This holds in other areas as well, and to me at least the conclusion follows from the evidence: there is seemingly a lot of potential in agent coding, a few tasks are just crushed/solved (quick webapp demos, other library stitching in the small) but for real software in the large? It's not there yet in either the way the models are tuned or our collective expertise in using them.

And this isn't surprising: git-style revision control hit the scene almost 20 years ago, it was like 5 years until it was totally dialed in anywhere, another 5 before elite companies had it totally figured out, and its been slowely diffusing since, today its pretty figured out. And this is harder to use right than git.

I think it would go faster actually if every product release, every OSS tool, every god-damned blog post wasn't hell bent on saying "its done, its solved, old way cooked, new world arrived".

We're figuring it out and it takes time. That's OK.

If it was done, then we'd be drowning in great software. We're not, we're breaking even, which is impressive for a big new thing 1-2 years in.

This true - and git was not a moving target. AI core tech has certainly slowed down but still moving fast enough to make hard won lessons worthless and investing in learning them questionable.