Re: AGI thoughts was[AGI Reproduction? (Safety)]

From: P K (
Date: Sat Feb 04 2006 - 00:29:30 MST

>From: "nuzz604" <>
>To: <>
>Subject: Re: AGI Reproduction? (Safety)
>Date: Fri, 3 Feb 2006 20:15:20 -0800
>I understand that you have good intentions with AGI. My worry involves
>-accidentally- creating an unfriendly AGI (And this worry applies to
>anybody who builds an AGI). You can have good intentions and still create
>an unfriendly AI because of some design flaw (even if it appears to be a
>great design).

>I am worried because no one really knows how a Seed AI will function when
>it is turned on, and whether it will be friendly or not.

The odds of randomly stumbling upon a working AGI are extremely small. AGI
programming will most likely be a very deliberate process. In other words,
if and when AGI is created, the builder(s) will most likely know exactly
what they are doing.

>There are so many things that can go wrong.

Yes, but for an AGI to work, allot of things would have to go right. Why
would builder(s) capable of overcoming the enormous technical challenges of
making a working AGI succeed on all the other points and fail on that
particular point; friendliness. I'm amazed at how, in SF, AGI creators are
smart enough to build it but give goal systems so stupidly flawed that...(I
know SF is just for entertainment, I'm trying to prove a point here.)

The complexity of the task (AGI) is naturally selecting for builder(s) that
have a clue.

Do you have any particular reason to believe that the FAI problem is more
complex than the AGI problem? Most people seem to believe that intuitively.
This is due to two reasons.
1) It is easier to argue about FAI because it doesn't require as much
technical knowledge. It is easier to grasp the complexity of the
Friendliness problem first hand. It looks like a big thing to solve.
2) General intelligence seems kind of straightforward because we do it all
the time however; doing it is definitely not the same as coding it. In fact,
people systematically underestimate how complex AGI really is. There have
been many that claimed to have the AGI solution. They have all failed
todate. If you ever try coding an AGI you will very likely realize it is
more complex than you originally thought.

These two reasons cause people to focus on the FAI problem more than on the
AGI problem. Which, in my opinion is a mistake at this stage.

There is another twist to this. The FAI or UFAI concepts are mostly useless
without AGI however; working on AGI will very likely help develop FAI
1) AGI theory will give a clearer picture of how FAI can be technically
2) AGI work can have semi-intelligent tools as offshoots that, when combined
with human intelligence, enhance it (ex: human + computer + Internet >
human). We could then work on FAI theory more efficiently (and AGI aswell).

>This is why I think that the system and its safety should be analyzed, and
>go through at least several phases of testing before activation is even

There will be plenty of testing all along the project. And there wont be
just a single activation where the coders put some jargon together, compile
and see what happens. (See above)
Poof-> FAI -> YAY! ->JK ->UFAI ->NOO! ->R.I.P.

>I would also feel better if these tests are conducted by a team of
>independent AGI researchers rather than just one firm, or RGE Corp. by

The AGI coder(s) will be pre-selected for competence by the complexity of
the task. The point is moot for evil coder(s) since they wouldn't agree for
inspection anyway.
How would the "independent AGI researchers" be selected? How would we know
they are trustworthy and competent? I think this introduces more uncertainty
than it removes.

> You can have many shots at creating a Seed AI, but you only get one shot
>at creating a friendly one.
> If this is the Seed AI that is successful, then I say make that friendly
>shot count.

If you are shooting randomly, there is a small chance you will hit the right
target, a slightly larger chance you will hit the wrong target, and an
overwhelmingly huge chance you will never hit anything. If you're one of the
few who have some talent, opportunity, and persistence, you can perfect your
archery and hit targets at will. We hope you aim at FAI for all.

>Since you want to be open with the technology, I think that it is an idea
>worth considering.

I am somewhat suspicious of AGI claims. The more advanced an AI is the less
public proving it needs. A seed AI could recursively self improve, start the
singularity ... trust me we would know. A superhuman AI that is a bit short
of the Singularity (could happen) could at least make its owners
millionaires, just let it play with the stock market. Even a slow AI could
do some amazing things. At least it would be immune to human biases. There
might be some areas where it could outperform humans. They would have
investors chasing them not them having to prove themselves. When I think of
some workshop format proving, what comes to mind is Eliza type AI. It looks
smart at first glance but is inferior to humans in practically every way.
I'm not trying to be rude, but the lack of a splash sort of indicates that
there isn't that much to see.

Take advantage of powerful junk e-mail filters built on patented Microsoft®
SmartScreen Technology.
  Start enjoying all the benefits of MSN® Premium right now and get the
first two months FREE*.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:55 MDT