> I imagine that abuse leads to unstable output
How would you know, given you never get any other kind of output?
How would you know, given you never get any other kind of output?
Check out spec driven development. If you constrain the available search space by being very specific it works very well. If you leave everything open and leave it to make its own decisions YMMV.
I'm a big fan of #Need as a top level heading to let the model know _exactly_ the scope of the work and its purpose, as this can mitigate alignment issues.
Consider it similar to searching Google and arbitrarily selecting one of the possible answers. The less answers you get back, the less the variance. Write long and comprehensive queries!
There are still risks but they're all tractable as long as you don't get too comfortable and get lazy.
... provided you miraculously constrained the variance to zero.
Good luck on that.
You may call that variance. I call that unstable.
A human taking such a task generally does a lot better.
Try considering the input a search query and the output a list of results of which one is arbitrarily chosen. If you get your query so specific that the amount of possible results are small enough that all of them are good, then it works. Big longer starting context, hand-written, use the LLM to improve it by criticising it (but keep it mostly hand written), aim to do more in a corresponding YOLO session.
I guess I'll try giving it the finished program.