# A Year of AI-Assisted Coding: What I Gained, and What I Lost > Source: > Published: 2026-09-18 00:00:00+00:00 [← All posts](https://bucketli.github.io/) # A Year of AI-Assisted Coding: What I Gained, and What I Lost I started using AI-assisted coding in October 2025. In January 2026, I told the team I lead that we needed to start using AI throughout our development work. So it has been almost a year. First, some context: what kind of software do we build? We are a small company that has spent more than a decade building tools for database operations and data management. We started in China and later expanded overseas. Our product is called CloudCanal in China and [BladePipe](https://www.bladepipe.com/) internationally. The brands are different, but the underlying product is the same. Later, we built a related product called CloudDM, a database management platform for teams. It covers access control, approval workflows, and CI/CD for SQL changes. Its open-source edition is called [open-cdm](https://github.com/ClouGence/open-cdm). That is enough background. AI coding tools turn out to be particularly capable at building standardized software like ours. These products are not especially close to the hardware, nor are they deeply tied to a company’s private business logic. This led us to an uncomfortable question: does buying software still make sense when users can vibe-code something with roughly the same features in a matter of minutes? We tried to reassure ourselves. Data management is not our customers’ core business. Building something quickly is one thing; maintaining it over time is another. Still, one thing seems undeniable: competition has intensified. Over the past days, every moment of excitement about the productivity gains from AI coding has been accompanied by anxiety about where all of this is heading. That anxiety has never really gone away. In the past, we would normally assign one engineer one major feature for a month, along with a few small bug fixes. Today, the same engineer can produce several times as much. My attitude toward AI has been simple: use it boldly, but judge its output carefully. That becomes harder as the scope of AI-generated changes grows. After a while, it is almost impossible to understand every change in detail. Often, the only practical way to decide whether the system still works is to rely on tests. Code review starts to feel like a ritual. Is that acceptable? At least for now, I am not convinced. Models have blind spots. But human trust in them tends to be binary: some people never trust them, while others trust them completely. The frightening part is that, as a product gradually fills with AI-generated logic, there may come a day when we realize we no longer truly understand or control it. There will be no warning telling us exactly when that line has been crossed. And what happens if access to OpenAI or Claude is cut off one day? After all, they are half a world away from us. I have considered stepping on the brakes—for example, banning AI coding in certain parts of the codebase. But I have never actually done it. For now, when the people on the team can still piece their knowledge together and understand more than 80 percent of the code, perhaps the right thing to do is keep watching. There is another kind of loss that is harder to measure. The quiet concentration of writing code, followed by the satisfaction of looking back at a piece of logic that feels rigorous and elegant—that feeling is mostly gone. For a technical person, especially one who enjoys mathematical reasoning and becomes emotionally invested in the work, this may mean losing something that once brought genuine excitement. Can money replace it? Can material rewards replace it? No. There is a saying that some people work not only for the result, but because they enjoy the process itself. That may not fit neatly with modern society’s obsession with efficiency, but those people exist, and so does their stubborn attachment to the craft. One era is fading, and another has arrived. Will I like it? I don’t know yet.