5
$\begingroup$

Is there empirical evidence that some approaches to achieving AGI will definitely not work? For the purposes of the question the system should at least be able to learn and solve novel problems.

Some possible approaches:

  1. A Prolog program
  2. A program in a traditional procedural language such as C++ that doesn't directly modify its own code
  3. A program that evolves genetically in response to selection pressures in some constructed artificial environment
  4. An artificial neural net
  5. A program that stores its internal knowledge only in the form of a natural human language such as English, French, etc (which might give it desirable properties for introspection)
  6. A program that stores its internal knowledge only in the form of a symbolic language which can be processed unambiguously by logical rules
$\endgroup$
7
  • 1
    $\begingroup$ To answer your second question, you could look at this and this (in particular, the computational theory of mind). So, I would clarify that your main question is "Are there any approaches to AGI that will definitely not work?" and actually focus on that interesting question (and remove the last one, which is different, and partially answered in the linked posts), which can also be improved, such as by asking "Is there any empirical evidence that a certain approach cannot be used to implement an AGI?". $\endgroup$
    – nbro
    Jan 26 '21 at 19:36
  • 1
    $\begingroup$ i would vote for ANN to do AGI, it's real biology $\endgroup$ Jan 27 '21 at 8:44
  • 1
    $\begingroup$ @persiflage Of course, most people nowadays would say "neural networks" (not to your question, but to the question "Which technique is the most promising to create AGI?"). However, yesterday, I was watching a recent video where 1 AI researcher and 1 cognitive scientist were discussing similar topics, and one of them said that he would be surprised if we could create an AGI only out of neural networks. He pointed out that we are mainly symbolic. We think in terms of symbols and concepts ("cat", "house", etc.). Neural networks are sub-symbolic (they do non-linear regression). $\endgroup$
    – nbro
    Jan 27 '21 at 16:23
  • 1
    $\begingroup$ The cognitive scientist, on the other hand, pointed out that we really need the ability to "learn how to learn" and we really need the neural network (or whatever) to be able to solve new problems without having really been programmed/trained to do so explicitly. Nowadays, most neural networks are actually programmed/trained to solve a specific task, and they would completely fail in all other tasks. This is absolutely no good. Unfortunately, neural networks trained with back-propagation actually may inherently suffer from this issue. $\endgroup$
    – nbro
    Jan 27 '21 at 16:28
  • 1
    $\begingroup$ That's why I also think that neural networks alone are probably not the path to AGI. $\endgroup$
    – nbro
    Jan 27 '21 at 16:28
1
$\begingroup$

Very interesting question. Assuming that the programming languages used are powerful enough (say Turing Complete), all of the above actually should lead to an AGI. The difference is in how efficiently they can do it, both in term of number of computations required and in the length of the program.

So the question could be rephrased as: which approach cannot lead to an AGI using less than X resources and being shorter than Y characters? The second question is basically asking the Kolmogorov complexity of the AGI in that language, which is uncomputable. Since we cannot find the shortest program, I don't think we can make conclusions about the maximum program efficiency either. In summary I don't see a way to rule out any of those approaches (but I would be very happy to be proven wrong).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.