SN942 more than random numbers, nondeterministic, noncomputable

[dusanmal]

I am a little behind on the show so, if you have already discussed this, well here it is...
I am speaking from experience of Physicist currently working on narrow, specific AI
system in Medical diagnostic field as Data Scientist (objective diagnosis of autism
from functional MRI brain scans).
1) Leo wrongly assumed that just because it is complex and powerful software
built on foundations of deterministic algorithmic hardware, AI is somehow
non-deterministic. No matter the complexity and power of software and hardware
it all must reduce to actions in that deterministic algorithmic foundation.
Deterministic.
2) Term not yet discussed is maybe more important: "noncomputable".
Incapable of being computed by any deterministic algorithm in any finite amount of time.
This goes beyond just inability to get random numbers, as Matematician and Physicist
R.Penrose Nobel Prize winner showed, mathematical properties of self awareness
require it to be noncomputable. That has consequences for AI, it can't ever be
self aware (yes, there are so called Strong AI theories which just assume that with
enough power and sophistication AI must develop self awareness but, by
Math of it in my opinion that is as stating that if we just had enough power and skill
we could fly by pulling ourselves up by our bootstraps).
How are we self aware?- It turns out that there are noncomputable processes in Nature,
Physics, Chemistry, Biology. There some of them known to be ongoing in our brains.
(Side note, just because something is quantum mechanical does not make it
noncomputable, quantum computer need not be!)

 

[PHolder]

Leo wrongly assumed ... AI is somehow non-deterministic.

I'm no expert, but I have investigated the basics enough to know that computer generated randomness is frequently part of the algorithms, both during training, and during execution of the user interaction. One can argue about the ability of a computer to be truly random, but it's my understanding that Intel CPUs have the RDRAND instruction which uses "true physical randomness" to at least seed the random number generation. Given that the algorithms use randomness to intentionally produce a different answer for the same input(s), I agree with Leo.