HTM seems to be a very power tool for machine intelligence. What is its limitation we have already known?
It has limited capabilities at this point in its evolution. The most useful application of HTM as it exists today in open source is anomaly detection on steaming scalar values.
Our approach promises more functionality as it evolves, mostly because we are keeping our research true to neuroscience. We should be able to systematically add more functionality of intelligence as we figure out how more of this works, especially the interaction between layers in a region.