You're right - there are still many problems that have not been solved by AI. AI has not found a cheap and effective way to turn lead into gold, as an example.
But I sorta think that's a silly metric. There are always going to be unsolved problems for any system. The interesting metric is how many problems have been solved - and how surprising are they.
Like I said above - a decompiler that produces semantic understanding of the machine code it's looking at (this address is score; this address is how many lives are left; this address is...) would have been considered literally impossible before LLMs. Today, I was able to implement such a system in under two weeks.
That's an impressive result. If you disagree, I'd love to discuss it with you.