# Why can't cognitive architectures achieve general intelligence?

Newbie here.

I recently read about cognitive architectures (see: https://en.wikipedia.org/wiki/Cognitive_architecture). They are supposed to be modeled after the human mind and represent a promising approach towards artificial general intelligence (AGI).

My question is, however, why haven't these cognitive architectures achieved AGI yet? What are the specific limitations and roadblocks that cognitive architectures face?

• IMO the problem is we don't know how to train giant recurrent neural networks over very long timescales. There is likely some giant recurrent neural network that could do the task of AGI, running on current hardware - but we don't know how to find the weights for it. Gradient descent suffers from the vanishing/exploding gradient problem. May 12 at 4:00
• Another way to look at it is, if you manually design a large software system such as a cognitive architecture, it's going to be full of bugs, because all large manually designed software systems are full of bugs, especially when they must interact with the real world. Thus what we need is some way of automatically ironing out the bugs. This is what gradient descent can do for neural networks, but it just doesn't do it well enough, and many cognitive architectures can't be backpropagated end-to-end anyway. May 12 at 4:06
• @causative: a possibility could be to generate the code of that cognitive architecture software. This was suggested by Jacques Pitrat in his Artificial Beings: the Conscience of a Conscious Machine book, and with others I am trying to reimplement that idea for the RefPerSys project. However, such an approach needs years of work. Feel free to email me at basile@starynkevitch.net May 12 at 5:54
• @Anonymous: I tend to believe that a full book could be written to answer your question. Did you consider starting your PhD on it? May 12 at 5:57