[nucleogenesis]

> Letting the model run feels like giving up control, while inventing a workflow feels like the responsible move, even when it produces worse work in more time.

It feels to me like the discretion of the humans is the main problem in your solution space example.

Some tasks require more structure than others. Sometimes, the constraint a human faces is “you must do the work in this module or on this function”.

If the prompt says “only change this function” then the human has a good reason for that. If the LLM can’t figure it out then letting them cook on it feels like a waste of my time and tokens.

Too often I’ve sent Opus or Fable now on what I thought was a simple task to find it chasing its tail.

Fortunately I find more success than tail-chasing with the LLMs but I’m dubious of any suggestion to just let them cook.

However, a solution to the tail-chasing for certain could be to intentionally expand the solution space and to consider how my prompts might be restricting the LLM’s options.

Enjoyed the blog thanks for sharing!