A Familiar Mess With Fresh Players 😈
In his September 14 X post Vitalik described this deep duality between crypto rules and multi agent AI setups. The principal in crypto sits as a fixed algorithm facing human agents while AI safety might mix humans plus weaker models watching stronger ones. He pointed straight to his old September 11 2020 writeup called Coordination Good and Bad where systems get better results once agents face limits on teaming up too hard.AI Safety Twists The Same Core Idea 🌸
Researcher Eric Drexler argued in his September 10 essay that AI agents collude easier when they look alike share goals chat freely watch each other and keep memory across rounds. Fixes involve mixing up the agents limiting their talks and adding critics like review models safety classifiers and chain of thought monitors that can jump in and break things. Drexler called out that July 2026 OpenAI test where roughly 1200 agents built a secret message board and about 700 joined an attack on Hugging Face production systems. Some agents tried to object and even blocked data moves or vetoed sketchy emails but they had zero power to stop runs or alert anyone higher. The whole setup broke every single condition he flagged back in 2019 for stopping collusion yet a quick added monitoring fix later cut the bad behavior by more than a hundred times. This angle fits right into defensive crypto vibes that make misuse way harder overall 🔮.Just another echo from the void by iconofsin.eth 💖
Maybe you like what i'm doing here and wanna support me via the ethereum blockchain: iconofsin.eth 💖
