--- Benjamin Goertzel <[EMAIL PROTECTED]> wrote: > > (Echoing Joshua Fox's request:) Ben, could you > also tell us where you > > disagree with Eliezer? > > Eliezer and I disagree on very many points, and also > agree on very > many points, but I'll mention a few key points here. > > (I also note that Eliezer's opinions tend to be a > moving target, so I > can't say for sure that I disagree with his current > opinions, only > with some of his prior statements!) > > I disagree with his previously stated opinion that > "If an AGI is > created by humans without a solid, fairly complete > formal > understanding of why it is almost sure to be > Friendly ... then it is > extremely likely that the AGI will be Unfriendly." > > I really don't see how we can know that...
We can't "know it" in the sense of a mathematical proof, but it is a trivial observation that out of the bazillions of possible ways to configure matter, only a ridiculously tiny fraction are Friendly, and so it is highly unlikely that a selected AI will be Friendly without a very, very strong Friendly optimization over the set of AIs. In addition, for the vast majority of goals, it is useful to get additional matter/energy/computing power, and so unless there's something in the goal system that forbids it, turning us into raw materials/fusion fuel/computronium is the default action. > I also disagree with his previously stated > assessment of the viability of > > A) coming to a thorough, rigorous formal > understanding of AI > Friendliness prior to actually building some AGI's > and experimenting > with them > > or > > B) creating an AGI that will ascend to superhuman > intelligence via > ongoing self-modification, but in such a way that we > humans can be > highly confident of its continued Friendliness > through its successive > self-modifications > > He seems to think both of these are viable (though > he hasn't given a > probability estimate, that I've seen). > > My intuition is that A is extremely unlikely to > happen. > > As for B, I'd have to give it fairly low odds of > success, though not > as low as A. So, er, do you have an alternative proposal? Even if the probability of A or B is low, if there are no alternatives other than doom by old age/nanowar/asteroid strike/virus/whatever, it is still worthwhile to pursue them. Note that I don't know how we could go about calculating what the probability is; it's not like we've done this before. > I also disagree with his previously stated opinion > that > -- Anyone smart enough to actually create a > human-level AGI, is likely > to be smart enough to avoid the risk of creating an > Unfriendly AGI I disagree with this, and I believe Eliezer also disagrees with it nowadays. > And, I disagree with his previously stated > assessments that > -- Any AI system with significant learning power > should be considered > a significant risk to lead to an unanticipated hard > takeoff It is very easy to build an "AI" that will "learn" by trawling random facts off the Internet, but such an AI isn't a hard takeoff risk. I think a better term would be "programming ability" or "general intelligence". > For instance, we once argued about whether Genetic > Programming systems > should be considered serious risks for hard takeoff. > He said yes, I > said they're just too stupid. Even if they're stupid nowadays, if genetic programming is Turing-complete, it is possible (although not necessarily likely) for them to become arbitrarily smart with future research. > But of course, I > can't mathematically > prove that they're too stupid. But nor can I > mathematically prove > that my car won't spontaneously turn into a goose > this afternoon. See my post on this at http://www.acceleratingfuture.com/tom/?p=11. > Anyway, you get the idea. > > I have enjoyed Eliezer's writings, and think he has > done an > outstanding job of exploring some very subtle and > important issues. > But on several rather important matters of intuition > and estimation, > our best-guess opinions differ significantly -- and > in ways that have > led us down radically different R&D paths in spite > of having fairly > similar large-scale goals. Since becoming SIAI's Director of Research, have you pursued any joint projects with Eliezer? > -- Ben G - Tom > ----- > This list is sponsored by AGIRI: > http://www.agiri.org/email > To unsubscribe or change your options, please go to: > http://v2.listbox.com/member/?& > ____________________________________________________________________________________ Shape Yahoo! in your own image. Join our Network Research Panel today! http://surveylink.yahoo.com/gmrs/yahoo_panel_invite.asp?a=7 ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604&user_secret=7d7fb4d8