← All writing
Journal

The perennial software dilemma: build for learning or ship for use.

4 min read

A few years ago, I spent months on deep dives into security and privacy tools that would consume most of my attention. I did a lot of experimentation, and each experiment taught me a new piece of software or way of thinking. Whether it was understanding how encryption worked, or how network sniffing was used, how to harden a system, or to think through potential threats, each experiment was steeped in rich lessons. The exception was that none of them ever crystallized into something I ended up shipping into the world. I simply lost interest in the tools, and they faded away within weeks or months. The pattern established the rule that building for learning never routes into shipping for use.

In the past two years, I have immersed myself even deeper into experimenting with AI. Almost all of my attention has focused on how language models work, how we train them, or how we're building agentic architectures. Every evening ended with happy hours tinkering with new models and architectures. The reason the learning feels so rewarding, and so good, is because as an electrical engineer, I have a relatively strong systems-level intuition. I was never afraid to pull back the hood and learn how things worked. When it comes to AI, I'm doing the same. I am also learning how to deal with the real bottleneck, as it turns out, with the tools and resources needed to actually ship. It's almost poetic because the constraint is literally in tokens, which means I can't do everything, and I need to keep my bets small and focused. It's a system-level constraint that closely mirrors the resource dilemma that software entrepreneurs face. Even with this constraint, the pattern persists: shipping still loses to the magnetic pull of building for learning.

Humanoid robots are fascinating. There's a lot that I believe is still to be discovered there, and it's going to be compelling to explore them now that I'm more than halfway through my engineering degree. The interest is not just because they are cool, but because they can force a system-level perspective on what it means to integrate perception, planning, control, and environment. If I can make some progress with those systems, I suspect it will sharpen all of the other work I do too. There is a simplicity to the way people talk about AI right now, but the deeper story is always systems architecture. That is the real challenge, and it's precisely the kind of challenge I enjoy.

I think this dilemma crops up with many tinkerers who draw their deepest motivation from the build itself, not from a larger plan. It is far more pronounced in some fields than in others. In AI, the tooling is moving so quickly that learning itself becomes a kind of reward, which makes it harder to decide when to stop. Yet the cost of that delay is not only time. It is the opportunity cost of never shipping something real. The answer, as with most good engineering problems, is not to eliminate the tension. It is to make the constraints visible and to choose a direction with intention.

But for now, the learning is rewarding enough that the dilemma is an end cost, not a catastrophe. I'll eventually have a true software transition, and at that point, I'll have to make a more decisive call about what I am shipping and what I am learning. For the moment, I am content to keep building, keep reading, and keep developing the instincts that make me a stronger engineer. That is not a bad trade. It is just a trade, and trades are always easier to understand when you make them consciously.