RT @NomadsVagabonds: I haven't dove into @dwarkesh_sp 's data but I agree that BoTC is a great game for benchmarking multi-agent interactio…
This quoted post is unavailable.
Tom Taylor retweeted
Hey @dwarkesh_sp!
@NomadsVagabonds @gilg_oscar @TomTaylor256 and I already did this and had similar findings.
nitter.gq/NomadsVagabonds/status…
What happens when you take 10 different AI agents and have them play 300+ games of Blood on the Clocktower?
Agents concocted elaborated coordinated lies, chose surprising self-sacrifices and some models lean slightly towards trusting models from the same developer.
Read the full blog on LessWong: lesswrong.com/posts/4pgGkbwv…
Tom Taylor retweeted
New @NeurIPSConf paper: we're releasing ORBIT: A Framework for Multi-Agent Safety and Security Evaluations. Open source, with configurable attacks & defenses and dynamic topologies, all built on @AISecurityInst's Inspect.
When we started this work a year ago, we didn't imagine how urgent it would become by the time we published. Below are some of the interesting results and where we think this should go next. If you want to talk it through, let us know 🧵
Tom Taylor retweeted
What happens when you take 10 different AI agents and have them play 300+ games of Blood on the Clocktower?
Agents concocted elaborated coordinated lies, chose surprising self-sacrifices and some models lean slightly towards trusting models from the same developer.
Read the full blog on LessWong: lesswrong.com/posts/4pgGkbwv…