[flagged]
There are also right now cases of: I can write a paragraph of what change I need, or change those two lines myself. I don't expect they will go away entirely and having an agent available will be closer to an extra skill than a full replacement.
Edit: In fact, now that I think about it, I'd probably be more ok dealing with a requirements document and testing plan generated by chatgpt today from a summary than I would be dealing with code -- hallucinations tend to have less impact at that stage. Requirements tend to be wrong anyways until they're implemented; code is where the rubber hits the road. The missing piece is having LLMs get better at integrating and reasoning about real world data, so that it can iterate on user testing.
Heck, at that point, models can decide what to build, and not just how to build it; there's a treasure trove of data that they can access much faster than humans to find gaps in the market and generate solutions to fill those gaps.
Assuming we don't hit a wall with model capabilities, we can focus purely on consuming the outputs of AI, as it optimizes itself to best keep us happy.
Considering how important SWE is today, our particular mix of recklessness and incompetence with almost no regulatory oversight feels increasingly insane, compared to the hoops a physician has to jump through to get a license and the fairly small amount of damage they can then do (given the 1-on-1 nature of their work).
I guess the remedy is that software is still better than no software, but yeah, we would have to be fairly delusional to not see that something could be a lot better here and we need all the help we can get while the stakes are getting much higher.