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.
Estimations from SAT scores imply that the IQ of teachers and education majors is belowaverage. 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.
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.
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:
Unmodified humans using larger displays are better at many tasks than humans using small displays (somewhat fluffy pdf research). It’ll be pretty surprising if being software doesn’t allow a better visual interface than a 30″ screen.
Unmodified humans who can touch-type spend less time and attention on the mechanics of human machine interface and can be more productive (no research close to hand). Who thinks that uploaded humans are not going to be able to figure better interfaces than virtual keyboards?
Argument maps improve critical thinking, but the interfaces are currently clumsy enough to discourage use (lots of clicking and dragging). Who thinks that being software won’t provide a better way to quickly generate argument maps?
In front of a computer loaded up with my keyboard shortcuts and browser plugins I have easy access to very fast lookup on various web reference sites. At the moment the lookup delay is still long enough that short term memory management (stack overflow after a mere 7±2 pushes) is a problem (when I need a reference I push my current task onto a mental stack; it takes time and attention to pop that task when the reference has been found). Who thinks I couldn’t be smarter with a reference interface better than a keyboard?
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”
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.
(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.)
*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?
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.
Non-self-modification is by no means safe, but it’s slightly less insanely dangerous than self-modification.
I think I see where you’re confused now. You think there’s only one of you. ;-)
But if you think about it, akrasia is an ample demonstration that there is more than one of you: the one who acts and chooses, and the one who reflects upon the acts and choices of the former.
And the one who acts and chooses also modifies itself all the frickin’ time, whether you like it or not. So if the one who reflects then refrains from modifying the one who acts, well… the results are going to be kind of random. Better directed self-modification than undirected, IMO.
(I don’t pretend to be an expert on what would happen with this stuff in brain simulation; I’m talking strictly about the behavior of embodied humans here, and my own experiences with self-modification.)
We’re talking about direct brain editing here. People who insist on comparing direct brain editing to various forms of internal rewiring carried out autonomously by opaque algorithms… or choice over deliberate procedures to follow deliberatively… well, don’t be surprised if you’re downvoted, because you did, in fact, say something stupid.
If by “direct” here you mean changing the underlying system—metaprogramming as it were, then I have to say that that’s the idea that’s stupid. If you have a system that’s perfectly capable of making changes on its own, debugged by millions of years of evolution, why on earth would you want to bypass those safeties?
If you have a system that’s perfectly capable of making changes on its own, debugged by millions of years of evolution, why on earth would you want to bypass those safeties?
You don’t need to bypass the safeties to do better. What you need is not a bigger hammer with which to change the brain, but a better idea of what to change, and what to change it to.
That’s the thing that annoys me the most about brain-mod discussions here—it’s like talking about opening up the case on your computer with a screwdriver, when you’ve never even looked at the screen or tried typing anything in—and then arguing that all modifications to computers are therefore difficult and dangerous.
To use an analogy, the kind of brain modifications we’re talking about would be the kind of modifications you’d have to do to a 286 in order to play Crysis (a very high-end game) on it.
If I’m not mistaken, as far as raw computing power goes, the human brain is more powerful than a 286. The question is—and this is something I’m honestly wondering—whether it’s feasible, given today’s technology, to turn the brain into something that can actually use that power in a fashion that isn’t horribly indirect. Every brain is powerful enough to play dual 35-back perfectly (if I had access to brain-making tools, I imagine I could make a dual 35-back player using a mere 70,000 neurons); it’s simply not sufficiently well-organized.
If your answer to the above is “no way José”, please say why. “It’s not designed for that” is not sufficient; things do things they weren’t designed to do all the time.
You don’t need to bypass the safeties to do better. What you need is not a bigger hammer with which to change the brain, but a better idea of what to change, and what to change it to.
But you do need a bigger hammer as well. And that bigger hammer is dangerous.
A brain emulation may want to modify so that when it multiplies numbers together, instead of its hardware emulating all the neurons involved, it performs the multiplication on a standard computer processor.
This would be far faster, more accurate, and less memory intensive.
Implementation would involve figuring out how the hardware recognizes the intention to perform a multiplication, represent the numbers digitally, and then present the answer back to the emulated neurons. This is outside the scope of any mechanism we might have to make changes within our brains, which would not be able to modify the emulator.
Cracking the protein folding problem, building nanotechnology, and reviving a cryonics patient at the highest possible fidelity. Redesigning the spaghetti code of the brain so as to permit it to live a flourishing and growing life rather than e.g. overloading with old memories at age 200.
I suppose you make a remarkable illustration of how people with no cosmic ambitions and brainwashed by the self-help industry, don’t even have any goals in life that require direct brain editing, and aren’t much willing to imagine them because it implies that their own brains are (gasp!) inadequate.
people with no cosmic ambitions and brainwashed by the self-help industry, don’t even have any goals in life that require direct brain editing, aren’t much willing to imagine them because it implies that their own brains are (gasp!) inadequate.
Is this your causal theory? Literally, that pjeby considered a goal that would have required direct brain editing, noticed that the goal would have implied that his brain was inadequate, felt negative self-image associations, and only then dropped the goal from consideration, and for no other reason? And further, that this is why he asked: “If you have a system that’s perfectly capable of making changes on its own, debugged by millions of years of evolution, why on earth would you want to bypass those safeties?”
I think that, where you are imagining direct brain editing done only with a formal, philosophically cross-validated theory of brain editing safety and only after a long enough delay to develop that theory, and where you imagine pjeby to be imagining direct brain editing done only with a formal, philosophically cross-validated theory of brain editing safety and only after a long enough delay to develop that theory, pjeby may be actually imagining someone who already has a brain-editing device and no safetiness theory, and who is faced with a short-range practical decision problem about whether to use the device when the option of introspective self-modification is available. pjeby probably has a lot of experience with people who have simple technical tools and are not reflective like you about whether they are safe to use. That is the kind of person he might be thinking of when he is deciding whether it would be better advice to tell the person to introspect or to use the brain editor.
(Also, someone other than me should have diagnosed this potential communication failure already! Do you guys prefer strife and ad-hominems and ill will or something?)
The x you get from
argmax_(x) U(x, y)
for fixed y is, in general, different from the x you get from
argmax_(x, y) U(x, y).
But this doesn’t mean you can conclude that the first argmax calculated U() wrong.
I suppose you make a remarkable illustration of how people with no cosmic ambitions and brainwashed by the self-help industry, don’t even have any goals in life that require direct brain editing, and aren’t much willing to imagine them because it implies that their own brains are (gasp!) inadequate.
Wow, somebody’s cranky today. (I could equally note that you’re an illustration of what happens when people try to build a technical solution to a human problem… while largely ignoring the human side of the problem.)
Solving cooler technical problems or having more brain horsepower sure would be nice. But as I already know from personal experience, just being smarter than other people doesn’t help, if it just means you execute your biases and misconceptions with greater speed and an increased illusion of certainty.
Hence, I consider the sort of self-modification that removes biases, misconceptions, and motivated reasoning to be both vastly more important and incredibly more urgent than the sort that would let me think faster, while retaining the exact same blindspots.
But if you insist on hacking brain hardware directly or in emulation, please do start with debugging support: the ability to see in real-time what belief structures are being engaged in reaching a decision or conclusion, with nice tracing readouts of all their backing assumptions. That would be really, really useful, even if you never made any modifications outside the ones that would take place by merely observing the debugger output.
you’re an illustration of what happens when people try to build a technical solution to a human problem
If there were a motivator captioned “TECHNICAL SOLUTIONS TO HUMAN PROBLEMS”, I would be honored to have my picture appear on it, so thank you very much.
If there were a motivator captioned “TECHNICAL SOLUTIONS TO HUMAN PROBLEMS”, I would be honored to have my picture appear on it, so thank you very much.
You left out the “ignoring the human part of the problem” part.
The best technical solutions to human problems are the ones that leverage and use the natural behaviors of humans, rather than trying to replace those behaviors with a perfect technical process or system, or trying to force the humans to conform to expectations.
(I’d draw an analogy with Nelson’s Xanadu vs. the web-as-we-know-it, but that could be mistaken for a pure Worse Is Better argument, and I certainly don’t want any motivated superintelligences being built on a worse-is-better basis.)
Wow what hubris the “brain is inadequate spaghetti code”. Tell me have you ever actually studied neuroscience? Where do you think modern science came from? This inadequate spaghetti code has given us the computer, modern physics and plenty of other things. For being inadequate spaghetti code (this is really a misnomer because we don’t actually understand the brain well enough to make that judgement) it does pretty well.
If the brain is as bad as you make it out to be then I challenge you to make a better one. In fact I challenge you to make a computer capable of as many operations as the brain running on as little power as the brain does. If you can’t do better then you are no better then the people who go around bashing General Relativity without being able to propose something better.
I look forward to it. (though I doubt I will ever see it considering how long you’ve been saying you were going to make an FAI and how little progress you have actually made)
But maybe your pulling a Wolfram and going to work alone for 10 years and dazzle everyone with your theory.
I don’t think there’s actually any substantive disagreement here. “Good,” “bad,” “adequate,” “inadequate”—these are all just words. The empirical facts are what they are, and we can only call them good or bad relative to some specific standard. Part of Eliezer’s endearing writing style is holding things to ridiculously impossibly high standards, and so he has a tendency to mouth off about how the human brain is poorly designed, human lifespans are ridiculously short and poor, evolutions are stupid, and so forth. But it’s just a cute way of talking about things; we can easily imagine someone with the same anticipations of experience but less ambition (or less hubris, if you prefer to say that) who says, “The human brain is amazing; human lives are long and rich; evolution is a wonder!” It’s not a disagreement in the rationalist’s sense, because it’s not about the facts. It’s not about neuroscience; it’s about attitude.
The post shows the exact same lack of familiarity with neuroscience as the comment I responded to. Examine closely how a single neuron functions and the operations that it can perform. Examine closely the ability of savants (things like memory, counting in primes, calender math...) and after a few years of reading the current neuroscience research comeback and we might have something to discuss.
What, you mean try to self-modify? Oh hell no. Human brain not designed for 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.
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...”
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”.
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”.
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.
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.
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:
Unmodified humans using larger displays are better at many tasks than humans using small displays (somewhat fluffy pdf research). It’ll be pretty surprising if being software doesn’t allow a better visual interface than a 30″ screen.
Unmodified humans who can touch-type spend less time and attention on the mechanics of human machine interface and can be more productive (no research close to hand). Who thinks that uploaded humans are not going to be able to figure better interfaces than virtual keyboards?
Argument maps improve critical thinking, but the interfaces are currently clumsy enough to discourage use (lots of clicking and dragging). Who thinks that being software won’t provide a better way to quickly generate argument maps?
In front of a computer loaded up with my keyboard shortcuts and browser plugins I have easy access to very fast lookup on various web reference sites. At the moment the lookup delay is still long enough that short term memory management (stack overflow after a mere 7±2 pushes) is a problem (when I need a reference I push my current task onto a mental stack; it takes time and attention to pop that task when the reference has been found). Who thinks I couldn’t be smarter with a reference interface better than a keyboard?
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”
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.)
*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.
I think I see where you’re confused now. You think there’s only one of you. ;-)
But if you think about it, akrasia is an ample demonstration that there is more than one of you: the one who acts and chooses, and the one who reflects upon the acts and choices of the former.
And the one who acts and chooses also modifies itself all the frickin’ time, whether you like it or not. So if the one who reflects then refrains from modifying the one who acts, well… the results are going to be kind of random. Better directed self-modification than undirected, IMO.
(I don’t pretend to be an expert on what would happen with this stuff in brain simulation; I’m talking strictly about the behavior of embodied humans here, and my own experiences with self-modification.)
We’re talking about direct brain editing here. People who insist on comparing direct brain editing to various forms of internal rewiring carried out autonomously by opaque algorithms… or choice over deliberate procedures to follow deliberatively… well, don’t be surprised if you’re downvoted, because you did, in fact, say something stupid.
If by “direct” here you mean changing the underlying system—metaprogramming as it were, then I have to say that that’s the idea that’s stupid. If you have a system that’s perfectly capable of making changes on its own, debugged by millions of years of evolution, why on earth would you want to bypass those safeties?
On that, I believe we’re actually in agreement.
To do better?
You don’t need to bypass the safeties to do better. What you need is not a bigger hammer with which to change the brain, but a better idea of what to change, and what to change it to.
That’s the thing that annoys me the most about brain-mod discussions here—it’s like talking about opening up the case on your computer with a screwdriver, when you’ve never even looked at the screen or tried typing anything in—and then arguing that all modifications to computers are therefore difficult and dangerous.
To use an analogy, the kind of brain modifications we’re talking about would be the kind of modifications you’d have to do to a 286 in order to play Crysis (a very high-end game) on it.
If I’m not mistaken, as far as raw computing power goes, the human brain is more powerful than a 286. The question is—and this is something I’m honestly wondering—whether it’s feasible, given today’s technology, to turn the brain into something that can actually use that power in a fashion that isn’t horribly indirect. Every brain is powerful enough to play dual 35-back perfectly (if I had access to brain-making tools, I imagine I could make a dual 35-back player using a mere 70,000 neurons); it’s simply not sufficiently well-organized.
If your answer to the above is “no way José”, please say why. “It’s not designed for that” is not sufficient; things do things they weren’t designed to do all the time.
But you do need a bigger hammer as well. And that bigger hammer is dangerous.
For what, specifically?
A brain emulation may want to modify so that when it multiplies numbers together, instead of its hardware emulating all the neurons involved, it performs the multiplication on a standard computer processor.
This would be far faster, more accurate, and less memory intensive.
Implementation would involve figuring out how the hardware recognizes the intention to perform a multiplication, represent the numbers digitally, and then present the answer back to the emulated neurons. This is outside the scope of any mechanism we might have to make changes within our brains, which would not be able to modify the emulator.
Cracking the protein folding problem, building nanotechnology, and reviving a cryonics patient at the highest possible fidelity. Redesigning the spaghetti code of the brain so as to permit it to live a flourishing and growing life rather than e.g. overloading with old memories at age 200.
I suppose you make a remarkable illustration of how people with no cosmic ambitions and brainwashed by the self-help industry, don’t even have any goals in life that require direct brain editing, and aren’t much willing to imagine them because it implies that their own brains are (gasp!) inadequate.
Is this your causal theory? Literally, that pjeby considered a goal that would have required direct brain editing, noticed that the goal would have implied that his brain was inadequate, felt negative self-image associations, and only then dropped the goal from consideration, and for no other reason? And further, that this is why he asked: “If you have a system that’s perfectly capable of making changes on its own, debugged by millions of years of evolution, why on earth would you want to bypass those safeties?”
I think that, where you are imagining direct brain editing done only with a formal, philosophically cross-validated theory of brain editing safety and only after a long enough delay to develop that theory, and where you imagine pjeby to be imagining direct brain editing done only with a formal, philosophically cross-validated theory of brain editing safety and only after a long enough delay to develop that theory, pjeby may be actually imagining someone who already has a brain-editing device and no safetiness theory, and who is faced with a short-range practical decision problem about whether to use the device when the option of introspective self-modification is available. pjeby probably has a lot of experience with people who have simple technical tools and are not reflective like you about whether they are safe to use. That is the kind of person he might be thinking of when he is deciding whether it would be better advice to tell the person to introspect or to use the brain editor.
(Also, someone other than me should have diagnosed this potential communication failure already! Do you guys prefer strife and ad-hominems and ill will or something?)
The x you get from
argmax_(x) U(x, y)
for fixed y is, in general, different from the x you get from
argmax_(x, y) U(x, y).
But this doesn’t mean you can conclude that the first argmax calculated U() wrong.
Wow, somebody’s cranky today. (I could equally note that you’re an illustration of what happens when people try to build a technical solution to a human problem… while largely ignoring the human side of the problem.)
Solving cooler technical problems or having more brain horsepower sure would be nice. But as I already know from personal experience, just being smarter than other people doesn’t help, if it just means you execute your biases and misconceptions with greater speed and an increased illusion of certainty.
Hence, I consider the sort of self-modification that removes biases, misconceptions, and motivated reasoning to be both vastly more important and incredibly more urgent than the sort that would let me think faster, while retaining the exact same blindspots.
But if you insist on hacking brain hardware directly or in emulation, please do start with debugging support: the ability to see in real-time what belief structures are being engaged in reaching a decision or conclusion, with nice tracing readouts of all their backing assumptions. That would be really, really useful, even if you never made any modifications outside the ones that would take place by merely observing the debugger output.
If there were a motivator captioned “TECHNICAL SOLUTIONS TO HUMAN PROBLEMS”, I would be honored to have my picture appear on it, so thank you very much.
You left out the “ignoring the human part of the problem” part.
The best technical solutions to human problems are the ones that leverage and use the natural behaviors of humans, rather than trying to replace those behaviors with a perfect technical process or system, or trying to force the humans to conform to expectations.
(I’d draw an analogy with Nelson’s Xanadu vs. the web-as-we-know-it, but that could be mistaken for a pure Worse Is Better argument, and I certainly don’t want any motivated superintelligences being built on a worse-is-better basis.)
Wow what hubris the “brain is inadequate spaghetti code”. Tell me have you ever actually studied neuroscience? Where do you think modern science came from? This inadequate spaghetti code has given us the computer, modern physics and plenty of other things. For being inadequate spaghetti code (this is really a misnomer because we don’t actually understand the brain well enough to make that judgement) it does pretty well.
If the brain is as bad as you make it out to be then I challenge you to make a better one. In fact I challenge you to make a computer capable of as many operations as the brain running on as little power as the brain does. If you can’t do better then you are no better then the people who go around bashing General Relativity without being able to propose something better.
I accept your challenge. See you in a while.
Awesome.
I look forward to it. (though I doubt I will ever see it considering how long you’ve been saying you were going to make an FAI and how little progress you have actually made) But maybe your pulling a Wolfram and going to work alone for 10 years and dazzle everyone with your theory.
I don’t think there’s actually any substantive disagreement here. “Good,” “bad,” “adequate,” “inadequate”—these are all just words. The empirical facts are what they are, and we can only call them good or bad relative to some specific standard. Part of Eliezer’s endearing writing style is holding things to ridiculously impossibly high standards, and so he has a tendency to mouth off about how the human brain is poorly designed, human lifespans are ridiculously short and poor, evolutions are stupid, and so forth. But it’s just a cute way of talking about things; we can easily imagine someone with the same anticipations of experience but less ambition (or less hubris, if you prefer to say that) who says, “The human brain is amazing; human lives are long and rich; evolution is a wonder!” It’s not a disagreement in the rationalist’s sense, because it’s not about the facts. It’s not about neuroscience; it’s about attitude.
While my sample size is limited I have noticed a distinct correlation between engaging in hubris and levelling the charge at others. Curious.
For calibration, see The Power of Intelligence.
“The Power of Intelligence”
Derivative drivel...
The post shows the exact same lack of familiarity with neuroscience as the comment I responded to. Examine closely how a single neuron functions and the operations that it can perform. Examine closely the ability of savants (things like memory, counting in primes, calender math...) and after a few years of reading the current neuroscience research comeback and we might have something to discuss.
Eliezer, replying to a comment by pjeby: “you did, in fact, say something stupid.”
Word.
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”.
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.
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.