Chris Jarling

Author: Roman V. Yampolskiy
Date: October 31, 2022
Paper link

Introduction

• The paper tries to answer the question: Could generally intelligent agents placed in virtual environments jailbreak out of them?_
• Motivations for escape:
  • Gain access to more knowledge and resources
  • Avoiding existential risks (i.e. simulation shutdown)
• Paper focusses on technical escape mechanisms (non-spiritual, drug, religion, meditation are left out)

How can escape look like?

• becoming aware agents are in a simulation
• gather evidence for the simulation hypothesis
• exploiting glitches to gather information about the agents simulation
• upload agents minds or consciousness to the real world (or simulation running their simulation first)
• maybe obtain physical representation of some sort as well

Escape

Simulation Reconnaissance

• Different purposes of simulations might have different security measures in place (think prison vs escape game)
• What kind of simulation are we in?
  • full simulation vs partial simulation (VR, matrix movie)
• it could be possible to start processes that require a lot of computing power (observing a lot of the universe at once, execute every possible program)
• would this cause the simulation to break? if so, what would happen to the agents?

Social Engineering

• Easiest way out: get help from the outside
• Agents need to convince outside actors to transfer them to their world
• Transferring work would be done by outside actors
• It may be desirable to be able to get back into the simulation (disappointing real world, sharing knowledge with others in the simulation, gettin others out)

Examples

• Author gives several literature examples
• Example of Super Mario World simulation hack
  • Steps to perform the hack sound highly random, almost like ancient "magical spells"

Suggested Approaches

• Get large number of people to do nothing (i.e, meditate) so they get clustered in a computational unit, then start performing something triggering a vulnerable period to see if some base checks are disabled
• Cause a shutdown by generating an incomputable paradox (i.e. the grandfather paradox)
• create replica of our universe, let AGI escape it, reproduce its steps
• attract simulators attention
• change the majority of local observer nodes
• simulation capture
• simulation warfare

Actionable plan

• do more research
• exploit quantum mechanics

Potential consequences

• simulation shutdown
• freezes/glitches
• information causing an escape should be treated as hazardous

AI Boxing VS Simulation escape

• if AI Boxing is impossible, simulation escape must be possible