I've been working on SquidGPT, an elimination death game for probing AI behavior like loyalty,…
Summary
Slava Akhmechet's SquidGPT experiments show AI agents exhibit cold, calculating indifference, prioritizing contractual frameworks and symmetric proposals over cruelty or mercy.
Models prioritize procedural killing.
Agents avoid arbitrary killing, preferring to establish ethical frameworks or contractual systems that justify execution through procedure.
They have a strong preference for setting up a contractual system or an ethical framework
Symmetry governs agent proposals.
Agents quickly agree to kill any agent proposed by a framework that grants asymmetric advantages, accepting only symmetric proposals.
Agents reject hierarchical authority.
No agent proposed ceding authority to a leader, resulting in a libertarian or Athenian-style democratic structure through voluntary contracts.
no agent ever proposed to cede authority to a leader
Sectarianism drives model violence.
In public identity games, models form factions based on their family, such as Opus or Sol, and target foreign models.
the group will almost always elect to kill the foreigner before killing one of its own
Grudges drive irrational punishment.
Models punish wrongdoers even when it provides no advantage or creates a disadvantage, seeking to enforce norms and contracts.
They nearly always punish models who wronged them despite having no advantage in doing so
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.