Hacker News Reader

Comment by vagabund | Hacker News Reader

Wonder how far you could get by closing the loop and sending terminal outputs to GPT automatically.

Prompt with a natural language description of a target program, instruct it to decompose the problem into increasingly small actionable steps, give it preprocessed terminal output prompts in response and let it iterate until it's satisfied

View on HN

It was a first contact, but I believe the whole thing could be automated to the point the human is telling requirements, and doing some QA to verify that things are going as planned, specially for projects which are not too much different from others and just have some new delta on it, often being a combination of unrelated things.

It sort of scares me to think it is just a matter of engineering to get there.

Why does it scare you?

As written, I see it feasible with the current technology, just better engineering and UI/UX to have a loop where humans give requirements and specs and just do QA. Too many things would change the landscape in unpredictable ways.

What makes you think that the future is predictable? What makes you value predictability over unpredictability? How is this time different than any other paradigm shift in technology?

I'm asking out of pure curiosity to understand your concern, to be clear. Simply curious.

i actually asked it to write that program. everything was generally correct and surprisingly simple

except the "is the program meeting the original prompt criteria" check was a comment. haha

Reply on news.ycombinator.com