Claude Code progressed from a young intern to an advanced junior programmer bordering on senior. Color me shocked.
@jrychter and you say that based on statistically significant data or just a single point that might have been a fluke?
@blotosmetek @jrychter there might be also some cognitive deterioration from too much use of the confabulation machines involved. (we see many examples of that, steve yegge being one of the most prominent.)

@blotosmetek I say that after two weeks of working with Claude Code on a large and complex code base, which allowed me to blaze through around 40 issues from my bugtracker. Some were small fixes, some were complex new features. My job is that of an architect at this point, guiding the programmer, which is exactly what I would be doing with a junior.

This isn't "statistically significant data", I guess, but it's not a single fluke. And I've worked with enough juniors to know this is better.

@blotosmetek The thinking is valid, code written is good. Most of it is what I would have written. It does need architectural guidance: use that way to access the data instead of this way, implement this using different abstractions, etc.

When it reviews code, it sometimes finds valid issues that I overlooked.

There really was a major change recently: I've been trying this regularly, and it was never good enough. It is now.

@blotosmetek Oh, and it also does adversarial end-to-end testing using a browser to run the app, click things and watch the results. The created test scenarios are good. Net result is that this work is what a really good tester would do. During the ~20 years that I was in a position to hire testers, I was never able to hire a tester this good, reliable and predictable. People who are this good at thinking do not want to work as testers.
@jrychter a lot of fun with Codex and/or GHCP - wrote several apps using them, somehowe with Codex on GPT-5 it started to be very usable
@jrychter and yes,
currently love using Opus 4.5