Broadly agree, but one point I think is (sadly) relevant:

> That LLMs do a better job if you know what you are asking for is old news.

Even a decade after Word Lens had demonstrated augmented reality live translation through a smartphone camera, I was amazing people by showing them the same feature in Google Translate.

Similar anecdotes about Shakuntala Devi, even in 2018 I was seeing claims about her mental arithmetic beating a supercomputer (claims that ignored that this happened in 1977 and the computer was already obsolete at the time), even though my mid-2013 MacBook Air could not only beat her by a factor of 150 million, it could also train an AI to read handwritten numbers from scratch in 0.225 seconds, and then perform inference (read numbers) at just over 6,629 digits per second*.

You say "old news", I say this discussion will be on repeat even in the early 2030s. And possibly even the 2060s.

* Uses an old version of python, you'll need to fix it up accordingly: https://benwheatley.github.io/blog/2018/03/16-10.44.18.html