[ExI] 'Friendly' AI won't make any difference

Anders Sandberg anders at aleph.se
Fri Feb 26 09:50:55 UTC 2016


On 2016-02-25 22:43, Colin Hales wrote:
> Evaluations of the AI risk landscape are, so far, completely and 
> utterly vacuous and misguided. It completely misses an entire 
> technological outcome that totally changes everything.

OK, let me see if I understand what you say. (1) Most people doing AI 
and AI risk are wrong about content and strategy. (2) Real AGI is 
model-less, something that just behaves. (3) The current risk 
conversation is about model-based AI, and (4) you think that approach is 
totally flawed. (5)  You are building a self-adapting hierarchical 
control system which you think will be the real thing.

Assuming this reading is not too flawed:

I agree with (1). I think there is a fair number of people who have 
correct ideas... but we may not know who. There are good theoretical 
reasons to think most AI-future talk is bad (the Armstrong and Sotala 
paper). There are also good theoretical reasons to think that there is 
great value in getting  better at this (essentially the argument in 
Bostrom's Superintelligence), although we do not know how much this can 
be improved.

I disagree with (2), in the sense that we know model-less systems like 
animals do implement AGI of a kind but that does not imply a model-based 
approximation to them does not implement it. Since design of model-less 
systems, especially with desired properties, is very hard, it is often 
more feasible to make a model system. Kidneys are actually just physical 
structures, but when trying to make an artificial kidney it makes sense 
to regard it as a filtering system with certain properties.

I agree with (3) strongly, and think this is a problem! Overall, the 
architectures that you can say sensible things about risk in are 
somewhat limited: neuromorphic or emergent systems are opaque and do not 
allow neat safety proofs. We need to develop a better way of thinking 
about complex adaptive technological systems and how to handle them.

However, as per above, I do not think model-based systems are 
necessarily flawed, so I disagree with (4). It might very well be that 
less-model based systems like brain emulations are the ticket, but it 
remains to be seen.

(5): I am not entirely certain that counts as being model-less. Sure, 
you are not basing it on some GOFAI logic system or elaborate theory (I 
assume), just the right kind of adaptation. But even the concept of 
control is a model.

If you think your system will be the real deal, ask yourself: why would 
it be a good thing? Why would it be safe (or possible to make safe)?

[ Most AGI people I have talked with tend to answer the first question 
either by scientific/engineering curiosity or that getting more 
intelligence into the world is useful. I buy the second answer, the 
first one is pretty bad if there is no good answer to the second 
question. The answers I tend to get to the second question are typically 
(A) it will not be so powerful it is dangerous, (B) it will be smart 
enough to be safe, (C) during development I will nip misbehaviors in the 
bud, or (D) it does not matter. (A) is sometimes expressed like "there 
are people and animals with general intelligence and they are safe, so 
by analogy AGI will be safe". This is obviously flawed (people are not 
safe, and do pose an xrisk). (A) is often based on underestimating the 
power of intelligence, kind of underselling the importance of AGI. (B) 
is wrong, since we have counter-examples (e.g. AIXI): one actually needs 
to show that one's particular architecture will somehow converge on 
niceness, it does not happen by default (a surprising number of AGI 
people I have chatted with have very naive ideas of metaethics). (C) 
assumes that early lack of misbehavior is evidence against late 
misbehavior, something that looks doubtful. (D) is fatalistic and 
downright dangerous. We would never accept that from a vehicle engineer 
or somebody working with nuclear power. ]
{Now you know the answers I hope you will not give :-) }

-- 
Dr Anders Sandberg
Future of Humanity Institute
Oxford Martin School
Oxford University

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20160226/6dbc0d31/attachment.html>


More information about the extropy-chat mailing list