Many people on this thread have noted how much better coverage is in Europe than in the US, though not much is said of why this might be. It's certainly partly a result of different goals and cost systems, but I think the biggest difference is one of population density. I am an American, though I'ved lived in Germany for about 7 years. When I first came to Europe I looked up the statistics for several different European countries. If memory serves, Belgium, Britain, Germany, Italy and the Netherlands each have population densities about 10 TIMES that of the US. This means for every square mile (or square kilometer), there are 10 times as many Europeans in those countries than in the US. For instance, Germany is only a little bigger than North and South Carolina put together in the US, but has some 90 million people. All of the US put together has about 270 million (if memory serves). France and Spain have somewhat lower densities but I believe they are still higher than the US. Probably most of Scandinavia is lower. What does this mean? It means it is relatively cheaper per-person (by as much as 10:1) to string up a cellular network with 100% coverage in Europe. It will take much longer and probably cost a lot more to do so in the US and the rest of North America. It would be interesting to correlate population density with cellular coverage.
Last night, I just finished reading one of the books mentioned in this link: Ray Kurzweil's, The Age of Spritual Machines. I have to say it is an exceptionally well-written book that I can definitely recommend, even though I don't fully agree with the author's conclusions. He gives a great overview of the current state of Artificial Intelligence (AI) research and practice and includes a lot of surprisingly useful paradigms for solving complex problems.
One of the central premises of the book is that Moore's Law (which says computing power doubles every 18-24 months for the same price) is part of a more general law that Kurzweil has traced backwards in time through analog computing back to the end of the 19th century. Apparently Moore's Law has not only been right since transistors were invented in the 1960s, but since 1900! He therefore assumes, quite reasonably IMHO, that technologies like DNA computing, quantum architectures, nanotube computing, etc. will continue the trend of Moore's law until at least 2100 (even though conventional transistor speeds will top out at around the year 2030).
One of the interesting corollaries to this general statement is that $2000 personal computers will have computing power equivalent to a human brain in about 25 years. The human brain has an aggregate capacity of 10^^14 computations per second. Since, today's $2000 PCs are around 1000 MIPS (or just under 10^^9 computations per second), computers 25 years from now will be 10,000 times faster (12.5 doublings in speed) according to Moore's law.
This means PC hardware will match human-brain computing hardware in around 2025. From this the author assumes further that this is about the time that computers will pass the Turing Test and have an "intelligence" close to human. 10 years later (in 2040), he expects a cheap computer to be 30 times smarter than an average human. So, if you believe his assumptions, it is easy to conclude that computer programs as a society will completely replace human society by 2100.
However, in my opinion, Kurzweil under-estimates the amount of time it will take to develop an executable simulation of the human brain. He assumes that we will be able to "scan" human brains into neural nets that execute on computers and that a functioning human brain could be bootstrapped into a machine brain virtually instantaneously with 2025 technology. This assumes MRI or some other technology acquires extremely high sensitivity - enough to scan the state of all neurons in the brain very quickly (destructively or non-destructively). I think this is a big leap of faith in technology. Based on the progress of software technology and the small numbers of AI researchers right now working on brain computability problems, I am very skeptical of this more "far out" prediction. On top of that, he assumes that such a "brain" template would be "searchable" and "well organized" (the knowledge encoded in the neurons could be easily catalogued and shared with other machine "intelligences"). This seems to contradict the widely held assumptions of both neural net theory and the "holographic" storage theory of brain information storage that knowledge is hard to pin-down and extract from a neural net or brain.
As for the idea that intelligent machine programs would also be "spiritual", as Kurzweil suggests, I think this is pretty far-fetched. He seems to over-anthropomorphize machine intelligences. (BTW, Kurzweil seems to define "spiritual" as being self-aware, concerned about personal welfare and primarily interested in research, constant sex and social relationships - which is a little differently than most people define the word "spiritual".;-) ) I think that even if such an intelligence originally used human thinking as a template, it would not necessarily follow that it will share human concerns or even be comprehensible to humans, as Kurzweil also assumes.
I don't really think we will see "true" machine intelligence for some time, because, although raw computing power follows exponential growth laws like Moore's Law, software development seems to follow a linear growth law, so I expect full human-level intelligence on machines to take longer than 25 years. (How much have Word and Excel and typical desktop OSes improved in the last 10 years? They have certainly not showed exponential improvements based on development time. In fact, judged by user utility, the growth law for software may be closer to "log t" rather than "t", where "t" is total aggregate development time.)
It may be that software improves so slowly that all that computing power 25 years from now will just be used to show more realistic 3D-multimedia advertisements for diet drinks, upcoming TV shows and get-rich schemes on my browser. In the meantime, I will be trying to get Word's slightly improved, but annoyingly-helpful grammar-checking wizard to stop automatically changing my sentences to say something I really don't want to say.;-)
I'd like to know the answer to this question, too. (BTW, this is my first Slashdot post -- sorry if I break some convention.;-) )
I've been looking for OSS speech recognition, text processing, and speech synthesis tools for quite some time now and haven't found anything that is royalty-free and subscription-free.
If I can't find it, maybe I'll try building it myself. I have been toying with the idea of starting a project to create a natural language toolkit. I envision a set of modules like the following:
Audio sampling
Speech to phoneme conversion (probably outputs IPA codes with extra duration and pitch information).
Multi-lingual dictionary (with spelling, pronunciation, language-neutral definition, etc.) project that is expandable to an arbitrary number of word senses and languages and dialects.
Software that matches phonemes to word senses in the dictionary. (Each sentence is mapped to a sequence of actions and actors in a multi-dimensional co-ordinate space.)
Software that matches written text to word senses.
Software that is able to ouput text or speech based on the internal language-neutral format.
Putting together these kinds of tools, things like bidirectional speech interfaces, real-time translators (on PDAs?) and a lot of other things are theoretically possible. (I would like to have my cell phone automatically translate any calls I get that aren't in German or English to English, automatically, for instance.) As an OSS project, it could make speech-enabled applications really cheap and ubiquitous.
I've been reading every book I have been able to find lately to understand the problems and solutions that have been tried. I think this is a good fit for an OSS project because the hardest parts (the dictionary and the grammar rule system) are so amenable to being done in parallel. I have done a few calculations and feel that the amount of work to do a project like this is in the hundreds of man-years to get hundreds of thousands of word senses and thousands of grammatical rules. This may have been what has put people off up to now.;-)
However, if the work were shared by 10-100 people, or even more, it could actually happen pretty quick. And any way, good interfaces may only need a few hundred words in certain specialized applications.
Does anyone have any idea how many people might be expected to contribute to a project like this? If it is on the order of hundreds, worldwide, then this could work. In fact, if it were thousands of contributors, it could gain enough critical mass to outpace other similar pure-commercial developments.
Slashdot Mirror
Many people on this thread have noted how much better coverage is in Europe than in the US, though not much is said of why this might be. It's certainly partly a result of different goals and cost systems, but I think the biggest difference is one of population density. I am an American, though I'ved lived in Germany for about 7 years. When I first came to Europe I looked up the statistics for several different European countries. If memory serves, Belgium, Britain, Germany, Italy and the Netherlands each have population densities about 10 TIMES that of the US. This means for every square mile (or square kilometer), there are 10 times as many Europeans in those countries than in the US. For instance, Germany is only a little bigger than North and South Carolina put together in the US, but has some 90 million people. All of the US put together has about 270 million (if memory serves). France and Spain have somewhat lower densities but I believe they are still higher than the US. Probably most of Scandinavia is lower. What does this mean? It means it is relatively cheaper per-person (by as much as 10:1) to string up a cellular network with 100% coverage in Europe. It will take much longer and probably cost a lot more to do so in the US and the rest of North America. It would be interesting to correlate population density with cellular coverage.
Last night, I just finished reading one of the books mentioned in this link: Ray Kurzweil's, The Age of Spritual Machines. I have to say it is an exceptionally well-written book that I can definitely recommend, even though I don't fully agree with the author's conclusions. He gives a great overview of the current state of Artificial Intelligence (AI) research and practice and includes a lot of surprisingly useful paradigms for solving complex problems.
One of the central premises of the book is that Moore's Law (which says computing power doubles every 18-24 months for the same price) is part of a more general law that Kurzweil has traced backwards in time through analog computing back to the end of the 19th century. Apparently Moore's Law has not only been right since transistors were invented in the 1960s, but since 1900! He therefore assumes, quite reasonably IMHO, that technologies like DNA computing, quantum architectures, nanotube computing, etc. will continue the trend of Moore's law until at least 2100 (even though conventional transistor speeds will top out at around the year 2030).
One of the interesting corollaries to this general statement is that $2000 personal computers will have computing power equivalent to a human brain in about 25 years. The human brain has an aggregate capacity of 10^^14 computations per second. Since, today's $2000 PCs are around 1000 MIPS (or just under 10^^9 computations per second), computers 25 years from now will be 10,000 times faster (12.5 doublings in speed) according to Moore's law.
This means PC hardware will match human-brain computing hardware in around 2025. From this the author assumes further that this is about the time that computers will pass the Turing Test and have an "intelligence" close to human. 10 years later (in 2040), he expects a cheap computer to be 30 times smarter than an average human. So, if you believe his assumptions, it is easy to conclude that computer programs as a society will completely replace human society by 2100.
However, in my opinion, Kurzweil under-estimates the amount of time it will take to develop an executable simulation of the human brain. He assumes that we will be able to "scan" human brains into neural nets that execute on computers and that a functioning human brain could be bootstrapped into a machine brain virtually instantaneously with 2025 technology. This assumes MRI or some other technology acquires extremely high sensitivity - enough to scan the state of all neurons in the brain very quickly (destructively or non-destructively). I think this is a big leap of faith in technology. Based on the progress of software technology and the small numbers of AI researchers right now working on brain computability problems, I am very skeptical of this more "far out" prediction. On top of that, he assumes that such a "brain" template would be "searchable" and "well organized" (the knowledge encoded in the neurons could be easily catalogued and shared with other machine "intelligences"). This seems to contradict the widely held assumptions of both neural net theory and the "holographic" storage theory of brain information storage that knowledge is hard to pin-down and extract from a neural net or brain.
As for the idea that intelligent machine programs would also be "spiritual", as Kurzweil suggests, I think this is pretty far-fetched. He seems to over-anthropomorphize machine intelligences. (BTW, Kurzweil seems to define "spiritual" as being self-aware, concerned about personal welfare and primarily interested in research, constant sex and social relationships - which is a little differently than most people define the word "spiritual". ;-) ) I think that even if such an intelligence originally used human thinking as a template, it would not necessarily follow that it will share human concerns or even be comprehensible to humans, as Kurzweil also assumes.
I don't really think we will see "true" machine intelligence for some time, because, although raw computing power follows exponential growth laws like Moore's Law, software development seems to follow a linear growth law, so I expect full human-level intelligence on machines to take longer than 25 years. (How much have Word and Excel and typical desktop OSes improved in the last 10 years? They have certainly not showed exponential improvements based on development time. In fact, judged by user utility, the growth law for software may be closer to "log t" rather than "t", where "t" is total aggregate development time.)
It may be that software improves so slowly that all that computing power 25 years from now will just be used to show more realistic 3D-multimedia advertisements for diet drinks, upcoming TV shows and get-rich schemes on my browser. In the meantime, I will be trying to get Word's slightly improved, but annoyingly-helpful grammar-checking wizard to stop automatically changing my sentences to say something I really don't want to say. ;-)
I'd like to know the answer to this question, too. (BTW, this is my first Slashdot post -- sorry if I break some convention. ;-) )
I've been looking for OSS speech recognition, text processing, and speech synthesis tools for quite some time now and haven't found anything that is royalty-free and subscription-free.
If I can't find it, maybe I'll try building it myself. I have been toying with the idea of starting a project to create a natural language toolkit. I envision a set of modules like the following:
Putting together these kinds of tools, things like bidirectional speech interfaces, real-time translators (on PDAs?) and a lot of other things are theoretically possible. (I would like to have my cell phone automatically translate any calls I get that aren't in German or English to English, automatically, for instance.) As an OSS project, it could make speech-enabled applications really cheap and ubiquitous.
I've been reading every book I have been able to find lately to understand the problems and solutions that have been tried. I think this is a good fit for an OSS project because the hardest parts (the dictionary and the grammar rule system) are so amenable to being done in parallel. I have done a few calculations and feel that the amount of work to do a project like this is in the hundreds of man-years to get hundreds of thousands of word senses and thousands of grammatical rules. This may have been what has put people off up to now. ;-)
However, if the work were shared by 10-100 people, or even more, it could actually happen pretty quick. And any way, good interfaces may only need a few hundred words in certain specialized applications.
Does anyone have any idea how many people might be expected to contribute to a project like this? If it is on the order of hundreds, worldwide, then this could work. In fact, if it were thousands of contributors, it could gain enough critical mass to outpace other similar pure-commercial developments.
Comments?