Multi-agent mechanistic interpretability and inferentialism

• Build philosophical framework to conceptualize what counts as a "conceptual perspective" in multi-agent phenomena and what types we could distinguish
• In tandem, build large scale LLM agent population/society where agents can take actions do things; similar style to concordia, sotopia, terralingua (won't be very hard lot of frameworks already exist). First we would use traditional multi-agent cooperation games as the environment. This allows for a more controlled setup. After that, ideally we would make it a big free-form action sandbox so agents can also take real world actions safely (like buying and selling stocks, setting up a company, etc.)
• Turn philosophical framework into multi-agent mechanistic interpretability method.
• Evaluate societal agent interactions using the multi-agent mechanistic interpretability techniques.

Involved:

• definitive: Ivar Frisch (me)
• likely: Mario Giulianelli (Professor UCL), Edward Hughes (ex-deepmind, cooperative AI, Inherent Labs), Arabella Sinclair (Professor UCL), Aron Vallinder (Independent) (im working on a similar paper with them now, this would be an extension of that).
• People I know who might be interested and worked with successfully in the past: Andrea Baronchelli (City St. George university), Annie Stephenson (ACS Research), Benjamin Bratton (philosopher of technology at UCSD)

Concrete output:

• one or multiple research papers detailing our findings; describing the multi-agent mechanistic interpretability methods, etc.
• Codebase to explore re-run our experiments and findings

(if interest for this) web playground where people can see interacting agents and apply real-time mechanistic interpretability interventions (like steering vectors) to see conceptual perspectives and external behavior change.

Timeline:

• Phase 1 (months 1–8): philosophical framework of conceptual perspectives + cooperation-game environments with 100-agent society (100-round multi-turn horizons).
• Phase 2 (months 9–16): multi-agent interpretability methods (cross-agent probes, SAEs, steering interventions) and the TrueBase vs. post-trained comparison; first paper submitted.
• Phase 3 (months 17–24): scale toward 1000 agents, free-form sandbox with simulated economic actions, and public web playground.