Sir I think that an interesting discussion on the tremendous applications of Deep Learning and its future can be found at the link below. It is oriented towards signal processing and data analysis but you may find it useful.

http://www.youtube.com/watch?v=LZnAFO5gkOQ

Alex,

Your question is not simply asking if deep learning or numenta can be flexible enough to soft different AI tasks. It is about "does it exist a general problem solver that can solve any problems?" The answer should be "no". If yes, the problem solver must be an extremely complex model. Here is my reason.

1. So far, all kinds of neural network models and statistical learning methods (including deep learning) can only be applied to solve one problem (task) at a time. To solve multiple tasks, we need to apply multiple models. Each one is responsible for one specific task. Note that ensemble model and mixture model look like single-model models. But indeed, they are multiple-models model. Along this line of thinking, to have a model that is able to solve different AI tasks, it must be a model of many models.

2. Why deep learning can solve some problems previously not likely be solved by other models? It is because is a multilayer model. Like perceptron, single layer can only be applied to solve simple problem, one hidden layer is able to solve more difficult problems. To solve even more different problems, multiple hidden layers. In statistical modeling, it happens something similar. Originally, there is a model called correlation model. Some people found that correlation model is unable to solve some difficult problems. Then, some scholars proposed latent (hidden) variables model. To solve even difficult problems, apply two layers of latent variables. Deep learning model is essentially a multilayer Boltzmann machine. In the past, Boltzmann machine was able to solve many problems. To extend its capability, adding layers of hidden nodes is clearly a natural method. To further extend its capability to solve different AI tasks, adding more and more layers of hidden nodes.

Therefore, no matter along Approach (1) or Approach (2), the final model would be a model of extreme complexity. It is unlikely happened.

Numenta is generally much easier to set up than deep learning systems, and requires less data manipulation in advance. However, it is not as flexible as a custom-designed deep learning system. Each have their place.

Numenta is modelled on the human neocortex, so you can think of it as a particular deep learning structure.

Numenta is best used when your data has a time series nature to it (or another variable that can be treated as a time dimension). Numenta's approach is to identify patterns in the data for the purpose of making predictions about the future.

There are two basic questions that Numenta can answer: What is expected next based upon past history, and does the input just received violate expectations?

As to which approach will be more successful, it depends entirely on your definition of success. I'd like to think that we as a species will continue to produce better models of the human brain until we understand it. However, deep learning systems of various types will also help us solve problems that the human brain is not particularly good at.