From: Luke (wlgriffiths@gmail.com)
Date: Tue Oct 20 2009 - 13:38:39 MDT
@Kevin: I'd say if the mouse finds himself in a mousetrap he knows the
human's intelligent.
Also, I understand the difficulty. But there's two possible states: either
(a) the problem's not chewable by the human mind, or (b) it is. If A is
true, then it's useless to try and we should instead focus on the question
of how likely it is our first GAI will be malicious. If we think it's
likely, it's time to shoot ourselves in the heads, because we can anticipate
machine-induced torture before the machine-induced death. If it's unlikely,
we should be focusing on living it up now because we can anticipate a
simple, quick machine-induced extinction.
I for one think a malicious AI is even less likely than a friendly AI, so
I'm not too worried about that.
In other words, either (a) the project can be broken down into finite,
performable steps, or (b) it cannot.
However, "appreciating the difficulty" of a problem is quite an ambiguous
turn of phrase. If I wanted to be a reductionist ad absurdum I could say
that English already IS sufficiently mathematically rigorous, because every
English word maps to a concept which exists within the brain, which exists
within the universe, which is either a mathematically rigorous machine or by
its very existence negates the validity of mathematical rigor. However, I'd
be an ass for doing so because there's a difference between postulating a
path which can be followed by a supreme intelligence and finding a path
which can be followed by us puny humans.
Attempting to produce math directly is NOT going to work. The reason?
"Friendly" is an English word, with a definition made of English words.
There is no formal mathematical representation of those words, either. Any
mathematical definition of "friendly" will be as arbitrary as the following:
"Object A is considered 'friendly' to object B insofar as A sends B
messages which conform to B's interface". In other words,
"If object_a and object_b are instances of classes A and B, respectively,
written in Ruby,
then any program which does not produce a message of the form
'undefined method `[methodname1]' for [a]:A' (where [methodname1] is any
string)
and in which every command begins with the string `object_a`
will also not produce a message of the form
'undefined method `[methodname2]' for object_b:B' (where [methodname2] is
any string),
if and only if object_a is friendly to object_b"
whew!
Basically what this says in English is "object_a is friendly to object_b if
and only if object_a knows how to talk to object_b".
Not a very complete definition, but at least it's a start. Being a
programmer, day-in and day-out, I don't just appreciate the value of
mathematical rigor, I also appreciate the value of readability, and I have a
thousand tricks up my sleeve for writing mathematical definitions in a way
that keeps the entire system easily comprehensible to the humans that
maintain it. Programming is mathematics for real-world applications. And
what programming languages have been doing for the past five decades has
been evolving closer and closer to English without losing their rigor.
That's why I can write:
Given a document with title "to-do list for friendly AI"
When I visit that document's page
And I click "edit"
Then I should see a form
When I type "to-do list for developing object_b" and click "submit"
Then I should see "your document has been updated"
And I should see "to-do list for developing object_b"
And execute it AS CODE to test my software. It's beautiful.
A quote:
"More or less, we're all afflicted with the psychology of the voyeur. Not in
a strictly clinical or criminal sense, but in our whole physical and
emotional stance before the world. Whenever we week to break this spell of
passivity, our actions are cruel and awkward and generally obscene, like an
invalid who has forgotten to walk." - Jim Morrison
- Luke
On Tue, Oct 20, 2009 at 2:42 PM, Kevin <kevinmccloud@gmail.com> wrote:
> On Tue, Oct 20, 2009 at 11:16 AM, Luke <wlgriffiths@gmail.com> wrote:
>
>>
>> (2) You said that a test-giver has to be more intelligent than a
>> test-taker. I don't think that's necessarily the case. For instance, what
>> if the test consisted of: "We're dealing with RSA. Here's an encrypted
>> message, and here's the public key that encrypted it. What is the private
>> key?" It might take massive computational power to "take" that test, i.e.
>> break the code. But it takes orders of magnitude less to both generate the
>> encrypted message, and confirm any answer the test-taker provides. This is
>> quite similar to the problem of theorem-provers mentioned above. Another
>> example of a test could be: "Here's a lab full of standard stock
>> ingredients. Create something that will make me trip. I will give you your
>> grade one hour after you deliver your answer."
>>
>> - Luke
>>
>>
>>
> http://lesswrong.com/static/imported/2008/05/21/mindscalereal.png
>
> How does a mouse administer a test to a human to gauge the human's
> intelligence? I could also write a program to crack the private key with
> brute force, but I'd hardly call that intelligent.
>
> I've been lurking on this list for a while and I haven't found this thread
> very useful or interesting. From my point of view, Luke, you aren't fully
> grasping the difficulties of creating a Friendly AI. I think your enthusiasm
> is to be admired, but there needs to be a lot more rigor in your material.
>
> - Kevin
>
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:05 MDT
