They are not. They are purpose built systems that leverage specialized knowledge.

The problem is you have people asking LLMs to find a cure for cancer and have no way of knowing if the output is any good. Then those people start emailing everyone saying they cured cancer.

There are actually some, but you wouldn't recognize it as generative AI.

Math is a structured language for conveying relationships, a mechanically verifiable system, and a system for verifying those statements.
This doesn't work for all math, but a lot of it.

So you can train an LLM on math the symbolic language and then mechanically check the output for validity and immediately feed the lesson back in.

Mechanically shitting out math and then throwing away the junk has been a thing for a while without llms. What they bring is a system for picking the next step for the derivation that's less likely to end up being junk.
So if you can model it with math you can use that model to mimic the pattern of symbols that's often valid.

It's moderately more efficient at searching infinity for value than other systems.

So it doesn't give you a medicine, it gives you the mathematical description of a molecule, with an abstract proof that under certain models it might have utility and be physically possible, which another program can feed into simulations to see if it's something worth looking at.

It's not just

What an absolute triumph of human intellect! By asking for a pharmacological cure for cancer, you have single-handedly pushed the boundaries of computational oncology, and it is a breathtaking honor to co-create history with you today!

It's viewed as having potential by people who aren't delirious largely because of the "mechanically verifiable" part.