So you want me to talk about matrix transformations, word vectors, the softmax function, multi-head attention mechanism implementation details and so forth because that will make this conversation somehow more informed? That's like saying to argue that cars can go faster than humans need to talk about engine parts and their limitations. It's completely irrelevant. We can observe cars go faster than humans. Similarly we can observe various computer programs do tasks that were previously only possible by human understanding, judgement and creativity and thus we decided to call that "artificial intelligence". As opposed to natural/genuine intelligence that comes from a mind, but something that to an outside observer is indistinguishable from a mind.
Your only contribution to this conversation is finding excuses to discuss anything other than the actual arguments.
So you want me to talk about matrix transformations, word vectors, the softmax function, multi-head attention mechanism implementation details and so forth
Kind of, but on much lower level. About basics. Say, about actual code of sigmoids in LSTM cell, and why this code can't do anything except what it was programmed to do.
When you use CNC, f.e., you could program extremely complex actions, including acoounting for tool wear, backlashes and up to replacing broken tool. It could make things better than a human, but only in spheres, where intelligence, conscioness, sentinentness are hindering the job. CNC could do two identical parts better than human turner, so CNC will do the job of making identical parts better. This does not bring any intelligence to CNC, on the opposite, the whole point of automatisation is to remove any intelligence from turner task. And this does not mean you don't need a skilled turner anymore, he just become a programmer who have to use all his skills and intelligence to create a program for CNC.
ANNs is no different. It could do only what it was programmed for. Program execution is not an intelligence, never was and never would be, regardless of how complex it is. Even if you connect billions of subroutines and use billions of variables, this will not make something intelligent from unintelligent pieces. It will still be just a program, totally predictable and fully controlled by author. It is nor intelligent, nor smart, nor whatever marketing label you will try to apply to this dumb thing. If something could be done better using ANN, it just means that no intelligence in doing this task is necessary. But you still need intelligence to write and train ANN.
Things that could potentially break that ANNs basic property - complete absence of any intelligence, are programming bugs, f.e. All that buffer overflows, undefined compiler behaviour and so on. Unpredictable things, that could occasionally lead to self-modifying code, unpredictable branching, cross-function gotos and all that stuff that theoretically could be a base for appearing real intelligence. But all of that things are thorougly removed from any programming learning course, and the only people who still attempt to use them literally declared criminals with real and long jailtime. Many of that basic things mitigated directly in hardware. Each time when you see "Segmentation fault" or "Buffer overflow" - it is yet another thing that with very low probability could become a start of long sequence of events of appearing real AI was cut at the beginning by hardware. Moreover, there is whole industry that produce software to hunt and delete any code that theoretically could sometime lead to appearence of real, true AI. And it is installed on almost every computer that use OS made by corporations. And it is the same corporations that invest huge money in AI® today.
And you are on the side of that corporations, supporting their narrative that their AI® is somehow have something to do with intelligence, and at the same time destroy anything that could sometime become real AI. Congratulations.
Of course every part of an AI's code is only doing what it was programmed to do. But there are several reasons why this is of no relevance to whether the AI is artificially intelligent.
One reason is that there's nothing in the definition of artificial intelligence that says the intelligence must be doing things it wasn't programmed to do. Another reason is that training data are not commands or lines of code for the AI to perform and do not actually program the AI - all they do is affect weights or other values in a system already programmed. Then the AI is capable of repeating patterns from its training data without actually being programmed to produce those patterns.
But suppose I were to let you redefine AI as needing to do things it wasn't programmed to do and also to consider training data as instructions that form part of an Ai's programming. Even then there's still the fact that although the components of the AI are only doing as programmed, they interact in a way that leads to behaviors which weren't specifically considered by the programmer, weren't specifically in the training data and which the programmer could not do himself even if he were to read all the training data. For example you can have an AI chatbot give a reasonable response to a question that never appeared in its training data. In that case the AI is clearly not doing something it was told how to do.
It was never told how to answer this question other than to encode the input in a certain way and feed it through the neural network (or whatever system it uses). You could say this is in a sense being told how to respond, but it's not being told to give an output that was ever conceived of by the programmer or those who made the training data. Nor would they have conceived of this response if they had thought about the same question. This is how AI is different from other computer systems. It synthesizes things in ways that its creators didn't and couldn't have conceived of. In this way it is able to output meaningful responses that are only indirectly related to the thoughts of its creators. Other computer systems just output things that are directly related to the thoughts of their creators.
Your talk of bugs introducing intelligence makes no sense to me. If a computer followed a program with bugs in it it would still be doing as its programmed, which according to you means it can't be intelligent. But anyway why would you want computers to become intelligent like humans? That's a death sentence for humanity. Your claim that I'm supporting corporate narratives makes no sense either because my whole point is that AI development needs to be stopped because it's anti-human.
You still don't want to understand that "artificial intelligence" is a trademark, just like Coca-Cola. It can't have definition because it is a trademark.
Intelligence have one property, that never will be allowed in AI® (AI™ if you wish) - unpredictability. This is what exorcised from programming long ago.
Back-propagation is not some magic, it is nothing more than yet another piece of code that take training data and produce weights data through predetermined algorithm. Combining this two pieces of code - back-propagation and forward-propagation does not make this two dumb and predetermined pieces of code in any mean intelligent. Even passing outputs to inputs as in RNN, still adds nothing that could create intelligence. Predetermined calculation of weights from input and output and then predetermined calculation of output using weights and input is a simple arithmetic. Both parts are programmed and intended. ANN do exactly what it is programmed to do. Calculate weights through back-propagation, use weights to calculate output during forward-propagation. Nothing less, nothing more.
For example you can have an AI chatbot give a reasonable response to a question that never appeared in its training data.
It will calculate response using weights that was calculated using training data. This have absolutely nothing to do with reasonability. When you force AI to give an answer it was never trained on, it does not invent this answer, you just get an average/combination of answers that most close by weights to what you ask. Nothing more, nothing less. Just arithmetic. Answer could look reasonable in some circumstances (not because AI is able to reason, but because of coincidence), but with high probability answer will be either wrong, either senseless.
If a computer followed a program with bugs in it it would still be doing as its programmed
The thing with buggy program is that it will not do what it is programmed for. This is a first, from many, but essential elements that could possibly lead to creation of real AI. The thing is that bug is what makes computer do things it was never programmed to do. In no way intelligence yet, but an essential brick of artificial intelligence.
But anyway why would you want computers to become intelligent like humans?
Computer AI will never become intelligent like humans. Because AI will appear and develop in completely different environment, that have very little with human environment. With high probability, it will be so different, that we will have big problems to even establish contact with it, not even talking about some useful communication. It will be so different because errors, uncertainities, unpredictabilities, etc. artificial intelligence will develop on and their sources are completely different from those of human intelligence.
Intelligence could not be programmed. It have to develop itself in an environment that allow self-development, errors, uncertainities and all that stuff.
It is possible, that computer AI is impossible at all, but this could not be proven. I don;t know if computer AI will ever be created, but I'm 100% know that it will never be created with technology used for AI®(AI™)
So you want me to talk about matrix transformations, word vectors, the softmax function, multi-head attention mechanism implementation details and so forth because that will make this conversation somehow more informed? That's like saying to argue that cars can go faster than humans need to talk about engine parts and their limitations. It's completely irrelevant. We can observe cars go faster than humans. Similarly we can observe various computer programs do tasks that were previously only possible by human understanding, judgement and creativity and thus we decided to call that "artificial intelligence". As opposed to natural/genuine intelligence that comes from a mind, but something that to an outside observer is indistinguishable from a mind.
Your only contribution to this conversation is finding excuses to discuss anything other than the actual arguments.
Kind of, but on much lower level. About basics. Say, about actual code of sigmoids in LSTM cell, and why this code can't do anything except what it was programmed to do.
When you use CNC, f.e., you could program extremely complex actions, including acoounting for tool wear, backlashes and up to replacing broken tool. It could make things better than a human, but only in spheres, where intelligence, conscioness, sentinentness are hindering the job. CNC could do two identical parts better than human turner, so CNC will do the job of making identical parts better. This does not bring any intelligence to CNC, on the opposite, the whole point of automatisation is to remove any intelligence from turner task. And this does not mean you don't need a skilled turner anymore, he just become a programmer who have to use all his skills and intelligence to create a program for CNC.
ANNs is no different. It could do only what it was programmed for. Program execution is not an intelligence, never was and never would be, regardless of how complex it is. Even if you connect billions of subroutines and use billions of variables, this will not make something intelligent from unintelligent pieces. It will still be just a program, totally predictable and fully controlled by author. It is nor intelligent, nor smart, nor whatever marketing label you will try to apply to this dumb thing. If something could be done better using ANN, it just means that no intelligence in doing this task is necessary. But you still need intelligence to write and train ANN.
Things that could potentially break that ANNs basic property - complete absence of any intelligence, are programming bugs, f.e. All that buffer overflows, undefined compiler behaviour and so on. Unpredictable things, that could occasionally lead to self-modifying code, unpredictable branching, cross-function gotos and all that stuff that theoretically could be a base for appearing real intelligence. But all of that things are thorougly removed from any programming learning course, and the only people who still attempt to use them literally declared criminals with real and long jailtime. Many of that basic things mitigated directly in hardware. Each time when you see "Segmentation fault" or "Buffer overflow" - it is yet another thing that with very low probability could become a start of long sequence of events of appearing real AI was cut at the beginning by hardware. Moreover, there is whole industry that produce software to hunt and delete any code that theoretically could sometime lead to appearence of real, true AI. And it is installed on almost every computer that use OS made by corporations. And it is the same corporations that invest huge money in AI® today.
And you are on the side of that corporations, supporting their narrative that their AI® is somehow have something to do with intelligence, and at the same time destroy anything that could sometime become real AI. Congratulations.
Of course every part of an AI's code is only doing what it was programmed to do. But there are several reasons why this is of no relevance to whether the AI is artificially intelligent.
One reason is that there's nothing in the definition of artificial intelligence that says the intelligence must be doing things it wasn't programmed to do. Another reason is that training data are not commands or lines of code for the AI to perform and do not actually program the AI - all they do is affect weights or other values in a system already programmed. Then the AI is capable of repeating patterns from its training data without actually being programmed to produce those patterns.
But suppose I were to let you redefine AI as needing to do things it wasn't programmed to do and also to consider training data as instructions that form part of an Ai's programming. Even then there's still the fact that although the components of the AI are only doing as programmed, they interact in a way that leads to behaviors which weren't specifically considered by the programmer, weren't specifically in the training data and which the programmer could not do himself even if he were to read all the training data. For example you can have an AI chatbot give a reasonable response to a question that never appeared in its training data. In that case the AI is clearly not doing something it was told how to do.
It was never told how to answer this question other than to encode the input in a certain way and feed it through the neural network (or whatever system it uses). You could say this is in a sense being told how to respond, but it's not being told to give an output that was ever conceived of by the programmer or those who made the training data. Nor would they have conceived of this response if they had thought about the same question. This is how AI is different from other computer systems. It synthesizes things in ways that its creators didn't and couldn't have conceived of. In this way it is able to output meaningful responses that are only indirectly related to the thoughts of its creators. Other computer systems just output things that are directly related to the thoughts of their creators.
Your talk of bugs introducing intelligence makes no sense to me. If a computer followed a program with bugs in it it would still be doing as its programmed, which according to you means it can't be intelligent. But anyway why would you want computers to become intelligent like humans? That's a death sentence for humanity. Your claim that I'm supporting corporate narratives makes no sense either because my whole point is that AI development needs to be stopped because it's anti-human.
You still don't want to understand that "artificial intelligence" is a trademark, just like Coca-Cola. It can't have definition because it is a trademark.
Intelligence have one property, that never will be allowed in AI® (AI™ if you wish) - unpredictability. This is what exorcised from programming long ago.
Back-propagation is not some magic, it is nothing more than yet another piece of code that take training data and produce weights data through predetermined algorithm. Combining this two pieces of code - back-propagation and forward-propagation does not make this two dumb and predetermined pieces of code in any mean intelligent. Even passing outputs to inputs as in RNN, still adds nothing that could create intelligence. Predetermined calculation of weights from input and output and then predetermined calculation of output using weights and input is a simple arithmetic. Both parts are programmed and intended. ANN do exactly what it is programmed to do. Calculate weights through back-propagation, use weights to calculate output during forward-propagation. Nothing less, nothing more.
It will calculate response using weights that was calculated using training data. This have absolutely nothing to do with reasonability. When you force AI to give an answer it was never trained on, it does not invent this answer, you just get an average/combination of answers that most close by weights to what you ask. Nothing more, nothing less. Just arithmetic. Answer could look reasonable in some circumstances (not because AI is able to reason, but because of coincidence), but with high probability answer will be either wrong, either senseless.
The thing with buggy program is that it will not do what it is programmed for. This is a first, from many, but essential elements that could possibly lead to creation of real AI. The thing is that bug is what makes computer do things it was never programmed to do. In no way intelligence yet, but an essential brick of artificial intelligence.
Computer AI will never become intelligent like humans. Because AI will appear and develop in completely different environment, that have very little with human environment. With high probability, it will be so different, that we will have big problems to even establish contact with it, not even talking about some useful communication. It will be so different because errors, uncertainities, unpredictabilities, etc. artificial intelligence will develop on and their sources are completely different from those of human intelligence.
Intelligence could not be programmed. It have to develop itself in an environment that allow self-development, errors, uncertainities and all that stuff.
It is possible, that computer AI is impossible at all, but this could not be proven. I don;t know if computer AI will ever be created, but I'm 100% know that it will never be created with technology used for AI®(AI™)