’ve been trying to write a persuasive essay about AI risk, but there are already a lot of those out there and I realize I should see if any of them are better before pushing mine. This also ties into a general interest in knowing to what degree persuasive essays really work and whether we can measure that.
So if you have time, I’d appreciate it if you did an experiment. You’ll be asked to read somebody’s essay explaining AI risk and answer some questions about it. Note that some of these essays might be long, but you don’t have to read the whole thing (whether it can hold your attention so that you don’t stop reading is part of what makes a persuasive essay good, so feel free to put it down if you feel like it).
Everyone is welcome to participate in this, especially people who don’t know anything about AI risk and especially especially people who think it’s stupid or don’t care about it.
I want to try doing this two different ways, so:
If your surname starts with A – M, try the first version of the experiment here at https://goo.gl/forms/8quRVmYNmDKAEsvS2
If your surname starts with N – Z, try the second version at https://goo.gl/forms/FznD6Bm51oP7rqB82
Thanks to anyone willing to put in the time.
Holy shit but that is a bad survey. The bias is real.Each essay is pretty dreadful in its own special way, but at least the test has the potential to prove it to the authors of said essays---if enough people outside the LessWrong filter bubble take the survey, that is.
It opens reasonably, by asking whether we think AI risk is worth studying -- I agree that it is, because it's a potential extinction-level event if an AI with sufficient power acts against our interests. But then the fucker concludes with a bunch of leading questions which all prejudice participants towards viewing AI as a threat.
Not to mention that the essays, even ones unrelated to AI risk, are pretty shit themselves.
Here's one:It's a stupid and intellectually dishonest argument, but note that it's formally identical to the original one used by Bostrom, Yudkowsky, and others. You know, the argument that was apparently good enough to convince super-genius Elon Musk of the unreality of our reality.The author moved the goal-posts on the initial premise from "it is easy to imagine simulating civilizations" to "it is easy to simulate civilizations" and apparently doesn't expect the audience to notice. He's changed the initial assumption from a reasonable and provable one to one which has been designed to be contradictory.Spoiler (click to show/hide)
Jesus, I hope none of these people are actually employed in the hard sciences, they've got a weaker grasp on experimental design, objectivity, and logic than I do, and I took a degree in Liberal fucking Arts. Like what the hell, they're not even trying to pretend that they're acting in good faith.Nah, they're not scientists in any real sense of the word: Big Yud is an autodidact Wunderkind whereas Scott is a doctor, and Bostrom and his colleagues are just generic hacks who have found a fertile niche in the transhumanist scene. I'm not sure all of them are always acting in bad faith, though: when you spend enough time within an insular subculture, you'll genuinely lose the ability to tell what makes a valid argument in the outside world.
As for Bostrom being a hack, I'd like you to actually explore the argument that he made in his book (rather than just some essay loosely based on his books) to come to that conclusion. That essay is not particularly representative of Bostrom's ideas.I swear to the robot-gods that one day I'll read Superintelligence from cover to cover. I've already tried a few times, but it's so riddled with goalpost-shifting shenanigans of the above type that I'm always overcome with RAEG before I get to page 5.
Holy shit but that is a bad survey. The bias is real.Each essay is pretty dreadful in its own special way, but at least the test has the potential to prove it to the authors of said essays---if enough people outside the LessWrong filter bubble take the survey, that is.
It opens reasonably, by asking whether we think AI risk is worth studying -- I agree that it is, because it's a potential extinction-level event if an AI with sufficient power acts against our interests. But then the fucker concludes with a bunch of leading questions which all prejudice participants towards viewing AI as a threat.
Not to mention that the essays, even ones unrelated to AI risk, are pretty shit themselves.Here's one:It's a stupid and intellectually dishonest argument, but note that it's formally identical to the original one used by Bostrom, Yudkowsky, and others. You know, the argument that was apparently good enough to convince super-genius Elon Musk of the unreality of our reality.The author moved the goal-posts on the initial premise from "it is easy to imagine simulating civilizations" to "it is easy to simulate civilizations" and apparently doesn't expect the audience to notice. He's changed the initial assumption from a reasonable and provable one to one which has been designed to be contradictory.Spoiler (click to show/hide)Jesus, I hope none of these people are actually employed in the hard sciences, they've got a weaker grasp on experimental design, objectivity, and logic than I do, and I took a degree in Liberal fucking Arts. Like what the hell, they're not even trying to pretend that they're acting in good faith.Nah, they're not scientists in any real sense of the word: Big Yud is an autodidact Wunderkind whereas Scott is a doctor, and Bostrom and his colleagues are just generic hacks who have found a fertile niche in the transhumanist scene. I'm not sure all of them are always acting in bad faith, though: when you spend enough time within an insular subculture, you'll genuinely lose the ability to tell what makes a valid argument in the outside world.
E:Spoiler: Related (click to show/hide)
As for the idea of "An AI would be able to predict every action of people blah blah blah". To be honest, I thought this to be true for a while. Putting it in a mathematical context, though, it's fundamentally impossible, assuming that the decision function of the human brain is a chaotic function, i.e. varies wildly for very close inputs (even if they appear close over small enough metrics of time close to 0), topologically mixes (covers every single possibility over a long enough period of time regardless of input) and has dense periodic orbits (for particular inputs the system might be predictable, but not everywhere). This system is, by its construction, impossible to mathematically model for all given inputs. No AI could "simulate" this system.A fun thing is that in the post-singularity setting Orion's Arm, the inability of AI to do this is considered one of their few absolute limits within the canon, alongside violating c and reversing entropy. It's worth noting that some of the AI in Orion's Arm are at a level where they are literally worshiped as gods even by people who understand what they are, on the basis that they fit the theological conception closer than any other demonstrable being.
Note that this does not mean that an AI can never exist as a chaotic system. It just means that the AI cannot approximate any chaotic system sufficiently far in time, including itself.
Good thing Elon Musk isn't the be-all end-all of intellect then, apparently, because looking at that I'm pretty sure it's just a word salad of the thousand year old ontological argument.
The only thing I know about the lesswrongosphere is that a bunch of them got convinced they were in robot hell unless they gave all their money to Elon Musk.
I got a part of the Waitbutwhy about superintelligence.Well, yes, this much is evident to anyone who has actually worked with data analysis (i.e. garbage in = garbage out, and no amount of clever algorithms will help produce non-garbage out of garbage), but there are also people - pretty popular people among the super-intelligence community - that say things like this (http://lesswrong.com/lw/qk/that_alien_message/):
I think the thing that bothers me most about the discussion is how easily people discard the distinction between intelligence and knowledge.
As long as man had intelligence, but hardly any culturally inherited knowledge, he wasn't significantly different from an animal.
All the scenarios where superintelligent AI kills all humans usually requires that it finds/knows a way to do this without significant effort. But how realistic is that? Extincting a specific pest is incredibly difficult for humans and requires a giant effort. Research is definitely slow and tedious even for a superintelligent AI, because AI can not change the scaling laws of common algorithms. The world is at least partly chaotic which means that prediction becomes exponentially difficult with time. There is nothing an AI can do about that.
A Bayesian superintelligence, hooked up to a webcam, would invent General Relativity as a hypothesis—perhaps not the dominant hypothesis, compared to Newtonian mechanics, but still a hypothesis under direct consideration—by the time it had seen the third frame of a falling apple. It might guess it from the first frame, if it saw the statics of a bent blade of grass.Basically, these people understand "super-intelligence" as "being omniscient", and "omniscient" as "having arbitrarily-powerful reality-warping powers" (https://wiki.lesswrong.com/wiki/Paperclip_maximizer).
One might think that the risk [..] arises only if the AI has been given some clearly open- ended final goal, such as to manufacture as many paperclips as possible. It is easy to see how this gives the superintelligent AI an insatiable appetite for matter and energy. […] But suppose that the goal is instead to make at least one million paperclips (meeting suitable design specifications) rather than to make as many as possible.Their "Bayesian" model of super-intelligence is so smart that it effortlessly takes over the world, yet so stupid that it can't even count. I'm fucking speechless.
One would like to think that an AI with such a goal would build one factory, use it to make a million paperclips, and then halt. Yet this may not be what would happen. Unless the AI’s motivation system is of a special kind, or there are additional elements in its final goal that penalize strategies that have excessively wide- ranging impacts on the world, there is no reason for the AI to cease activity upon achieving its goal. On the contrary: if the AI is a sensible Bayesian agent, it would never assign exactly zero probability to the hypothesis that it has not yet achieved its goal. […]The AI should therefore continue to make paperclips in order to reduce the (perhaps astronomically small) probability that it has somehow still failed to make at least a million of them, all appearances notwithstanding. There is nothing to be lost by continuing paperclip production and there is always at least some microscopic probability increment of achieving its final goal to be gained. Now it might be suggested that the remedy here is obvious. (But how obvious was it before it was pointed out that there was a problem here in need of remedying?)
We could and should be doing so many better things like not all dying in the climate crisis and maybe trying to actually get a funded transhuman project off the ground instead of trying to invent AI Jesus to just give it to us. Or in this case, having abstract silly vacuous arguments about what we can do to keep AI Jesus from sending us to AI Hell.>.> Even the most outlandish model runs don't suggest anything which could possibly involve us "all dying in the climate crisis" being a thing for centuries man, where does this come from? Did I miss the part where people start to spontaneously combust if the planet were to become half a Kelvin warmer? How is panic over potential kilodeath-scale outcomes over the next couple hundred years via sea level rise/heat waves/aridification/etc any less silly than concern over potential gigadeath-scale outcomes via unexpected superintelligence excursion a matter of decades from now?
As for the idea of "An AI would be able to predict every action of people blah blah blah". To be honest, I thought this to be true for a while. Putting it in a mathematical context, though, it's fundamentally impossible, assuming that the decision function of the human brain is a chaotic function, i.e. varies wildly for very close inputs (even if they appear close over small enough metrics of time close to 0), topologically mixes (covers every single possibility over a long enough period of time regardless of input) and has dense periodic orbits (for particular inputs the system might be predictable, but not everywhere). This system is, by its construction, impossible to mathematically model for all given inputs. No AI could "simulate" this system.I've been told my jerk circuits have a strangely attracting property, but I don't think I've ever been called a fractal before. My self-similarity is pretty low, after all.
We've already spoken at length about the dangers of the climate crisis, and how it's here now not in centuries, in other threads. Though the topic of this thread is an actually fake thing, let's not ruin it by repeating ourselves.We could and should be doing so many better things like not all dying in the climate crisis and maybe trying to actually get a funded transhuman project off the ground instead of trying to invent AI Jesus to just give it to us. Or in this case, having abstract silly vacuous arguments about what we can do to keep AI Jesus from sending us to AI Hell.>.> Even the most outlandish model runs don't suggest anything which could possibly involve us "all dying in the climate crisis" being a thing for centuries man, where does this come from? Did I miss the part where people start to spontaneously combust if the planet were to become half a Kelvin warmer? How is panic over potential kilodeath-scale outcomes over the next couple hundred years via sea level rise/heat waves/aridification/etc any less silly than concern over potential gigadeath-scale outcomes via unexpected superintelligence excursion a matter of decades from now
All the scenarios where superintelligent AI kills all humans usually requires that it finds/knows a way to do this without significant effort. But how realistic is that? Extincting a specific pest is incredibly difficult for humans and requires a giant effort.Back a bit, but... the latter is incredibly difficult for humans and requires a giant effort because we still more or less require the same environment to live, and need it to mostly be there when the pest is gone. We could probably wipe out, say, mosquitoes relatively easily at this point, ferex (tailored diseases, genetic muckery, etc.), but we don't because of the various knock-on effects (biosphere disruption, potential mutation in diseases or whatev') aren't worth anything involved with it. Unfortunately, most of the knock-on effects of wiping out humanity are, uh. Pretty positive. Particularly if you can still use our infrastructure and accumulated knowledge without actually needing the fleshsacks walking around crapping on everything >_>
Of course. However, stretch that period out to 10 years from now. It's very possible that a series of tragic events could leave you completely mentally broken and that that set of circumstances would enter the phase space.Assuming the chaotic decision function is a reasonable model, of course.
As I said, the phase space is actually rather small initially, depending on the event in question (sufficiently potent events, i.e. torture, war, etc, could represent the more "extreme" divergences between possible events and push towards what would not normally be considered in the phase space), and an AI could reasonably predict over a more immediate time frame one's decisions based on that immediate event. The decisions made as a result of this event, say, a year in the future, could not be predicted with any such accuracy, because the error in the initial assumptions from reality increase exponentially over a period of time.
In other words, why is it implausible that an AI could run a sufficiently advanced simulation that the simulation thinks that it is the person?
We do have weird progress towards this end goal though: http://browser.openworm.org/#nav=4.04,-0.03,2.8In other words, why is it implausible that an AI could run a sufficiently advanced simulation that the simulation thinks that it is the person?
Because it wouldn't know how. This should give you an idea of the problems we're currently banging our collective heads against. (http://blogs.sciencemag.org/pipeline/archives/2015/10/30/simulating-the-brain-sure-thing) Direct to the source. (https://mathbabe.org/2015/10/20/guest-post-dirty-rant-about-the-human-brain-project/) It won't know anything we can't tell it, and the idea that it could somehow divine the answers to these problems out of sheer processing power brings us into omniscient AI Jesus territory.
In other words, why is it implausible that an AI could run a sufficiently advanced simulation that the simulation thinks that it is the person?Ah, you're looking at a different problem, intelligence is messy anyways, but the important thing is that there is no simple way for me to prove to you that I am sitting next to real!Ispil and we're watching the outputs on the machine running sim!Ispil, i.e. the you I am speaking with on this forum.
Assuming the chaotic model here, it's really rather quite simple.
The AI would have to simulate the entire universe to exact detail. Sure, you might argue that that could be possible on sufficiently potent software design and hardware architecture.
However, such an AI would necessarily need to simulate itself.
It would need to simulate itself simulating itself.
And so on.
As to why it would need to simulate the entire universe, any chaotic model requires an exact simulation to get exact results; the system is deterministic. However, any error increases exponentially over time, so the simulation must be exact or else risk serious errors coming up. No one thing can be neglected from the simulation, due to the nature of mathematical chaos.
And yes, this is all assuming that the chaotic decision function is a reasonable model. If it isn't, then either the decision model is stochastic, or both deterministic and polynomial-scaling (or less) in perturbation scaling for any error in the input.You omit that it could be chaotic and deterministic with perfect initial information. Figuring out what will happen when you start a chaotic system is totally possible if you know how you started it.
In other words, humans are either chaotic, random, or predictable over the entirety of the phase space (in this case, the Oxford comma is in use; the "entirety of the phase space" only applies to predictable). There of course exists plenty of given inputs for particular decision functions with particular priors that particular decisions are predictable; those are the periodic orbits of the decision function.
I'm optimistic about cyborg technology outpacing AI.Ah-ha, but you see, you would be Ozymandias: Which is easier, for your computer chips to do their programmed tasks or to brainwash you into wanting to nuke the world, thus curing cancer! Oh, you poor deluded innocent, thank god we have enlightened folk like me to rationalize your utility functions in these matters. /s
By the time that AI starts evolving independent of human intervention, we should all have computer chips throughout our brains allowing us to match the AI's in thinking speed, and mega-man fists that can shoot lasers at any bots that try to attack us.
I wonder if it can be argued that if a system is Turing complete, then it exhibits chaos.There's the halting problem (https://en.wikipedia.org/wiki/Halting_problem), which, while isn't technically chaos, means that, in general, the only way to certainly predict the execution length of a program is to run it. Since a program's output can be made a function dependent on executing length, it means that, in general, the only way to certainly predict the output of a program is to run that exact program.
I'm optimistic about cyborg technology outpacing AI.The trick is, just like our current functionally!cyborg technology, the chips probably won't be in our brain. We do have the occasional bit of internal or grafted cybertech at the moment, but it looks a lot like most of our development there is going to be like it currently is -- via external peripherals. It's a lot safer, probably a fair bit more efficient, and certainly currently a hell of a lot easier to just... make stuff that interfaces with the wetware via the wetware instead of implanted hardward. Smartphones, glasses, guns... bluetooth, developing AR software, etc., etc., etc. Conceptually we could probably wire some of those directly to our brain, even at the moment (if with likely fairly shoddy results -- results, but not particularly decent ones), but.. why, when you can get the same effect laying it on the palm of your hand or building it into your eyeware?
By the time that AI starts evolving independent of human intervention, we should all have computer chips throughout our brains allowing us to match the AI's in thinking speed, and mega-man fists that can shoot lasers at any bots that try to attack us.
Their "Bayesian" model of super-intelligence is so smart that it effortlessly takes over the world, yet so stupid that it can't even count. I'm fucking speechless.
As an aside, I found two papers; one demonstrates that gliders in Rule 110 exhibit topological mixing, and the other demonstrates that gliders in Conway's Game of Life exhibit topological mixing. This means that both of those exhibit chaos, which means that all Turing-complete systems exhibit chaos.I think you need a more explicit proof that human decision making is actually Turing-complete. I see this stated but have yet to actually find proof of it, I figured for sure that Wolfram would have been responsible for this given his combination of brilliance and fetish for showing everything is computers on turtles on turtles made of computers all the way down, but I don't think he has gone that far yet.
So my initial premise, that the human decision function is chaotic, is correct by virtue of it being Turing-complete.
QuoteTheir "Bayesian" model of super-intelligence is so smart that it effortlessly takes over the world, yet so stupid that it can't even count. I'm fucking speechless.
I thought the idea there was a sort of solipsism-for-computers: the AI can't be 100% certain it has made sufficient paperclips yet, so it'll keep working to diminish the chance. After all, it might have a camera feed to count the number of paperclips rolling of the factory floor, but who'se to say the video feed isn't a recording/simulation made by those dastardly humans! As part of a test to check the AI's behaviour perhaps, or because they thought a reverse matrix would be hilarious. Or maybe a small software glitch made the computer miscount by 1, so better make more paperclips just to be a little extra sure.