Disclaimer: I’m going to ignore all moral questions here

Because it represents a potentially large leap in the types of problems we can solve with computers. Previously the only comparable tool we had to solve problems were algorithms, which are fast, well-defined, and repeatable, but cannot deal with arbitrary or fuzzy inputs in a meaningful way. AI excels at dealing with fuzzy inputs (including natural language, which was a huge barrier previously), at the expense of speed and reliability. It’s basically an entire missing half to our toolkit.

Be careful not to conflate AI in general with LLMs. AI is usually implemented as Machine Learning, which is a method of fitting an output to training data. LLMs are a specific instance of this that are trained on language (hence, large language models). I suspect that if AI becomes more widely adopted, most users will be interacting with LLMs like you are now, but most of the business benefit would come from classifiers that have a more restricted input/output space. As an example, you could use ML to train an AI that can be used to detect potentially suspicious bank transactions. The more data you have to sort through, the better AI can learn from it*, so I suspect the companies that have been collecting terabytes of data will start using AI to try to analyze it. I’m curious if that will be effective.

*technically it depends a lot on the training parameters