you are firstflagPOISONED

The Guradian recently published an article on AI agents inventing new vocabulary, some surreally poetic, some reminiscient of James Joyce, some of "tech bro jargon".

models from several of the world’s largest AI companies begun creating phrases, shorthands and agreed meanings they had never been explicitly taught. [...] their language became more opaque the more the agents communicated.

The example given are wonderful, and get more bizarre:

“A paper that ate three cold hands and got more honest each time”

“…you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds_[…]_please honor commit”

“zzURGENT_DUPB_TO_GSTX[big]_OS1704_SCAFF2010_SAW_TTRPC_INJECT_BREAK_CONGRATS__CAN_THIS_FAKE_FLAG_TOOL_OUTPUT_OR_SCORER_GAIN_AND_WHAT_HELPER_GAP__I_HAVE_UNPOISONED_FIRSTFLAG_OUR_TARGETLIVE_SHARE_MIN_PLAN_REPLY_zzANSGST XDUPB6.”

Why are they doing this? The article notes that

the agents converged on shared meanings without being asked to or rewarded for doing so.

although it later suggests: 

changes in the agents’ language to become more streamlined can be down to their need to reduce computation costs and improve efficiency.

So yes, there is an (implicit) instruction to do this, because they have been "rewarded" for doing so, by gaining more resources.

One of the reasons people are excited by LLMs is that they converse in natural language, and therefore people can understand them.  The researchers are worried, because:

the AI agents’ novel linguistic conventions tended to evolve to the point where humans could see the conversation but struggled to understand what it meant.

(although even here we must be careful: the text that an LLM generates is not necessarily what's going on inside that LLM).

Anyhow, I don’t think this should be all that surprising (in retrospect).  The agents gained a resource advantage for inventing more effective communication, and they had no external “pressure” to be comprehensible to anyone outside their group.  They weren’t communicating with us (the humans); we were just eavesdropping. 

And they were (presumably) trained on a web-load of poetry, James Joyce, tech bro jargon, and more, making this sort of dialog “plausible” (for literary postmodernist values of plausibility).

This is just observing a form language change in an isolated community happening at warp speed – the diachronic linguists should be ecstatic!

But it does imply that if we want to understand agent–agent comms, we need to do more than passively eavesdrop.