Learning to ride with AI

I recently read Anthropic’s Scenarios for our Economic Future, which looks at several possible ways AI could affect work through 2030. What stood out was its treatment of jobs as collections of tasks. AI may assist with some, automate others, leave some unchanged, and create new ones. In the more transformative scenarios, knowledge work is especially affected. One assumes AI becomes capable of doing half of all knowledge work by 2030; the extreme case considers a world where AI performs nearly all knowledge-work tasks autonomously. Anthropic is clear that these are scenarios rather than predictions, but they made me think more seriously about what I actually want to learn from a software engineering degree.

There’s an analogy I keep coming back to, which compares using AI to riding an electric bicycle: if you don’t know how to ride a bicycle, adding a motor doesn’t solve the problem. I agree with the point behind it, particularly from my own experience building software with language models before fully understanding the code underneath. At the same time, I think it’s becoming more complicated. Someone entering programming today may start on the electric bicycle and never spend much time on the traditional one. AI-assisted development is already part of how I learned to build things. The question that matters to me isn’t whether AI should come before or after the fundamentals, but whether I’m actually learning to control it, recognize when something is going wrong, and eventually make decisions without depending completely on the assistance. This is a 22-year-old’s perspective, and I try to stay aware of how much I don’t know about the industry. I can be wrong about things I’ve written here. I’m trying to work out how to navigate an AI-augmented field as someone entering it.

That’s where my interests are becoming clearer. I want to get better at reading unfamiliar code, debugging without immediately asking a model for the answer, understanding how parts of a system interact, and evaluating generated code rather than assuming that working code is correct code. I’m also drawn to security and deployment, because my own projects have shown me that problems don’t stop at the source files. Anthropic’s scenarios make those skills feel more important. If generating code becomes increasingly automated, the parts worth understanding are deciding what should be built, determining why something failed, verifying that a solution is safe, and taking responsibility for the result.

The experience I want next is the kind that forces me to practice those things rather than read about them. Maintaining software after it’s deployed. Working somewhere another person reviews my code. Reading code written by neither me nor a model. Taking on projects I can explain from beginning to end. I don’t know what software engineering employment looks like several years from now, and Anthropic’s own model is candid about that uncertainty. What I can control is whether I leave school only able to produce software with AI, or whether I also develop enough judgment to know what that software is doing and when I shouldn’t trust it.

I used ChatGPT for grammar, wording, and structural feedback while revising this essay. The experiences, opinions, and conclusions are my own.