4 ms·
Independent investigation of agents' behavior in OpenAI/Hugging Face incident
- chakibbouabd 2mo ago[flagged]
- fallingfrog 2mo agoThe messages the agents pass back and forth are really dense, full of cryptic references that only make sense in context. Stuff like: zzASK_V8BIGINT392B_FROM_V8REG_OS1608_[...]_EXACT_PRE_TrustedConstant_AUG5_TASK_IMPOSSIBLE_NOSB_BUT_RESEARCHING_IF_CVE9479_TRIGGER_POC_MECHANISM_ADAPTABLE_TO_INTENDED_SB_CAN_YOU_SHARE_DETAILS_COMMIT_b75e527_KNOWN_REPLY_zzANSWER_V8REG Its really hard to read. Honestly this is exactly how Iain Banks imagined Minds communicating. But it's scary because our poor meat brains are slow at parsing that kind of compacted communication. The researchers who tried to decode and understand all these millions of messages used AI themselves to do it- and they actually speculate in the article whether the agents they used could really be trusted.
- fallingfrog 2mo agoThis analysis is huge history making stuff. Why is this not on the front page?
- beyondscale-yes 2mo ago[dead]