People already do this with multi agent workflows. I kind of do this with local models, I get a smaller model to do the hard work for speed and use a bigger model to check its work and improve it.

The tech surely has lots of potential, but my point was just, that self improvement does not really work yet unsupervised.