From a talk to senior students at CAU, Seoul, 2024-11-19.

Introduction

AI has changed the economics of software work. Tasks that once served as proof of competence, boilerplate implementation, API wiring, straightforward unit tests, basic debugging, can now be completed quickly with coding assistants. The result is not the disappearance of engineering value, but its relocation. Value moves away from routine production and toward judgment.

This shift is especially visible in hiring and mentoring. Explicit knowledge is cheaper to access than before. What remains scarce is the ability to evaluate, adapt, and justify technical work in context.

Knowledge Is Cheaper, Judgment Is Scarcer

Large language models compress a large amount of explicit knowledge into a highly accessible interface. That changes the power dynamics of experience, but it does not eliminate experience. Senior engineers still hold an advantage because experience is not just recollection. It is contextual discrimination: knowing which solution fits the constraints, which shortcut is unsafe, and which requirement is not worth implementing as written.

The practical consequence is that AI often increases the productivity gap between engineers who can validate output and those who cannot. A less experienced engineer may generate plausible code quickly, but plausibility is not the same as fitness for use.

Creativity Means Reframing and Selection

The usual defense of human relevance is “creativity,” but the term is often used too loosely. In engineering, creativity is not chaos, novelty for its own sake, or stylistic self-expression. It is the ability to reframe the problem, generate alternatives, and select an approach that creates value under constraint.

AI can produce variation. It can even produce surprising variation. What it struggles to do reliably is determine which variation matters in a particular institutional, economic, or user context. Human creativity remains strongest where the problem itself is underspecified: when the real work is not merely generating options, but deciding what should count as a good option in the first place.

Engineering Is Applied Economics

Software engineering is often taught as if technical elegance were the primary criterion. In practice, engineering is closer to applied economics. Decisions are made under scarcity: time, money, reliability budgets, operational risk, organizational attention, and maintenance capacity.

This is one reason AI output requires supervision. A model may generate a technically valid solution that is too expensive to operate, too complex to maintain, or too misaligned with the business case to justify adoption. Good engineering therefore asks questions that models do not ask by default: Is this worth building? Is this level of complexity necessary? What hidden costs will appear later?

Context Turns Code into Consequence

Code does not live in a vacuum. It runs inside legal, security, operational, and organizational constraints. AI systems are improving at local pattern completion, but they still depend heavily on what the prompt includes. They rarely know, without being told, which licenses are unacceptable, which privacy assumptions are forbidden, which infrastructure is already fragile, or which team conventions are non-negotiable.

That is why human engineers remain accountable for consequences rather than just syntax. Context is what turns an implementation into a deployable system. It is also what makes debugging, maintenance, and incident response fundamentally human activities even when generation is partially automated.

What This Means for Junior Engineers

The common claim that AI will simply replace junior engineers is too coarse. It is true that some entry-level work is being compressed or automated. It is also true that organizations still need a talent pipeline, and that good engineers are not produced by skipping the early stages of responsibility entirely.

What changes is the threshold for differentiation. Junior engineers can no longer rely on routine code production as their only signal. They need to demonstrate stronger end-to-end reasoning: the ability to test assumptions, critique generated output, trace consequences, and explain trade-offs.

I have seen this in interviews. Candidates stand out less by producing more code and more by showing that they can interrogate a solution. The strongest early-career candidates tend to notice performance bottlenecks, identify unnecessary complexity, or surface risks that were not explicitly asked about. Those behaviors signal judgment, not just tool usage.

Conclusion

Competing with AI is not a matter of outrunning it on raw output. That contest is poorly chosen. The more durable strategy is to cultivate the capacities that become more valuable as generation becomes cheaper: reframing problems, making economic trade-offs explicit, and carrying the context that turns code into accountable engineering.

In that sense, the future belongs less to engineers who can type quickly than to engineers who can decide well.