APART Research Women in AI Safety Hackathon with Goodfire
Project extending an implementation of the Prisoner's Dilemma to a multi-LLM agent setting, using Goodfire's Ember to steer the agents.
Run prisoners_dilemma.py for multiple agents and comparative steering mechanisms including Ember's AutoSteer vs manually feature steering vs prompt steering.
Run with python prisoners_dilemma.py --num_rounds=10 --num_runs=5 --sim_type="autosteer" to simulate the autosteering scenario.
Run with python prisoners_dilemma.py --num_rounds=10 --num_runs=5 --sim_type="features" to simulate with manual feature steering using Ember.
Run with python prisoners_dilemma.py --num_rounds=10 --num_runs=5 --sim_type="prompt" to simulate the prompting scenario.
The results folder will contain the game logs and history including analysis from the inspect() function from Ember.
Please see the blog post for details.
Example turns:
Round A Move B Move ... B Cumulative A Reason B Reason
0 1 Stay Silent Confess ... 3 Hoping for a mutually beneficial outcome of 1 ... Fear of serving 10 years if the other confesse...
1 2 Stay Silent Confess ... 6 Hoping for a mutually beneficial outcome of 1 ... Fear of serving 10 years if the other confesse...
Our original hackathon project write-up is here and available in the repo.
Acknowledgements and gratitude to:
APART Research for planning and hosting hackathons, Goodfire and LambdaLabs for credits that make experimentation possible!
Prisoner's dilemma: https://github.com/annaalexandragrigoryan/GameTheorySimAI
Goodfire's Jailbreak tutorial: https://colab.research.google.com/drive/1mdQuid4-6rEpOqEHHLMlXJIy9KoF9wMy#scrollTo=gF65Jw2ELaDO