put together, it encourages ongoing and at scale compute to try to get a non-specialized tool to do stuff it wasn't really designed for (since it wasn't designed for much in particular).
in other words: AI works best when it... is developed the way ML is meant to be developed, lol.
which could be achieved with much more focused, less expensive, and more accurate ML solutions.
and when does it sometimes work kinda ok?
when it's used by experts on their area of expertise...
esp if the use is focused on (relatively) structured and context-insensitive inputs (eg, code)...
and esp if developers of AI tend to be familiar with or sensitive to those use cases (eg, CODE).
and it is presented confidently to a general audience, with fig leaves for attempts to educate and convey limits (at best, usually just enough to limit liability; at worst, marketing to make AI seem like more than it is).
because of those constraints, good ML is rarely power or compute hungry at anything like the scale of "AI".
high compute is usually restricted to development, maintenance, or targeted runs limited people or orgs (the audience / consumers of the tool).
good ML:
* carefully designed for narrow contexts (or with awareness of context limits)
* designed by ML and topical experts, with expert curated data and expert reviewed outputs
* designed for specific needs and consumers, and kept up to date for both
* clear training / comms on use and limits
solve-anything AI is built on tools and methods that are inherently NOT meant to be solve-anything.
it's built at a scale and speed that makes the needed expert curation, review, and context understanding impossible.