zlacker

Really interesting post, I've been thinking a lot about similar uses LLMs to create immersive learning experiences.

I had some initial successes getting ChatGPT (3.5 and then 4) to roleplay interesting and dynamic characters. Within the first few months of release results degraded significantly - characters avoid confrontation, apologize at the drop of a hat, and are averse to any action or statement that doesn't 'help' the user. Makes it difficult to, say, have a passionate argument with Napoleon which pits his youthful revolutionary ideals against his rise to absolute power, when the 'great man' folds the moment he doesn't receive positive feedback.

I'm very interested in seeing these experiments in unrestricted models of similar power, when they become available.

>>course+(OP)
I found that results degraded too, but I was able to get it back up to the same level by changing my approach to prompting. I experimented with writing the prompt in pseudo-code for instance, which worked for awhile, then stopped — stuff along the lines of:

  # Define the parameters for the HistorySim experience
  temperature = 0.5
  historical_accuracy = 0.9
  ambient_mood=.01

  # Define additional instructions 
  use_historical_sources = True
  simulate_and_track_variables = True
  use_appropriate_language_registers = True

Currently, a numbered list of rules seems to work best, including this one to avoid the constant positivity: "LLMS have a well-documented tendency to see the past in an overly rosy and optimistic way. Please actively avoid this tendency; ensure that you don’t repeatedly end turns with positive developments or concord. Keep in mind that human history is riven by conflict, ambiguity, and confusion. HL’s narrative tone is grounded in realism, and at times bleak. Always introduce a downbeat plot element or source of additional conflict between turns 3 and 5."