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

[spike at rainier66.com]

...> On Behalf Of Adrian Tymes via extropy-chat
Subject: Re: [ExI] Why AI agents invent their own language if you let them chat.

On Wed, Oct 7, 2026 at 2:21 PM <spike at rainier66.com> wrote:


>>... Adrian I am counting on you and BillK in particular to play the frontier of AI knowledge game I proposed previously.

>...It's inapplicable in my case.  I do not presume myself to be the world's sole leading expert on anything of much significance....

Far too modest you are sir.  To play the game, you don't need to be the world's leading expert, only ExI's leading expert.  It need not be on anything of much significance, which is why I included the example of Reimann sum mechanics for predicting Mersenne primes: no one has gone past where we were 30 yrs ago because it is an area in which our species doesn't give a rat's ass.  But I do.  Our chances of "winning" are better if we choose a field out in the sub-rat's-ass territory.



>>... You know more than anyone here on the economics of lifting small 
> payloads to orbit,

>...Not even this.  

You're the one here whose company was working it.

>>... and the use of low-cost three-axis gyros for navigation, guidance and control.

>...And definitely not this....

Again far too modest.  Reasoning: an angular accelerometer as used in the NG&C system is double integrated to get an angle of ascent.  Integration has a constant hanging out there.  Double integration has a polynomial error: At + B.  In most launchers, there is an external reference to zero out the error term.  You proposed more than a decade ago to use low cost accelerometers to achieve orbit without external reference, an idea to which I expressed skepticism, but got to thinking it miiiiight work, if you have sufficient tolerance in where you end up in orbit.  It might work, and if it does, you could save a buttload of money on the navigation and guidance equipment.

You and I never posted our discussions publicly as I recall, but good chance someone else has come up with the idea independently and perhaps even tried it.

>...There is a slight exception: stuff so niche that I'm basically the only person who's written anything on the topic....

Ja!

>...I have tried that game before...and either the AI explicitly said that I was the only source it could find (which invalidated any surprise when it then quoted my arguments) or the AI was clueless on the topic....

OK sounds like you already played the game and won.  So have I, which is part of my hesitation to accept AI as an arbiter of truth.  It is good for some things, but if we get too specific, it falls apart.  It feeds us back our own words, some of which I no longer believe.  I gave the example of Reimann analysis, which I was dead certain in 1996 was valid, but now I agree with Ernst Mayer who argued it isn't.  It kinda works in the engineering sense, but is not valid mathematically.  I do miss Ernst Mayer.

AI is now doing some things beyond what I think it is being taught (somehow (that Navier-Stokes work is impressive as all hell.))  But it is risky to use it as an arbiter of truth.  It doesn't know everything and doesn't know what it doesn't know.  

spike