IQ TESTS
IQ tests were designed to assess individual capabilities for performing highly specialized tasks/jobs in society. it is the outcrop of a very narrow definition of intelligence which ignores large a swath of many other capabilities of the human mind. I guess Shakespeare would score high on verbal fluency and low on numerical skills.
The tests are not bad in themselves as they test some real but narrow mental abilities, but the societal effects have been destructive. Apart from the fact that they make some people feel inferior to others, like artistic people will do poorly on some of the tests that are meant for the human calculators, the most destructive is that it gave us a totally wrong idea of what the word “intelligence” means. This has affected the way AI research has gone on since it’s inception.
I personally think that the reason why we don’t a have full AGI is not because of any lack of human ingenuity, I think problem statement is even more important than actually trying to solve “a problem”. In my opinion, we have been hacking for decades on the wrong problem definition, we have been trying to replicate a machine that is capable of doing really good on human IQ tests.
Alan Turing for all his contribution to cracking the Enigma code and the foundations of computer science, led the world in a very wrong direction when he proposed the Turing test, that simple proposition has led many people into the wrong lane trying to solve the wrong problem.
When I read a typical AI textbook, my mind always goes like: How different is this from normal computer Algorithms? All I see is computer algorithms! What were we meant to use computers for? To automate human routine tasks! So in order to automate some task like inspecting product quality in a production system, we develop a computer algorithm for doing that! It could involve computer vision system for seeing; a pattern recognition for detecting anomalies in the product; some expert system for inferring what kind of anomaly we are detecting, etc. All these sub-systems are usually what we define as Artificial intelligence but indeed they are computer algorithms designed to solve specific problems.
The intelligence that we seek to create is not these specific algorithms for achieving specific goals. Unfortunately, this is what we have been doing for decades now, and whether it is a new deep learning model or some new kind of probabilistic programming, etc. we are still looking to solve very narrow domains.
The intelligence we should be seeking to create is that which can actually invent deep learning and actually implement it when maybe proposed with a high-level task like invent an algorithm for recognizing images! And boom out comes deep learning neural network output as computer code that you could run.
This is the kind of intelligence we should be seeking to create, not SIRI or whatever talking bots are out there. I am not against developing talking bots nor deep learning not logic oriented systems, etc. what I am saying here is that we need to change our perspective on what we are trying to achieve, rather than trying to hack one aspect of human intelligence after the other we should be trying to build a generic system that can assist us with the solution to problems and this, as I have said in this book, will come from pure research into network structures.
This is what AGI could really look like, an oracle that can solve any human proposed problem! Trying to build programs that understand human emotions be it as it is expressed in text or in facial or bodily gestures is a good thing in the field of algorithms research but the problem of AGI is not really about making a simulacrum of an entire human being, which is what most current AGI research is focused on. We are supposed to be trying to build substrate free synthetic intelligence that we can apply to any task.
The AGI system could be asked a simple question like design a more economical jet engine! And it will do all the research that it can with all the information that is available online or in any repository and come up with a full CAD design of an engine that will be impossible for a human mind to conceive.
I am proposing a tool based intelligence that we can apply like a tool, no different from the way we would turn on the coffee maker to make us some coffee. This is a much more focused proposal than trying to create a God.
Back to the issue of IQ tests, the big issue with such tests is that it leads us in a very wrong path. Intelligence tests can help filter people that are not suited for particular tasks, for instance, people with high numerical literacy will be much suited to work in the financial markets and people with good spatial reasoning abilities will work better as pilots.
Most of our examinations in school are actually structured like IQ tests or in the worse case like actually memory championships. The difference between a typical examination and a full-on IQ tests is that the IQ test tries to be more generic and claims to be measuring “general intelligence”, which is actually wrong because what it is actually testing specific cognitive abilities.
Examinations will remain a filter that society uses to weed out those that are supposedly not suited for work in certain fields, but I doubt that Einstein had a very high IQ in the regular sense of the word. He was actually a slow learner and had trouble with mathematics till later years. It is reported that Hermann Weyl helped clean up his tensor maths that was required for some of his discovery in general relativity.
Einstein was no human calculator! He had an enormous ability to recognize patterns in vastly different sections of physics and also the synthetic ability to put together these large ideas together in the theory he invented. He stood on the shoulders of giants, feeding on work by Lorentz and others and integrating it and then synthesizing something new. That is why he came up with the grand ideas that he invented and revolutionized physics.
Now the question is, could we build a computer that actually reads all the physics papers, analyze all the experimental data out there and come up with some kind of Grand Unified Theory of the universe? In Theory, this is what an AGI should be capable of doing. It could even propose new experiments, to obtain new data, it could even detect bad experiments and reject the data, it could do the required mathematics and maybe in a few years of “thinking” through these information come up with novel theory after the other till it achieves the goal which it set out to do: Come up with a Grand Unified theory of all physics. We shouldn’t be surprised if it comes up with a contradictory claim like, there is no Grand Unified theory! And goes ahead to write a proof of why it can’t be possible. And we could say, keep searching or we keep searching whatever the case may be.
This is the kind of machine we should be aiming to build and had we not been misdirected by a wrong understanding of intelligence we would not be unconsciously trying to build a machine with a very high IQ which is typically called a “rational” agent.
IQ tests were designed to assess individual capabilities for performing highly specialized tasks/jobs in society. it is the outcrop of a very narrow definition of intelligence which ignores large a swath of many other capabilities of the human mind. I guess Shakespeare would score high on verbal fluency and low on numerical skills.
The tests are not bad in themselves as they test some real but narrow mental abilities, but the societal effects have been destructive. Apart from the fact that they make some people feel inferior to others, like artistic people will do poorly on some of the tests that are meant for the human calculators, the most destructive is that it gave us a totally wrong idea of what the word “intelligence” means. This has affected the way AI research has gone on since it’s inception.
I personally think that the reason why we don’t a have full AGI is not because of any lack of human ingenuity, I think problem statement is even more important than actually trying to solve “a problem”. In my opinion, we have been hacking for decades on the wrong problem definition, we have been trying to replicate a machine that is capable of doing really good on human IQ tests.
Alan Turing for all his contribution to cracking the Enigma code and the foundations of computer science, led the world in a very wrong direction when he proposed the Turing test, that simple proposition has led many people into the wrong lane trying to solve the wrong problem.
When I read a typical AI textbook, my mind always goes like: How different is this from normal computer Algorithms? All I see is computer algorithms! What were we meant to use computers for? To automate human routine tasks! So in order to automate some task like inspecting product quality in a production system, we develop a computer algorithm for doing that! It could involve computer vision system for seeing; a pattern recognition for detecting anomalies in the product; some expert system for inferring what kind of anomaly we are detecting, etc. All these sub-systems are usually what we define as Artificial intelligence but indeed they are computer algorithms designed to solve specific problems.
The intelligence that we seek to create is not these specific algorithms for achieving specific goals. Unfortunately, this is what we have been doing for decades now, and whether it is a new deep learning model or some new kind of probabilistic programming, etc. we are still looking to solve very narrow domains.
The intelligence we should be seeking to create is that which can actually invent deep learning and actually implement it when maybe proposed with a high-level task like invent an algorithm for recognizing images! And boom out comes deep learning neural network output as computer code that you could run.
This is the kind of intelligence we should be seeking to create, not SIRI or whatever talking bots are out there. I am not against developing talking bots nor deep learning not logic oriented systems, etc. what I am saying here is that we need to change our perspective on what we are trying to achieve, rather than trying to hack one aspect of human intelligence after the other we should be trying to build a generic system that can assist us with the solution to problems and this, as I have said in this book, will come from pure research into network structures.
This is what AGI could really look like, an oracle that can solve any human proposed problem! Trying to build programs that understand human emotions be it as it is expressed in text or in facial or bodily gestures is a good thing in the field of algorithms research but the problem of AGI is not really about making a simulacrum of an entire human being, which is what most current AGI research is focused on. We are supposed to be trying to build substrate free synthetic intelligence that we can apply to any task.
The AGI system could be asked a simple question like design a more economical jet engine! And it will do all the research that it can with all the information that is available online or in any repository and come up with a full CAD design of an engine that will be impossible for a human mind to conceive.
I am proposing a tool based intelligence that we can apply like a tool, no different from the way we would turn on the coffee maker to make us some coffee. This is a much more focused proposal than trying to create a God.
Back to the issue of IQ tests, the big issue with such tests is that it leads us in a very wrong path. Intelligence tests can help filter people that are not suited for particular tasks, for instance, people with high numerical literacy will be much suited to work in the financial markets and people with good spatial reasoning abilities will work better as pilots.
Most of our examinations in school are actually structured like IQ tests or in the worse case like actually memory championships. The difference between a typical examination and a full-on IQ tests is that the IQ test tries to be more generic and claims to be measuring “general intelligence”, which is actually wrong because what it is actually testing specific cognitive abilities.
Examinations will remain a filter that society uses to weed out those that are supposedly not suited for work in certain fields, but I doubt that Einstein had a very high IQ in the regular sense of the word. He was actually a slow learner and had trouble with mathematics till later years. It is reported that Hermann Weyl helped clean up his tensor maths that was required for some of his discovery in general relativity.
Einstein was no human calculator! He had an enormous ability to recognize patterns in vastly different sections of physics and also the synthetic ability to put together these large ideas together in the theory he invented. He stood on the shoulders of giants, feeding on work by Lorentz and others and integrating it and then synthesizing something new. That is why he came up with the grand ideas that he invented and revolutionized physics.
Now the question is, could we build a computer that actually reads all the physics papers, analyze all the experimental data out there and come up with some kind of Grand Unified Theory of the universe? In Theory, this is what an AGI should be capable of doing. It could even propose new experiments, to obtain new data, it could even detect bad experiments and reject the data, it could do the required mathematics and maybe in a few years of “thinking” through these information come up with novel theory after the other till it achieves the goal which it set out to do: Come up with a Grand Unified theory of all physics. We shouldn’t be surprised if it comes up with a contradictory claim like, there is no Grand Unified theory! And goes ahead to write a proof of why it can’t be possible. And we could say, keep searching or we keep searching whatever the case may be.
This is the kind of machine we should be aiming to build and had we not been misdirected by a wrong understanding of intelligence we would not be unconsciously trying to build a machine with a very high IQ which is typically called a “rational” agent.
Comments
Post a Comment