Reading posts on HN, I notice that even across different programming domains, surprisingly similar problems emerge. Seeing this, I can't help but think that programming might ultimately be a matter of organizational theory. Conway's Law speaks to this as well. Code structure and approach shift significantly depending on how an organization is structured. From that perspective, it makes sense why Gen AI coding yields such starkly divided opinions.

Zooming out, the debate largely boils down to the downsides of severed junior pipelines and lost tacit knowledge on one hand, versus the upside of being able to build everything you envisioned on the other.

Honestly, I suspect that programming in the future will become something like the art of bonsai, a form of sculptural cultivation. Just as you dig furrows for water to flow through, and once the water emerges it follows those furrows, Gen AI will pour out an immense volume of code, and we will shape that code into form. It is like when you build with IoC based frameworks. If you do not adopt hexagonal architecture, the code inevitably converges toward the patterns the community recommends anyway. Ultimately, it is less about individual code fragments and more about hoping those fragments do not go wrong, contained by guardrails that ensure they do not exceed certain thresholds. If that is the case, programmers might end up focusing on questions like, "How much loss is acceptable?"

Many programmers say, "You need to know the entire codebase." But as someone who has delivered over 30 projects with codebases of 60,000 lines or more, I know that once my own code alone surpasses 40,000 lines, it becomes nearly impossible to hold it all in memory. So I document it in broad categories. And the next day, I have to retrace where I left off and refresh my memory, which takes a long time. This is one area where Gen AI genuinely excels.

Personally, what I have noticed when looking at other people's code is that those who work at a low level often struggle with high level integration, and vice versa. When cognitive resources become concentrated, specialization deepens in one area while others are neglected. And how one manages those cognitive resources ultimately diverges depending on the domain and the specific problem at hand. Just as the mindset required for architectural design differs from the mindset required for diving deep into implementation.

My workflow, broadly speaking, has not changed. My main job has always been to grasp a legacy codebase, open up the black box, connect debug logs one by one, find safe footholds, and add features. The only difference now is that the black box called "legacy codebase" has simply become "Gen AI code." It is not like I could ever claim my own code was bug free to begin with. Ultimately, what mattered was which bugs to contain and how far to wrap them. Open source programmers, however, are different. Their code blocks serve as their cognitive solutions and their reputation, so their approach differs from delivery oriented code. That is because code quality itself contributes to their prestige. For me, it is the brands and the number of companies I have delivered to that build my reputation. That distinction seems to divide programmers at a fundamental level.

In that sense, I think future programming work will split into roughly three categories. The first is the exceptionally talented group that creates the datasets Gen AI will consume. In other words, the renowned open source groups we already admire. In AI labeling terms, they will be the ones producing the gold standard answer sheets. I believe they will survive and remain admired, just as they are now, though their numbers will shrink even further.

The second group will be those who make Gen AI code flow through industries and organizations. When Gen AI pours out code like water, they will be the ones managing the channels to prevent it from flooding. Most people will migrate into this role.

The third group is QA, and I think they will survive. They will execute Gen AI code and relay to the channel managers when the watercourses are going wrong.

In truth, things are already moving in this direction. It is like how no one thinks about how OOP inheritance maps to vtables anymore. People just treat objects as abstract concepts and move on.

Personally, I think the overall flow in Gen AI coding will shift toward implementing system state in the form of state transition diagrams, a return to the primitive form of Parnas's theory of information hiding. Parnas's theory states that "modules should hide design decisions that are likely to change." Humans interpreted this as something like private variables in OOP, but then leaked internal implementations to the outside anyway. When business policies end up scattered across controllers and services, information hiding has effectively failed. The decisions most likely to change are not hidden in one place. They are dispersed throughout the system. But in the Gen AI era, ironically, information hiding might actually become easier to achieve. You just define the state transition from "payment pending" to "payment completed," and that is it. Ultimately, we might end up circling back to earlier approaches, do you not think?

Anyway, I am not asking for someone else's answer. I need to find my own way. Because my way and someone else's way are different, just as my programming and someone else's programming are different.

Programming is ultimately the work of constructing a mental model, and that mental model is inherently personal.

I just hope I can survive into the next era.

Reflections on software engineering in the age of AI