RE: About "safe" AGI architecture

From: Ben Goertzel (
Date: Sun Jun 13 2004 - 11:20:41 MDT

> As I pointed out in my other note, this is not true. The Java syntax
> does not give human programmers control over pointers or memory
> management. But the binaries produced still have pointers and memory
> management. It is just that they were programmed into the
> Java runtime
> system programmers rather than the Java source code
> programmers. This
> just moved the problem, it technically didn't solve anything.

It means that, in verifying a Java program, you don't have to deal with
pointer errors.

However, you have a point in that the Java interpreter/compiler and the
operating system, themselves, still need to be checked for correctness.

An AGI, in principle, could find a way to exploit flaws in the OS to
allow it to break the controls that its software framework places upon
it. In fact, if the AGI were implemented in Windows (which Novamente is
not), this would be a pretty easy problem for an AGI....

For even greater safety, we'd want to have a formally verifiable
*operating system*, where the OS code could be formally proved to
fulfill its mathematical specifications. There has been work done on
this as well, but I'm not as familiar with it. The most interesting
work in this direction, that I've read, pertained to funky alternative
chip designs in which the basic chip operations were based on graph
rewriting rather than numerical arithmetic. Such chips would be
beautiful for Novamente, which relies heavily on
functional-programming-type ideas.

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:47 MDT