things i've found without education: Training a Model to Make Choices
- you can backpropagate loss around a decision, by -> weighting different outcomes with the likelihood of choosing them, and summing them <- then the loss can propagate to the impact of the weight on the final sum you can even do it in random minibatches with small samples from the outcome space. guessing that rl ppo does something analogous this briefly worked for me a little to automatically tune prompts might need some further review or rephrasing (and/or education) to refine and reduce inhibiton around
