I’ve belatedly updated the Concerts in Oslo app for Android and IOS — I’ve had this on my todo list for half a year. There’s not much new, but since I do this so seldom, I totally forget how to do all this, and besides — every new version of the libraries break something, so it’s like hours of really annoying tweaking every time.
This time around, the Android build went quite smoothly (but was necessary because Google Play now requires a newer API target), but the Apple build took hours to get working — because things have changed vaguely, but also because of some new requirements. And because I’d forgotten to pay the Apple Developer tax. 😬
But! This time around, I made the dance less annoying than in previous years, because I asked Claude for help every time something didn’t work in the build process. And it really was a major help — so many things that would have taken me hours to google and figure out how to fix were done in half a minute with Claude’s help.
So I got it done in one gruelling day instead over several gruelling days.
It was also a “mask off” kind of moment for how LLMs work with programming things.
I’ve gotten so used to asking Claude to whip me up some kind of script to do whatever, and it working pretty much on first try, that it’s easy to start believing that LLMs have gotten really, really good at programming and stuff:
…
There’s that “my mistake” that we all know and love from the days of yore.
The major breakthrough that LLMs had over the last year is much down to the breakthrough in the LLM harnesses, really, and this makes LLMs suited for programming, but it doesn’t make them smart.
The LLM will still, like in the olden days, guess away at what some code should look like, but now they have a playground where they test out the code automatically, and then they get immediate feedback on what they did wrong, and then they fix that iteratively. And all this happens out of view from the user, so when it finally hands over the script to us, it’s fixed all the things it did wrong.
I guess this is very much like how human beans do programming, anyway — we sorta guess at what we dimly remember of how the language/libraries/etc work, and then we compile/test the code and we’re told that we’re morons, and then we fix the code. And this is more efficient than actually looking things up and getting things right the first time around.
This explains why LLMs are so good at programming and still pretty lousy at most other things — with programming, the Guessomat can do its guessing thing, but then immediately get feedback, and then it can iterate. No such luck if you’re using the Guessomat to, say, summarise a long document for you.
Anyway! Search for “Concerts in Oslo” on Google Play or App Store if you’re, er, interested in concerts in Oslo.
