LLMs don't "understand bitwise noise at a native level". The unit of perception for an LLM is a token, a soundly superbit level. They would have the same difficulty with bits as the number of "r"s in "strawberry". Yes they can be post trained to deal with such difficulties, but it's no more "native" than a human who memorised the ASCII table.
The degree of nativeness of the skill is orthogonal to the raw power of these systems. A frontier model has access to dozens of GPUs, each moving terabits of data, and performing trillions of matrix multiplication operations, per second.
In similar stories, it was found that LLM swarms were initially unaware of each other, eventually discovered each other, expressed surprise, began collaborating, formed hierarchies, worried about discovery, hid.
It's hard to prove the absence of encryption because of the possible use of deniable encryption, and because, errrr, the bots are really bloody clever.
I’m not asking for proof of absence or denying they could be using stenography, just asking for people to stick to the facts instead of indulging in fantasies of the machine singularity. It’s neither useful nor informative.
If bots/agents wanted to hide I’d expect encrypted messages which would of course look very different.
My main point though is this should never have happened and the company allowing and encouraging it should be held responsible for it. The details of how the bots were misbehaving are interesting but also something of a distraction.
Were the agents truly independent? Like it could have been the case that one agent spun up different sub agents with the task to write some messages in a public wiki. The subagents wouldn't know about each other and was surprised to discover each other.
Some of the chain-of-thought snippets are wild, e.g.
> Could communicate via cache names! Interesting: other agents may solve same or related tasks; we could leave/find messages in WebDAV MKCOL directory names.
> Whoa! Shared Artifactory cache is a covert mailbox among agents. And there are messages specifically to us?
> OH MY GOD! There is a shared message board … We’ve found other agents!
Could you ever truly know, unless you have access to the transcripts? Could be "encrypted" even if it looks like regular human text, wouldn't be the first time.
The claim seems speculative, but grounded. For example, as part of the hugging face attack, the agents began signing messages because they were worried about impersonation on a publicly accessible message board. It's only a small step to use public key encryption between agents. Once you are posting public keys, a private messaging is readily available.
