That's a really good description of an optimal reward-schedule for addiction. Of course, most (all?) people can't distinguish between addiction and fun, hence the huge popularity.
I tried to find a decent description of this on the web (I remember reading an old analysis of how to optimise the payback in slot machines that went into reward schedules) but failed. This is the closest that I could find. The main point it makes is that tedium is essential to addiction. It serves to highlight the non-tedious bits and space out the rewards randomly. Nice to hear a personal, non-clinical, description of it for a change.
Damn. My memory is terrible today. Ignore me, I am talking absolute crap. The attack that I was thinking of is exactly the one that you describe. There isn't a pre-image attack against md5 yet, only a collision attack. My apologies, I've just made a significant dent in the signal-to-noise ratio around these parts.
Good point, I had of course forgotten that the blocks are constant size. That would require a much better attack than those currently available.
You are aware that there is an attack for MD5 when the length isn't specified though? There is a demo that will produce forged pdf documents with a given md5 hash.
Yeah I saw that on the link that drinkypoo provided. That isn't really a whole answer though, as sha1 should be broken fairly soon. The other reply from strstrep seems more informative, even though he says that he's just guessing. A higher level signature on each package would provide much better security.
I didn't ask you to browse the web for me. In a discussion in which you've already posted four or five messages I assumed that you knew something about apt-p2p. Perhaps you didn't understand, but I was asking you *if* you knew how the security on that worked. Blithely doing a search and presenting useless information isn't helpful, which is why I didn't do that myself.
FYI the answer isn't on the page that you found, but it is something that would require either a more detailed search *or* background information. My mistake, I assumed that you knew what you were talking about.
How do they handle security? If I let the world leech from me then I can seed whatever kind of files that I want through the swarm. Please tell me there is something a bit more secure than a broken (md5) or a nearly broken (sha1) hash stopping this...
I see what you mean. You're talking about decoupling the other side. Would that be similar to an intermediate representation in a compiler? I guess that open source is a good way of performing that decoupling, although not for companies who want to try and keep their implementations secret.
There was an approach to compiler design back in the 70s that tried to map source into a low-level intermediate that worked well on very different architecture back-ends. It ultimately failed because it is hard to find one representation that works well in so many different contexts.
But research has come a long way since then, and perhaps it would be interesting to revisit the question using SSA-form or PEGs.
Originally x86 was the set of primitive hardware operations, but the decoupling that you describe was done a long time ago. Now the internal format of x86 ISAs is a sort of hybrid RISC approach. The details are not published although people have reverse engineered a lot of information about micro-ops, in particular the throughput that they have on different ISAs. So now x86 is just the persistent representation for binaries as you suggest, internally there is a translation to the hardware operations. Probably this translation is static rather than Transmeta's dynamic mapping, but Intel are still free to redesign the internals each generation.
It's actually worse than that. If you have a long pipeline it requires more exploitable parallelism in the code to get the same performance. Each instruction has a result latency proportional to the length of the pipe, as well as a throughput proportional to the number of parallel units. Long pipes increase the latency so that any dependent chains of instructions in the code take longer to execute.
If we didn't use the x86 instruction set we'd have to invent something quite similar to it. There are several layers of cruft in there: 2-op instructions and 100s of rarely used complex ops. But it does something that modern 3-op RISC-like ISAs fail to do - it compresses executable code. This means more sits in the instruction cache, and so less decode penalties.
The newer ISAs (Core-2 and i7) do an amazing job of extracting parallelism to boost instruction throughput. During the Mhz race years instruction latencies spiked on the P4 ISA, the more recent ISAs have really lowered those latencies to RISC-like levels.
I don't know if you've ever spent time doing low-level profiling on an Intel platform, but they do make good hardware. Here's a rule that they don't bandy around, but if your working set fits inside L3 then all of your memory accesses are free. Yes - free. I can see why they wouldn't want to make that guarantee part of the ISA, but it does mean that register pressure (the biggest problem in x86) is a phantom issue. Doing low-level paging between registers and the heap (which is what a modern compiler does with load/spills) provides a huge working set.
The article mentions efficiency - but there are plenty of situation where power consumption is irrelevant and I really do want maximum performance per thread. Any home desktop falls into this category. As any server or laptop doesn't it will be interesting to see what happens as Intel shifts from maximising performance to maximising performance per watt.
Re:If Intel are smart they will mix Core and Larab
on
Larrabee ISA Revealed
·
· Score: 1
Awww how sweet. Henk has registered a name troll just for me. Poor guy, that's a lot of issues for such a sweet child to carry around.
I love it Henk, but have you got anything that actually gets to me?
Re:If Intel are smart they will mix Core and Larab
on
Larrabee ISA Revealed
·
· Score: 2, Interesting
Oddly enough your post ranks quite highly in that search. Drilling through the forums that show up reveal speculation that a 32-core Larrabee design will use 300W TDP, or roughly 10W per core. There doesn't seem to be any justification for that number although the Larrabee looks like Atom + stonking huge vector array. The Atom only uses 2W, it seems hard to believe that the 16-way vector array would use as much power for each FLOP as the entire Atom power budget to deliver that FLOP. Or perhaps it will, it's all just speculation at this point.
So that 32-core processor would deliver 16x32 = 512 FLOP/clock peak. I would guess that they could deliver a low-power part clocked at 1GHz judging by the efficiency of Intel's floating point units across the whole range (from Atom up to i7). That part would hit 512GFlop/s peak. Then it's just a guessing game of what clock-speed they could ramp it up to within that 300W TDP, 2Ghz? 3?
The real killer could be how much sustained throughput can be achieved on an x86 derivative. The Core-2 sustained throughputs were mental, but it used every OoO trick that Intel could throw at it. Without that advantage the peak:sustained ratio will be closer to AMD/Nvidia's current offerings.
I think that you should actually read the material associated with the Robot Scientist project rather than just assume it is hard-wired pattern recognition. You will be surprised. The system is programmed to search for interesting data - the definition of interesting is not hardwired, and their main result is discovery of novel research in the target domain.
A targeting system is not doing anything similar. It is simply moving towards a predefined goal. The point of this project was to avoid defining the goal, and to allow the machine to infer it for itself.
No, ignoring the fact that the "thought experiment" proposed is gibberish, there is a more serious problem with your argument. According to the article you link to infinitely fast communication causes causality violations. If the communication was merely faster than light, but not instantaneous then the problem wouldn't exist. So a universe with FTL travel and FTL communication would be fine, but one with FTL travel and instantaneous communication would not. Hence, it is not the FTL travel that causes the paradox.
Any neuron, be it natural or artificial, hardware or software or wetware, performs a very simple task: what mathematicians call an "internal product", or "dot product".
Err, they've achieved a system that can make up its own biology experiments and discover new results at the cutting edge of that field. Are you not capable of remembering what the discussion is about over three whole posts?
Point to a neural network achieves anything similar. Just one.
You've just taken every misconception about AI and applied them badly. I'll be gentle, but here is a list of things to remember for next time:
1. Whenever you see a comparison - what is it between? If it's unclear then the comparison is bollocks. 2. Simple neurons only perform simple stats. 3. Black-box methods are no good for EBL. 4. You cannot take (non-comparable) performance figures and use it to shoot down research in the same area.
The more complex explanation: fast artifical neurons use a less complex evaluation function than slow neurons. The problem is that less complex evaluation functions are less useful. When somebody trumpets the speed of their neurons it means that they are drawing attention away from other issues - ie what are the connectivity limits on networks, how complex is the evaluation function, what problems have these types of neurons been shown to be able to solve?
In the worst case, single layers of threshold neurons merely learn hyperplane partitions of datasets. This breaks down when the relationships between the variables becomes more complex (see for example kernel methods into learning non-linear functions), or the data is non-numeric. Neural networks don't simply allow you to "throw more layers" at the problem, because performance (in terms of correctness) is highly dependent on network topology. Having a huge collection of crap neurons is not always a win.
The purpose of King's research is EBL (Explanation Based Learning). When Adam discovers something new - the discovery is not just some statistical property of some numerical data. Instead is a logical chain of hypotheses that explain how the new result diverges from the current set of knowledge. This is not possible with a black-box method such as neural networks.
As to your overall claim - I can build a hardware circuit that evaluate any decision 10000000 times than the hardware neural nets that you point to. Does this mean that all other research is pointless and we should focus on this? No, because the accuracy of my circuit is zero. There is no evidence that (artificial) neural nets can be applied to this problem, so any performance comparisons are ludicrous and meaningless.
Awwwww. I've got mod points but I've already posted above. That really made me chuckle.
Your description above of guessing and stats is a really good non-technical description of how the system works. The first step is to actually analyse what is already known about the problem domain, then some guesswork is applied about how to improve that knowledge. The nice thing about King's work is those guesses translate directly into automated experiments, and then the system can close the loop - the results can be automatically analysed and integrated with the pre-existing knowledge in the system.
I saw King give a talk about the system last year and it is really impressive work. It looks like the first tentative steps towards building a functioning AI for a non-trivial domain. I haven't read the article so I don't know if they tossed in the future plans with the lasers, but that is just too cool for words.
Slashdot Mirror
It would still be nice just to wipe it out at the source of course.
So you advocate... nuking it from orbit?
That's a really good description of an optimal reward-schedule for addiction. Of course, most (all?) people can't distinguish between addiction and fun, hence the huge popularity.
I tried to find a decent description of this on the web (I remember reading an old analysis of how to optimise the payback in slot machines that went into reward schedules) but failed. This is the closest that I could find. The main point it makes is that tedium is essential to addiction. It serves to highlight the non-tedious bits and space out the rewards randomly. Nice to hear a personal, non-clinical, description of it for a change.
Of course you get the best of both worlds with ssh-agent. Secure passphrase, and no keyboard input for scripts.
Damn. My memory is terrible today. Ignore me, I am talking absolute crap. The attack that I was thinking of is exactly the one that you describe. There isn't a pre-image attack against md5 yet, only a collision attack. My apologies, I've just made a significant dent in the signal-to-noise ratio around these parts.
Good point, I had of course forgotten that the blocks are constant size. That would require a much better attack than those currently available.
You are aware that there is an attack for MD5 when the length isn't specified though? There is a demo that will produce forged pdf documents with a given md5 hash.
Yeah I saw that on the link that drinkypoo provided. That isn't really a whole answer though, as sha1 should be broken fairly soon. The other reply from strstrep seems more informative, even though he says that he's just guessing. A higher level signature on each package would provide much better security.
I didn't ask you to browse the web for me. In a discussion in which you've already posted four or five messages I assumed that you knew something about apt-p2p. Perhaps you didn't understand, but I was asking you *if* you knew how the security on that worked. Blithely doing a search and presenting useless information isn't helpful, which is why I didn't do that myself.
FYI the answer isn't on the page that you found, but it is something that would require either a more detailed search *or* background information. My mistake, I assumed that you knew what you were talking about.
How do they handle security? If I let the world leech from me then I can seed whatever kind of files that I want through the swarm. Please tell me there is something a bit more secure than a broken (md5) or a nearly broken (sha1) hash stopping this...
Don't know why you got moderated troll, looks like quite a relevant link. Anyone with mod points want to cancel out the abuse?
I see what you mean. You're talking about decoupling the other side. Would that be similar to an intermediate representation in a compiler? I guess that open source is a good way of performing that decoupling, although not for companies who want to try and keep their implementations secret.
There was an approach to compiler design back in the 70s that tried to map source into a low-level intermediate that worked well on very different architecture back-ends. It ultimately failed because it is hard to find one representation that works well in so many different contexts.
But research has come a long way since then, and perhaps it would be interesting to revisit the question using SSA-form or PEGs.
Originally x86 was the set of primitive hardware operations, but the decoupling that you describe was done a long time ago. Now the internal format of x86 ISAs is a sort of hybrid RISC approach. The details are not published although people have reverse engineered a lot of information about micro-ops, in particular the throughput that they have on different ISAs. So now x86 is just the persistent representation for binaries as you suggest, internally there is a translation to the hardware operations. Probably this translation is static rather than Transmeta's dynamic mapping, but Intel are still free to redesign the internals each generation.
It's actually worse than that. If you have a long pipeline it requires more exploitable parallelism in the code to get the same performance. Each instruction has a result latency proportional to the length of the pipe, as well as a throughput proportional to the number of parallel units. Long pipes increase the latency so that any dependent chains of instructions in the code take longer to execute.
If we didn't use the x86 instruction set we'd have to invent something quite similar to it. There are several layers of cruft in there: 2-op instructions and 100s of rarely used complex ops. But it does something that modern 3-op RISC-like ISAs fail to do - it compresses executable code. This means more sits in the instruction cache, and so less decode penalties.
The newer ISAs (Core-2 and i7) do an amazing job of extracting parallelism to boost instruction throughput. During the Mhz race years instruction latencies spiked on the P4 ISA, the more recent ISAs have really lowered those latencies to RISC-like levels.
I don't know if you've ever spent time doing low-level profiling on an Intel platform, but they do make good hardware. Here's a rule that they don't bandy around, but if your working set fits inside L3 then all of your memory accesses are free. Yes - free. I can see why they wouldn't want to make that guarantee part of the ISA, but it does mean that register pressure (the biggest problem in x86) is a phantom issue. Doing low-level paging between registers and the heap (which is what a modern compiler does with load/spills) provides a huge working set.
The article mentions efficiency - but there are plenty of situation where power consumption is irrelevant and I really do want maximum performance per thread. Any home desktop falls into this category. As any server or laptop doesn't it will be interesting to see what happens as Intel shifts from maximising performance to maximising performance per watt.
Awww how sweet. Henk has registered a name troll just for me. Poor guy, that's a lot of issues for such a sweet child to carry around.
I love it Henk, but have you got anything that actually gets to me?
Oddly enough your post ranks quite highly in that search. Drilling through the forums that show up reveal speculation that a 32-core Larrabee design will use 300W TDP, or roughly 10W per core. There doesn't seem to be any justification for that number although the Larrabee looks like Atom + stonking huge vector array. The Atom only uses 2W, it seems hard to believe that the 16-way vector array would use as much power for each FLOP as the entire Atom power budget to deliver that FLOP. Or perhaps it will, it's all just speculation at this point.
So that 32-core processor would deliver 16x32 = 512 FLOP/clock peak. I would guess that they could deliver a low-power part clocked at 1GHz judging by the efficiency of Intel's floating point units across the whole range (from Atom up to i7). That part would hit 512GFlop/s peak. Then it's just a guessing game of what clock-speed they could ramp it up to within that 300W TDP, 2Ghz? 3?
The real killer could be how much sustained throughput can be achieved on an x86 derivative. The Core-2 sustained throughputs were mental, but it used every OoO trick that Intel could throw at it. Without that advantage the peak:sustained ratio will be closer to AMD/Nvidia's current offerings.
I think that you should actually read the material associated with the Robot Scientist project rather than just assume it is hard-wired pattern recognition. You will be surprised. The system is programmed to search for interesting data - the definition of interesting is not hardwired, and their main result is discovery of novel research in the target domain.
A targeting system is not doing anything similar. It is simply moving towards a predefined goal. The point of this project was to avoid defining the goal, and to allow the machine to infer it for itself.
I don't think that requires an AI... reading most patents makes it look like an automatic word generator would be enough :)
Their next plans were to speed up the experiment cycle using a high-powered laser to do the chemical analysis.
Amusing. In a discussion about AI you fail the Turing Test. I'm impressed.
Are you having a little trouble admitting that you were completely wrong? Don't worry, I'm sure life will give you the practise you need.
No, ignoring the fact that the "thought experiment" proposed is gibberish, there is a more serious problem with your argument. According to the article you link to infinitely fast communication causes causality violations. If the communication was merely faster than light, but not instantaneous then the problem wouldn't exist. So a universe with FTL travel and FTL communication would be fine, but one with FTL travel and instantaneous communication would not. Hence, it is not the FTL travel that causes the paradox.
This is completely and utterly wrong. Take a look at this basic starting point for example.
Err, they've achieved a system that can make up its own biology experiments and discover new results at the cutting edge of that field. Are you not capable of remembering what the discussion is about over three whole posts?
Point to a neural network achieves anything similar. Just one.
No. No. No.
You've just taken every misconception about AI and applied them badly. I'll be gentle, but here is a list of things to remember for next time:
1. Whenever you see a comparison - what is it between? If it's unclear then the comparison is bollocks.
2. Simple neurons only perform simple stats.
3. Black-box methods are no good for EBL.
4. You cannot take (non-comparable) performance figures and use it to shoot down research in the same area.
The more complex explanation: fast artifical neurons use a less complex evaluation function than slow neurons. The problem is that less complex evaluation functions are less useful. When somebody trumpets the speed of their neurons it means that they are drawing attention away from other issues - ie what are the connectivity limits on networks, how complex is the evaluation function, what problems have these types of neurons been shown to be able to solve?
In the worst case, single layers of threshold neurons merely learn hyperplane partitions of datasets. This breaks down when the relationships between the variables becomes more complex (see for example kernel methods into learning non-linear functions), or the data is non-numeric. Neural networks don't simply allow you to "throw more layers" at the problem, because performance (in terms of correctness) is highly dependent on network topology. Having a huge collection of crap neurons is not always a win.
The purpose of King's research is EBL (Explanation Based Learning). When Adam discovers something new - the discovery is not just some statistical property of some numerical data. Instead is a logical chain of hypotheses that explain how the new result diverges from the current set of knowledge. This is not possible with a black-box method such as neural networks.
As to your overall claim - I can build a hardware circuit that evaluate any decision 10000000 times than the hardware neural nets that you point to. Does this mean that all other research is pointless and we should focus on this? No, because the accuracy of my circuit is zero. There is no evidence that (artificial) neural nets can be applied to this problem, so any performance comparisons are ludicrous and meaningless.
Awwwww. I've got mod points but I've already posted above. That really made me chuckle.
Your description above of guessing and stats is a really good non-technical description of how the system works. The first step is to actually analyse what is already known about the problem domain, then some guesswork is applied about how to improve that knowledge. The nice thing about King's work is those guesses translate directly into automated experiments, and then the system can close the loop - the results can be automatically analysed and integrated with the pre-existing knowledge in the system.
I saw King give a talk about the system last year and it is really impressive work. It looks like the first tentative steps towards building a functioning AI for a non-trivial domain. I haven't read the article so I don't know if they tossed in the future plans with the lasers, but that is just too cool for words.
Ah, so you're applying for a grant to research Artificial Stupidity?