite significant technological advancements over the past few decades, fully reliable self-driving cars remain out of reach. Even the most advanced systems cannot completely avoid errors in extreme scenarios. This "last mile" problem is not only a technical challenge but also a test of developers' patience and attention to detail.
In 2023, Roose, in his coverage of chatbots, underestimated the severity of the "hallucination" issue. Recent studies show that even advanced models like GPT-4.5 still produce outputs with "hallucinations" in over one-third of standardized tests. As a developer, I am well aware of the complexity behind this phenomenon. AI still has fundamental limitations in understanding and generating content, especially when dealing with ambiguous or open-ended problems, where error rates rise significantly. While AI-generated code can quickly accomplish basic tasks, its error rate in complex scenarios is equally concerning. These errors may persist for decades, particularly when building innovative applications. The precision issues in AI programming not only affect development efficiency but also pose potential security risks.
Writing code is just the first step; debugging and maintenance are the real challenges. For non-programmers, fixing AI-generated code is nearly impossible. For instance, if the "Lunchbox Buddy" image recognition system malfunctions, users could face significant frustration. In scenarios like automated email writing, code errors could lead to serious consequences. Non-programmers struggle to ensure the stable operation of these systems, let alone resolve complex issues like backend network calls. Roose's article overlooks the difficulties of debugging and maintenance, making his perspective overly optimistic and even resembling commercial promotion. As a developer, my experience with tools like Cursor AI is mixed. While they can improve development efficiency, the actual workload for debugging and maintenance often far exceeds expectations. Roose's examples are too simplistic to reflect the real challenges of AI programming. The effort required to debug and maintain complex systems is much greater than he suggests.
Many parents and educators may discourage children from learning to code, believing that AI can already replace human developers. However, the reality is that large-scale projects still require experienced software architects, and this demand will not change for decades. In 2016, Geoff Hinton predicted that deep learning would surpass radiologists within five years. Yet, nine years later, radiologists have not been replaced; instead, many regions face a shortage of them. This case reminds us that the practical application of AI technology is far more complex than theoretical predictions. As tech writer Mayo Olshin pointed out, "If we blindly trust AI's output due to a lack of expertise, the long-term harm will far outweigh the short-term productivity gains." This insight profoundly reveals the risks of over-reliance on AI.
As a developer, I believe the advancements in AI are undoubtedly exciting, but we cannot ignore the challenges of improving precision and the complexities of debugging and maintenance. Blind optimism and hype not only mislead the public but could also hinder the healthy development of the technology. In practice, developers—especially "Vibe Coders"—value flexibility and creativity in their work. AI-generated code often falls short in meeting these needs. Vibe Coders prefer manual debugging and optimization to ensure code quality and performance, rather than relying on automated tools. While this approach is time-consuming, it better addresses complex scenarios and personalized requirements. For Vibe Coders, the human touch in coding remains irreplaceable, and AI is seen more as a supplement than a replacement.
