> tools that let users build personalized mini apps tailored to their needs.

We’ve reinvented Java Applets. I mean, I do like the idea behind that sort of stuff, it’s just that all sorts of little things break along the way. For example, I asked Claude to put together a specific recipe, it could do that, I got my Artifact/cooking widget/whatever. It even let me switch between metric and imperial (and didn’t save that preference) and let me change the quantity and updated the ingredient amounts (except the phone going to sleep led to it all resetting).

Sometimes I feel like we are very much stuck in being able to produce things but they simply aren’t high enough quality (which might take years or decades more of model training and efficiency improvements) and also that maybe we’re doing things a decade too soon. Imagine trying to build AI data centres with 2010 or 2000 hardware and how limited the models you’d be able to run would be. Maybe that’s also why the current outcomes are sometimes shitty. The other theory is that there’s simply not enough high quality data to train truly good models and we’ll plateau and model collapse in training will be common.