Technical ability != building something people want, will use on a daily basis, and provides economic value or enjoyment for humans. Models have been trained to be superhuman at the former, not the latter. Taste is what allows humans to do the latter.

Therefore, until taste can be programmed as a reward function in RL, or we have demonstrated that scaling RL improves model generalisation to such an extent that we don't need to create that reward function,

> Coding is solved, unless you are doing something truly novel.

Agreed. And even then, AI are sometimes better at humans at finding solutions to problems humans can't (e.g. Navier Stokes equations). But that still doesn't mean all tasks of the job are done (though obviously the definition of the job have and will continue changing as the tech improves).