It seems to me generative AI is doing great at the thing it does, which is provide responses that sound like a human wrote them! Look at how naturally written those responses are!
Now if we could just stop treating it like it's learning and interpreting factual information.
You get the impression that training this LLM did not include defining the phrases “yes, you can” or “no, you cannot”.
Or it was trained to draw them at random.
$10 says there was no "training," it's literally just a modified ELIZA with a larger database of prompts and responses. It's an FAQ with the illusion of interactivity.
Tired: writing unit tests to ensure your software does exactly what you intend it to do.
Wired: letting your software just make things up, exposing your business to legal problems and financial ruin.
it's a little fascinating sometimes why it gets it wrong, like it's because it is matching up with the volume of questions of people asking if they have to respect someone's pronouns, and failing because that impacts higher than the context provided in the question. chatbots aren't ready for this
