michael vassar vs. weird things, round two
Here it is, the second episode of my exchange with Michael Vassar, the president of the Singularity Institute. If you caught the first edition, get ready to switch gears a bit and dive into theoretical computer science as we explore the ideas of mind uploading, simulating human brains with supercomputers and what that may mean for scientists while encountering a scenario that could make Descartes shudder in terror…
“Please Leave Your Brain Where It Is” expresses a fairly detailed argument that brains are unlike computers. This is obviously true. The brain doesn’t resemble a computer very closely, certainly not closely enough for such a resemblance to convince us that computers could think. Turing said so quite clearly in his seminal article “Computing Machinery and Intelligence” and I addressed precisely this misconception in my recent article in Forbes online.
The post itself dealt with Ray Kurzweil’s idea of uploading human minds to a computer by mid-century which he expressed in The Singularity Is Near, giving it a description very reminiscent of the classic anime Ghost In The Shell. In order to upload a brain anywhere, you need a system that actually works like a brain so whatever you upload will actually function. Since computers don’t match the criteria, an upload seems like an idea that’s completely unrealistic in implementation.
In paragraphs four and five of “Ray Kurzweil’s Digital Pipe Dreams” you seem to say that Ray is promoting some process of sucking information out of a person while discarding all the rest of what they are. I can’t really make any sense out of such a proposal, but I am confident no one I know of is making it.
So is Ray alone in making this claim? Here’s a Wikipedia’s cliff notes on his aforementioned book, for free, public reference. In the description of Ray’s vision for the 2030s, we find the mention of mind uploading, the exact idea that post and it’s follow up tackle. Let’s remember that Kurzweil’s goal is to cheat death with cutting edge technology so simply trying to copy his mind to a supercomputer wouldn’t get the job done. He’d need a full blown brain to machine transfer.
People do propose simulating brains, which will of course require gathering a lot of information from them. The first two talks at the 2009 Singularity Summit will discuss technical details relating to brain simulations, but the important claim is simply that the brain is a physical system and it is possible for a computer to simulate any physical system. If a physical brain is interested in steak or in sex, as in your examples, a simulation of that brain which produces brain-like behavior will also be interested in steak or sex, or at least will transform inputs into outputs as if it was.
Ah but it’s not that simple. You need stimuli and a way to virtually control those urges. In effect, you would be a puppet master running through the brain’s routine as understood by the developers who write the software by which it functions. But really, that’s beside the point when it comes to mind uploading because the way that a simulated brain will work, will be very different from the way a human brain does, as we both agree.
In general, we currently lack a robust theory of consciousness. Most Singularitarians do think that a simulation that behaves exactly like them must be conscious, but the truth [...] of this claim doesn’t have any bearing on the practical impact of simulated humans.
Actually it does. If a simulated human brain is conscious, is aware and is capable of reasoning, anything you do to it must follow the same ethical guidelines as any other person. If you were to run a test on a conscious brain in a computer and your test causes a critical system crash, then you would have technically committed homicide. The laws and rights for human beings are based on the ability to reason and our sapience. If your creation has a consciousness and aware of the environment around it, it should have the same legal right as a human. The testing and experiments you go on to mention may be severely restricted by ethical guidelines and rightfully so. However, I don’t see any reason why a simulated brain would be capable of consciousness since it would simply visualize chemical and electrical signals in our brain by solving formulas.
And this is where I have to put up another objection when it comes to using this simulated brain in a medical experiment or for clinical research. The brain would be built by software designers and developers and thus, it will be based on their understanding of the brain and how it works. But that understanding might be wrong so a doctor trying to figure out something about the brain would of course be skeptical and would need to confirm whether the brain truly works the way it does in the simulation. Even if everything works fine, the fact that it’s a simulation and not an actual brain would restrict the applicability of the research, like a cosmological model has to be supported by astronomical observations before it becomes a full blown theory.
[ illustration by Steven Wen ]
“In order to upload a brain anywhere, you need a system that actually works like a brain so whatever you upload will actually function. Since computers don’t match the criteria, an upload seems like an idea that’s completely unrealistic in implementation.”
OK, computer science then. The most basic principle is that a computer can, as my linked essay points out, simulate things other than computers, such as airplanes, hurricanes, or brains. To everyone who uses the term seriously “upload” means create a computer model of the brain with high enough fidelity that it produces the same simulated behaviors as the real person would in an equivalent environment. Of course, simulating the creation of a new scientific theory or a new novel IS creating a new scientific theory or a new novel.
“So is Ray alone in making this claim?”
No. Ray isn’t making that claim. No-one at all prominent is. Cliff notes aren’t a good substitute for a book. Given that Kurzweil explicitly considers simulating his father as reconstructed from Ray’s and other people’s memory plus DNA samples to constitute bringing his dead father back to life (see http://www.rollingstone.com/news/story/25939914/when_man__machine_merge/6) if’s a VERY safe claim that he thinks a higher fidelity simulation would constitute survival.
I can’t make any sense of the distinction made here
“Let’s remember that Kurzweil’s goal is to cheat death with cutting edge technology so simply trying to copy his mind to a supercomputer wouldn’t get the job done. He’d need a full blown brain to machine transfer.”
I couldn’t make much sense of the next few paragraphs either. Honestly, it seems to me that Greg is confused about what science is currently confused about. A simulated brain may or may not be conscious, as we don’t have a theory of consciousness, but it would definitely be capable of reasoning, as reasoning is a process of transforming inputs into the same outputs that a human would transform them into in cases where such a transformation is deterministic and not chaotic, which is just what a simulation of a human would do given those same inputs or it’s not a very good simulation.
Then we get into ethics and we don’t disagree much once again. Hopefully such experiments would be seriously restricted by ethical guidelines, but of course, that will only be able to happen if a great deal of prior debate takes place, as currently you would be laughed out of court for prosecuting someone for the homicide of a computer simulation.
Software modelers will simulate brains with wrong models of how brains work MANY times before getting it right of course, but those simulations won’t predict realistic human behaviors. Once you have a simulation that predicts realistic human behaviors in a wide variety of circumstances you will have validated your model to a high degree and can rely on it fairly confidently in new domains.
Really though, I would rather have Greg’s elaboration on this one sentence than on anything else.
“”I don’t see any reason why a simulated brain would be capable of consciousness since it would simply visualize chemical and electrical signals in our brain by solving formulas.”"
By that, Greg, do you mean that you don’t see any reason why a simulated brain would be able to create scientific theories? Create art? Respond appropriately to all manner of drugs and modifications? If not, why not? How about convince a human that it was conscious? Convince a human that it was conscious if it was given control of a suitable real-world body? If not, why not?
If the simulation was accurate, how would we know that it was or wasn’t conscious?
“To everyone who uses the term seriously ‘upload’ means create a computer model of the brain with high enough fidelity…”
That’s not what it means from an implementation standpoint. And as noted, if you create a computer model of a brain, that doesn’t mean you’ll necessarily get the brain you want. Just a model of one.
“No. Ray isn’t making that claim. No-one at all prominent is. Cliff notes aren’t a good substitute for a book.”
The cliff notes were, as noted, for public reference. In the book itself, Ray talks about actual mind uploading, the transfer of consciousness into a machine. As I said, it’s all very Ghost In The Shell.
“A simulated brain may or may not be conscious, as we don’t have a theory of consciousness, but it would definitely be capable of reasoning…”
Well that’s a pretty wild claim. A simulated airplane (to use one of your favorite analogies) isn’t actually landing or fighting with crosswinds. Instead, a routine in the computer simulates the conditions and calculates the proper response that should be given in a simulation. Likewise, a simulated brain wouldn’t really reason or think on its own, just activate the areas associated with a particular type of reasoning when you enter a command.
Human brains work by reacting to stimuli and as bits and bytes on the inside of a hard drive, there’s really no stimulation at all. An operator would have to activate the senses and feed them with information. Then, formulas, routines and subroutines would have to calculate a response and return a reaction.
“Once you have a simulation that predicts realistic human behaviors in a wide variety of circumstances you will have validated your model to a high degree and can rely on it fairly confidently in new domains.”
This is basically a more complicated Turing Test which is a good way to examine our ideas of what is and isn’t intelligent at first glance, but the test itself always had a caveat. It’s not whether it meets out expectations, it’s how it does it. Previously on this blog, I’ve given a specific case where I would consider the actions of a computer to be indicative of intelligence and creativity. Machines doing what they’re expected and how, doesn’t impress me much. Machines actually innovating and doing things never programmed into them by a human… that would get me all excited.
And this brings us back to one of the basic points of scientific research that slips if we build a predictable model of the human brain in a computer. Scientists constantly find new things about the brain and how it works. Because this knowledge is being constantly updated, it wouldn’t be in the model since it’s not a functional requirement during dev. If you’d want to map out neuron by neuron how a well known and understood reaction takes place, the computer model is perfect. If you want to truly discover brand new functions of the brain, you’re better off with a real mind to study.
“By that, Greg, do you mean that you don’t see any reason why a simulated brain would be able to create scientific theories?”
It could be guided into making some sort of hypothesis, yes. And you could uncover all sorts of interesting things in the process of creating a simulated mind. But would the computer mind be able to do it all on its own? No.
“Create art?”
No. We don’t even know how we create art. Programming an artistic subroutine into a simulated mind would be hit and miss experiment with the odds stacked against the dev team.
“Respond appropriately to all manner of drugs and modifications?”
If by respond you mean that the right areas would light up according to the software that controls the output, then yes. But of course that would be more of the dev team’s achievement than the simulated brain’s.
“Convince a human that it was conscious if it was given control of a suitable real-world body?”
I think it could depending on the programming and the stimulus. We base the idea of consciousness on our interactions. Whether that consciousness is seen as genuine or just being put on for show is for the human to decide.
“If the simulation was accurate, how would we know that it was or wasn’t conscious?”
That’s a very tricky question. If the simulation merely follows the routines, I think we can safely say that it’s not conscious. If it starts acting of its own free will and tries to run other software packages or otherwise acting as it wants to, then there’s a fair chance that it is a conscious, self-aware construct. Though considering how this software would have to be built, the chances against that are pretty significant.
What do you mean by “That’s not what it means from an implementation standpoint. “? That’s exactly what it means to Kurzweil, as described in his book. He also thinks that this process would constitute the transfer of consciousness, or the duplication of consciousness, depending on whether the original still existed. If you disagree with that, your disagreement is metaphysical.
A good enough simulated airplane responds to crosswinds exactly as a real one does. A simulated world and simulated body would provide the inputs and accept the outputs. VR systems should be good enough to provide the simulated world in the not terribly distant future. Simulations constantly do things we don’t expect them to do. That’s why we bother to make them. To find out what they would do.
Obviously we need to figure out the rest of how the brain works completely on at least one level of description in order to adequately stimulate it. Without nanotech I think that would take most of a century, but with nanotech it should happen in half that time.
I don’t actually think you disagree with “The Singularity Is Near” in any important respect other than whether a simulation of your brain constitutes ‘you’.
“[Kurzweil] thinks that this process would constitute the transfer of consciousness, or the duplication of consciousness…”
Ok, when you’re uploading something, you’re engaging in a process of transferring data from one device to another. When you’re copying, you’re creating a duplicate of the data to exist on another device. You either transfer or duplicate the data because you can’t do both at the same time. This is something that a person with Ray’s impressive credentials in comp sci should know very well.
This is hardly a metaphysical issue. It’s an issue of what exactly you’re transferring, where and how.
“A good enough simulated airplane responds to crosswinds exactly as a real one does.”
And yet again, that’s completely besides the point. The question is how. In our world, an airplane would respond to crosswinds because there’s actual fluid friction of the air acting on it as it tries to land or take off. In the virtual world, a programmer creates a virtual crosswind the behavior of which is calculated by a series of applicable Navier–Stokes equations and the aircraft model runs a subroutine which would calculate how it would respond to the forces produced by the friction and respond.
There’s a whole lot of calculations and fine tuning that go under the hood rather than in the real world where things just happen and we come up with equations to quantify the events we see. And sometimes we get them wrong which means that while our model airplane looks right, it’s actually violating real world principles we don’t know yet.
“Simulations constantly do things we don’t expect them to do.”
Yes, in no small part because of bugs from the development process. The ones that work correctly and do something we never expected, are usually a result of complex calculations we know and understand being taken to extremes and grand scales we couldn’t do by hand. Let’s keep in mind that all these simulations are ran on powerful supercomputers which are basically designed to be giant calculators. If you don’t know the basic rules of what you’re trying to simulate, you get a GIGO model: garbage in, garbage out.
The term upload is being used in it’s standard within-field jargon sense for transhumanists, who have been using it Ray’s way for 20 years in this context. Energy means “pep” but to physicists its standard use is force times distance. Words develop alternative meanings. In this jargon sense, when you use the data in a person’s brain to make a copy high enough in resolution to retain their memories and claim their identity credibly, you have ‘uploaded’ them. In any event, any non-quantum data that can be uploaded can also be copied.
Are you saying that maybe an upload’s behavior would be good enough to fool us but not exactly the same as the real person’s behavior would have been? Surely that’s true. It’s a legitimate concern. This wouldn’t prevent uploads from completely changing the world though. Here’s the classic article on that. Note that consciousness is irrelevant. http://hanson.gmu.edu/uploads.html
Who do you imagine would disagree with the statement below. We all know the point of simulating things.
“The [simulations] that work correctly and do something we never expected, are usually a result of complex calculations we know and understand being taken to extremes and grand scales we couldn’t do by hand. Let’s keep in mind that all these simulations are ran on powerful supercomputers which are basically designed to be giant calculators. If you don’t know the basic rules of what you’re trying to simulate, you get a GIGO model: garbage in, garbage out.”
If I might interject ex post facto: I think there’s a definitional problem going on here.
gfish is using “simulation” as if it means “a high-level simulation of the physical processes of a generic human brain, run in a way roughly analogous to the way computer programs we’re familiar with are run.
Michael Vassar is using “simulation” as if it means “a reproduction of the laws of physics at whatever level turns out to be necessary to make a digital replica of an actual human’s brain function the same way it does outside the computer, on a universal turing machine with inputs and outputs that may not resemble computer programs we’re familiar with.”
It’s probably too late to impact this particular debate, but that might help in the future.