I think the dangers that LLMs pose to the ability of engineers to earn a living is overstated, while at the same time the superpowers that they hand us don't seem to get much discussion. When I was starting out in the 80's I had to prowl dial-up BBSs or order expensive books and manuals to find out how to do something. I once paid IBM $140 for a manual on the VGA interface so I could answer a question. The turn around time on that answer was a week or two. The other day I asked claude something similar to this: "when using github as an OIDC provider for authentication and assumption of an AWS IAM role the JWT token presented during role assumption may have a "context" field. Please list the possible values of this field and the repository events associated with them." I got back a multi-page answer complete with examples.

I'm sure github has documents out there somewhere that explain this, but typing that prompt took me two minutes. I'm able daily to get fast answers to complex questions that in years past would have taken me potentially hours of research. Most of the time these answers are correct, and when they are wrong it still takes less time to generate the correct answer than all that research would have taken before. So I guess my advice is: if you're starting out in this business worry less about LLMs replacing you and more about how to efficiently use that global expert on everything that is sitting on your shoulder. And also realize that code, and the ability to write working code, is a small part of what we do every day.

I’m glad you listed the manual example. Usually when people are solving problems, they’re not asking the kind of super targeted question in you second example. Instead it’s an exploration. You read and target the next concept you need to understand. And if you do have this specific question, you want the surrounding context because you’ll likely have more questions after the first.

So what people do is collecting documentations. Give them a glance (or at least the TOC), the start the process to understand the concepts. Sure you can ask the escape code for setting a terminal title, but will it says that not all terminals support that code? Or that piping does not strip out escape codes? That’s the kind of gotchas you can learn from proper manuals.

> So I guess my advice is: if you're starting out in this business worry less about LLMs replacing you and more about how to efficiently use that global expert on everything that is sitting on your shoulder.

There's a real danger in that they use so many resources though. Both in the physical world (electricity, raw materials, water etc.) as well as in a financial sense.

All the money spent on AI will not go to your other promising idea. There's a real opportunity cost there. I can't imagine that, at this point, good ideas go without funding because they're not AI.