67 Should People Understand Code Logic?
A few days ago, my good friend said that a colleague in their group vibe coded a bunch of code and submitted it, and when problems arose, he was asked to investigate and fix them. He said, "I don't know, didn't you submit the code? It was all written by AI, I don't know." This is really absurd. He complained to me, and then we chatted a bit more, and the topic narrowed down to: after people use AI to write code, do they need to understand the specific logic of that code? We had differing opinions: he thought yes, I thought no, and I'll explain why later. But we agreed on one point: people are responsible for the code. The situation I just described is an example of irresponsibility, which has nothing to do with AI and is directly related to the person. Regardless of whether the code was written by AI, if it's submitted under my identity and my account, then I should be responsible. Conversely, a responsible person, even if all the code they submit is written by AI, and even if problems arise and you go to them, they can investigate it themselves or have AI investigate it, and solve the problem. Then everything is fine, and cooperation is pleasant, right? So the core is responsibility. Responsible colleagues, whether they use vibe coding or not, are very nice people. Returning to the original disagreement: now that people write all code with AI, is it still necessary to understand the specific implementation details, to know your code modules inside and out, to be able to recite them with your eyes closed? Is it still necessary? My view is that it's not necessary. Half a year ago, I proposed a viewpoint: people control AI, and AI controls code. That is, as long as I use my AI to develop the code well, pass tests, and deploy it, and then even if there are problems, I solve them through my AI, then I am a responsible and qualified programmer. As for whether I need to be able to understand and know the implementation logic and specific details without AI, I think it's not necessary. That's a lower level of abstraction, and I don't need to consider it. Here's a concrete example: if you come and ask me about a specific logic in my module, I might not be able to answer because all my code is written by AI. But give me two minutes, and I'll command my AI to understand the context and logic, and then I'll reply to you clearly. Ignoring the two minutes in between, do you think there's any problem? Of course not. That's my view: people only need to control AI, so that AI can reliably accomplish various tasks. That level is sufficient. If you have time and are willing to understand, you can go deeper and grasp the details; if you don't have time or don't have a strong desire to understand the details, you don't have to. As long as you can quickly understand the current situation accurately through AI, that's enough. That's the first thing. Second, I recently saw an interesting saying: understanding is for participation. It gave an example: what's the difference between a programmer who understands code and one who doesn't? From a purpose standpoint, programmers understand code to participate in the software development activity and make decisions. For people to participate and make decisions, they need to know a lot of information, so they need to understand code. I think this viewpoint is correct. First, let me say what's right about it: I think if I'm going to participate in decision-making, of course I need to know the current state of the code, and based on the current state and the goal, find a path from the current state to the goal. That's the main decision-making action people do. Now, returning to the act of understanding itself: I need to understand code to participate in decision-making. OK, so how do I understand code? Can I understand through AI? Of course. Is there a difference between reading a book and asking someone who has carefully read the book and getting the same conclusion? From the result, there's no difference. If you say that from the process, you think you understand more deeply and thoroughly by reading line by line, I disagree. I think current AI can understand more thoroughly than humans. Humans are far behind AI in reading code and sorting out the current state. We have to accept that, admit it, and embrace it. Another point: about "understanding is for participation," it sounds reasonable at first, but on closer inspection, it doesn't point to a specific thing. The question I'm thinking about is: after understanding the code, what is it for? What is it for in the software development process? What do you do when you participate? I've been thinking about this for a few days and haven't figured out the answer. I think: is it to set a goal for AI, or to make technical direction decisions and judgments, or to give AI some empirical judgment when investigating difficult problems, or to give AI a market direction or strategic goal that I think is reliable for AI to pursue? People participating in any of these different aspects counts as participation. So what am I understanding code for? I don't have a good answer. The reason is: if you say participating in technical decisions, I would only tell AI to be simple and reliable. Is that a decision? It's not. It can decide on its own according to that standard. If you say participating in market direction and strategic formulation, I'm just guessing. AI can do market research, data analysis, and simulations in a sandbox. So its market direction judgment is certainly more reliable than mine. If you say troubleshooting difficult problems, what can I participate in? What I actually tell AI is some methodology, like using the process of elimination: first diverge to explore various possibilities, then verify and eliminate one by one, and finally converge on the most likely cause. That's my role in troubleshooting. But does that require me to participate? No, because it's a fixed methodology. You can just write it into the rules and tell AI. So my confusion is: after understanding the code and the project, what exactly should people participate in, and in which aspect can they play a significant and decisive role? That's what I think people should really think about and do. I haven't figured it out yet. A rough answer might be: people are responsible for organizing the whole activity, connecting everything, setting up the entire workflow and environment, and then they just sit aside. Do I understand or not? Do I participate or not? It's not that critical. The shopkeeper's wife sitting at the entrance of a small store with a big fan, is she participating in the retail activity? If you say yes, then she is. Analogous to this scenario, I think that's roughly the meaningful thing people can do at this stage, similar to the role of the shopkeeper's wife: if there's not enough money to stock up, I go get money; if there are expired items that can't be sold, I find acquaintances and connections to sell them cheaply to other channels, things like that. The rest of the store runs itself: customers come, self-checkout, and leave. Order transaction data—I can look at it or not. What's the point of my participation? Current AI programming is roughly in this state.

67 Should People Understand Code Logic?

67 Should People Understand Code Logic? A few days ago, my good friend said that a colleague in their group vibe coded a bunch of code and submitted it, and whe

67 Should People Understand Code Logic?

Comments

Leave a comment