The sense that it exists in is substance, but not in form. I think that the point that needs repeating is this is not a technical issue, it's a marketing issue. It is not about lacking the ability to tag different parts of wikipedia, or show subsets of the database. It's purely about branding and perception. So the problem with Wikipedia1.0 is that it keeps the two sides fighting over what is wiki and what is not. For a lot of these deletion debate the quality of finish is orthogonal to whether or not it should exist in the first place.
Bingo. One of the original motivations for the project was "Wiki is not paper". Originally they wanted the inclusive h2g2 approach. The current debate is one of image - if it is diluted by lots of borderline articles. Only a wikipedian could think that dodgy articles would somehow damage the "reputation" of the site, but that's a digression.
As the problem is simply one of image, create two brands, say "Wikipedia Core" and "Wikipedia Fringe". Keep everything, but only elevate articles into the core on some sort of vote / consensus. Keeps both sides happy. The inclusionists get every bit of trivia every recorded, and the deletionists get their pristine image of a "pure" encyclopaedia. Given that the project was initiated as a response to the problems of paper-based encyclopaedias I'm surprised nobody within the project has suggested this.
Seems vaguely reminiscent of slashdot around the time they introduced moderation. Reading the Fringe could someday be seen as browsing at -1.
There is a more general lesson here. When someone comes out of the workwork and says: "Look I've invented a new type of input device" then generally it will be interesting regardless of how well it works. When someone tries to flog the same dead horse that hasn't worked for twenty years then you know that it will suffer all the same failures. I'm sure everyone in this crowd has used voice activated interfaces and knows just how much they suck.
When a business analyst / investment "consultant" starts hyping up marginal advances as revolutionary and talking about coming "paradigm shifts" then you know that the bullshit is in full flow. Accelerometer interfaces are nice, they do feel more natural - I worked on one for an educational games project seven years ago. But the key point that you've captured is they are intrusive. Until the accuracy is high enough that we can make a twitch interface they are not a replacement for the traditional tools of mouse and keyboard.
What really pissed me off about the article was the insistence that these interfaces were a "direct manipulation" of images on screen. No, if you reach in and move an image (somehow) then that would be direct manipulation. If your physical gestures are translated into screen motion by accelerometers rather than a mouse then it is still an indirect interface. It is at most a minor increment on the user interface technology that we have already, the term "Paradigm Shift" is thrown about with abandon by too many suits without a understanding of what it implies.
Nah, the docking computer is for girls. It takes up a ton of space that is essential for loading up with consumer goods for the Sol to Barnard's Star cargo run. Definitely not worth it until you get a much bigger ship...
That's much more simple that what I thought you meant. It's a really nice idea, never come across anyone proposing it before but it would software a hell of a lot more robust. You should really spend some time pushing this scheme onto the right people because it would be a great deal for software reliability.
Erm WTF? The inherent difference that you are missing is that a domain name is an alias for an IP address whereas a Uniform Resource Locator is the address of a resource offered by a particular domain. Just because most URLs don't feature a path, contain only a domain name and resolve implicitly to the root of the namespace on the server does make them equivalent to domain names.
Given that the entire context of this discussion is that someone tried to hide a resource by not publishing the URL,and we can infer from this that the obscurity was in the path/filename part rather than the host, then you are talking complete shit.
It sounds like a very interesting project. The idea that screams out from the wiki summary is static analysis and verification. There is a really good rundown in one of the wiki links. The really big difference from previous work is not just the use of managed code, but splitting the entire system into either trusted, or verified code. The trusted component is a tiny core, which they are working on verifying. The design of the rest of the kernel and the SIPs is a good one: instead of doing arbitrary verification, change the language design so that you can only write verifiable code. Then see how much of an O/S you can write. The progress is astounding.
For the IPC they have made some strange choices, receiving is synchronous (as in process calculi) but sending is asynchronous. As they are writing the lowest level parts (such as the schedular) in this code it may be an implementation difficulty with synchronous sends. The cheapness of the IPC seems to be routed in the transfer of ownership that communication implies. In essence you can't alias, you can only pass by value - but the low-level runtime can modify this to pass more efficiently by reference because it can verify there are no dangling references. This would (if it works over a large enough code base) solve the performance issue with IPC in a microkernel. It is (as another reply pointed out) similar to providing the semantics of heavy-weight communication to the programmer in a way that can be implemented with cheap co-routines.
Having done some (well, little) work in this area I'm really impressed by what they've achieved already.
How far back are you talking? I'd always heard the phrase rootkit as the code you downloaded after an exploit to hide your tracks and make sure that you keep root. Of course on windows getting root in the firtst place isn't such a big deal so some language drift is understandable...
You are assuming that there is only one variable (resolution) that can be adjusted. Actually the quality of the scene is a function of two variables: resolution and scene complexity. When the complexity of scenes was low, rasterization produced much better results than raytracing for the same effort. Now that scene geometry has increased so much we are reaching the point where raytracing will produce the same (or better quality) for less effort. The main issue is that rasterization is O(n) in scene complexity while raytracing is O(log n). Of course there are lots of other issues and tradeoffs otherwise we would be using raytracing in games already.
If you're interested there is a detailed comparison available here.
"Unfortunately we're not as clever as those Intel chaps, how will we make it work?" "Hmmm....." (long pause) "What about rendering really small scenes on a big stonking server and then using some sort of 'Network' to make the images appear?" "That sounds like some kind of magic!"
If you are interested then there is a brilliant paper on the "Gold Effect"; propagation of errors within a body of literature. Unfortunately I can't find an online copy anywhere, but the details are "Journey to the centre of uncertainty", R. A. Lyttleton, in Science and Uncertainty ISBN 0-905927-96-6. He uses a series of interesting examples where the errors are now more commonplace than the theories they have displaced; the "iron core" of the earth (which I was taught in school), the existence of the Oort cloud (tenuous theory not backed up by observation) and others.
You may not realize it, but you can still make a valid point without swearing.
Very true, and it's odd but as I was posting the reply it did occur to me that we use different standards of behaviour in real life and in online forums.
I do understand what you were getting at, and the specific semantics of the term email aside it does make sense. I partly agree with your point, as it does seem absurd for the existence of certain information to be illegal. But I still partly disagree as within the context of the court case it did seem quite clear what was meant by the term.
An email is not simply some text. It is the actual message that is sent containing the text. If I write a letter to somebody and save it on my harddrive it is not an email. But if I attach an SMTP header and transmit it then it becomes an email. These messages only became emails when they were sent, and so their very existence is infact illegal.
If you are going to be anally pedantic on the use of language then have the fucking decency to apply some actual thought to your argument.
Indeed they are, and after reading your somewhat familiar brain dump I am curious enough to ask you: Is your mental illness of the manic/psychotic persuasion, or is it a type of schizophrenia? Your writing shows lots of high-level understanding of the subject and parts of your cognitive processes are obviously still functional. But you are completely irrational and your gibberish shows that you have lost the ability to filter and discard faulty hypothesis, and as a result you are unable to see where your unique view of reality has diverged from those around you. Medication could really change your life for the better.
Yes, this being slashdot I realise that I didn't (skim-)read the paper properly, or in fact, err *cough*, your comment:) Two application for this spring to mind, one is the stolen laptop scenario that everyone has mentioned. If your laptop has TruCrypt for example then this is an interesting way around it that other approaches (with physical access) would have trouble doing. The other application (which I think they mention) is forensics. If the police break into a house and confiscate a running machine with an encrypted harddrive then this would seem to be the easiest approach to gain access to the data. Where this is weaker than other physical access approaches is the assumption that you're not going to give the hardware back (quickly)...
Oh yes, I'd forgotten they were going to market it as a FPGA replacement. Although it can be used to replace a FPGA that's not such a good description of what it does. You're probably aware of Occam and the old transputer design. The Xmos chip is the logical sucessor to that. Point to point connections between processes are helped by hardware acceleration to provide to low-latency, low cost communications. Each core is multi-threaded, with multiple cores fabricated onto a single chip. I can't remember what interconnect they are using between chips, but internally it is very fast. Orders of magnitude faster to do MPI than on something like an x86. The reason that it can act as a FPGA replacement is because the new I/O is so lightweight and predictable.
I think they're still thrashing out what kind of programming model can exploit it the best, but expect something similar to the process model in Erlang. When David May has given presentations about the architecture he's mentioned that building a million-core machine is something they would like to do just to see what new applications it opens up. The challenges for the tool-chain, or the operating system are quite interesting when you hit that kind of scale. Even simple problems like how best to fork / replicate a thread across the entire machine become quite tricky.
AI is more of a hobby for me really. My real research is in compilers and languages, hence the interest in the low-level nitty gritty of how to implement neural networks effectively - all compilers need interesting test cases.
Depending on what type of NN model you are using, you may want to look at Xmos. Their processor is designed to scale up to the size of machine that you are talking about. I think that they are going to build a box like that just to see what it can do. There is no floating point unit on the chip - but you can do a lot with fixed-point math and a clever compiler.
We played around putting simple neural networks on a PIC-16f84. Much simpler than the normal discrete neurons that people use, let alone something biologically plausible. Communication between the them would have been a serious issue - but the Xmos processor is designed from the ground up to allow ligtweight communication on really large scale clusters. At some point I'm going to get around to seeing what the state-of-the-art is for biologically plausible neurons as the simulation code is probably a good test case for my new compiler.
If you get the cash from investors then you should think about contacting Xmos. They could probably build you a machine that isn't too far from Danny Hillis' vision.
From the reports that they've published I thought their simulation was quite realistic - but I'm a computer scientist rather than a neurobiologist. What are the major differences between their simulation and the real thing?
I hadn't come across that work before. They've made interesting progress in the past ten years. Although there aren't any simulations floating across the net they have reimplemented the network in silicon to drive their bio-buggy. Doesn't that count?
So that said, what type of virtual NN do you need to be supported? There is a local company coming of stealth mode at the moment that will provide the hardware that you say is necessary. I have some background in simulating NNs on simple hardware, and compiler design. What do you need?
Although I have mod points at the moment it seems that your comments have already been modded up anyway so I guess I'll reply. We've actually argued about this before (probably last time Kurzweil was mentioned on slashdot), but as you say in a different comment this discussion is always fun:)
Slight nitpick, currently we are at one billion transistors on a chip, not two, but that doesn't really change the point you are making.
A bigger issue that I have with what you've described is that simulating a brain is not the same as "solving" AI. The problem that Kurzweil has is that he refuses to accept that there is a difference. Sure, if they are the same then strong AI is inevitable and it's merely a question of building fast enough hardware. But why assume that they are the same thing?
Twenty years from now we may have hardware that can simulate an entire human brain; and yet we may be no closer to solving any of the problems in understanding how to solve the many problems in AI. The mental sleight-of-hand that Kurzweil applies to this argument is: Once we can simulate a brain we have AI, therefore the AI can design he next generation, therefore we will reach the singularity. This argument is a logical fallacy because it assumes being able to run the system, and knowing how to design the system / how the system works are equivalent.
Everything that we know complex and dynamic systems tells us that this is not so; given a simulation of the brain it is reasonable to assume that intelligence is the ultimate emergent property of the system. Understanding this property and how to refine it is the hardest problem that mankind has ever undertaken. Currently we don't really know how to pose the question, let alone how to arrive at an answer. To assume that some kind of standard engineering methodology will solve this in 20 years is wild speculation.
As always with AI, the hardware will be available but nobody yets knows if we can write the software to run on it.
Slashdot Mirror
The sense that it exists in is substance, but not in form. I think that the point that needs repeating is this is not a technical issue, it's a marketing issue. It is not about lacking the ability to tag different parts of wikipedia, or show subsets of the database. It's purely about branding and perception. So the problem with Wikipedia1.0 is that it keeps the two sides fighting over what is wiki and what is not. For a lot of these deletion debate the quality of finish is orthogonal to whether or not it should exist in the first place.
Bingo. One of the original motivations for the project was "Wiki is not paper". Originally they wanted the inclusive h2g2 approach. The current debate is one of image - if it is diluted by lots of borderline articles. Only a wikipedian could think that dodgy articles would somehow damage the "reputation" of the site, but that's a digression.
As the problem is simply one of image, create two brands, say "Wikipedia Core" and "Wikipedia Fringe". Keep everything, but only elevate articles into the core on some sort of vote / consensus. Keeps both sides happy. The inclusionists get every bit of trivia every recorded, and the deletionists get their pristine image of a "pure" encyclopaedia. Given that the project was initiated as a response to the problems of paper-based encyclopaedias I'm surprised nobody within the project has suggested this.
Seems vaguely reminiscent of slashdot around the time they introduced moderation. Reading the Fringe could someday be seen as browsing at -1.
He's taking the piss out of you because it's the polite thing to do when a geek-rapture crazy comes out of the woodwork.
There is a more general lesson here. When someone comes out of the workwork and says: "Look I've invented a new type of input device" then generally it will be interesting regardless of how well it works. When someone tries to flog the same dead horse that hasn't worked for twenty years then you know that it will suffer all the same failures. I'm sure everyone in this crowd has used voice activated interfaces and knows just how much they suck.
When a business analyst / investment "consultant" starts hyping up marginal advances as revolutionary and talking about coming "paradigm shifts" then you know that the bullshit is in full flow. Accelerometer interfaces are nice, they do feel more natural - I worked on one for an educational games project seven years ago. But the key point that you've captured is they are intrusive. Until the accuracy is high enough that we can make a twitch interface they are not a replacement for the traditional tools of mouse and keyboard.
What really pissed me off about the article was the insistence that these interfaces were a "direct manipulation" of images on screen. No, if you reach in and move an image (somehow) then that would be direct manipulation. If your physical gestures are translated into screen motion by accelerometers rather than a mouse then it is still an indirect interface. It is at most a minor increment on the user interface technology that we have already, the term "Paradigm Shift" is thrown about with abandon by too many suits without a understanding of what it implies.
Nah, the docking computer is for girls. It takes up a ton of space that is essential for loading up with consumer goods for the Sol to Barnard's Star cargo run. Definitely not worth it until you get a much bigger ship...
That's much more simple that what I thought you meant. It's a really nice idea, never come across anyone proposing it before but it would software a hell of a lot more robust. You should really spend some time pushing this scheme onto the right people because it would be a great deal for software reliability.
Erm WTF? The inherent difference that you are missing is that a domain name is an alias for an IP address whereas a Uniform Resource Locator is the address of a resource offered by a particular domain. Just because most URLs don't feature a path, contain only a domain name and resolve implicitly to the root of the namespace on the server does make them equivalent to domain names.
Given that the entire context of this discussion is that someone tried to hide a resource by not publishing the URL,and we can infer from this that the obscurity was in the path/filename part rather than the host, then you are talking complete shit.
It's not. If I understand what you mean properly then you're describing Sun's approach right through to the Niagra.
It sounds like a very interesting project. The idea that screams out from the wiki summary is static analysis and verification. There is a really good rundown in one of the wiki links. The really big difference from previous work is not just the use of managed code, but splitting the entire system into either trusted, or verified code. The trusted component is a tiny core, which they are working on verifying. The design of the rest of the kernel and the SIPs is a good one: instead of doing arbitrary verification, change the language design so that you can only write verifiable code. Then see how much of an O/S you can write. The progress is astounding.
For the IPC they have made some strange choices, receiving is synchronous (as in process calculi) but sending is asynchronous. As they are writing the lowest level parts (such as the schedular) in this code it may be an implementation difficulty with synchronous sends. The cheapness of the IPC seems to be routed in the transfer of ownership that communication implies. In essence you can't alias, you can only pass by value - but the low-level runtime can modify this to pass more efficiently by reference because it can verify there are no dangling references. This would (if it works over a large enough code base) solve the performance issue with IPC in a microkernel. It is (as another reply pointed out) similar to providing the semantics of heavy-weight communication to the programmer in a way that can be implemented with cheap co-routines.
Having done some (well, little) work in this area I'm really impressed by what they've achieved already.
How far back are you talking? I'd always heard the phrase rootkit as the code you downloaded after an exploit to hide your tracks and make sure that you keep root. Of course on windows getting root in the firtst place isn't such a big deal so some language drift is understandable...
What about: "Would you stop doing something that you weren't doing if somebody accused you of it with no real proof?
You are assuming that there is only one variable (resolution) that can be adjusted. Actually the quality of the scene is a function of two variables: resolution and scene complexity. When the complexity of scenes was low, rasterization produced much better results than raytracing for the same effort. Now that scene geometry has increased so much we are reaching the point where raytracing will produce the same (or better quality) for less effort. The main issue is that rasterization is O(n) in scene complexity while raytracing is O(log n). Of course there are lots of other issues and tradeoffs otherwise we would be using raytracing in games already.
If you're interested there is a detailed comparison available here.
"Unfortunately we're not as clever as those Intel chaps, how will we make it work?"
"Hmmm....."
(long pause)
"What about rendering really small scenes on a big stonking server and then using some sort of 'Network' to make the images appear?"
"That sounds like some kind of magic!"
Fantastic research.
If you are interested then there is a brilliant paper on the "Gold Effect"; propagation of errors within a body of literature. Unfortunately I can't find an online copy anywhere, but the details are "Journey to the centre of uncertainty", R. A. Lyttleton, in Science and Uncertainty ISBN 0-905927-96-6. He uses a series of interesting examples where the errors are now more commonplace than the theories they have displaced; the "iron core" of the earth (which I was taught in school), the existence of the Oort cloud (tenuous theory not backed up by observation) and others.
Very true, and it's odd but as I was posting the reply it did occur to me that we use different standards of behaviour in real life and in online forums.
I do understand what you were getting at, and the specific semantics of the term email aside it does make sense. I partly agree with your point, as it does seem absurd for the existence of certain information to be illegal. But I still partly disagree as within the context of the court case it did seem quite clear what was meant by the term.
No.
An email is not simply some text. It is the actual message that is sent containing the text. If I write a letter to somebody and save it on my harddrive it is not an email. But if I attach an SMTP header and transmit it then it becomes an email. These messages only became emails when they were sent, and so their very existence is infact illegal.
If you are going to be anally pedantic on the use of language then have the fucking decency to apply some actual thought to your argument.
Indeed they are, and after reading your somewhat familiar brain dump I am curious enough to ask you: Is your mental illness of the manic/psychotic persuasion, or is it a type of schizophrenia? Your writing shows lots of high-level understanding of the subject and parts of your cognitive processes are obviously still functional. But you are completely irrational and your gibberish shows that you have lost the ability to filter and discard faulty hypothesis, and as a result you are unable to see where your unique view of reality has diverged from those around you. Medication could really change your life for the better.
Yes, this being slashdot I realise that I didn't (skim-)read the paper properly, or in fact, err *cough*, your comment :) Two application for this spring to mind, one is the stolen laptop scenario that everyone has mentioned. If your laptop has TruCrypt for example then this is an interesting way around it that other approaches (with physical access) would have trouble doing. The other application (which I think they mention) is forensics. If the police break into a house and confiscate a running machine with an encrypted harddrive then this would seem to be the easiest approach to gain access to the data. Where this is weaker than other physical access approaches is the assumption that you're not going to give the hardware back (quickly)...
All of your comments are answered in the article that you haven't read.
In particular it is not simply a research issue - they have a fully automated attack on a usb flash drive.
Oh yes, I'd forgotten they were going to market it as a FPGA replacement. Although it can be used to replace a FPGA that's not such a good description of what it does. You're probably aware of Occam and the old transputer design. The Xmos chip is the logical sucessor to that. Point to point connections between processes are helped by hardware acceleration to provide to low-latency, low cost communications. Each core is multi-threaded, with multiple cores fabricated onto a single chip. I can't remember what interconnect they are using between chips, but internally it is very fast. Orders of magnitude faster to do MPI than on something like an x86. The reason that it can act as a FPGA replacement is because the new I/O is so lightweight and predictable.
I think they're still thrashing out what kind of programming model can exploit it the best, but expect something similar to the process model in Erlang. When David May has given presentations about the architecture he's mentioned that building a million-core machine is something they would like to do just to see what new applications it opens up. The challenges for the tool-chain, or the operating system are quite interesting when you hit that kind of scale. Even simple problems like how best to fork / replicate a thread across the entire machine become quite tricky.
AI is more of a hobby for me really. My real research is in compilers and languages, hence the interest in the low-level nitty gritty of how to implement neural networks effectively - all compilers need interesting test cases.
Depending on what type of NN model you are using, you may want to look at Xmos. Their processor is designed to scale up to the size of machine that you are talking about. I think that they are going to build a box like that just to see what it can do. There is no floating point unit on the chip - but you can do a lot with fixed-point math and a clever compiler.
We played around putting simple neural networks on a PIC-16f84. Much simpler than the normal discrete neurons that people use, let alone something biologically plausible. Communication between the them would have been a serious issue - but the Xmos processor is designed from the ground up to allow ligtweight communication on really large scale clusters. At some point I'm going to get around to seeing what the state-of-the-art is for biologically plausible neurons as the simulation code is probably a good test case for my new compiler.
If you get the cash from investors then you should think about contacting Xmos. They could probably build you a machine that isn't too far from Danny Hillis' vision.
From the reports that they've published I thought their simulation was quite realistic - but I'm a computer scientist rather than a neurobiologist. What are the major differences between their simulation and the real thing?
I hadn't come across that work before. They've made interesting progress in the past ten years. Although there aren't any simulations floating across the net they have reimplemented the network in silicon to drive their bio-buggy. Doesn't that count?
Extraordinary claims require extraordinary proof.
So that said, what type of virtual NN do you need to be supported? There is a local company coming of stealth mode at the moment that will provide the hardware that you say is necessary. I have some background in simulating NNs on simple hardware, and compiler design. What do you need?
Although I have mod points at the moment it seems that your comments have already been modded up anyway so I guess I'll reply. We've actually argued about this before (probably last time Kurzweil was mentioned on slashdot), but as you say in a different comment this discussion is always fun :)
Slight nitpick, currently we are at one billion transistors on a chip, not two, but that doesn't really change the point you are making.
A bigger issue that I have with what you've described is that simulating a brain is not the same as "solving" AI. The problem that Kurzweil has is that he refuses to accept that there is a difference. Sure, if they are the same then strong AI is inevitable and it's merely a question of building fast enough hardware. But why assume that they are the same thing?
Twenty years from now we may have hardware that can simulate an entire human brain; and yet we may be no closer to solving any of the problems in understanding how to solve the many problems in AI. The mental sleight-of-hand that Kurzweil applies to this argument is: Once we can simulate a brain we have AI, therefore the AI can design he next generation, therefore we will reach the singularity. This argument is a logical fallacy because it assumes being able to run the system, and knowing how to design the system / how the system works are equivalent.
Everything that we know complex and dynamic systems tells us that this is not so; given a simulation of the brain it is reasonable to assume that intelligence is the ultimate emergent property of the system. Understanding this property and how to refine it is the hardest problem that mankind has ever undertaken. Currently we don't really know how to pose the question, let alone how to arrive at an answer. To assume that some kind of standard engineering methodology will solve this in 20 years is wild speculation.
As always with AI, the hardware will be available but nobody yets knows if we can write the software to run on it.