Comment by vintermann
3 hours ago
What are the hard things AI can't do yet, then? Are you sure we shouldn't rather spend time getting better at, say, playing the piano?
Assuming we're allowed to prioritize that in a world where the rich have all the money.
(Looking at the local flow of money, I can't avoid the conclusion that my wages come indirectly mostly from the building of an extreme luxury product for an extremely rich person.)
Honest answer, building and maintaining a large complex project over many years. We have plenty of examples of one-shots like the 1000x slower gcc, but those don't see updates after they achieved their advertisement goal.
Maybe a little more philosophical, novelty. You can ask an LLM to write in the style of <famous author> but afaik the results are bad when you ask it to write in a novel style no-one has seen before.
Fair enough, but I don't think that's something fundamental. Agentic AI hasn't been around for many years, and many of its most enthusiastic adopters have, shall we say, attention issues, so I don't think it's odd that we don't have long term maintained wholly AI-coded projects yet.
Lack of novelty in creative tasks is largely a product of how we deliberately train them to not surprise us. AIs from a few years ago, of the sort Janelle Shane wrote about, were puny compared to today's. But they weren't instruction tuned, so they had plenty of chaotic creative energy.