Have the models come full circle?

It feels like the big LLMs have come full circle. At first they were too naive to implement meaningful features autonomously and took so much handholding to get anything done it was quite a pain babysitting them. Then they got good enough to do meaningful software engineering work and there was (I now recognize in hindsight) a sweet spot at some point early this year where they got meaningfully better and good enough to take non-trivial work off my hands without causing dramatically more work. But now they're far too eager and require constant babysitting again. […2 min read]