On 09 Oct 2012, at 20:39, Stephen P. King wrote:

On 10/9/2012 12:28 PM, Bruno Marchal wrote:

On 09 Oct 2012, at 13:22, Stephen P. King wrote:

On 10/9/2012 2:16 AM, meekerdb wrote:
On 10/8/2012 3:49 PM, Stephen P. King wrote:

Hi Russell,

Question: Why has little if any thought been given in AGI to self-modeling and some capacity to track the model of self under the evolutionary transformations?

It's probably because AI's have not needed to operate in environments where they need a self-model. They are not members of a social community. Some simpler systems, like Mars Rovers, have limited self-models (where am I, what's my battery charge,...) that they need to perform their functions, but they don't have general intelligence (yet).


Could the efficiency of the computation be subject to modeling? My thinking is that if an AI could rewire itself for some task to more efficiently solve that task...

Betting on self-consistency, and variant of that idea, shorten the proofs and speed the computations, sometimes in the "wrong direction".

Hi Bruno,

Could you elaborate a bit on the betting mechanism so that it is more clear how the shorting of proofs and speed-up of computations obtains?

The (correct) machine tries to prove its consistency (Dt, ~Bf) and never succeed, so bet that she can't do that. Then she prove Dt -> ~BDt, and infer interrogatively Dt and ~BDt. Then either she adds the axiom Dt, with the D corresponding to the whole new theory. In that case she becomes inconsistent. Or, she add Dt as a new axiom, without that "Dt" included, in that case it is not so complex to prove that she will have infinitely many proofs capable to be arbitrarily shortened. I might explain more after I sump up Church thesis and the phi_i and the W_i. That theorem admits a short proof. You can find one in Torkel's book on the use and misuse of Gödel's theorem, or you can read the original proof by Gödel in the book edited by Martin Davis "the undecidable" (now a Dover book).

On almost all inputs, universal machine (creative set, by Myhill theorem, and in a sense of Post) have the alluring property to be arbitrarily speedable.

    This is a measure issue, no?


Of course the trick is in "on almost all inputs" which means all, except a finite number of exception, and this concerns more evolution than reason.


Evolution is basically computation + the halting oracle. Implemented with the physical time (which is is based itself on computation + self-reference + arithmetical truth).


So you are equating selection by fitness in a local environment with a halting oracle?

Somehow. Newton would probably not have noticed the falling apple and F=ma, if dinosaurs didn't "stop" some times before. The measure depends on 'computation in the limit' (= computation + halting oracle) because the first person experience is invariant of the UD's delays.



You received this message because you are subscribed to the Google Groups 
"Everything List" group.
To post to this group, send email to everything-list@googlegroups.com.
To unsubscribe from this group, send email to 
For more options, visit this group at 

Reply via email to