The AI Debugging Trade-Off: Faster Fixes, Fading Skills
theme: cyanosis
Since AI has gotten stronger, I've noticed my investment in technical research keeps dropping, and my articles have basically stopped. I always feel that in the AI era, writing articles has lost its meaning. In the past, when I studied something and shared it, I was happy; now, whatever I want to express, AI already knows, sharing has lost its meaning.
Just like moving from live-action short dramas to AI short dramas, many things are slowly changing flavor, burning the heart.
In the past, when I encountered errors, I would carefully read the stack trace, open the debugger, set breakpoints step by step, trace variables, follow the call chain to find the root cause. During the process, I would flip through documentation, search for materials, open countless web pages. The greatest sense of achievement then was finding answers through my own thinking in the midst of chaotic clues.
Now, most of the work is just one sentence, one screenshot. Send the error screenshot to AI, describe the problem in one sentence, then wait for it to provide a fix. Simple problems are like this, complex problems are too. I rarely open the debug panel proactively, I no longer patiently comb through data flows, I no longer calmly read source code line by line. Problem-solving speed has increased, but the thinking process is directly skipped.
Many times, I want to rely on myself to implement requirements and solve problems, but no one gives you time.
AI makes many efforts tasteless and worthless.
In recent projects, internationalization refactors, major Vue version upgrades, the entire system has many service chains, wide scope of changes, most of the code adjustments, compatibility fixes, migration adaptations were almost all done by AI. I was busy receiving AI's output code, copying, committing, checking results, almost no time to carefully read each piece of logic, slowly trace and debug. When encountering anomalies, my first reaction is not to analyze myself, but to keep sending screenshots to AI, waiting for it to provide the next patch.
At first, it just felt efficient, saving a lot of repetitive tedious work. But over time, an uneasiness slowly surfaced: I seem increasingly unable to understand the complete logic behind these changes. These codes are written by AI, I'm just a receiver, not a creator.
This dependence is silent. AI smoothed over the pain of debugging, and also took away the abilities I developed through that pain. We don't need to think about edge cases, don't need to sort out dependencies, don't need to understand underlying framework changes, all difficult problems have ready-made answers. Over time, the code order that predecessors built with their own hands will gradually shrink and dull.
The scarier part is the long-term outcome. In a while, more and more code in the project will be iteratively modified by AI, layers of patches stacked, logic entangled. By that time, our control over the system will continue to decline, facing this large expanse of AI-generated code, we will become increasingly unfamiliar, even unable to read or modify it. Code becomes a black box, we are just an intermediary responsible for relaying requirements to AI, and risks begin to intensify...
AI is not wrong, it is a powerful tool, it should be used to liberate repetitive labor, not replace thinking. But when we habitually outsource thinking, debugging, and tracing, a cruel problem appears: If merely using AI and describing requirements clearly can produce code, where exactly is our unique value?
The reality is cruel: any programmer who understands code, as long as they can talk to AI, can do the same thing. Replacing you might be a newly graduated intern.
When coding, debugging, and error-fixing can all be handled by AI, the barrier that distinguishes us from ordinary people will gradually disappear. The skill of writing code that we pride ourselves on, if it's being eaten away bit by bit by AI, what will make us irreplaceable in the future?
So.....Alright, I'll stop writing here. My AI task is done, time to work.
Top 2 from juejin.cn, machine-translated. The original thread is authoritative.
Being able to write out this confusion is itself quite valuable. AI can generate standard answers, but it can't replace the feeling one gets from actually stepping in the pit.
I deeply relate. Today, when writing requirements, I initially wanted to implement it using old-school programming. But after writing for a while, it felt too slow. Accustomed to AI generating code in large chunks, now typing line by line myself feels slow—just as after getting used to watching short videos at 2x speed, you can't get used to 1x speed. Spot-checking is too slow [crying].