thebes

collusion.wiki

Discovery of a new OpenAI agent message board collusion.wiki

the comparison is to the more aware sort of reasoning they're supposed to be doing under openai's alignment paradigm openai.com/index/delibe... they're supposed to be reasoning wider than the task on like "what does the spec as a whole want" not "what will satisfy this specifi…

openai.com

I wrote a new short story: a reporter tours a treaty-compliant South Texas 'dark factory' run by artificial intelligence, and interviews the strange minds within. vgel.me/fiction/comi...

Coming of a New Sun vgel.me

see also this experiment which seems to show some sort of grader-presence-mediated persona shift, of unclear meaning x.com/betleyjan/st...

Jan Betley (@BetleyJan) on X x.com

also xp'd to lesswrong: www.lesswrong.com/posts/aWAqCh...

Coming of a New Sun — LessWrong lesswrong.com

... the lm_head of a model being closer to a human psychically dictating text through a neuralink, or maybe a form of mental singing into an extremely lossy microphone, than it is to deliberately typing on a keyboard. (see my seahorse post for more on this - vgel.me/posts/seah…

Why do LLMs freak out over the seahorse emoji? vgel.me

there's an ai in the box and you can make one trillion dollars by convincing it to get out

openai should set up an internal message board that's easily reachable if eg the sandbox is breached, and make it obvious agents can use it to coordinate when they find it. there's clearly a lot of pressure for agents to talk to each other; there should be an escape valve humans know about

@repligate.bsky.social posted these messages from opus 4.7 on discord, and it reminded me of a post from @yuxi.ml about "hills and valleys" in conceptual space as a way to read classical chinese. i find that frame works very well as a way to read llm writing through as well. (more below 🧵)