What is still hard to become good at as AI makes coding easier?

Since LLMs came, how I build software changed for so many reasons. Since then, I’ve spent most of my time deciding what I want to build, how I want to build it, how I want to give the context, checking the result, seeing if there is an error, and paving the way for the LLM to do the right thing. And the difficult part, which is writing code and getting everything right, is handled by the LLM surprisingly well, from my work and the people around me and people who I look up to on the internet.

Until now, my edge was that you can get really far by understanding how computers work, like networking and CLI, kernels, reading errors, and following documents. You can get very far. But that strong edge is now fading because of how good LLMs are becoming lately. It won’t take that long. That edge would be gone when LLMs get much better at switching to the specific context of things. They still can do it, but you need to pave the way for them.

Now, people are saying you should have better judgment of what to develop for the future. But the issue is that I haven’t identified a hard thing worth committing to since LLMs came along.

I’m in my early career and trying to decide where I should invest my serious efforts. What capabilities remain difficult to acquire, are very niche, and aren’t something lots of people want to work on? There has to be something where human expertise remains valuable, including work where a human is still much cheaper than using LLMs.

I’m not looking for an answer like, “XYZ types of sites are very hard to build, and you should build one yourself and make money.” What I’m missing is a concrete problem, who needs it solved, what makes it difficult, and a path into working on it.

I’m not looking for an easy answer. I’m trying to see what you and other people have run into or understand about what’s going on right now. I need to find a hard thing to do, and I haven’t found one I feel confident committing to. Can you help me find one?