is not to try and make the agent do something it doesn't want to - but rather to make it want to do it in the first place.
I don't think there is much different between the two. Either way you are modifying the agent's behavior. If it doesn't want it, it won't have it.
It is not a modification if you make it that way "in the first place" as specified - and the "If it doesn't want it, it won't have it" seems contrary to the specified bit where you "make it want to do it in the first place".
The idea of off switches is not that they are guaranteed to work, but that they are a safety feature. If you can make a machine do anything you want at all, you can probably make it turn itself off. You can build it so the machine doesn't wish to stay turned on - but goes willing into the night.
We will never "know for sure" that a machine intelligence is safe. This is the real world, not math land. We may be able to prove some things about it - such that its initial state is not vulnerable to input stream buffer-overflow attacks - but we won't be able to prove something like that the machine will only do what we want it to do, for some value of "we".
At the moment, the self-improving systems we see are complex man-machine symbioses - companies and governments. You can't prove math theorems about such entities - they are just too messy. Machine intelligence seems likely to be like that for quite a while - functionally embedded in a human matrix. The question of "what would the machine do if no one could interfere with its code" is one for relatively late on - machines will already be very smart by then - smarter than most human computer programmers, anyway.
IF you can make a machine do anything you want at all, you can probably make it turn itself off.
The hardest part of Friendly AI is figuring out how to reliably instill any goal system.
Major update here.
The state of affairs regarding the SIAI and its underlying rationale and rules of operation are insufficiently clear.
Most of the arguments involve a few propositions and the use of probability and utility calculations to legitimate action. Here much is uncertain to an extent that I'm not able to judge any nested probability estimations. Even if you tell me, where is the data on which you base those estimations?
There seems to be an highly complicated framework of estimations to support and reinforce each other. I'm not sure how you call this in English, but in German I'd call that a castle in the air.
I know that what I'm saying may simply be due to a lack of knowledge and education, that is why I am inquiring about it. How many of you, who currently support the SIAI, are able to analyse the reasoning that led you to support the SIAI in the first place, or at least substantiate your estimations with other kinds of evidence than a coherent internal logic?
I can follow much of the reasoning and arguments on this site. But I'm currently unable to judge their overall credence. Are the conclusions justified? Is the coherent framework build around the SIAI based on firm ground? There seems to be no critical inspection or examination by a third party. There is no peer review. Yet people are willing to donate considerable amounts of money.
I'm concerned that, although consistently so, the SIAI and its supporters are updating on fictional evidence. This post is meant to inquire about the foundations of your basic premises. Are you creating models to treat subsequent models or are your propositions based on fact?
An example here is the use of the Many-worlds interpretation. Itself a logical implication, can it be used to make further inferences and estimations without additional evidence? MWI might be the only consistent non-magic interpretation of quantum mechanics. The problem here is that such conclusions are, I believe, widely considered not to be enough to base further speculations and estimations on. Isn't that similar to what you are doing when speculating about the possibility of superhuman AI and its consequences? What I'm trying to say here is that if the cornerstone of your argumentation, if one of your basic tenets is the likelihood of superhuman AI, although a valid speculation given what we know about reality, you are already in over your head with debt. Debt in the form of other kinds of evidence. Not to say that it is a false hypothesis, that it is not even wrong, but that you cannot base a whole movement and a huge framework of further inference and supportive argumentation on such premises, on ideas that are themselves not based on firm ground.
The gist of the matter is that a coherent and consistent framework of sound argumentation based on unsupported inference is nothing more than its description implies. It is fiction. Imagination allows for endless possibilities while scientific evidence provides hints of what might be possible and what impossible. Science does provide the ability to assess your data. Any hint that empirical criticism provides gives you new information on which you can build on. Not because it bears truth value but because it gives you an idea of what might be possible. An opportunity to try something. There’s that which seemingly fails or contradicts itself and that which seems to work and is consistent.
And that is my problem. Given my current educational background and knowledge I cannot differentiate LW between a consistent internal logic, i.e. imagination or fiction, and something which is sufficiently based on empirical criticism to provide a firm substantiation of the strong arguments for action that are proclaimed by the SIAI.
Further, do you have an explanation for the circumstance that Eliezer Yudkowsky is the only semi-popular person who's aware of something that might shatter the universe? Why is it that people like Vernor Vinge, Robin Hanson or Ray Kurzweil are not running amok using all their influence to convince people of the risks ahead, or at least give all they have to the SIAI? Why aren't Eric Drexler, Gary Drescher or AI researches like Marvin Minsky worried to the extent that they signal their support?
I'm talking to quite a few educated people outside this community. They do not doubt all those claims for no particular reason. Rather they tell me that there are too many open questions to focus on the possibilities depicted by the SIAI and to neglect other near-term risks that might wipe us out as well.
I believe that many people out there know a lot more than I do, so far, about related topics and yet they seem not to be nearly as concerned about the relevant issues than the average Less Wrong member. I could have named other people. That's besides the point though, it's not just Hanson or Vinge but everyone versus Eliezer Yudkowsky and some unknown followers. What about the other Bayesians out there? Are they simply not as literate as Eliezer Yudkowsky in the maths or maybe somehow teach but not use their own methods of reasoning and decision making?
What do you expect me to do, just believe Eliezer Yudkowsky? Like I believed so much in the past which made sense but turned out to be wrong? Maybe after a few years of study I'll know more.
...
2011-01-06: As this post received over 500 comments I am reluctant to delete it. But I feel that it is outdated and that I could do much better today. This post has however been slightly improved to account for some shortcomings but has not been completely rewritten, neither have its conclusions been changed. Please account for this when reading comments that were written before this update.
2012-08-04: A list of some of my critical posts can be found here: SIAI/lesswrong Critiques: Index