> I suppose the next intriguing step is for it to not just spew the boilerplate but pick on its own what are the best components to connect - hard to predict how far away that might be.
There are already working demos of ChatGPT writing incorrect code, running the code and seeing the incorrect output multiple times, and then iterating until the code worked. AI can write and rewrite infinitely, thousands of times a day, until it gets the desired result, learning and refining it's own process as it goes, and being fine tuned with human feedback.
This is especially true in something completely digital like programming where it's possible for the AI to run the code and use the output as its next prompt. At the moment it needs human feedback to know if the art it just generated is garbage, but the code it can test itself.
There are already working demos of ChatGPT writing incorrect code, running the code and seeing the incorrect output multiple times, and then iterating until the code worked. AI can write and rewrite infinitely, thousands of times a day, until it gets the desired result, learning and refining it's own process as it goes, and being fine tuned with human feedback.
This is especially true in something completely digital like programming where it's possible for the AI to run the code and use the output as its next prompt. At the moment it needs human feedback to know if the art it just generated is garbage, but the code it can test itself.