notesthoughts

What Should Developers Learn in the Age of AI?

As AI lowers the cost of implementation, where does the value of learning come from? Caught between pressure to be productive and uncertainty about which skills will last, I've been thinking less about what to learn and more about which problems deserve my time.


Can We Give Up Productivity for the Sake of Learning?

Rich Hickey sees AI not as a tool but as automation that takes over human work. I largely agree with his concern that, in exchange for productivity, we may lose the chance to think and learn for ourselves. But reality adds another question: can we actually afford to give up productivity for learning? Companies demand more output with AI, and there is no guarantee that the time saved will come back as time for learning. Using AI less in order to learn might simply mean working overtime. A company is not a school.

At the same time, I can’t be confident that existing skills will keep the same economic value. If AI automates not only implementation but also testing, review, deployment, and incident response, it is too simple to say the human role just shifts from “implementation” to “verification.” Knowledge that is useful to people and knowledge that companies will pay for are not the same thing. So the moral argument that we should preserve our craft cannot, on its own, explain the ROI of learning. This is less a matter of individual attitude than a structural problem, where labor organization and the boundaries of automation are tangled together.

Discovering Abstractions

Still, I see real potential in things like React Query and Clojure. TanStack Query bundled the recurring problems of server-state management into a single abstraction, and Clojure proposed a way of handling complexity through state, identity, and immutable data. Reproducing a known design is different from discovering what should be grouped into one concept. There is no guarantee that AI can’t do the latter too, and being rare doesn’t mean it will be rewarded. Even so, learning an API is different from understanding the problem it was built to solve and the principles behind it, and the latter can be applied across many technologies.

Starting from the Problem You Want to Build

What an AI-native developer should learn is still an open question. So I want to tie learning not to vague preparation for the future, but to problems I actually want to solve. Say I’m building a tool that treats images and video through a single sampling model, or one that applies the concept of music sequencing to visual media. I might use AI to implement it quickly, but I can still explore for myself how to abstract the differences between media, and how to model the time axis and real-time interaction. This kind of learning doesn’t guarantee a high reward, but at least it doesn’t leave its value entirely to uncertain future demand.

In the end, what I’m curious about is not only which skills will survive. What are we trying to gain through learning, and who should bear its cost? When the responsibility for reskilling is pushed onto individuals, developers end up studying after work, and that is hard to solve through personal attitude alone. On the other hand, assuming that every skill will become useless and giving up on learning is hasty. What I can say for now is that no skill is permanently safe, and that choosing for ourselves which problems to care about and which experiences to spend our time on is becoming more important. The more AI lowers the cost of implementation, the more the question of what we want to build will matter, as much as how to build it.