= 27d567bc2d48d9f40e9ccc20ea370b15)
Eliezer_Yudkowsky comments on The Sword of Good - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (292)
What, you mean try to self-modify? Oh hell no. Human brain not designed for that. But you would have a longer time to try to solve FAI. You could maybe try a few non-self-modifications if you could find volunteers, but uploading and upload-driven-upgrading is fundamentally a race between how smart you get and how insane you get.
The modified people can be quite a bit smarter than you are too, so long as you can see their minds and modify them. Groves et al managed to mostly control the Manhattan project despite dozens of its scientists being smarter than any of their supervisors and many having communist sympathies. If he actually shared their earlier memories and could look inside their heads... There's a limit to control, you still won't control an adversarial super intelligence this way, but a friendly human who appreciates your need for power over them? I bet they can have a >50 IQ point advantage, maybe even >70. Schoolteachers control children who have 70 IQ points on them with the help of institutions.
Is it relevant that IQ is correlated with obedience to authority?
And how dumb do you think schoolteachers are? Bottom of those with BAs. I'd guess 100. And correlated with their pupils.
Estimations from SAT scores imply that the IQ of teachers and education majors is below average. Conscientious, hardworking students can graduate from most high schools and colleges with good grades, even if they are fairly stupid, as long as they stay away from courses which demand too much of them, and there are services available for those who are neither hardworking nor conscientious.
Education major courses are somewhat notorious for demanding little of students, and it is a stereotypically common choice for students seeking MRS degrees.
I'd like to imagine that the system would at least filter out individuals who are borderline retarded or below, but experience suggests to me that even this is too optimistic.
I don't buy the conversion in the first link, which is also a dead link. That Ed majors have an SAT score of 950 sounds right. That is 37th percentile among "college-bound seniors." If this population, which I assume means people taking the SAT, were representative of the general population, that would be an IQ of 95, but they aren't. I stand by my estimate of 100.
I doubt you have much experience with people with an IQ of 85, let alone the borderline retarded.
What makes you doubt I have much experience with either? IQ 85 is one standard deviation below average; close to 14 percent of the population has an IQ at least that low. The lower limit of borderline retardation, that is, the least intelligent you can be before you are no longer borderline, is two standard deviations below the mean, meaning that about one person in fifty is lower than that.
As it happens, I've spent a considerable amount of time with special needs students, some of whom suffer from learning disabilities which do not affect their reasoning abilities, but some of whom are significantly below borderline retarded.
At the public high school I attended, more than 95% of the students in my graduating year went on to college. While the most mentally challenged students in the area were not mainstreamed and didn't attend the same school, there was no shortage of <80 IQ students.
An average IQ of 100 for education majors would be within the error bars for the aforementioned projection, but some individuals are going to be considerably lower.
Those two sentences are not very compatible.
The rates at which students progress to college have a lot more to do with parental expectations, funding, and the school environment than the intelligence of the students in question. My school had very good resources to support students in the admissions process, and students who didn't take it for granted that they were college bound were few and far between.
It seems unrealistic to assume that we'll be able to literally read the intentions of the first upload; I'd think that we'd start out not knowing any more about them than we would about an organic person through external scanning.
You won't be able to evaluate their thoughts exactly, but there's a LOT that you should be able to tell about what a person is thinking if you can perfectly record all of their physiological reactions and every pattern of neural activation with perfect resolution, even with today's knowledge. Kock and Crick even found grandmother neurons, more or less.
I'd still expect it to be hard to tell the difference someone between thinking about or wanting to kill someone/take over the world and someone actually intending to. But I can imagine at least being able to reliably detect lies with that kind of information, so I'll defer to your knowledge of the subject.
Eliezer, I'm with you that a properly designed mind will be great, but mere uploads will still be much more awesome than normal humans on fast forward.
Without hacking on how your mind fundamentally works, it seems pretty likely that being software would allow a better interface with other software than mouse, keyboard and display does now. Hacking on just the interface would (it seems to me) lead to improvements in mental capability beyond mere speed. This sounds like mind hacking to me (software enhancing a software mind will likely lead to blurry edges around which part we call "the mind"), and seems pretty safe.
Some (pretty safe*) cognitive enhancements:
All of which is just to say that I don't think you've tried very hard to think of safe self-modifications. I'm pretty confident that you could come up with more, and better, and safer than I have.
* Where "pretty safe" means "safe enough to propose to the LW community, but not safe enough to try before submitting for public ridicule"
*blinks* I understand your "oh hell no" reaction to self modification and "use the speedup to buy extra time to solve FAI" suggestion.
However, I don't quite understand why you think "attempted upgrading of other" is all that much better. If you get that one wrong in a "result is super smart but insane (or, more precisely, very sane but with the goal architecture all screwed up) doesn't one end up with the same potential paths to disaster? At that point, if nothing else, what would stop the target from then going down the self modification path?
Non-self-modification is by no means safe, but it's slightly less insanely dangerous than self-modification.
Ooooh, okay then. That makes sense.
Hrm... given though your suggested scenario, why the need to start with looking for other volunteers? ie, if the initial person is willing to be modified under the relevant constraints, why not just, well, spawn off another instance of themselves, one the modifier and one the modifiee?
EDIT: whoops, just noticed that Vladimir suggested the same thing too.
If insane happens before super-smart, you can stop upgrading the other.
Well, fair enough, there is that.
Perhaps you mean to say that we're not particularly trustworthy in our choices of what we modify ourselves to do or prefer?
Human brains, after all, are most exquisitely designed for modifying themselves, and can do it quite autonomously. They're just not very good at predicting the broader implications of those modifications, or at finding the right things to modify.
We're talking about direct explicit low level self modification. ie, uploading, then using that more convenient form to directly study one's own internal workings until one decides to go "hrm... I think I'll reroute these neural connections to... that, add a few more of this other kind of neuron over here and..."
Recall that the thing doing all that reasoning is the thing that's being affected by these modifications.
Yes, but that would be the stupidest possible way of doing it, when there are already systems in place to do structured modification at a higher level of abstraction. Doing it at an individual neuron level would be like trying to... well, I would've said "write a property management program in Z-80 assembly," except I know a guy who actually did that. So, let's say, something about 1000 times harder. ;-)
What I find extremely irritating is when people talk about brain modification as if it's some sort of 1) terribly dangerous thing that 2) only happens post-uploading and 3) can only be done by direct hardware (or simulated hardware) modification. The correct answer is, "none of the above".
Well, we're talking about the kind of modifications that ordinary, non-invasive, high-level methods, acting through the usual sensory channels, don't allow. For example, no amount of ordinary self-help could make someone unable to feel physical pain, or can let you multiply large numbers extremely quickly in the manner of a savant. Changing someone's sexual orientation is also, at best, extremely difficult and at worst impossible. We can't seem to get rid of confirmation bias, or cure schizophrenia, or change an autistic brain into a neurotypical brain (or vice versa). There are lots of things that one might want to do to a brain that simply don't happen as long as that brain is sitting inside a skull only receiving input through normal human senses.
Lists like that have a good chance of canceling out. That is, there are a bunch of ways people disagree with you because they're talking about something else.
You can make volunteers out of your own copies. As long as the modified people aren't too smart, it's safe keep them in a sandbox and look through the theoretical work they produce on overdrive.
AI boxes are pretty dangerous.
(I agree that "as long as the modified people aren't too smart" you're safe, but we are hacking on minds that will probably be able to hack on themselves, and possibly recursively self-improve if they decide, for instance, that they don't want to be shut down and deleted at the end of the experiment. I'm pretty strongly motiviated not to risk insanity by trying dangerous mind-hacking experiments, but I'm not going to be deleted in a few minutes.)