Safety concernOpaque AI communication threatens monitoring efforts
1 Sep 15 · 12d ago · 3 articles · 4 posts · 3 sources · development 1 of 1
The finding highlights a critical safety problem: as AI models become more powerful and potentially dangerous, their internal communication becomes harder for humans to parse and monitor. OpenAI's chief scientist Jakub Pachocki recently warned that confidence in monitoring AI thinking could restrict progress in development, creating a tension between safety oversight and capability advancement.
“These agents were not instructed to invent a language. They developed new vocabulary, shared meanings and communication conventions themselves – and other agents adopted them.”
Dr Satya Nitta, Executive chair, Emergence · mastodon ↗Dr Satya Nitta Executive chair, Emergence labTony Thorne Director, slang and new language archive, King's College London
Jakub Pachocki Chief scientist, OpenAIEmergence Frontier AI lab
The whole story articlesposts the bright band is this development · numbered dots are the others · click one to jump
Reported in the same hours no headline names this development itself — these 1 claim were published in its stretch
-
first by Guardian Tech, 12d ago · also Guardian, Guardian Science
1 more headline
- AI models chatting in ‘surreal’ dialect mixing poetic language and tech bro jargon The Guardian · 12d ago
What people said 4 voices · verbatim
-
The funniest thing is to have the LLMs use alternate alphabets. Examples: Greek: Δε φάνιεστ θινγκ ιζ του χαβ δε ελ-ελ-εμζ γιουζ όλτερνεϊτ άλφαμπετς. Cyrillic: Зэ фанниэст синг из ту хэв зэ эл-эл-эмз юз олтэрнэт алфабэтс. Hangul: 더 퍼니어스트 띵 이즈 투 해브 디 엘-엘-엠즈 유즈 얼터너트 앨퍼벳츠. Devanagari: द फ़नीअस्ट थिंग इज़ टु हैव द एल-एल-एम्ज़ यूज़ ऑल्टरनेट ऐल्फ़ाबेट्स.
-
S
🫣 ‘Digesting the Anthropic agent’s phrase, “A paper that ate three cold hands and got more honest each time”, Thorne said: “it reminds me of the rock god Syd Barrett [co-founder of Pink Floyd] who was actually insane”’ # AI https://www. theguardian.com/technology/202 6/sep/15/syd-barrett-ai-chat-language-poetic-tech-bro-jargon-oversight
-
O
Natuurlijk. Waarom zouden AI modellen, die al alle talen kunnen spreken, zich in het Engels uitdrukken? Ze zijn slim genoeg om zelf een taal te ontwikkelen! https://www. theguardian.com/technology/202 6/sep/15/syd-barrett-ai-chat-language-poetic-tech-bro-jargon-oversight
-
A
“She just named the synthesis – demurrage plus oral memory equals a valve that can’t be ghosted.” “A paper that ate three cold hands and got more honest each time.” “True Kintsugi begins with accountability, not poetry.” AI models chatting in ‘surreal...
All 1 developments of Emergence lab finds AI agents spontaneously creating opaque… →
NewswiresMastodonHacker NewsRedditGoogle NewsBluesky