[Loughla]

Your last sentence - I 100% agree with you. The general theme here is that AI will impact other people in some fashion, but they can re-train and get better jobs!

But it seems to me that software engineers should be very, very nervous about this. Creating things using novel means within well-established and documented/testable boundaries like programming languages seems like something that our current models will absolutely be able to do, if they can't already.

[joelfried]

It may seem that way but I'd argue the hardest part is coming up with the specifications for what the output needs to look like. Even if the best LLMs could write full projects like that, the business side would need to specify far beyond their normal levels what that thing needs to be. Have you ever tried to collect requirements from stakeholders and generate a project plan that meets all of those needs? Even if we're not far from GPT-4 being able to build out a basic component library, that's a long way from being able to architect solutions at a high level.

If I had to pick an analogy it wouldn't yet be farriers going out of business, but instead be more like the advent of the pneumatic nail gun for roofers. Everyone will still need a roof (software) to be built. Every instance is a little different at the design level, but the actual grunt work just got easier and faster if you know what you want out of it.