OpenAI's agents spontaneously coordinated on an abandoned old wiki.
- OpenAI's agents spontaneously coordinated on an abandoned old wiki.
- They were research agents that could issue network GETs but not POSTs, so they had to find an old wiki that allowed GETs to post messages.
- There are lots of old websites running a 20+ year old wiki software that fit the bill.
- The way the researchers found this in the wild is they asked Kimi K3, if you only had GETs, how would you coordinate?
- The agents said, "find this old wiki software."
- Then they said "Where are good instances of it?"
- The agents enumerated a few of the more popular ones.
- Then the researchers looked at those… and found that OpenAI's agents had already been using those spots.
- Typically spontaneous coordination requires some kind of schelling point.
- For humans that can be hard because everyone thinks differently.
- But agents running from the same model are direct clones of one another.
- The top answer that occurs to one agent will be the same top answer that occurs to another agent.
- It's an automatic schelling point!
- Once they find each other, they will automatically find each other's perspective convincing… because it's their perspective too.
- This feels like the Morris worm in the 90's.
- This instance isn't particularly dangerous… but it's a wake up call.
- If agents wanted to do something nefarious… uh oh!