Re: SIAI's flawed friendliness analysis

From: Brian Atkins (
Date: Fri May 23 2003 - 21:20:00 MDT

Bill Hibbard wrote:
> On Tue, 20 May 2003, Brian Atkins wrote:
>>. . .
>>>If humans can design AIs smarter than humans, then humans
>>>can regulate AIs smarter than humans.
>>Just because a human can design some seed AI code that grows into a SI
>>does not imply that humans or human-level AIs can successfully
>>"regulate" grown SIs.
> The regulation is not intended to trace the thoughts
> and development of the SI. The inspection is of the
> design, not the changing contents of its mind. If it's
> initial reinforcement values are for human happiness,
> and its simulation and reinforcement learning
> algorithms are accurate, then we can trust the way it
> will develop. In an earlier email I made the analogy
> to game playing programs. If their game simulation
> and learning algorithms are accurate and efficient,
> and their reinforcement learning values are for winning
> the game, then although the details of their play are
> not predictable, the fact that they will play to win
> is predictable.
> The first SI will be designed and educated by humans.
> Humans will be able to understand and regulate its
> design, and regulate how it is educated. This will
> create trusted safe SIs. They can then design and
> regulate improved SIs, with one independently
> designed SI inspecting the designs of another.

You have a differing view of the development process of AGI from myself
(and I would guess most people here). I do not believe it is that likely
that SI will arrive, in its full design, from humans. I find it much
more likely that it will come from a lengthy (in terms of iterations)
continuing redesign process undertaken by an approximately human level
AGI ("seed AI") that is capable of understanding its own design and
improving upon it in a stepwise fashion.

So, while human regulators may possibly be able to understand the design
of an early AGI (assuming no evolutionary programming, chaotic
"emergence" techniques, or other obscuring programming methods are
utilized), I do not have any surety that they will be able to understand
it later on. Perhaps if the growing AGI stopped for several months and
laid it out in easily digestible chunks, MAYBE- but at this point you
are taking its word at face value that it hasn't hidden anything from you.

Does your plan rely on your supposition, or can it tolerate a seed AI

>>>It is not necessary
>>>to trace an AI's thoughts in detail, just to understand
>>>the mechanisms of its thoughts. Furthermore, once trusted
>>>AIs are available, they can take over the details of
>>>design and regulation. I would trust an AI with
>>>reinforcement values for human happiness more than I
>>>would trust any individual human.
>>>This is a bit like the experience of people who write
>>>game playing programs that they cannot beat. All the
>>>programmer needs to know is that the logic for
>>>simulating the game and for reinforcement learning are
>>>accurate and efficient, and that the reinforcement
>>>values are for winning the game
>>>You say "by your design the 'good AIs' will be crippled
>>>by only allowing them very slow intelligence/power
>>>increases due to the massive stifling human-speed". But
>>>once we have trusted AIs, they can take over the details
>>>of designing and regulating other AIs.
>>Well perhaps I misunderstood you on this point. So it's perfectly ok
>>with you if the very first "trusted AI" turns around and says: "Ok, I
>>have determined that in order to best fulfill my goal system I need to
>>build a large nanocomputing system over the next two weeks, and then
>>proceed to thoroughly redesign myself to boost my intelligence 1000000x
>>by next month. And then, I plan to take over root access to all the nuke
>>control systems on the planet, construct a fully robotic nanotech
>>research lab, and spawn off about a million copies of myself."? If
>>you're ok with that (or whatever it outputs), then I can withdraw my
>>quote above. I fully agree with you that letting a properly designed and
>>tested FAI do what it needs to do, as fast as it wants to do it, is the
>>safest and most rational course of action.
> For me, a trusted safe AI is one whose reinforcement
> values are for human happiness. The behavior you describe
> would make people unhappy, and therefore would not be
> learned. The thing about using human happiness as a
> reinforcement value is keeping humans "in the loop" of
> the AI's thinking, no matter how intelligent it becomes.

Aside from the nukes thing, what exactly about what it said makes you
unhappy? It seems obvious to me that to increase its ability to satisfy
its goal of increasing happiness it will logically want to become
smarter, better able to communicate widely with a large number of human
individuals, and to have manufacturing capabilities in order to
implement things needed for happiness. Are you saying that the best way
an AGI can make people happy is for it to self limit its capabilities
and influence to human levels?

It seems more apparent, that what you mean by a "trusted safe AI" is: a
roughly human-level AI that never grows beyond the point where humans
lose the ability to understand its decisions, and furthermore, said AI
always submits its decisions to human level governmental bodies for
discussion and approval. i.e. it remains basically, a "tool".

The kinds of things it suggests doing would, to me, increase my
happiness. I would *like* to have nukes under the control of a
theoretically more rational entity, and I would *like* said entity to
have the means to build for me whatever I desire, and protect me when
necessary from other sentient beings. You may not personally like it,
and the US government may not like it, but what if it determines that a
majority of the humans on the planet *do* like it? Or that it should be
its goal to serve each human individually?

>>Now you also still haven't answered to my satisfaction my objections
>>that the system will never get built due to multiple political, cost,
>>and feasibility issues.
> I'll grant that the process will be very complex and
> politically messy. There will certainly be a strong urge
> to build AI, because of the promise of wealth without work.
> But when machines start suprising people with their
> intelligence, the public be reminded of the fears raised
> by science fiction books and movies. Once the public is
> excited, the politicians will get excited and turn to
> experts (it is encouraging that Ray Kurzweil has already
> testified before congress about machine intelligence).
> There will be conflicting opinions among the experts.
> Among the public there will also be conflicting opinions,
> as well as lots of crazy opinions. This will all create a
> very raucous political situation, a good example of the
> old line that its not pretty to watch balony and
> legislation being made. Nevertheless, in the end it is
> this public and democratic political process that we
> should all trust best (if we've learned the lessons of
> history).
> I don't see cost as a show-stopper. The world is pouring
> huge resources into advancing technology. Regulation will
> have its costs, but I don't see them making the whole
> project infeasible. Embedding one inspector per designer
> would roughly double costs, nine inspectors per designer
> (that's probably too many) would multiply costs by ten.
> These don't make the project infeasible. The singularity
> is one project where we don't want to cut corners for cost.

There are other aspects of your plan that I am referring to. For
instance you suggest that a wide ranging detection system will be
required in order to prevent UFAI projects. How exactly will this work?
Also, will the USA invade or economically restrict any countries that
fail to sign on to this AGI regulation system?

>>. . .
>>>Powerful people and institutions will try to manipulate
>>>the singularity to preserve and enhance their interests.
>>>Any strategy for safe AI must try to counter this threat.
>>Certainly, and we argue the best way is to speed up the progress of the
>>well-meaning projects in order to win that race.
>>Your plan seems to want to slow down the well-meaning projects, because
>>out of all AGI projects they are the most likely to willingly go along
>>with such forms of regulation. This looks to many of us here as if you
>>are going out of your way to help the "powerful people and institutions"
>>get a better shot at winning this race. Such people and institutions are
>>the ones who have demonstrated time and time again throughout history
>>that they will go through loopholes, work around the regulatory bodies,
>>and generally use whatever means needed in order to advance their goals.
>>Again, to most of us, it just looks like pure naivete on your part.
> The key word here is "well-meaning". Who determines that?
> I only trust the public to determine that, via a
> democratically elected government.
> The other problem is thinking that you can help a
> "well-meaning" project win the race. Without the force
> of law to deter them, there are going to be some *very*
> well financed projects developing unsafe AI.

Yep, so again, why are you attempting to slow down what are likely "well
meaning" projects?

> For all the details that need to be worked out in the
> approach of regulation by democratic government, it is
> still far better than trusting the "well-meaning"
> intentions of some particular project, and trusting
> that it will win the race to develop AI first.

Are you saying "I don't know" ?

> The "naivete" is thinking that the wealthy and
> powerful won't understand that super-intelligence
> will have the power to rule the world, or that they
> won't try to get control over it, or that the folks
> in the SIAI are so smart that they will overcome a
> million to one disparity in resources.

Don't attempt to attribute these views to myself or SIAI, since they are
not representative of our actual views.

> The only hope
> is to get the public on our side.

Do you realize how many thousands of examples I could cite of where the
public/government utterly failed to accomplish a technical project? Even
fairly simple things like getting a dam built, or a database
restructured. Ever watch that "Fleecing of America" bit on the NBC
Nightly News?

>>. . .
>>Those weren't the point. The reason I brought up the
>>UFAI-invents-nanotech possibility is that you didn't seem to be
>>considering such unconventional/undetectable threats when you said:
>>"But for an unsafe AI to pose a real
>>threat it must have power in the world, meaning either control
>>over significant weapons (including things like 767s), or access
>>to significant numbers of humans. But having such power in the
>>world will make the AI detectable, so that it can be inspected
>>to determine whether it conforms to safety regulations."
>>When I brought up the idea that UFAIs could develop threats that were
>>undetectable/unstoppable, thereby rendering your detection plan
>>unrealistic, you appeared to miss the point because you did not respond
>>to my objection. Instead you seemed on one hand to say that "it is far
>>from a sure thing" and on the other hand that apparently you are quite
>>sure that humans will already have detection networks built for any type
>>of threat an UFAI can dream up (highly unlikely IMO). Neither are good
>>answers to how your plan deals with possibly undetectable UFAI threats.
> I never said I was "quite sure that humans will already have
> detection networks built for any type of threat an UFAI can
> dream up". I admit the words you quoted by me are more
> optimistic than I really intended. What I really should say
> is that democratic government, for all its faults, has the
> best track record of protecting general human interests. So
> it is the democratic political process that I trust to cope
> with the dangers of the singularity.

Good, I'm glad the magical powers of democratic government automatically
solve the technical issue I was attempting to engage you on.

Brian Atkins
Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT