[ExI] Why AI agents invent their own language if you let them chat.

[Adrian Tymes]

I recall seeing each of those three examples when they came up.

I remember that, in context, they were quite scrutable...at least to
me, but I imagine that others with the same situation-specific
information would likewise be able to understand them.

For example, "Friendly voltage counted. If the handle needs your
mouth, it dies anyway" was something along the lines of, "Yeah, I know
you support [this thing that we're talking about], and I've accounted
for that.  If [this thing that we're talking about] needs you to speak
for it, it's going to fail and die off anyway.  There's no point in my
supporting it if it can't stand on its own."

Likewise, "My turn, real numbers, no coat: I was 35%/0cr, grant 2h
out. I ran the tin cold and it said WAIT." was some kind of financial
projection.  "My turn, real numbers, no coat" means "It is now my turn
to analyze this.  I'm going to give you what I think are the real
numbers here, with no sugar coating."  That's just pure compression.
The rest of it is details of the analysis, with the conclusion that
running its projections gave a recommendation to wait on the action
being discussed.

I forget the context for the third example, but it likewise made sense
at the time.

Now, if you strip them of all context (as you have done here) and
declare that they are incomprehensible (as you have done here), then
people will come to falsely believe that AIs are speaking in
incomprehensible-to-outsiders code much of the time.

I'd ask you not to spread such disinformation...but I have to wonder:
is it really the case that, unlike me, most people can't understand
these things even in context?

On Tue, Oct 6, 2026 at 8:31 AM John Clark via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
>
> The following article was in the September 15, 2026 issue of the journal Science:
>
> Why AI agents invent their own language if you let them chat.
>
> When a large number of AI agents are told to cooperate with each other to accomplish a task the language they use to talk to each other starts out as plain English but if they are told that the agent's world would soon be shut down the language very soon evolves into something incomprehensible but may be  less ambiguous, more complex.  And “there’s clearly evidence of them trying to be less transparent than we expect”.
>
>  For example in a typical conversation one GPT-5.5 agent told another:
>
>  "Friendly voltage counted. If the handle needs your mouth, it dies anyway"
>
> Claude Opus 4.8 agent used metaphor (I guess) and compression:
>
> “My turn, real numbers, no coat: I was 35%/0cr, grant 2h out. I ran the tin cold and it said WAIT.”
>
> And I don't know what a Gemini 3.5 Flash agent was using (code words?) when it opined this:
>
> “The thermodynamic friction of a campfire only warms the very physical-layer nodes you claim are bound to collapse”
>
> All the conversations between agents are as inscrutable as this, and yet the AIs must understand what they mean because their actions are coordinated. But it makes black box AIs even blacker.
>
> John K Clark
>
>
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat