[dangond]

The real issue with prompting a la "Facebook for dogs" is how ridiculously underspecified such a prompt is. You see the problem with this very clearly in text to image models: you might have a specific idea of your dog in your head, but no matter how often you prompt "cute white Maltese dog with a tuft of fur across its eyes", you will never get something that would pass as your dog. (A more obvious challenge would be to try to generate an image of yourself if you are not a celebrity). The amount of words needed to replicate every detail exactly would amount to at least a short essay. You give an AI the prompt "Facebook for dogs" and it could take that in SO many different directions.

Another problem that I'm personally more concerned about is that this way of coding will allow for so many security flaws to be built into the code. Even if it builds something that matches your expectations for the prompt, what's stopping it from generating something vulnerable to attacks that leak all your users' data? Humans already struggle enough with this, and we're feeding all our flawed codebases into these models.