> I don't accept the argument that correct and LLMs errors is the same as correcting a junior devs errors because the junior dev would (presumably) learn and grow to become a senior. The technology doesn't exist for an LLM to do the same today and there's no viable path in that direction.
But the technology does exist. The proof is in the models you can use today, on two lines:
First, what you describe is exactly what the labs are doing. We went from "oh, look, it writes poems and if you ask for code it almost looks like python" 3 years ago. Since then, the models can handle most programming tasks, with increasing difficulty and increasing accuracy. What seemed SF 3 years ago is literally at your fingertips today. Project scaffolding, searching through codebases, bug finding, bug solving, refactorings, code review. All of these are possible today. And it all became possible because the labs used the "signals" from usage + data from subsidising models + RL + arch improvements to "teach" the models more and more. So if you zoom out, the models are "learning", even if you or I can't teach them in the sense you meant.
Secondly, when capabilities become sufficiently advanced, you can do it locally, for your own project, with your own "teachings". With things like skills, you can literally teach the models what to do on your code base. And they'll use that information in subsequent tasks. You can even use the models themselves for this! A flow that I use regularly is "session retro", where I ask the model to "condense the learnings of this session into a skill". And then those skills get invoked on the next task dealing with the same problem. So the model doesn't have to scour the entire code base to figure out where auth lives, or how we handle migrations, and so on. This is possible today!