Image generation is fun, and LLMs can be a great way to find a starting point for learning something already known by humanity as a whole, but not known by one in particular. Because they are statistical association machines, they are practically perfect for answering the 'what word am I looking for?' question when you can only 'talk around' the concept.

However, that's not what they are being used for, and the user cost does not match the externalized cost. If users had to pay the real cost today, the AI companies would die tomorrow. (This is probably true of a great many companies but we're talking AI ones here.)

One of the concepts I keep returning to is 'X was cool, but then the idiots got it.' Early internet? Absolute nerdity; the only people on there were highly educated, usually intelligent as well, and the new people came at a pace the community could absorb. Then the idiots came, including business majors, kids, and eventually just everyone. Early mass media? Libraries of printed books. It was still expensive, so no one bothered making and distributing 3,000,000 copies of Ted from the pub's musings on redheads, but as it became cheaper, and eventually even cheaper in electronic form, gates were no longer kept, and the idiots got in.

In this same way, AI in the form of statistical analysis tools has always been fascinating, and kind of cool. AI assisted radiology is great. Data analysis tools are great. But the idiots have the controls now, and they're using them to put shrimp Jesus on their deep fake pizza, at the top of GPT-generated 'articles,' and we're all paying the price for their fun in the form of uncountable subsidies, environmental damage, and societal damage.