Comment by Abdur Razzaq Sindhu on "AI can build anything except an understanding of what you are building"

This really hits on something a lot of people miss when they hype AI as “able to build anything.”

It can generate code, systems, even entire products—but that doesn’t mean it understands why those things should exist or how they behave outside the patterns it’s seen. In a way, it feels like we’ve optimized for output before truly solving understanding.

What stood out to me is how similar this is to how modern AI is trained—more like “growing” something than engineering it step by step, which explains why it can be incredibly capable yet still unpredictable in deeper reasoning.

I think the real edge right now isn’t just using AI to build faster, but pairing it with strong human judgment—especially in defining problems, not just solving them.

Curious how others see this: do you think better tooling will close this “understanding gap,” or is it something fundamentally different from how current AI works?

The real question is whether better tooling / future AI will close this “understanding gap.” I believe the answer depends on the scale of the application:

For small, relatively simple applications (roughly up to 20-30 user stories), tooling and current AI can already replace most traditional development. The domain is shallow enough that deep modeling isn’t critical. For larger, long-lived enterprise systems, I don’t think better tooling alone will be enough. The gap isn’t just a matter of more context or better prompts. Building a good domain model requires imagination, conceptual invention, and iterative discovery through friction, the very process that current AI architectures bypass. AI generates structure quickly, but it doesn’t hold intent or accumulate nuanced understanding the way evolutionary modeling does.

Until AI can actively perform real domain modeling (discovering natural concepts, responsibilities, and boundaries), the most effective approach remains strong human engineering judgment paired with AI as a powerful accelerator for implementation. So yes, tooling helps a lot. But it hasn’t replaced engineering in complex domains, and I’m not convinced it will anytime soon.

Search Hashnode